United States
Environmental Protection
Agency
Industrial Environmental Research
Laboratory
Research Triangle Park NC 27711
EPA-600/2-78-004u
August 1978
Research and Development
Source Assessment:
Analysis of
Uncertainty -
Principles and
Applications

-------
                 RESEARCH REPORTING SERIES


Research reports of the Office of Research and Development, U.S. Environmental
Protection Agency, have been grouped into nine series. These nine broad cate-
gories were established to facilitate further development and application of en-
vironmental technology. Elimination  of  traditional grouping was consciously
planned to foster technology transfer and a maximum interface in related fields.
The nine series are:

    1. Environmental Health Effects Research

    2. Environmental Protection Technology

    3. Ecological Research

    4. Environmental Monitoring

    5. Socioeconomic Environmental Studies

    6. Scientific and Technical Assessment Reports (STAR)

    7. Interagency Energy-Environment Research and Development

    8. "Special" Reports

    9. Miscellaneous Reports

This report has been assigned to the  ENVIRONMENTAL PROTECTION TECH-
NOLOGY series. This series describes research performed to develop and dem-
onstrate instrumentation, equipment,  and methodology to repair or prevent en-
vironmental degradation from  point and non-point sources of pollution. This work
provides the new or improved technology required for the control and treatment
of pollution sources to meet environmental quality standards.
                       EPA REVIEW NOTICE
This report has been reviewed by the U.S. Environmental Protection Agency, and
approved for publication. Approval does not signify that the contents necessarily
reflect the views and policy of the Agency, nor does mention of trade names or
commercial products constitute endorsement or recommendation for use.

This document is available to the public through the National Technical Informa-
tion Service, Springfield, Virginia 22161.

-------
                                EPA-600/2-78-004u
                                        August 1978
    Source Assessment:
 Analysis of Uncertainty -
Principles and  Applications
                   by

  R.W. Serth, T.W. Hughes, R.E. Opferkuch, and E.G. Eimutis

          Monsanto Research Corporation
              1515 Nicholas Road
              Dayton, Ohio 45407
            Contract No. 68-02-1874
             ROAPNo. 21AXM-071
           Program Element No. 1AB015
         EPA Project Officer: Dale A. Denny

      Industrial Environmental Research Laboratory
       Office of Energy, Minerals, and Industry
         Research Triangle Park, NC 27711
                Prepared for

     U.S. ENVIRONMENTAL PROTECTION AGENCY
        Office of Research and Development
             Washington, DC 20460

-------
                            PREFACE
The Industrial Environmental Research Laboratory (IERL)  of EPA
has the responsibility for insuring that pollution control tech-
nology is available for stationary sources to meet the require-
ments of the Clean Air Act, the Federal Water Pollution Control
Act and solid waste legislation.  If control technology is
unavailable, inadequate, or uneconomical, then financial support
is provided for the development of the needed control techniques
for industrial and extractive process industries.  The Chemical
Processes Branch of the Industrial Processes Division of IERL
has the responsibility for investing tax dollars in programs to
develop control technology for a large number of operations (more
than 500) in the chemical industries.

Monsanto Research Corporation (MRC)  has contracted with EPA to
investigate the environmental impact of various industries which
represent sources of pollution in accordance with EPA's respon-
sibility as outlined above.  Dr. Robert C. Binning serves as MRC
Program Manager in this overall program entitled "Source Assess-
ment," which includes the investigation of sources in each of
four categories:  combustion, organic materials, inorganic
materials, and open sources.  Dr. Dale A. Denny of the Industrial
Processes Division at Research Triangle Park serves as EPA Pro-
ject Officer.  Reports prepared in this program are of three
types:  Source  Assessment Documents, State-of-the-Art Reports
and Special Project Reports.

Source Assessment Documents contain data on emissions from spe-
cific industries.  Such data are gathered from the literature,
government agencies and cooperating companies.  Sampling and
analysis are also performed by the contractor when the available
information does not adequately characterize the source emis-
sions.  These documents contain the information necessary for
IERL to decide whether a reduction in emissions is required.

State-of-the-Art Reports include data on emissions from specific
industries which are also gathered from the literature, govern-
ment agencies and cooperating companies.  However, no extensive
sampling is conducted by the contractor for such industries.
Results from such studies are published as State-of-the-Art
Reports for potential utility by the government, industry, and
others have specific needs and interests.
                               111

-------
Special projects provide specific information or services which
are applicable to a number of source types or have special
utility to EPA but are not part of a particular source assess-
ment study.  This special project report, "Source Assessment:
Analysis of Uncertainty, Principles and Applications," was pre-
pared to describe a procedure for analyzing the uncertainty
associated with data and other information contained in source
assessment studies.  The general principles and procedures
used in the analysis are illustrated by means of four applica-
tions to decision making in the area of environmental control.
The application of these principles to the study of air emissions
in the Source Assessment Program is described in detail.
                               IV

-------
                           ABSTRACT
This report provides the results of a study that was conducted
to analyze the uncertainties involved in the calculation of the
decision parameters used in the Source Assessment Program and
to determine the effect of these uncertainties on the decision-
making procedure.

A general procedure for performing an analysis of uncertainty
is developed based on the principles of error propagation and
statistical inference.  It is shown that this simple and
straightforward method represents an approximation to standard
statistical techniques.  The approximate method is illustrated
by application to four problems in the area of environmental
control.

The general procedure is used to establish guidelines for
conducting air emissions studies in the Source Assessment
Program.  In particular, guidelines are established for preci-
sion in field sampling and analytical work, and for setting
critical values of decision parameters.

This report was submitted in partial fulfillment of Contract
No. 68-02-1874 by Monsanto Research Corporation under the
sponsorship of the U.S. Environmental Protection Agency.  The
study described in this report covers the period November
1976 to March 1978.
                               v

-------
                            CONTENTS
Preface	ill
Abstract  	   v
Figures	ix
Tables  	   x
Abbreviations and Symbols	xi

   1.   Introduction  	 1
   2.   Summary		4
            Background  	 4
            Source severity 	 4
            National emission burden  	 5
            Offset calculations 	 5
            Reasonably available control technology 	 6
   3.   Statistical Basis for the Procedure 	 8
   4.   Source Assessment Methodology 	  13
            Definition of indices 	  13
            Expanded discussion of source severity  	  15
            Representative source 	  17
            Source severity distributions 	  19
            Stochastic approach to source severity  	  19
   5.   Source Assessment:  Source Severity 	  22
            Introduction  	  22
            Uncertainties in parameters 	  23
            Deterministic decision approach 	  25
            Stochastic decision approach  	  28
            Tests of hypothesis	30
            Comparison of alternate decision approaches ....  35
            Guidelines for source assessment program  	  35
            Allowable random uncertainty in emission factor .  .  39
            Effect of random uncertainty in emission factor .  .  40
            Operating characteristics of the test	42
            Summary	46
   6.   Source Assessment: Emissions Burden 	  47
            Introduction  	  47
            Uncertainty in Ng	48
            Statistical test of hypothesis	50
            Summary	51
   7.   Offset Calculations: Plant Expansion Problem  	  52
            Introduction  	  52
            Working equations	52
            Uncertainty in (Q2 - Qt)	53
            Test of hypothesis	55
            Numerical Example 	  56
            Summary	57

                                vii

-------
                       CONTENTS (continued)
   8.  Comparison of Alternative Controls:  Reasonably
         Available Control Technology 	  59
            Introduction  	  59
            Governing equations .	59
            Test of hypothesis	61
            Numerical example 	  62
            Summary	63
   9.   Conclusions and Recommendations	64
            Conclusions	64
            Recommendations 	  66

References	68
Appendices
   A.  Accuracy, error, and uncertainty 	  74
   B.  Use and interpretation of error propagation formulas .  .  77
   C.  Derivation of source severity equations  	  98
   D.  Plume rise	110
   E.  Alternative methods for estimating "acceptable"
         concentration for noncriteria pollutants 	 116
   F.  Construction of operating characteristic curves  .... 120
   G.  Relationship of sampling and analysis bias to
         systematic errors  	 	 127
   H.  Source severity simulation and probabilistic
         sensitivity analysis 	 129
   I.  Quantification of uncertainty in source severity . . .  .139
   J.  Example calculations for source assessment:  carbon
         black manufacture	151

Glossary	154

Conversion Factors and Metric Prefixes  	 156
                                Vlll

-------
                              FIGURES
Figure                                                         Page

  1.   Deterministic source severity distribution for
          carbon monoxide emissions from carbon black
          plants	20

  2.   Impact of alternative hypothesis test approaches
          on S*	36

  3.   Variation in critical source severity as a function
          of emission factor data quality and the accept-
          able number of days for severity to exceed 1.0
          when treating the true source severity as a
          random parameter 	   36

  4.   Schematic representation of the fiducial statis-
          tical approach	42

  5.   Operating characteristic curves for Case A with
          br = 0.10  .  . .	43

  6.   Effective operating characteristic curve for
          Case A with b  = 0.10	44

  7.   Effective operating characteristic curves for
          Case A as a function of random uncertainty in
          emission factor  	   45

  8.   Operating characteristic curve for the stochastic
          decision test approach 	   45

  9.   Effective operating characteristic curve for
          example problem  	   57
                                IX

-------
                              TABLES


Number

  1.    Measure/Analysis Aspects of Source Severity 	  5

  2.    Measure/Analysis Aspects of the National
          Emissions Burden  	  6

  3.    Generalized Error Bounds for Source Assessment  .... 37

  4.    Contribution of Individual Uncertainties to Total
          Uncertainty in Source Severity (Deterministic
          Decision Approach)   	 40

  5.    Contribution of Individual Uncertainties to Error
          Bounds on Source Severity (Deterministic
          Decision Approach)   	 41
                                 x

-------
                    ABBREVIATIONS AND SYMBOLS
SECTIONS 1 THROUGH 5

A        — estimated value of average production capacity, kg/s

a        — random uncertainty associated with A, kg/s

ar       -- ar/A

B        — measured value of emission factor, g/kg

b        — bias

b,,, b    — bounds on bias
 JG   LI
b        — random uncertainty associated with B, g/kg

b        — uncertainty in emission factor due to imprecision in
 ri
 r2
            sampling, g/kg
b        — uncertainty in emission factor due to imprecision in
            analysis, g/kg
b        — uncertainty in emission factor due to process
            variation, g/kg

b        — uncertainty in emission factor due to imprecision in
            measurement of production rate, g/kg
si
su'
bsu
c
CAP
CO
Cr
ฃr
— br/B
" bs/B
. — upper and lower systematic uncertainties associated
s with B, g/kg
— b /B
su'
— estimated value of effective emission height,
— production capacity, kg/s
— carbon monoxide
- random uncertainty associated with C, m
-- cr/C
m



c  , c . — upper and lower systematic uncertainties associated
 su   s*    with C, m

c        — c  /C
 su          su7
D        — estimated value of "acceptable" concentration, g/m3

                                xi

-------
               ABBREVIATIONS  AND  SYMBOLS  (continued)
 ds*       -  ds/D
 d  ,  d .  —  upper  and  lower  systematic  uncertainties  associated
  su    siL     with D,  g/m3
 d        —  d   /D
  su          su7                     ^
 e        —  natural  base  logarithm,  =  2.72
 E (x)      —  expected value of % (underbar denotes  random  variable)
 EF        —  emission factor, g/kg
 e        —  random uncertainty  associated with  Sr
 *l        ~  er/SC
 /s
 e  ,  e .  —  upper  and  lower  systematic  uncertainties  associated
  su    SJ6     with Sc
 >  A    "  esu/SC
 e , e.    —  upper  and  lower  total relative uncertainties  associ-
  u           ated with  SG
 F        —  "acceptable"  pollutant concentration,  g/m3
 flf f2    —  weighting  factors
 G        —  factor which  converts TLV to "equivalent" PAAQS
 h        —  physical stack height, m
 H        —  effective  emission  height,  m
 HQ        —  null hypothesis
 HA        —  alternate  hypothesis
 I        —  calculated value of decision index
 I        —  true value of decision index
 k        —  constant,  s/m
 m        —  97.5%  point of distribution of y
 M        —  annual mass emissions of given criteria pollutant
  n           from all stationary sources nationwide, metric tons
 M        —  annual mass emissions of given criteria pollutant
             from given source type, metric tons
 M        —  annual mass emissions of given criteria pollutant
  s           from all stationary sources within  a given  state,
             metric tons
 n, nj, n2—  sample size
 n        —  fraction of a year
NO       — nitrogen oxides
PAAQS    — primary  ambient air quality standard,  g/m3

                              xii

-------
               ABBREVIATIONS  AND SYMBOLS  (continued)

 Q        — mass emission rate, g/s
 r        — number of days per year
 s,  BI,  82— sample standard  deviation
 S        — source severity
 S_,        — "calculated"  source severity
 S*        — critical  test value of  calculated  source  severity
  w
 SO        — sulfur oxides
  X
 S        — "true" source severity
 t        — averaging time,  min;  or student  t  value
 t        — weighted  average Student t value
  ave
 TLV      — threshold limit  value,  g/m3
 t        — short-term averaging  time, 3 min
 V(x)      — variance  of x
 u        — wind  speed, m/s
 a        — level  of  test
 Y        — correction factor  for Gaussian plume equation
 AH       — plume  rise, m
 y,  yi, y2— population mean
 a        — population standard deviation
 X        — estimate  of population  mean
 x"'  x"i i X2—  sample mean
 Xmax     —  maximum time-averaged ground level concentration, g/m3
 v
 Ameas,
  X_red  — measured  and  predicted  values of ground level concen-
   p         tration,  g/m3
 Z        — number of days per year
 SECTION 6
A        — measured  value of CAP ,  kg/yr
 *                                 +•
 a        —  relative  random  uncertainty associated with A
 B        — measured  value of EF^,  g/kg
 *
 b        —  relative  random  uncertainty associated with B
 j-,   f b   —  upper  and lower  systematic uncertainties  associated
  su   ai     with B, g/kg
 CAP      —  total  production capacity of given source type, kg/yr
 D        — measured  value of M  , kg
 /s                               ^
 d        — relative  random  uncertainty associated with D

                              xiii

-------
              ABBREVIATIONS AND SYMBOLS  (continued)

d   , d  0 — upper and lower systematic uncertainties  associated
 su   s*    .with D, kg
EF       — representative emission factor  for given  source  type,
  R
e        — relative random uncertainty associated with  (NB)
 XT             .                                              ^f
e   , e  „ — upper and lower systematic uncertainties associated
 su   sJi    with  (NB)C
e , e\   — relative upper and lower total uncertainty in N
H        — null hypothesis
M        — annual mass emissions of a given criteria pollutant
 n          from all stationary sources nationwide, kg
MTTTP^O    — estimate of M  obtained from 1972 National Emissions
 1*EDS       Report, kg   n
M        — annual mass emissions of a given criteria pollutant
 "          from a given source type, kg
N_.       — national emissions burden
 13
(N,,)-,    — calculated value of N,,
  B C-                            o
(NB)T    ~ true value of NB
(N*)_    — critical test value of  (N_)n
  D t-                                D L.
a        — level of test
SECTION 7
b            — random uncertainty associated with B , g/s
b N-         — random uncertainty in Q  . , g/s
b            — random uncertainty associated with BI , g/s
b   i         — random uncertainty associated with BI ' ,  g/s
b   .         — random uncertainty in Ql . , g/s
b   , .         — random uncertainty in Q  , . , g/s
b            — random uncertainty associated with B^, g/s
b  , b       — upper and lower systematic uncertainties
 un             associated with BN/ g/s
b   .,  bOM.    — upper and lower systematic uncertainties
 UNI   AMI      associated with Q ^, g/s
b  , bff      — upper and lower systematic uncertainties
 u              associated with BI , g/s
b   , ,  b. ,    — upper and lower systematic uncertainties
 ul              associated with Bj1, g/s

                               xiv

-------
              ABBREVIATIONS AND SYMBOLS  (continued)

b   ., ba,.    — upper and  lower systematic uncertainties  in Q. . ,
 ui]   *ID      g/s                                           ID
b   ,., b,, ,.  — upper and  lower systematic uncertainties  in Q ,.,
     3       D    g/s                                           l  D
b  ., bป.      — upper and  lower systematic uncertainties
 u              associated with B., g/s
B             — measured value of  Q , g/s
BI            — measured value of  Ch > g/s
BI,           — measured value of  Qi,, g/s
B.            — measured value of  A i
e             — random uncertainty associated with difference
                BN - BA, g/s
e   , e .,      — upper and  lower systematic uncertainties
 su   S         associated with B  - B. , g/s
e  , e         — upper and  lower total uncertainties associated
 U            .  with BN -  V g/S
Ha            — alternative hypothesis
 A
HO            — null hypothesis
Q             — emission rate from new unit, g/s
Q  .           — emission rate from ith emission point in  new
                unit, g/s
Qi            — original emission  rate, g/s
Qi,           — emission rate from original unit after expansion,
                g/s
Q  .   .        — emission rate before expansion from jth emission
  ^             point in original  unit, g/s
Q  ,.          — emission rate after expansion from jth emission
   •*            point in original  unit, g/s
Q2            — emission rate after expansion, g/s
(Q2 - QI),-,    — calculated value of Q2 - QI = B  - B., g/s
(Q2 - QI)T    — true value of Q2 - QI/ g/s
AQi           — Q2 - QIii  - reduction in emission rate from
                original unit, g/s
(AQj).        — difference in emission rate before and after
     •^          expansion  from jth emission point in original
                unit, g/s
a             — level of test
b             — random uncertainty in Qj, g/s
                                XV

-------
 SECTION  8
              ABBREVIATIONS AND  SYMBOLS  (continued)
 rl    n!
  etc.
br2
bui' bฃi
bui = bui/Bi
  etc.
B
esu'
V e*
HA
HO
Qz
a
 (ฃ2)c
 (ฃ2 -
SECTION 9
/\
br
TLV

APPENDIX B

a
A
— relative random uncertainties
-- random uncertainty in Q , g/s
— upper and lower systematic uncertainties in Q  , g/s

-- relative systematic uncertainties
— upper and lower systematic uncertainties in Q  , g/s
— measured value of Ql , g/s
-- measured value of Q , g/s
— upper and lower systematic uncertainties in
   (ฃ2 - ei)
— upper and lower total uncertainties in  (ฃ2 ~ e1)
— alternative hypothesis
— null hypothesis
— uncontrolled emission rate, g/s
— controlled emission rate with presently installed
   device, g/s
— level of test
— lower and upper systematic uncertainties in e2
-- random uncertainty in e2
-- control efficiency of presently installed device
— control efficiency of RACT
— nominal value of RACT control efficiency
— nominal value of (e2 - e^)
— true value of (e2 - e^
  relative random uncertainty in emission factor
  threshold limit value
  uncertainty associated with A
  nominal value of X  or x
                               xvi

-------
              ABBREVIATIONS AND SYMBOLS  (continued)
ai, a2      —  lower and upper systematic uncertainties associ-
                ated with A
b           —  uncertainty associated with B
B           —  nominal value of X2 or x2
bj, b2      —  lower and upper systematic uncertainties associ-
                ated with B
c           —  uncertainty associated with C
C           —  nominal value of y or y
GI, C2...C  —  constants
dn-;i        —  Behrens-Fisher statistic
f           —  effective degrees of freedom
f'(x)       —  first derivative of f(x)
f(xj, x2)   —  arbitrary function of xi and x2
k, k        —  constants
n  , n       —  sample sizes
                 !1 if x > 0
                -1 if x < 0
                 0 if x = 0
s           — pooled estimate of standard deviation
s           — sample standard deviation of x
 J\.
tj_  ,       — (1 - a/2) percentage point of t-distribution
 i-a/2, v      with v degrees of freedom
V , V,       — variance factors
xi, x2...x  — independent variables
xj, x2      — mean values of xj and x2
y           — dependent variable
y           — mean value of y
Z,    ,       — (1 - a/2) percentage point of standard normal
  ~a/2         distribution
a           — one minus confidence level
04, a2ป
  3lป 32    — factors defined in Table A-3
eA, EB...
            — errors associated with A, B. .  -N, and AB

        •   — maximum and minimum values of error in product AB
       mm                                            r
                               xvii

-------
              ABBREVIATIONS AND SYMBOLS  (continued)

 a2       — variance of x
  H

 APPENDIX C
 A       — area of annular  region,  m2
 A'       — area of sector,  m2
 a, b    — parameters  in  relationship  for  a
 c, d,  f  — parameters  in  relationship  for  a
                                           z
 Dp       — population  density, persons/m2
 F       — "acceptable" concentration, g/m3
 H       — effective emission height,  m
 P       — affected population, persons
 PAAQS    — primary ambient  air quality standard,  g/m3
 Q       — mass emission  rate, g/s
 S       — source  severity
 t       — averaging time, min
 TLV      — threshold limit value, g/m3
 t       — short-term averaging time,  3 min
 u       — wind speed, m/s
 x       — downwind distance from emission source, m
 xi,  X2   — roots of equation x/F =  1-0, m
 y       — horizontal distance from plume centerline, m
 a       — lateral dispersion coefficient, m
 a       — vertical dispersion coefficient, m
  z
 X        — short-term  (3 min) time-averaged ground level concen-
           tration, g/m3
 X        — short-term maximum ground level concentration, g/m3
 max
 X        ~~ maximum time-averaged ground level concentration,  g/m3
 max
 APPENDIX D
C        — estimated value of effective emission height, m
C-.       — heat capacity at constant  pressure of effluent,
'P
            kcal/g-ฐK
                      AHcalc>
                              XVlll

-------
               ABBREVIATIONS AND  SYMBOLS  (continued)

 c   ,  c    —  upper  and  lower systematic error bounds  for effective
             emission height, m
 Csu       -  Csu/(Hs +  AHcalc>
 D         —  stack  diameter, m
 H         —  physical stack height, m
 s
 M         —  molecular  weight of  effluent, g/g mole
 P         —  atmospheric pressure
 R         —  gas constant = 8.314 x 105 dyne-m/g mole-ฐK
 T         —  ambient air temperature,  ฐK
 a
 T         —  stack  gas  temperature, ฐK
 S
 u         —  wind speed, m/s
 V         —  stack  gas  exit velocity, m/s
 s
 AH        —  plume  rise, m
 AH  ,     —  calculated value of  plume rise, m
  calc
 AH,      -- true value of plume rise, m

APPENDIX  E
 b         - exponent in averaging time correction factor
C       — pollutant exposure concentration, g/m3
D       — estimated value of "acceptable" concentration, g/m3
d.,  d   — lower and upper uncertainties associated with D, g/m3
 A/   U
F       — "acceptable" concentration, g/m3
G       — conversion factor
LD50    — dosage which results  in mortality to 50% of exposed
           population,  g/m3 or yg/m3
P       — probability of a lifetime response to pollutant
           exposure
p*      — "acceptable" probability of a lifetime response to
           pollutant exposure
PAAQS   — primary ambient air quality standard, g/m3 or yg/m3
PAAQS2i+ — estimated 24-hr primary ambient air quality standard,
           g/m3 or yg/m3
t       — averging time for PAAQS, hr
 3. V CJ
TLV     — threshold limit value, g/m3 or yg/m3

                               xix

-------
              ABBREVIATIONS AND  SYMBOLS  (continued)


PAAQS    — primary  ambient air quality standard, g/m3 or  yg/m3

PAAQS2it  — estimated  24-hr primary ambient air quality  standard,
           g/m3 or  yg/m3

t        — averging time  for PAAQS, hr
 clVCJ
TLV      — threshold  limit value, g/m3 or yg/m3

APPENDIX F
/\
b        — relative random uncertainty associated with  emission
           factor
s*
e        — relative random uncertainty associated with  S

e   , e  0 — upper and lower systematic uncertainties  associated
 su   s*   with s_
/s    ^            w
e   , e  . — upper and lower relative  systematic  uncertainties
 su   s     associated with S_
A   A                        V^
e  , e.   — upper and lower total relative uncertainties  associ-
 u   *      ated with S,,
                       \*
F(z)     — area under normal curve between minus  infinity and  z

P        — probability

Sc       — calculated source severity

S        — true source severity
z        — standard  normal deviate

a        — standard  deviation

y        — mean

APPENDIX G

B    , B.    — measured  and true values of emission  factor,  g/kg
 meas    true
A    A
b  , b  .     — upper and lower  relative systematic uncertainties
 su   s         in  emission factor
B , 3~       — bounds on positive and negative  bias  in sampling
                and analysis

APPENDIX H

A          — coefficient in expression for a
           — parameters in expression for  a
% Be       — fraction of beryllium in coal
CONS       — coal consumption, g/s
                                xx

-------
              ABBREVIATIONS AND SYMBOLS (continued)


 d          — stack diameter,  m
 EF         — emission factor, g/kg
 h          — physical emission height,  m
 H          — h + Ah = effective emission height,  m

 P
 p,          — barometric pressure,  mb
 Q          — mass emission rate,  g/s
 S          — source severity

 S~          — mean source severity
  J\,
 So.gs       — 95%  point  of  source  severity  distribution
 T          — ambient temperature,  ฐK
  a
 TLV         — threshold  limit value,  g/m3
 T          — stack  gas  temperature,  ฐK
  S
 u          — wind speed, m/s
 V          — stack  gas  exit  velocity, m/s
  S
 x          — downwind distance, m
 Ah          — plume  rise, m
 a          — lateral dispersion coefficient, m
 a          — vertical dispersion coefficient, m
  z
 Y  Y   ,     — short-term average ground level concentration,  g/m3
  ,  peaK
 X2. .        -- 24-hr  average ground  level concentration,  g/m3

APPENDIX I
 S*
a             — relative random uncertainty in average production
  r              capacity
 ^ป
b             — relative random uncertainty in emission  factor
 A. 27   /\
b p , b        — lower and upper relative systematic uncertainties
  s    su        in emission factor

CO            — carbon monoxide
 s\
c             — relative random uncertainty in average emission
 *r   *          height
c . , c        — lower and upper relative systematic uncertainties
                in average  emission height
/\    /\
d ., d        — lower and upper relative systematic uncertainties
  s    su        in "acceptable" concentration

                               xxi

-------
               ABBREVIATIONS  AND  SYMBOLS  (continued)

 (F  ).         —  "calculated" value of  "acceptable" concentration
                 for pollutant  i,  g/m3
 (F  ).         —  "true" value of  "acceptable"  concentration  for
     1            pollutant  i, g/m3
 G             —  factor which converts  TLV to  "equivalent" PAAQS
 G             —  average TLV  conversion factor
 G.            —  TLV conversion factor  for pollutant  i
 m             —  97.5% point of distribution of y
 NO            —  nitrogen oxides
  H
 PAAQS         —  primary ambient  air quality standard, g/m3
 s             —  sample standard  deviation
 SO            —  sulfur oxides
  X
 t             —  averaging  time, min
 TLV           —  threshold  limit value, g/m3
 (TLV).         —  threshold  limit value  of pollutant i, g/m3
 t             —  short-term averaging time, min
 x             —  sample mean
 I    _^
 3,3         —  bounds on  positive and negative bias in sampling
                 and analytical procedures
 Y             —  correction factor for  Gaussian plume equation
 y             —  population mean
 a             —  population standard deviation
 Y             —  maximum time-averaged  ground  level concentration,
 max             g/m3
 X      ~x   a  —  measured and predicted values of ground level
 meas,   pred     concentration, g/m3
APPENDIX J
 /s
a        — relative random uncertainty in average production
            capacity
 /\
b        — relative random uncertainty in emission factor
 /\ JL    s\
b ., b    — lower and upper relative systematic uncertainties in
            emission factor
 /\
c        — relative random uncertainty in emission height
e ,  e    — lower and upper uncertainties in  source severity
POM      — polycyclic organic material

                               xxii

-------
              ABBREVIATIONS AND SYMBOLS  (continued)
S        — calculated source severity



S*       — critical test value of calculated source severity



S        — true source severity



TLV      — threshold limit value
                              XXlll

-------
                            SECTION 1

                          INTRODUCTION


The  Industrial Environmental Research Laboratory  (IERL) of EPA
uses emissions data to generate information essential to environ-
mental decision making.  These data are always subject to a
degree of uncertainty because of inherent errors  in the methods by
which they are obtained.  When projections based  on the data are
employed in decision making, this uncertainty, which extends into
the projections, must be considered.

This report develops a decision analysis procedure for dealing
with the uncertainty in source assessment information in a system-
atic and logical fashion.  The procedure, as developed, makes
allowance, on a mathematically sound basis, for the degree of
uncertainty introduced into environmental decision making by
uncertainty in the data.

The procedure described in this report applies specifically to
the evaluation of technical information used in determining the
need to reduce emissions from a source.  It does  not, however,
take into account value judgments, bias in projecting biological
tests from test animals to humans, or unpredictable variations in
magnitude of emissions due to process upsets, spills, or accidents.

The decision analysis procedure developed in this report consti-
tutes an approximation to standard statistical methods.  Such an
approximation is necessary, in general, due to the complexity of
the types of problems considered.  The procedure  is simple and
straightforward, requiring no detailed knowledge  of statistical
distribution theory.  Application of the decision analysis method
requires only the manipulation of a few simple formulas (tabu-
lated in Appendix B) with which most engineers and scientists
have some familiarity.  Therefore, lack of training in statis-
tical methods should not prove a deterrent to utilization of
this procedure.

The decision analysis procedure is illustrated by means of four
examples related to environmental control.  The first two exam-
ples relate to decision problems associated with  lERL's Source
Assessment Program involving air emissions.  The  third example
involves compliance with a given emissions offset policy concern-
ing industrial plant expansions.  The fourth example involves a
comparison between two alternative emission control techniques.

-------
The general decision analysis procedure is outlined in Section 3
of the report.  The approximations involved in the procedure and
its relationship to standard statistical methods are described.
In Section 4,  the decision indices employed in the Source Assess-
ment Program are defined and their use in decision analysis is
discussed.

Sections 5 and 6 present applications of the general decision
analysis procedure to two indices used in the Source Assessment
Program, namely:  source severity and national emissions burden.
Source severity combines emissions data, dispersion modeling, and
health effects information to yield a measure of the pollution
hazard represented by a given source of emissions.

Section 7 addresses a problem involving compliance with local air
pollution regulations.  The situation considered is that in which
an industrial plant expansion must be accomplished without an
increase in total pollutant emissions in order to prevent deteri-
oration of ambient air quality.  Section 8 considers a problem
associated with environmental improvement by means of reasonably
available control technology (RACT).  The decision to be made is
whether to require adoption of RACT in preference to an existing,
installed control device.

The major results presented in the report are summarized in
Section 2.  Conclusions and recommendations are presented in
Section 9.

The decision analysis procedure utilizes standard concepts of
statistical hypothesis testing.  However, there are several areas
where the report used a robust application of statistical theory
in order to make the decision analysis procedure useful to de-
cision makers who are unfamiliar with detailed statistical
methods.  Two aspects of the test of hypothesis as used in
this report that are not commonly employed in statistical analy-
ses involve:   1)  treatment of systematic errors separately from
random errors, and 2)  estimation of the risks of making incorrect
decisions.  Those who rigorously apply statistical methods in a
puristic fashion may take exception to the approximate statis-
tical methods employed in the report.  The discussion presented
below is designed to identify areas where approximate statis-
tical methods have been employed.

Treatment of Systematic Errors

Systematic errors in data are treated separately from random
errors rather than being combined with them.  This treatment is
designed to make the systematic errors (bias)  more visible to
the decision maker so that he can deal with the uncertainty
attributed to systematic errors in a conscious fashion.  This
treatment leads to a family of operating characteristic (OC)

-------
curves for the test of hypothesis and to the associated concept
of an "effective" OC curve for the test.  As explained in
Section 5, the "effective" OC curve represents an entire family
of OC curves and contains a horizontal section, or plateau, in
the center.  While such an OC curve may be anathematic to
statisticians, this procedure permits systematic errors to be
treated in a logical and straightforward manner without recourse
to arguments involving statistical distributions.  It permits
the distinction between bias and imprecision to be maintained
throughout the analysis.  In addition, the authors believe that
this procedure is at least as valid as one which treats system-
atic and random errors as nonseparable entities.

Estimating the Risk of an Incorrect Decision

In developing the estimated risks of making an incorrect de-
cision,  the decision alternatives were handled in a conservative
fashion.  Normal decision analysis methods divide the decision
into three regions:  the acceptance region; the indifference
region;  and the rejection region.  Further, usual practice in
decision analysis methods is to treat the indifference region
as though it is in the acceptance region.  The conservative
approach taken in the report treats the indifference region as
though it is in the rejection region.  This approach provides
the decision maker opportunity to ensure that a decision is not
counter to his charter.  Specifically, it enables IERL the
opportunity to ensure that pollution control technology is avail-
able to meet the requirements of environmental legislation.  The
decision analysis technique does not and is not intended to
ensure that pollution control technology is applied.  The
decision to apply control technology requires input for trade-
off (economic, social, political, etc) considerations to be
made and those considerations are beyond the scope of lERL's
source assessment program.

-------
                            SECTION 2

                             SUMMARY
BACKGROUND

The objective of completing a source assessment in the Source
Assessment Program is to provide sufficient information to aid
IERL in determining the need for reducing pollution from station-
ary sources.  A source is an entire industrial, commercial, or
municipal operation which is national in scope.  An assessment is
the evaluation of a source's pollution based on all available
information on process emissions, discharges, and pollution con-
trol.  The product of a source assessment is a Source Assessment
Document.  The ultimate result is an EPA decision regarding the
need for further study of the source.

This report describes a decision analysis procedure that can be
used as an aid in evaluating the information contained in Source
Assessment Documents.  This procedure, which constitutes an
approximate statistical analysis, employs the concepts of error
propagation, confidence intervals, and hypothesis testing.  The
statistical concepts cannot be applied in a rigorous manner due to
both the complexity of the problems considered and the subjective
nature of environmental decision making.

The general decision analysis principles are illustrated by four
examples in the areas of emissions assessment and control describ-
described in the following subsections.

SOURCE SEVERITY

Source severity is defined as the ratio of two concentrations:
1) concentration of pollutant to which the human population may
be exposed and 2)  concentration of pollutant which represents an
"acceptable" concentration.  The exposure concentration is the
time-averaged maximum ground level pollutant concentration as
determined by Gaussian plume dispersion methodology.  The "accept-
able" concentration is estimated in two ways.  It is the primary
ambient air quality standard for criteria pollutants, or it is a
surrogate primary ambient air quality standard for noncriteria
pollutants.

The values of source severity which IERL uses as a guide for
screening sources are shown in Table 1.  The value of S = 0.05 as

-------
      TABLE  1.  MEASURE/ANALYSIS ASPECTS OF SOURCE SEVERITY
Measure
a'b'c	Analysis
    P
S  E —  >  0.05   There is sufficient cause to consider a source as
                 a candidate for further study.
    P
S  = —  <  0.05   There is not sufficient cause to consider a source
                 as a candidate for further study.


 S = calculated source severity.
 C = concentration to which the population is exposed.

 F = potentially hazardous concentration.

a milepost was obtained by evaluating the uncertainties in sampl-
ing and  analysis results, atmospheric dispersion models, and
health effects data.

NATIONAL EMISSIONS BURDEN

The national emissions burden is the mass of each criteria pol-
lutant emission from a source divided by the national mass of
each criteria pollutant emission.  This index uses engineering
and emissions data for a source to develop an emissions inventory
for that source.  The values which IERL uses as a guide for
screening sources are shown in Table 2.  The value of BN = 0.05
as a milepost was obtained by evaluating the uncertainties in
sampling and analysis data and the uncertainties in national
emission inventories such as those generated by EPA's National
Emissions Data System (NEDS).  In short, if emissions from a
source amount to more than 0.05% of the U.S. total emission rate
for a given criteria pollutant, the source is considered as a
candidate for further study.

OFFSET CALCULATIONS

As part  of EPA's overall program to achieve and maintain national
ambient  air quality standards, the states are required (through
state implementation plans) to determine the amount of emissions
reduction necessary to offset the probable impact of increased
population, industrial activity, motor vehicle traffic, and other
growth factors.  Of particular concern is the policy regarding
increased emissions due to increased industrial activity through
expansions of existing plants.

One way  to offset emissions from a plant expansion is to require
the plant to make a corresponding emissions reduction from its
existing unit.   Such a requirement entails the associated problem
of determining a posteriori whether the plant is complying with

-------
              TABLE 2.  MEASURE/ANALYSIS ASPECTS OF
                        THE NATIONAL EMISSIONS BURDEN
    Measure * *	Analysis
BN = Trrr-(lOO) > 0.05   There is sufficient cause to consider a
       N                 source as a candidate for further study.

     IMp
BN = y-ri—(100) < 0.05   There is not sufficient cause to consider
       N                 a source as a candidate for further
                         study.


 B  = national criteria pollutant emissions burden, percent.

 Mp = mass of criteria pollutant emissions from a source.
 M^ = national mass of criteria pollutant emissions.

the offset policy.  Such a determination is not entirely straight-
forward because the emission rates are not known exactly.  This
uncertainty should be taken into account in the decision analysis
process.

To quantify the uncertainty in the emission rates before and
after plant expansion, error bounds on the following quantities
are required:

   • Emission rates from each emission point in the new unit.

   • Emission rates before and after expansion from each
     emission point in the original unit which has different
     emissions after expansion.

A simple,  approximate statistical test is used to determine
whether the plant may be in compliance with the offset policy.
In order to ensure that air quality is not degraded, the pro-
cedure requires that emissions from the original plant be reduced
by an amount sufficient to compensate not only for the emissions
from added capacity, but also for the uncertainty in emissions
data.

REASONABLY AVAILABLE CONTROL TECHNOLOGY

In order to meet environmental regulations, a plant may be
required to install reasonable available control technology  (RACT)
based on a comparison of two alternative control technologies.
For this analysis, it is assumed that a plant currently employs a
given control technology having a nominal  (measured) control effi-
ciency which is less than that of RACT for a given application.
In order that the plant be required to adopt RACT, it should be
established that RACT is indeed superior to the installed control
technology.  Since neither of the two control efficiencies can be

-------
known with certainty, this determination should not be based on
the nominal efficiencies alone.

A simple, approximate statistical test can be employed to compare
the effectiveness of RACT with that of the installed technology.
The essence of the procedure is to ensure that the incremental
emissions reduction obtained by RACT more than compensates for
the uncertainties in the control efficiencies for the existing
control device and RACT.

-------
                            SECTION 3

               STATISTICAL BASIS FOR THE PROCEDURE
The general procedure set forth in this report is based on the
statistical concepts of confidence intervals and hypothesis test-
ing.  In this initial effort, no attempt has been made to apply
alternative decision-theoretic precepts such as the minimax prin-
ciple or the principle of maximum expected utility (1, 2) .

The general procedure can be summarized as follows:

   • Define a numerical index to be used as an aid in decision
     making.  Source severity, defined in Section 4,  is such
     an index.  The value of the index may depend upon experi-
     mental data, a mathematical model, and/or a semiempirical
     relationship, each of which introduces an element of
     uncertainty into the problem.

   • Perform an error analysis to determine the uncertainty in
     the value of the index due to uncertainties in the data,
     the mathematical model, etc.

   • Based on the results of the error analysis, formulate a
     statistical test of hypothesis as the basis for using the
     index in the decision-making process.

   • Set the values of controllable parameters (e.g., the
     level of the test, the critical test value, the accept-
     able random uncertainty in the data)  to keep the proba-
     bility of making an incorrect evaluation of the  index
     within acceptable limits.

The above procedure may be an iterative one.   For example,  after
completing the above steps, the decision index may be redefined
in order to reduce the risk of incorrect decisions.  In this
manner,  the entire decisionmaking process, from initial program
planning through data collection to the final decision, is
focused toward and guided by the needs of the decision maker.
Specifically, the decisionmaking procedure is based on the con-
cept of controlling the risk of making an incorrect decision.
(1)  Savage, L. J.  The Foundations of Statistics.  John Wiley &
    Sons, Inc., New York, New York, 1954.

(2)  Lindley, D. V.  Making Decisions.  John Wiley & Sons, Ltd.,
    London, 1971.

-------
 At  the heart of  the  above  procedure  is the  second  step—error
 analysis.   This  analysis makes  use of formulas  for the propaga-
 tion  of random and systematic errors presented  in  Appendix B.
 The error  analysis procedure can be  summarized  as  follows:

    •  Use formulas for propagation of random errors to obtain
      the random  component  of uncertainty in the index.

    •  Use formulas for the  propagation of systematic errors to
      obtain the  systematic component of uncertainty in the
      index.

    •  Combine the results of the above two steps to obtain the
      total  uncertainty in  the index; i.e.,  the  interval of
      uncertainty associated with the index.

 The above error  analysis procedure represents an approximate sta-
 tistical technique,  as shown in Appendix B.  The last two steps
 of  the  procedure, involving the test of hypothesis, employ stand-
 ard statistical  procedures.  However, since the test of hypothe-
 sis is  based on  the  results of the error analysis, the entire
 procedure constitutes an approximate statistical technique.

 The nature  of  the above approximation and the circumstances which
 necessitate  it are discussed in the  following paragraphs.  The
 discussion  involves  the calculation  of confidence  intervals using
 the "Student t"  statistic.  This subject is covered in most books
 on  elementary  statistics,  such as Reference 3.

 Let IG  denote  the ("calculated") value of the decision index
 obtained from  experimental data, and let IT denote the ("true")
 value which  would be obtained if there were no  errors in the data,
 the mathematical model employed, etc.  Ic can be regarded as a
 value assumed by a random  variable,  I^/ which is an estimator or
 the parameter  I-p.3   The value, Ic / isHref erred  to  as a point esti-
mate of  IT/  and  the  simplest approach to decision making is to
base the decision on this  value.  However,  in order to take
account of the uncertainty in the data and  thereby control the
 risk of making an incorrect decision, it is desirable to obtain
an  interval  estimate (i.e., a confidence interval) for IT-

The standard statistical approach to this problem  is to determine
the probability distribution of the  random variable IG_, and to
use this distribution to construct a confidence interval for IT
and/or  to set up a statistical test  of hypothesis  concerning IT-
The confidence interval thus obtained or, alternatively,  the out-
come of the  statistical test,  then determines the decision to be
made.
 Underbars denote random variables.
(3)  Freund, J. E.  Mathematical Statistics.  Prentice-Hall, Inc.,
    Englewood Cliffs, New Jersey, 1962.  390 pp.

-------
The above statistical approach is not generally applicable to the
type of problem considered  in this report  for the  following
reasons:

    • The estimator, IQ, is  in general biased due to biases in
     the experimental methods, and the value of the bias  in
     I_C is not known.

    • lฃ_ is generally a complex function of other random vari-
     ables, so that its distribution function cannot be
     determined analytically.

The approach used in this report utilizes  the principles  of error
propagation to obtain an approximate confidence interval  for IT.
Although an approximation,  the resulting confidence interval is
generally conservative, thereby permitting the establishment of
an upper bound on the risk  of an incorrect decision.

To aid in clarifying the relationship between the  approximate pro-
cedure used in this report  and the rigorous statistical approach,
two simple examples are given below.  Both examples involve varia-
tions of the fundamental problem of estimating the mean,  y, of a
normally distributed population having variance, a2, given a
random sample of size n_from that population.  In  the usual proce-
dure, the sample mean, \, and sample standard deviation,  s, are
used to obtain a confidence interval for y:


                            y = x ฑ —                          (1)
                                   /n

where t represents the appropriate value of Student's statistic.

Example A

Suppose now that instead of )(, we use a different  statistic, X'
to estimate the mean, y.  Further, suppose the random variable
ฃ has the same distribution as )(, except that ฃ is biased by an
amount, b.   That is, the expected value of ฃ is given by:

                          E(ฃ) = y - b                         (2)

We would then obtain a confidence interval for y as follows:


                         y  - b = ฃ ฑ —                        (3)
                                     /n


                         y = x + b ฑ —                        (4)
                                     /n


                   X + b - — 
-------
 Suppose,  however,  that  we  do  not  know  the value  of  b, but  that we
 have bounds  on  it  as  follows:
                                                               (6)
 Combining  Inequalities  5  and  6  gives :

 Inequality  7  does not  represent a confidence  interval  in  the
 usual  sense,  because we cannot associate a  specific confidence
 level  with  it.  However,  if  Inequality  6 is strictly true,  then
 Interval  7  contains Interval  5.  Hence, the confidence  level
 associated  with Interval  7 must be at least as great as that
 associated  with Interval  5.   For example, if  the  latter repre-
 sents  a 95% confidence interval, the confidence level  associated
 with the  former must be greater than or equal to  95%.   Further-
 more,  Inequality 7 is  the smallest interval that  could  be con-
 structed  with the given information for which the confidence
 level  is  at least 95%.

 In the approximate procedure  used in this report, formulas  for
 the propagation of systematic errors are used to  obtain bounds on
 the bias  as in Inequality 6.  These bounds are then used  to
 construct an  approximate confidence interval as in Inequality 7.

 Example B

 Consider  next another variation of the standard problem in  which
 it is desired to estimate the sum of means from several normal
 populations based on independent random samples from each of
 these populations.  (For simplicity, we consider  the sum  of two
 means only.)  Thus, we wish to estimate (y^ + y2) based on
 samples of size nj and n2 from two normal populations whose
 variances, aj2 and aj2, are unknown.  In this case, there is no
 generally applicable statistic corresponding to the t  statistic
 which can be used to obtain a confidence interval for y^ +  y2.
 The method reported here uses formulas for the propagation  of
 random errors to generate an approximate confidence interval
which is based on a weighted average of the two individual
 t-values.   The (approximate) interval corresponding to  Interval 1
 is :

                  ฑ t    s               (8)
where                  s = I - + - J                         (9)
                           Vn,    n, /
                               11

-------
                    tave
                      fl = -                    (11)
                           Si2/r\i + s22/n2

                                s22/n2
                      f2 = -                    (12)
                           Sl2/nl + s22/n2

In using the error propagation formulas, Interval 8 is obtained
directly without the necessity of computing either t    or s.
                                                    a, V G

In the general case, the problems illustrated in examples A and B
both occur, and it is necessary to use both random and systematic
error propagation formulas.  The two calculations are made
separately, and the results are then combined to obtain an approx-
imate confidence interval.  This interval can be used directly or,
alternatively, it can be used to set up an approximate statisti-
cal test for decision making.
                                12

-------
                            SECTION 4

                  SOURCE ASSESSMENT METHODOLOGY
DEFINITION OF  INDICES

The  study of air emissions in the Source Assessment Program com-
prises a screening procedure designed to identify the need for
application of or development of new control technology for sta-
tionary sources of air pollution.  Five numerical indices are
used in the program to identify those industries, or source types,
which are considered to be potential candidates for emissions
reduction.

Source Severity

Source severity, S  , is defined as
where Xmax ^s tne maximum time-averaged ground level concentra-
tion of a given pollutant emitted from the source type in ques-
tion, and F represents an "acceptable" concentration of that
pollutant.  The source type is considered to be a potential can-
didate for emissions reduction if S ฃ 1.0.  Since the true
value of severity is unknown, however, the decision must be
based on an estimated, or "calculated," value of source severity,
SG.  In recognition of the uncertainty associated with the esti-
mated severity, the cut points for decision making were original-
ly  (June, 1975) set up as follows:

         If S  > 1.0:  Source type is a potential candidate
                       for emissions reduction.

   If 0.1 < Sc < 1.0:  Source type may be a potential candidate
                       for emissions reduction.

         If S  < 0.1:  Source type is not a potential candidate
                       for emissions reduction.

National Emissions Burden

The national emissions burden is defined as the ratio M /M_
where Mp denotes the total annual mass of emissions of each

                               13

-------
 criteria pollutant from a given source  type,  and  Mn  denotes  the
 total annual mass of emissions  of  that  pollutant  from all  station-
 ary sources nationwide.   When the  national  emissions burden  from
 a specific source type exceeds  0.001  (0.1%),  the  source  type is
 considered to be  a potential  candidate  for  emissions reduction.

 State Emissions Burden

 The state  emissions  burden is defined as the  ratio Mp/Ms where Mp
 now denotes the annual mass of  emissions of each  criteria  pollut-
 ant from a given  source  type within a given state, and Ms  is the
 annual mass of emissions of that pollutant  from all  stationary
 sources  within that  state.  When the state  emissions burden
 exceeds  0.01 (1%),  the source type is considered  to  be a poten-
 tial candidate for emissions  reduction.

 The preceding are the  three dominant screening criteria and  com-
 prise the  basic components of the decisionmaking  process.  Since
 the first  index is directly proportional to the emission rate
 from a given source, a discrepancy arises when comparing a source
 type having a few large  plants  with a source  type having many
 small plants.  The second and third criteria, which  deal with the
 total mass of emissions,  are designed to overcome this problem.

 The two  remaining decision indices are  used to support or  modify
 decisions  based on the first three indices.

 Affected Population

 Affected population  designates  the average  number of persons
 exposed  to high concentrations  (e.g., those for which )(/F>1.0) of
 a given  emission  from  a  given source.   This quantity is useful in
 characterizing emissions  because a given source may  exceed the
 first three  criteria;  yet, if it is located in a  sparsely  popu-
 lated area,  it may have  a  relatively small  impact on human health
 compared with a source located  in a densely populated area.   In
 addition,  a  source may have a large value of  source  severity due
 to  a  low emission  height.  Again, its impact  on human health may
 be  small because  the low  emission height results  in  pollutants
 being dispersed over a very small area  in the immediate vicinity
 of  the source.  The  calculation of affected population is
 described  in  Appendix C.

 Growth Factor

 The growth  factor  is determined from the ratio of known to pro-
 jected emissions  from a source type.   For example,

           Growth  factor = Projected emissions in 1978
                                Emissions in  1973             v   '

Other  5-yr periods  (e.g.,  1975 and 1980) could also  be used
depending on  available data.  The main  purpose of this index is

                                14

-------
 to  eliminate  from consideration those sources whose emissions are
 expected  to decrease greatly in the near future; e.g., due to the
 implementation of new emission controls or to a process being
 phased out of use.

 In  the subsequent analysis of uncertainty, the discussion centers
 mainly on the first decision index, source severity.  The nation-
 al  emissions burden is also considered in more detail in Section 6.

 EXPANDED  DISCUSSION OF SOURCE SEVERITY

 Basic Definition

 As  noted  previously in Equation 13, the source severity, SC' fฐr a
 given pollutant emitted from a given source is defined as
                            C - -f

where Xmax equals the maximum time-averaged ground level concen-
tration of pollutant which would result from the source emitting
into a standard receiving atmosphere, and where F equals an
"acceptable" concentration.  The standard receiving atmosphere is
defined as one having Class C stability and a mean wind speed of
4.5 m/s.  The averaging time for the mean concentration is the
same as that used in the specification of F.  For the Source
Assessment Program, the "acceptable" concentration is specified
as follows:

              F = PAAQS  for criteria pollutants

                = (TLV)G for noncriteria pollutants (4)

where G is a factor which converts the TLV into an "equivalent"
PAAQS.  The averaging time used in the PAAQS for each criteria
pollutant is shown in Table C-4 of Appendix C.  The averaging
time used for noncriteria pollutants is 24 hr.

Appendix E develops a correlation between the Primary Ambient
Air Quality Standard for criteria pollutants and their corre-
sponding TLV's.  The analysis indicates that  "G" for criteria
pollutant TLV's ranges from 0.0260 to 0.158 with an average value
of 0.0467.
 PAAQS represents primary ambient air quality standard.
 (4) TLVsฎ Threshold Limit Values for Chemical Substances and
    Physical Agents in the Workroom Environment with  Intended
    Changes for 1976.  American Conference of Governmental
    Industrial Hygienists, Cincinnati, Ohio, 1976.  94 pp.
                               15

-------
Calculation of Source Severity

Ideally, SQ would be determined by placing the source in the
standard receiving atmosphere and measuring the resulting maximum
ground level concentration.  If the measurement could be made
with no error, the measured value would yield the true source
severity.  Obviously, such a procedure is impossible in practice.
Therefore, S^ must be estimated by measuring the source emission
rate and by using dispersion modeling to relate emission rate to
Ymax-  In the Source Assessment Program, the simple Gaussian
plume equation is used to calculate source severity as follows
(5, 6) :


              S  =  2 Q   /3\ฐ-17 = 2 CAP EF/3\ฐ-17           (15)

               C   TreuFH2\t/        ueuFH2  \t/

where     Q = mass emission rate, g/s
          H = effective emission height, m
          u = average wind speed (4.5 m/s)
          e = natural base logarithm = 2.72
        CAP = production capacity,  kg/s
         EF = emission factor; i.e.,  mass of emissions generated
              per unit of product produced, g/kg
          t = averaging time for mean concentration, min

The concentration obtained from the Gaussian plume equation cor-
responds to an averaging time of approximately 3 min (5).  The
factor (3/t)ฐ-17 in Equation 15 corrects this value for averaging
times between 3 min and 24 hr (5, 7).  Equation 15 is used for
all pollutants with the exception of nitrogen oxide (NOX), for
which the primary standard averaging time is 1 yr.  Since the
above correction factor is not valid for averaging times of this
magnitude, a modified approach, which is described in Appendix C
is necessitated.
(5) Turner, D. B.  Workbook of Atmospheric Dispersion Estimates.
    Public Health Service Publication No. 999-AP-26, U.S. Depart-
    ment of Health, Education, and Welfare, Cincinnati, Ohio,
    1969.  84 pp.

(6) Pasquill, F.  Atmospheric Diffusion, Second Edition.  John
    Wiley & Sons, Inc. (Halsted Press), New York, New York, 1974,
    429 pp.

(7) Cheremisinoff,  P. N., and A. C. Morresi.  Predicting Trans-
    port and Dispersion of Air Pollutants from Stacks.  Pollu-
    tion Engineering, 9 (3): 3, 26, 1977.
                               16

-------
The conversion factor, G, for converting TLV to an equivalent
PAAQS is specified as follows:


                   G = 300" = 2? 100" = ฐ-0033

Theoretically, the factor 8/24 adjusts the TLV from an 8-hr work
day to continuous  (24-hr) exposure, and the factor of 1/100 is
designed to account for the fact that the general population
constitutes a higher risk group than healthy workers.  Alterna-
tive methods of estimating "acceptable" concentrations for non-
criteria pollutants are considered in Appendix E.

REPRESENTATIVE SOURCE

In assessing the potential pollution problems associated with an
industry or source type, it is necessary to take into account
plant-to-plant variations in physical parameters such as stack
height, production capacity, or process technology.  In the
Source Assessment Program, this problem is handled in two ways:
1) through the concept of a representative source and 2) through
the use of source severity distributions.  The representative
source concept is discussed in this subsection; source severity
distributions are discussed in the following subsection.

The concept of a representative source constitutes a simplified
approach to plant-to-plant variability in which an industry or
source type is characterized by means of a "typical" (or repre-
sentative)  source or sources.  In general, the representative
source is a hypothetical plant having physical parameters which
are typical or average values for the source type in question.
The methodology employed in defining a representative source is
described in the following discussion (8).

Simple Case

The representative source, for a simple case, can be defined as
one having an average plant capacity, average stack height, aver-
age county population density, and an average emission factor.

The following definitions can be used in calculating required
average parameters.

   • Average plant capacity is the total capacity of the
     industry divided by the number of plants.

   • Average stack height is the sum of all known stack heights
     for that industry divided by that number of stacks.
(8) Hughes, T. W., R. B. Reznik, R. W. Serth, and Z. S. Kahn.
    Source Assessment Methodology.  Contract 68-02-1874, U.S.
    Environmental Protection Agency, Research Triangle Park,
    North Carolina, January 7, 1976.  33 pp.
                              17

-------
   • Average county population density is the sum of the
     population densities of all counties within which
     plants are located divided by the number of sites.

   • Average emission factor is the sum of the known emis-
     sion factors divided by the number of factors known.

These definitions are valid when there is no correlation between
plant size, emission factor, stack height, and population density
and when emission factor data are taken from a random sample of
the population.  One simple way to detect possible correlations
is to graphically depict the variables in question; e.g., plant
size versus population density.

Special Cases

Small Number (Fewer than 10) of Sources—
When a small number (fewer than 10) of sources comprises the
source type being studies, a representative source is defined as
the largest source within the group.  It may also be desirable to
treat each particular source individually.

Large Number of Sources—
For source types with a large number of sites, it is impractical
to determine all of the average parameters.  This problem can be
handled in several ways depending on the data base available.
Some possibilities include

   • Consideration of 30 plants at random.  If the parameters
     are normally distributed, this is sufficient to find an
     average.  Random election of 30 plants will give fairly
     accurate estimates of the industry's mean parameters while
     describing the variation within the industry.

   • Use of a simulation technique.

   • Use of state average population density.

Nonhomogeneous Sources--
When qualitative differences in subgroups exist, a representative
source is defined by subdividing into homogeneous subgroups and
considering each group separately.

Log-Normal Distributions—
In some sources, the plant capacity is log-normally distributed.
Most of the plants are small, but the few large plants account
for a major portion of the production.  Under these circum-
stances, it would be better to subdivide into size groups or to
treat only the larger plants as a worst case condition.
                               18

-------
Correlation of Averaging Parameters—
Sources which exhibit a correlation between the averaging param-
eters  (emission factor, plant size, stack height, population den-
sity) should be treated on a case-by-case basis.  If the emission
factor or population density depends on plant size, a weighted
average should be calculated.  For other situations, calculating
the distribution of severities using a simulation technique may
be appropriate.

SOURCE SEVERITY DISTRIBUTIONS

A source severity distribution depicts the distribution of source
severities among individual plants within an industry.  It thus
provides more detailed information about potential air pollution
problems associated with a given industry than does the repre-
sentative source severity.  When sufficient information is avail-
able to compute a source severity for each plant in an industry,
a deterministic source severity distribution can be constructed
for the industry.  An example of a deterministic severity distri-
bution is shown in Figure 1  (9).  When such detailed information
is not available, a simulation technique can be employed to
generate an approximate (simulated) source severity distribution
based on data from a sample of plants.  The methodology used
to simulate source severity distributions is described in
Reference 10.

STOCHASTIC APPROACH TO SOURCE SEVERITY

In the Source Assessment Program, a deterministic approach to
calculating source severity has been adopted by employing the con-
cept of a fixed receiving atmosphere.  The main advantage of this
approach is its simplicity:  It permits a wide variety of source
types to be analyzed and compared on a consistent basis with a
minimum of experimental and computational effort.

An alternative approach is to allow meteorological parameters
(i.e., stability class or wind speed) to vary stochastically.
This approach generates a frequency distribution of source sever-
ities rather than a single value, as is obtained in the
 (9) Serth, R. W.,  and T. W. Hughes.  Source Assessment:  Carbon
     Black Manufacture.  Contract 68-02-1874, U.S. Environmental
     Protection Agency, Research Triangle Park, North Carolina.
     (Preliminary document submitted to the EPA by Monsanto
     Research Corporation, December 1975.)  145 pp.

(10) Eimutis, E. C., B. J. Holmes, and L. B. Mote.  Source Assess-
     ment:  Severity of Stationary Air Pollution Souces--A Simu-
     lation Approach.  EPA-600/2-76-032e, U.S. Environmental
     Protection Agency, Research Triangle Park, North Carolina,
     July 1976.  119 pp.

                               19

-------
             10.0
             9.0
             8.0
             7.0
             6.0
             5.0

             4.0

             3.0
             2.0
             1.0
             0.9
             0.8
             0.7
             0.6

             0.5

             0.4
          s
          l/l
             0.1
            0.09
            0.08

            0.06
            0.05

            0.04

            0.03


            0.02
            0.01
            0.009
            0.008
            0.007
            0.006
            0.005

            0.004

            0.003
            0.002
           0.001
CO BOILER AND THERMAL INCINERATOR
                i  i  i
                                ii

                0.01 C.05'0.1 0.2  0.5 1  2    5      20  30  40 50 60 70  BO  90  95   98  "   W-6

                     PERCENT OF PLANTS HAVING SOURCE SEVERITY LESS THAN OR EQUAL TO OROINATE


Figure  1.   Deterministic  source  severity  distribution  for carbon

               monoxide emissions  from  carbon black  plants   (9).
                                          20

-------
deterministic approach.  For making decisions concerning the need
for emissions reduction, a value of source severity corresponding
to a relatively high  (e.g., 99 or 99.9) percentage point of this
distribution would be appropriate.  For example, for severities
based on 24-hr mean concentrations, values higher than the 99%
point would be expected to occur about four times per year.  Thus,
a low value of severity at the 99% point would indicate a low risk
of potentially hazardous conditions arising from the source in
question.  The same would not be true of a low value of the mean
severity, for example, since severities much higher than the mean
might occur many times over the course of a year.

The stochastic approach to the calculation of source severity is
investigated in Appendix H.  A Monte Carlo simulation and proba-
bilistic sensitivity analysis are carried out in which wind speed,
stability class, lateral and vertical dispersion coefficients,
peak-to-mean concentration ratio-, etc., are allowed to vary sto-
chastically.  The output from the simulation is a probability dis-
tribution of values of source severity for beryllium emissions
from a coal-fired powerplant.  In this example, the upper 95%
point of the severity distribution is approximately 1.9, and the
upper 99% point is approximately 3.0.

By comparison, the deterministic value of severity (corrected for
plume rise)  for the above example is 3.3.  This value corresponds
roughly to the 99% point of the stochastic severity distribution.
Thus, on the basis of this example, it appears that the determin-
istic approach is consistent with the stochastic approach.

The simulation described in Appendix H is based on the general-
ized Gaussian dispersion equation (Equation H-7).  Hence, the
results are restricted to meteorological conditions under which
the Gaussian model is valid.  A complete solution to the problem
would have to include dispersion models applicable to other mete-
orological conditions, such as fumigation (inversion breakup) and
trapping, and the frequency with which the various meteorological
conditions occur.

Although ground level concentrations under trapping and fumiga-
tion conditions can be considerably higher than those predicted
by the Gaussian model (coning conditions), these conditions gener-
ally persist for relatively short periods (30 min to 4 hr).
Therefore, these conditions may have a relatively small effect on
24-hr mean concentrations, which are of primary interest in rela-
tion to source severity.  The above reasoning provides justifica-
tion for the exclusive use of the Gaussian model as an approxima-
tion in the simulation procedure.

The example given in Appendix H also indicates that the determin-
istic approach to source severity yields a worst case value in
the sense that higher values would be expected to occur only
infrequently, on the order of once per year.

                               21

-------
                            SECTION 5

               SOURCE ASSESSMENT:  SOURCE SEVERITY
INTRODUCTION

In this section, the general procedure outlined in Section 3 is
applied to the source severity decision index.  It was noted in
the previous section that the true value of source severity is
always unknown and that decisions must be based on an estimated,
or "calculated," value of severity.  The uncertainty associated
with the estimated value of severity is related to uncertainties
in individual parameters (i.e., production capacity or emission
factor) used to calculate severity.  In the following subsections,
a mathematical expression for this relationship is obtained and
used to establish overall guidelines for decision making in the
Source Assessment Program.

Only the source severity of a representative source will be con-
sidered.  In addition, it will be assumed that the source type in
question is homogeneous with respect to emission factor, so that
the representative source emission factor can be obtained by sam-
pling at a single plant.9  This assumption may appear to pose a
severe restriction on the applicability of the analysis; in prac-
tice, however, economic constraints often dictate that sampling
be conducted at a single plant.  Use of the measured emission
factor to calculate the severity of the representative source
then involves the implicit assumption of homogeneity with respect
to emission factor.

For the purpose of this analysis, it is assumed that the true
source severity, S , is given by the following equation:

                                        0.17
where y is a correction factor for the simple Gaussian plume
equation.  Since we are dealing with the representative source,
CAP and H now represent arithmetic mean production capacity and
effective emission height, respectively, for the source type in
 This assumption is made to simplify the analysis and does not
 represent an inherent limitation of the general procedure. .

                               22

-------
question.  The emission factor, EF, in Equation 16 is the emis-
sion  factor for the plant at which sampling was performed.

The following discussion will analyze Equation 16 by defining
the uncertainties contained in each of the variables, by showing
how the uncertainties are propagated, by developing alternate
decision tests  (tests of hypothesis), by quantifying and compar-
ing the alternate decision tests, and by developing guidelines
for the Source Assessment program.  Two alternate decision tests
will  be developed analytically - a deterministic approach and a
stochastic approach.  The deterministic approach assumes that
the true value of source severity is a fixed parameter and
treats large uncertainties in Equation 16 as systematic errors.
The stochastic approach assumes that the true value of source
severity is a random parameter (i.e., changes daily) and treats
all uncertainties as random errors.

UNCERTAINTIES IN PARAMETERS

The uncertainties associated with each of the individual param-
eters appearing in Equation 16 are discussed below.  The concepts
of random and systematic uncertainties are discussed in Appendices
A and B.

CAP (Average Production Capacity)

If CAP is determined from a sample of plants in the industry or
source type, there is a random sampling uncertainty which shall
be equated with the 95% confidence interval about the sample mean
plant capacity.  If CAP is determined from a knowledge of all
plants in the industry, then this uncertainty is zero.

H (Average Effective Emission Height)

The average effective emission height, H, is comprised of two
components, H = h+AH, where h is the physical stack height and
AH is the plume rise.  A statement similar to the one made for
CAP applies to h as well.   In addition, there is uncertainty in
estimating plume rise, AH.  The error in plume rise can be
assumed to be a randomly distributed uncertainty over the popu-
lation of all sources and atmospheric conditions, provided AH is
indeed included in H.

In the development of the deterministic decision test later,
source severity is computed for a fixed source and a fixed re-
ceiving atmosphere.    Thus,  the estimated severity, S_, can be
                                                     \*,
 The receiving atmosphere was defined by specifying stability
 class and wind speed.  These two parameters do not suffice to
 define a unique state of the atmosphere.  However, it may be
 assumed that the additional meteorological parameters necessary
 to define a unique atmospheric state are maintained at fixed,
 albeit unspecified, values.

                               23

-------
considered as a sample value from a population of severities in
which both source type and receiving atmosphere are fixed.  In
this use, the error in plume rise is treated as a bias in Sr since
the error is amplified by squaring H in Equation 16.  In the
development of the stochastic decision test developed later,
source severity is computed for a fixed source and a variable
receiving atmosphere.  In this use, the error in plume rise
represents a random uncertainty in S_,.

EF  (Emission Factor)
There is an uncertainty in EF which is composed of process varia-
tion, sampling uncertainty, analysis uncertainty, and uncertainty
in production rate during sampling.  The total uncertainty has
both a random and a systematic component, the latter resulting
from possible bias in the sampling and analytical procedures.
The random component is given by:
                                 a
                   ซ_ 9 	 i 9  .  i_9   . i_9   , i_ 9
                   D  = D   T D   T D   ~r D


where   b  = random uncertainty in emission factor, g/kg

       b   = uncertainty in emission factor due to imprecision
             in sampling, g/kg

       b   = uncertainty in emission factor due to imprecision
             in analysis, g/kg

       b   = uncertainty in emission factor due to process
             variation, g/kg

       b   = uncertainty in emission factor due to imprecision
             in measurement of production rate, g/kg

Y  (Correction Factor)
There is an uncertainty in the correction factor, y/ which is a
reflection of the uncertainty inherent in the Gaussian dispersion
equation itself.  For the purpose of this analysis, values of y
are assumed to be log-normally distributed over source types and
atmospheric conditions with mean of 1.0 and a variance of 0.264
(standard deviation = 0.514).  The 2.5% and 97.5% points of the
distribution are taken to be 1/m and m (where m = 3).

In the deterministic decision test developed later, the uncer-
tainty in Y is treated as a systematic uncertainty with m = 3.
In the stochastic decision test, the uncertainty in Y is treated
as a randam uncertainty with a variance of 0.264.
 If the source type is not assumed to be homogeneous with respect
 to emission factor, then the random uncertainty contains an addi-
 tional term due to plant-to-plant variability of emission  factor,
 In general, the plant-to-plant distribution of emission factors
 is not a normal distribution, which further complicates the
 analysis.

                               24

-------
 (t /t)ฐป17  (Averaging Time Correction Factor)

For the purpose of the present analysis, the uncertainty associ-
ated with the averaging time correction factor  (t /t)ฐ-^7 is
assumed to be included in the uncertainty associated with y•
This assumption is discussed further in the following section.

F (Acceptable Pollutant Concentration)

There may be an uncertainty associated with the "acceptable"
pollutant concentration, F, depending upon interpretation.  It
can be argued that there is no pollutant concentration, however
small, which is entirely innocuous.  Therefore, specification of
an acceptable concentration represents a subjective judgment on
the part of the decision maker.  Hence, the "true" value of F may
be regarded as whatever value the decision maker specifies.  In
this case then, there is no uncertainty in F since the "true"
value equals the "calculated"  (specified) value.

Alternatively, it can be argued that the PAAQS represents a pol-
lutant concentration at which no adverse effects on human health
can be detected.  Therefore, the "true" value of F can be equated
with the PAAQS.  In this case, there is no uncertainty in F for
criteria pollutants.  For noncriteria pollutants, however, there
is uncertainty associated with the conversion factor, G, which
converts the TLV into an equivalent PAAQS.  This uncertainty can
be assumed to be randomly distributed over the population of all
pollutants.  However, since source severity is computed for a
given pollutant, this uncertainty should be treated in the analy-
sis as a systematic uncertainty.

DETERMINISTIC DECISION APPROACH

In developing a deterministic decision test it is assumed that
the true value of source severity is a fixed parameter.  The test
is developed for a fixed source in a fixed atmosphere.  It treats
small uncertainties as random errors and large uncertainties as
systematic errors.  The approach employs the development of formu-
lae for the propagation of errors.

Propagation of Random Uncertainties

The total uncertainty in source severity is determined by the
propagation of uncertainties in CAP, H, EF, y, and F through
Equation 16.  For convenience, let A, B, C, and D represent
estimated or measured (as opposed to true) values as follows:

                             A = CAP
                             B = EF
                             C = H
                             D = F
                               25

-------
The  "measured" value of y is taken as unity.  Furthermore,  let
                   k = -Mir1)     = constant                 (18)
                       ireuyt /

Considering the random uncertainty first,
                  s
                   C
where ar, br, cr, and er are the random uncertainties  in A, B, C,
and SQ, respectively, and Sc is the nominal, or  "calculated,"
source severity; i.e.,
                            S_ =    .                          (20)
                             C   DC2

Thus, if we are working at the 95% confidence level,  (A  ฑ ar)
represents a 95% confidence interval for the industry mean produc-
tion capacity, CAP.
Now let
                            ar = arA
                               = brB
                               - crC
                            e  = e S_
                             r    r C
i.e., ar, br, cr, and er are the relative random uncertainties
associated with A, B, C, and Sc-  Using the error propagation
formula for multiplication  (see Appendix B),
(A  ฑ  aWB  ฑ
                            = AB ฑ
                            = AB ฑ WB2A2ar2 + A2B2br2
                            = AB ฑ AB  ar2 + br2              (21)


Using the error propagation formula for exponentation,

                 i cj 2 = C2 ฑ 2 Ccr = C2 ฑ 2 C2cr            (22)
                                26

-------
Using the division  formula,
AB ฑ AB Ja 2 + b 2   fln   1fc*B2(ฃ 2 + ฃ2)
        * JiT     JL  _ t\D  ,  ป  •  ^ i

                         "
                                                          c 2
         C2  ฑ 2  C28        C2           C1*               C8
                        =  AB  ฑ  AB /j 2  + ฃ 2  + 4 ฃ 2
                          c2    c2 T  r     r       r
Thus,
              S   ฑ  e   =      ฑ       a 2  + ฃ 2  + 4  ฃ2          (24)
              C    r    C2D    C2D f r     r       r
Therefore.

and
                	 = e  = Ja  2  +  b  2  +  4  3  2             (26)
                kAB/C2D    r   ' r     r      r

Propagation of Systematic Uncertainties

In general, the systematic components  of uncertainty  are  unsymmet-
rical.  Therefore, let

                  +b     +c     +d           +e
                 B_bSU, C_CSU, D_dSU,  and  S-  SU
                    C 0     C 0     O 0         	n

be systematic error bounds on EF, H, F,  and  S-p,  and let bsu,  bsฃ,
etc., denote the corresponding relative  uncertainties.  The upper
and lower bounds on y are taken as m and 1/m,  respectively.
Since all variables in Equation 16 are non-negative,  the  upper
bound on source severity can be obtained by  substituting  upper
bounds in the numerator and lower bounds in  the  denominator.
Thus,

                                   m(B
                                (E - O
               q  j. Q   =  flrA\	i	  '	             O~l\
               O —. T C   —  I JX ฃ\ I       	                  I Z / )
                C    SU        '       "  '       ^  -                '
Rearranging this equation yields
                               27

-------
                   su
                                             -  1
                                                   (28)
The lower bound is obtained in a similar manner.
                           m l
                                                              (29)
Uncertainty Interval for Source Severity

The total uncertainty associated with source severity  is  the  sum
of the random and systematic components  (see Appendix  B).   Symbol
ically,

                          A    A    A
                          e  = e  + e
                           u    r    su

                          s\    /\    s\
                          e* = er + esฃ
Using Equations 26, 28, and 29 for er, esu, and eSฃ yields:


                                                     • - 1
         •     /A ',   T
         u = Var  +
                                  ฃsu)
        -V
a_> + b..2 + 4 cj- + 1 -
                                                    su
                                                   (30)
                                                              (31)
The uncertainty interval for source severity is then given by  the
following inequality:

                                                              (32)
STOCHASTIC DECISION APPROACH

In developing the stochastic decision test approach the true value
of source severity is assumed to be a random parameter.  The test
is developed for a fixed source in a random receiving atmosphere.
It treats all uncertainties as random errors.  The approach differs
from the deterministic decision approach in that the deterministic
approach fixes the receiving atmosphere.  The approach also differs
from the one presented in Appendix H in that the latter uses vari-
able emission sources as well as a random receiving atmosphere.
                               28

-------
Uncertainties in the Stochastic Decision Approach

In the deterministic decision approach, S   is  treated as  a fixed
parameter by treating y as a constant.  It  is  intuitively reason-
able to attempt to describe ST in a  stochastic manner since
meteorological change day to day.  The calculated  source  severity
is:
                                       /t  \ ฐ • l 7
                       _  _ 2 (CAP) (EF) To]                     {   .
                       SC      eyFH'   \~]                     (33)

which may be rewritten as.

                        q  _ v  (CAP) (EF)
                        sc - K -Z -                       (34)
where K is a constant.  Recalling Equation  16, we  have

                                        /t \ ฐ . * 7
                          _  2 (CAP) (EF)  ^o)
                       ST - Y  eyFH^ - (—/

which may be rewritten as

                                 (CAP) (EF)                      ,
                        ST ~ YK       ^                         (35)
Based on the discussions presented earlier,  ST can be  treated as  a
log-normally distributed random parameter.   S_ is the  product of  a
constant and a log-normally distributed variable  (y) •   Treatment
of the above equations is best accomplished  by taking  logarithms.


    In S_ = In K = In  (CAP) + In  (EF) - In y - In F  -  2 In  H
        \—

The expected value of Sc (E(log S_) ) is
                   E(log Sr) = log
.(CAPHEF! .                ,
The variance of log S_  (V(log S )) is
                     V-         V,
                Vdog Sc) = i
    S
Uncertainty Interval for Source Severity
The uncertainty interval for ST is derived as  follows:
      In Sc - ZaVVdog SG) < log ST < In SG +  Zaป/V(log  SG)     (38)
                                29

-------
 TESTS OF HYPOTHESIS

 The above discussion leads  to  the  development  of  two  approaches  to
 hypothesis testing for source  severity.   The deterministic  decision
 approach treats  S  as a fixed  parameter  and tests whether S  > 1.0.
 The stochastic approach treats S_  as  a random  parameter; it tests
 the number of days for which S_ 2  1 againsts an  "acceptable"  number
 of such occurances.   The following discussion  developes  the hypo-
 thesis testing procedure for both  decision approaches.

 Hypothesis Testing for Deterministic  Decision  Approach

 Inequality 32 can  be used to set up a formal test of  hypothesis.
 The hypothesis that  we wish to test  (the null  hypothesis, H0) is
 that the true severity is greater  than or equal  to  unity; i.e.,

                          H0:   ST  > 1.0


 This hypothesis  is tested against  the alternative hypothesis, HA>

                          HA:   ST  < 1.0


The  statistic which  is used  to make the  test is Sc-   If SQ
exceeds  a  specified  critical value, Sฃ,  H0 is  accepted and  it is
concluded  that ST  >  1.0;  if  SQ is  less than Sฃ, H0  is rejected
and  it is  concluded  that  ST  <  1.0.

The  basic  idea behind the test is  that it should  be concluded
that ST  >  1.0 whenever the upper end  point of  the uncertainty
interval  (Interval 32) exceeds unity.   Thus, the  critical value,
Sฃ,  is given by

                        S*l + e   =  1.0


                           S* = -~-                        (39)
                                1  + eu

With any such statistical test, two types of errors can occur.
These errors, usually designated Type I  and Type  II,  are defined
as follows:

Type I Error;  Reject H0 when it is,   in  fact,   true.   That is, we
               conclude that Sm <  1.0 when it  is, in  fact,
               greater than unity.

Type II Error; Accept Hg when it is,   in  fact,   false.  That  is,
               we conclude that ST >   1.0 when  it  is,  in fact,
               less than unity.
                                 30

-------
These possible errors can  be represented schematically as
follows:
                        HO  is true
                                           HO is false
        Accept HO

        Reject HO
Correct decision
Type I error
Type II error
Correct decision
Within the context of the  Source Assessment Program,  the effects
of the two types of incorrect decisions are as follows:   a
Type I error may results in a needed emissions reduction not
being accomplished; a Type II eror may result in an emissions
reduction which is not really needed.

In order to see how the risk of an incorrect decision can be
controlled, it is convenient to use Inequality 32:  as repeated
here.
                                                           (32)
Now assuming that 95% confidence  intervals  (i.e., A ฑ ar, etc.)
were used to construct the approximate  Interval  (32), and further
assuming that the confidence level  associated with Interval 32 is
greater than or equal to the nominal level, the probability that
the true value of ST will fall  in Interval 32 is greater than or
equal to 95%.  In other words,  the  probability of obtaining a
value of ST which falls on either side  of Interval 32 is less
than or equal to 2.5%.

Thus,
                      •fe,  < —•
                        1.0.
Then we have                                         T  ~
      K <  rh; s> HHS< < r?r \-^
                                                      2.5%
Using Equation 39, we obtain:


                   P|SC < Sc'ST - 1'
                                        2'5%
(41)
                             31

-------
But Sc < 5$ is precisely the condition for rejecting H0.  There-
fore, the probability of rejecting H0 when it is true, i.e., of
making a Type I error, is less than or equal to 2.5%.  In general,
if we work at the (1 - 2a)(100%)  confidence level in constructing
the uncertainty interval (32) , then the probability of making a
Type I error will be less than or equal to 100 1/d - eฃ) ,  the probabil
ity of a Type II error will be less^than 2.5%.  However, H0 is
accepted whenever Sc > Sฃ = 1/(1 + eu).   Thus, for values of Sc
in the interval
                                                              (42)
there will be a  (relatively) large chance of a Type II error.3
Therefore, in order to minimize the number of Type II errors in
repeated applications of the test procedure, the Interval 36
should be made as small as possible; i.e., the smaller the un-
certainty in source severity, the fewer the number of incorrect
decisions that are expected to be made.

Within the context of the Source Assessment Program, the width of
the Interval 36 can be controlled primarily through the random
uncertainty in emission factor, br.  This uncertainty can (theo-
retically) be made as small as desired by analyzing a sufficient-
ly large number of samples.  The other uncertainties which
contribute to Interval 36 cannot readily be controlled in this
program.

In order to Specify an allowable value for b , a width must be
specified for Interval 36.  Although this laiter specification is
arbitrary, an upper bound on br can be obtained by setting the
upper limit in Interval 36 equal to infinity.  Roughly speaking,
 An alternative procedure isAto reject H0 when  G
                                              Sc < 1/(1 + eu) ,
accept HO when S  > 1/(1 - eป) ,  and make no decision when SG
falls in the Interval 36.  This  procedure is considered in more
detail below.  For the present,  it is assumed that a decision
(accept or reject H0) is always  required.
                               32

-------
this condition means that we can never be sure that S  > 0, no
matter how large the calculated severity, SQ, may be.  This can
be seen by setting  (1 - ep) = 0 in Interval 32.  Stated another
way, when this condition prevails, there will always be a large
probability of a Type II error, regardless of how large the value
of SQ may be.  From Interval 36, this upper bound on 6r is found
by solving the equation

                            ej, = 1.0                          (43)

Hypothesis Testing for Stochastic Decision Approach

The hypothesis that we wish to test is not whether S  > 1.0 but
whether the number of times per year S  > 1.0 is more than some
acceptable number of times; i.e., Ho: ST > 1 on more than r days
per year.  This hypothesis is tested against the alternative
hypothesis, HA,

             H :  S  > on less than r days per year
              ฃ\    J.


Obviously this hypothesis would be most relevant when the "measure
of harmful effect" is proportional to the number of days of
threshhold exceedance.  If the harm done depends in a more compli-
cated way on S_, for example on the actual amounts by which S_
exceeds 1.0, tnen other hypotheses may be more appropriate.

The hypothesis H may be rephrased as:

     Ho:  The expected number of days on which S^ > 1, exceeds r.

or
     H0:  E(Z) > r

where E denotes the mean (expected value), but, E(Z) = 365 P\S >l|
so that                                                     l 1  '

     H0:

where n = r/365.

The hypothesis becomes,

     Ho:
     H0:  p|lny>lnS >>n

The hypothesis becomes,

     Ho:  E(ln Sc)>V(lnY)Z-
                               33

-------
where V(lny)  is  the  variance of log y (V(lny)  = 0.3136 from
Appendix  I)  and  Z— is  the  n-percentage point of the standard
normal  distribution.   The  level a test for H0  is to reject H0 if
In  S>, < In  S* where
                  Pjln  Sc  <  In  S*|W(ln Y)Z- = aj
so that
or
                        in  S*  -  fV(ln Y)Z-= z                (44)
                            V(ln Sc)a
          log  S* =  (Z-)[v(log  Y)]*5 +  (Za)  [v(log SG)] h      (45)
The critical value of  S-  for  an  a-level  test is


            S* = e(ZE)tv(log  Y)^ +  (Za}  CV(lo9 SC^h       (46)


As an example, assume  that  r  = 1 day  per year (current basis for
several of the criteria pollutants),  and that a = 0.05 is desired,
We have :

     r = 1
     n = 1/365 = 0.00274
    Z- = -2.786
     n
    Z  = Zo .05 = -1.645

Hence, we would reject Ho if,
            c       ._..,„,. x^y  Y)  +  1.645/V(log S_))         .._.
            Of-, *- C  \                               *— /         \^ I I

Remembering that

                        V(log  Y)  =  0.3134

or

                        [V(log  v)]35  =  0.560

we would reject H0 if
                   < e-/1.555 +  1.645  V{log
                              (48)


34

-------
COMPARISON OF ALTERNATE DECISION APPROACHES

The above approaches to hypothesis testing utilize different start-
ing assumptions  (S_ is a "fixed" versus a "random" parameter) and
they treat uncertainties in different manners  (large uncertainties
treated as systematic errors versus all uncertainties treated as
random errors).  Yet, the critical values of source severity for
decision making  are quite similar.  This is best illustrated in
Figure 2 where S* is plotted as a function of  the alternative
decision appraches and uncertainty in emission factor.  The most
frequently occuring range of b   (emission factor uncertainties)
is 0.6 to 1.0.   The bias (b  )rin available sampling and analysis
technologies ranges from 0.6uto 0.5.  The values S* vary from 0.5
to 0.8 for the general cases displayed in Figure 2 regardless of
the decision approach used.

The following observations of Figure 2 are readily apparent.
Given the current (1979) capabilities in measurement technologies,
both decision approaches give the same S* values.  The determin-
istic decision approach is insensitive to improvements in measure-
ment technologies.  The stochastic decision approach is quite
sensitive to measurement technology improvements.

Figure 3 shows how S* changes as r increases for the stochastic
decision approach.

GUIDELINES FOR SOURCE ASSESSMENT PROGRAM

The above test of hypothesis can be applied to the decisionmaking
process by computing the critical severity, Sฃ, for each pollut-
ant emitted from each source type; i.e., the test can be perform-
ed for each individual pollutant and each source type studied.
An example of this approach is given in Appendix J.  Alternative-
ly, the test procedure can be used to derive general guidelines
for decision making in the Source Assessment Program.  The latter
approach is followed in this section.

In order to establish overall guidelines for the Source Assessment
Program, generalized error bounds have been estimated as shown in
Table 3.  These estimates are derived in Appendix I.  The values
listed for^ar and cr represent 95% confidence  limits.  In this
analysis,  br is considered to be an independent variable.  The
values given for br in Table 3 indicate the range in which br can
be expected to fall in most cases.  Three sets of values are
listed for the uncertainty in "acceptable" concentration, F.  In
Case A, there is no uncertainty associated with F (cf. discussion
on page 30).   In Case B\, the "acceptable" concentration is
considered to be uncertain, and the TLV conversion factor, G, is
equal to the mean value (0.047)  for criteria pollutants  (see
Appendices E and I).  Case B2 is the same as B\ except that
G = 0.0033,  the value currently used in the Source Assessment
Program.
                                35

-------
o
QC
O
in
 -0.2
^x     ASSUMING ST IS PERMITTED TO
   xv   EXCEED 1.0 ONLY 1 DAY PER YEAR
                                       ST IS A "FIXED PARAMETER"
                                	ST IS A "RANDOM PARAMETER"
                                       MOST FREQUENTLY OCCURING RANGE
                                     OF br ON SOURCE ASSESSMENT PROGRAM
              ,7
                       OBSERVED RANGE OF BIAS IN AVAILABLE
                       SAMPLING AND ANALYSIS TECHNOLOGIES.
         0.1     0.2    0.3     0.4    0.5    0.6    Q.I    0.8     0.9
                       UNCERTAINTY IN EMISSION FACTOR, br

         Figure  2.   Impact of alternative  hypothesis
                       test  approaches on S*
                                                             i.o
  
-------
    TABLE 3.  GENERALIZED ERROR BOUNDS FOR SOURCE ASSESSMENT

Uncertainty Value
A
a

A
C

A
b

b

b
A
C

A
C
A
d
A
d
m

0.05
r

0.05
r

0 to 1.0
r
0.5
su
0.1
O A/
0.1
su

~ o 0.5
O A/
0 4.2 71.6
A Bl B2
sฃ 0 0.81 0
3.0
For  the  Deterministic  Decision Approach
Substituting values from Table 3 into Equations 30 and 31 yields
the following error bounds for source severity:

Case A
ea =
                      eu = ^tปr2 + 0.01 + 17
                               +0.01  +0.75
                                                              (49)
Case
                                + 0.01 + 94
                        = Jbr2 +0.01 + 0.95
                                                              (50)
                                37

-------
Case B?_


                             "  * +  0.01 +  17
                                                              (51)
The corresponding uncertainty interval  for  source  severity is
obtained by substituting Equation 49 , 50 / or  51  into  Inequal-
ity 32.  For example, taking br = 1.0 as a  worst case,3

Case A                   0 < ST < 19 Sc                      (52)


Case Bi                  0 < ST < 96 Sc                      (53)


Case B?                  0 < S  < 19 S                        (54)
                                      c
At the other extreme, when b  =  0,
Case A                0.15 SG  < ST  <  18  SG                    (55)


Case B!                  0 < S^ < 95  Sn                       (56)
~-                          ""   ฑ "*     \*


Case B?                  0 < S  < 18  S                        (57)
   T ~n --rn"-'~-                    ™~    ^
The critical value of the  "calculated"  severity  is  given by Equa-
tion 39.  Substituting for eu from Equations  49,  50,  and 51
yields the following results:


Case A     .           S* = 	 1                           (58)

                           18 + Vbr2 +0.01


Case BT               S* = 	 l                           (59)

                           95 + Vbr2 +0.01


Case B7               S* =	                   (60)
                           18 + Vb  2 +  0.01
 Note that source severity cannot be  negative,  so the lower bound
 cannot be less than zero.


                               38

-------
As br varies from 0 to 1.0, SJ varies from 0.055 to 0.053 in
Cases A and B2 and from 0.0105 to 0.0104 in Case BX.  It follows
that the critical test value should be set at 0.05 for Cases A
and B2 and at 0.01 for Case BX.

For the Stochastic Decision Approach

Substituting values from Table 3 into Equation 26 yields the
following value for V(log S ), regardless of how F is treated.
                     V(log Sp) = Vb^ + 0.01
                            \~r      I.


The corresponding uncertainty interval for source severity is
obtained by substituting Equation 61 into Equation 48.  For ex-
ample, taking b  = 1.0 as a worst case, we get,


                    0.0404 Sc < ST < 24.7 SG                  (62)


The corresponding uncertainty interval for source severity is ob-
tained by substituting by taking b  =0.0.


                     0.179 SG < ST < 5.58 Sc                  (63)

ALLOWABLE RANDOM UNCERTAINTY IN EMISSION FACTOR

An upper bound on the allowable value of br is obtained by solv-
ing Equation 44.  Substituting for e  from Equations 49, 50, and
51 gives the following results:

                         Deterministic            Stochastic
                           Approach                Approach
Case A                    br < 0.23    (59)

Case BI                  No solution

Case B2                  No solution
infinity
In Cases BI and B2 , the restriction imposed by Equation 44 can-
not be met even with 6r = 0.  Hence, in these cases, there will
always be a large risk of a Type II error in the screening
procedure.

Strictly speaking, Equations 58 through 61 are valid for a test
level, a, of 0.025.  However, these results are nearly independ-
ent of a.  This is a reflection of the fact that the total
uncertainty in source severity is dominated by the systematic
component of uncertainty (see below).

                               39

-------
In principle,  then,  for Case A  (no uncertainty in "acceptable"
concentration),  Inequality 61 can be used  with standard proce-
dures to estimate the number of samples  required in a sampling
program.   In most situations involving environmental sampling,
however, the restriction imposed by Equation 61 will result  in  an
impractically  large  number of samples.   In practice then, there
is generally little  that can be done to  control the risk of  a
Type II error  when using the present screening procedures.   This
is again a  reflection of the fact that the random uncertainty in
emission factor  makes a relatively small contribution to the
total uncertainty in source severity.

EFFECT OF  RANDOM UNCERTAINTY IN EMISSION FACTOR

Comparison  of  Inequalities 52 through 53 with Inequalities 55
through 57  shows that the random uncertainty in emission factor
has a relatively small effect on the total uncertainty in source
severity.   The total uncertainty is dominated by the systematic
uncertainties, as illustrated in Table 4.   This table lists  the
reduction  in the width of the uncertainty  interval  (Inequality  41,
42, or 43)  which would be obtained if each of the individual
uncertainties  was eliminated while leaving the others unchanged.
The table was  constructed using the values of the individual
uncertainties  listed in^Table 3, together  with br = 1.0, to  cal-
culate the  width (eu + ejj,) of the Interval 32.  Each of the
individual  uncertainties was then set to zero while leaving  the
others unchanged, and the resulting reduction in the width of the
interval was computed.

       TABLE 4.   CONTRIBUTION OF INDIVIDUAL UNCERTAINTIES
                  TO  TOTAL UNCERTAINTY IN SOURCE SEVERITY
                  (DETERMINISTIC DECISION APPROACH)   .

                                         Reduction in width of uncertainty
                                         interval for source severity, %^
            Uncertainty eliminated	Case A	Case BI	Case  62
1.
2.
3.
4.
5.
6.
7.
Random uncertainty in emission factor.
Systematic uncertainty in emission factor.
Total uncertainty in emission factor (1 + 2) .
Uncertainty in dispersion equation.
Uncertainty in plume rise.
Total uncertainty in dispersion modeling (4 + 5) .
Uncertainty in "acceptable" concentration.
10
30
40
63
69
87
0
2
33
35
65
74
90
80
10
30
40
60
68
83
1.






3
 Uncertainty interval computed assuming b  = 1.0.
 K
 No uncertainty in acceptable concentration.
 Acceptable concentration uncertain; TLV conversion factor equal to geometric mean value.
 Acceptable concentration uncertain; TLV conversion factor equal to 1/300.


Contrary to what might be expected,  Case B2, in  which the accept-
able  concentration is "intentionally underestimated,"  has a nar-
rower uncertainty interval than does Case Bj.  In general,
underestimating the acceptable  concentration decreases both the


                                 40

-------
 upper bound  (which tends to decrease  the  total uncertainty)  and
 the lower bound  (which tends to increase  total uncertainty)  on
 source severity.   However, in Case BI  the lower bound is already
 zero, so it cannot be decreased further.   Therefore, the only
 effect of underestimating the acceptable  concentration is  to
 decrease the upper bound, which reduces the total uncertainty in
 source severity.

 In addition, since eu can vary from zero  to infinity while ej,
 can vary from zero to one, the above  calculation tends to  give
 more weight to the upper bound than to the lower bound.  Although
 underestimating the "acceptable" concentration does not reduce
 the uncertainty associated with acceptable concentration,  it does
 shift it to the lower bound where it  is less conspicuous.   There-
 fore, in Table 4,  Case Bj is more indicative of the relative
 magnitude of the uncertainty associated with acceptable concen-
 tration than is Case  B2.

 The upper and lower error bounds on source severity are tabulated
 in Table 5 for each of the cases listed in Table 4.  These  error
 bounds correspond  to  the  situation in which ar = cr = 0, so  that
 the random component  of uncertainty becomes
                   er = War^  + br2 + 4 c^ = br                 (64)


 TABLE 5.  CONTRIBUTION OF INDIVIDUAL  UNCERTAINTIES TO  ERROR BOUNDS
           ON  SOURCE SEVERITY3  (DETERMINISTIC DECISION  APPROACH)
Lower bound;

1

2.
3.
4.

5.
6.
7.


8.
'Uncertainty eliminated
None.

Random uncertainty in emission factor.
Systematic uncertainty in emission factor.
Total uncertainty in emission factor
(2 + 3).
Uncertainty in dispersion equation.
Uncertainty in plume rise.
Total uncertainty in dispersion modeling
(5+6).

Uncertainty in "acceptable" concentration.
b
Case A
0.25 -

0.25
0.28 -

0.28
0.74 -
0.30 -

0.90 -

0.25 -
h
r

br


b
br

h
r
br
, 1 •
Case BI
0.


0.


0.
0.

0.

0.
05 -

0.05
05 -

0.05
14 -
06 -

17 -

25 -
h
r

br


b
br

h
r
br
' "l
Case 82 '
0.00

0.00
0.00

0.00
0.01 - b
0.00

0.01 - b
r
0.25 - br
Upper bound , 1 H
Case A Case BiC
18 +

18
12 +

12
6 +
4.5 +

1.5 +

18 +
b 95 + b
r r
95
b 63 + b

63
b 32 + b
r r
b 24 + b
r r

b 7.9 + b
r r
b 18 + b
h e
u
A
Case 82
18 + b
r
18
12 + b

12
6 + b
r
4.5 + b

1.5 + b
r
18 + b
      * - T   c    u
  No uncertainty in acceptable concentration.
  Acceptable concentration uncertain; TLV conversion factor equal to geometric mean value.
 d
  Acceptable concentration uncertain; TLV conversion factor equal to 1/300.
 Q
  Since source severity is nonnegative, the lower bound cannot be less than zero.


From Table  5,  it can be  seen  that in Case B2  the  uncertainty
associated with  acceptable concentration has  no effect on the
upper bound  since the values  for Case B2 are  identical to those
for Case A  (no uncertainty in acceptable concentration).   On the



                                  41

-------
other hand, the lower bound is dominated by the uncertainty in
acceptable concentration in Case B2.  In fact, the lower bound is
essentially zero for Case Bฃ in all instances except Item 8,
which corresponds to no uncertainty in acceptable concentration.

Table 5 also shows how each of the individual uncertainties con-
tributes to the departure from the ideal situation in which both
the upper and lower bounds are equal to unity.  (In this situa-
tion, Inequality 32 becomes
< ST <
                    .e.
                                                   = SQ.)  The
ideal situation is approached only in Item 7, Case A  (no uncer-
tainty in dispersion modeling or acceptable concentration) ^and
then only when the random uncertainty in emission factor, br,
is small.

OPERATING CHARACTERISTICS OF THE TEST

In order to fully describe the risk of a Type II error in a  test
of hypothesis such as that described in this section, it is  neces-
sary to construct an operating characteristic curve for the  test.
The probability of a Type II error can be expressed as a function
either of the "true" value  (ST in this case) or  the estimated
value  (S^;) of the parameter in question.

In the Neyman-Pearson theory of statistical  inference, the  true
value of the unknown parameter, ST, is regarded  as a  fixed  entity,
and the measurements, SG, are visualized as  distributed about
this value.  This situation is illustrated in Figure  4.
                  to
                  5
                  o
                 CO
                 i
                 0.
                              TRUE VALUE

                            -MEASURED VALUES -
             Figure  4.
Schematic representation of the
fiducial statistical approach.
Deterministic Decision Approach
Operating characteristic  (OC)  curves are given in Figure 5 for
Case A  (no  uncertainty  in  "acceptable"  concentration)  with
br = 0.1. a  The curves  give  the  probability that ST < 1 . 0 as a
function of the  true  severity,  ST.   For ST < Scd - e^) ,  the
curves give the  probability  of  not  making a Type I error; for
ST > Sc(l + eu) , they give the  probability of making a Type II
error.  A family of OC curves is  obtained with the systematic
error in source  severity as  a parameter.   In Figure 5, three

Construction  of these  curves  is  discussed in Appendix F.
                                42

-------
 members of this family of curves are shown corresponding  to  no
 systematic error, systematic error equal to esu, and  systematic
 error equal to esฃ.  The latter two curves represent  the  extreme
 cases; i.e., all OC curves for this problem lie between these two
 curves.

 Note that in Figure 5 the left-most curve  (systematic error
 equal to esu) represents the worst case with respect  to Type I
 errors, while the right-most curve  (systematic error  equal to es&)
 represents the worst case with respect to Type II  errors.  Hence,
 a convenient means of representing the family of OC curves is a
 combination of the upper half of the left-most curve  and  the
 lower half of the right-most curve, as shown in Figure 6.   (The
 other halves of the curves are of lesser interest  since they
 correspond to error probabilities of 50% or greater.)  The curve
 in Figure 6 is termed an "effective" OC curve since it corre-
 sponds to treating the Interval 32 as an exact 95% confidence
 interval.
     1.0

     0.9

     0.8

   5-0.7
   en
-rr'-c 0.6
 O 
-------
consequence whether the worst case probability of  a  Type  II  error
is, for example, 0.5 or 0.9.  The important point  is that there
is a large risk of a Type II error in this region, and  this  is
clearly indicated by the "effective" curve.)  Thus,  the graph
shows that the worst case probability of a Type  II error  cannot
be reduced to less than 0.5 in this region by reducing  the random
uncertainty alone.  This result can only be achieved by reducing
the systematic component of uncertainty.
            1.0

            0.9

           _ 0.8
         V o
           B
           a 0.4
            0.3

            0.2 -

            0.1 -
             0
             0.01
                       0.1
                                  1.0
100
        Figure 6.   "Effective"  operating characteristic
                    curve  for  Case  A with ฃ>r = 0.10.
In Figure 7, "effective" OC curves are given  for  Case  A  (no
uncertainty in "acceptable" concentration) as a function  of  br,
the random uncertainty in emission factor.3   (The  curves  for
Cases BI and B2 are similar except that the horizontal portions
of the curves extend to infinity for all values of br. )   The
curves show that when br > 0.23, the worst case probability  of a
Type II error is greater than or equal to 0.5 for  all  values of
the estimated severity, Sc (cf. Equation 48,  inequality 59,  and
the related discussions).
 The left-hand branch of the curve is insensitive  to  the  value  of
 br due to the dominant effect of the systematic component  of
 uncertainty.
                               44

-------
           V__
           n
1.U
0.8
|ฐ0.7
|fl.6
ฐ 0.5
|aซ
ฃ 0.3
0.2
0.1
0
0

.
-
•
NO RANDOM ERROR. ar = t>r = Cr = 0
IN ALL OTHER CASES. ฃ. = ^=0.05



•
r" 	 •
i <
i b iO.23
\V b -0-0

\\^^
\^^____
01 0.1 1.0 10 10
ST _
         Figure  7.   "Effective operating characteristic
                    curves for Case A as a function of
                    random uncertainty in emission factor.


Stochastic Decision Approach

The operating characteristic curve for the stochastic decision
approach is given in Figure 7.  OC curves for r = 1 and r = 183
are shown; corresponding curves for other values of r are parallel
to those shown in Figure 8.  Since the stochastic decision test
approach treats all uncertainties as random errors, there is no
plateau in the curves as found in the previous OC curves.
            Figure 8.  Operating characteristic curve
                       for the stochastic decision test
                       approach .
                                45

-------
SUMMARY

The results of the above analysis can be summarized as follows:

   • The stochastic decision test approach should be used
     in preference to the deterministic decision test
     approach.

   • The critical value, Sฃ, for the calculated source
     severity should be set at 0.05 to 0.08.

   • In general, there is no justification for taking a
     large number of samples to obtain a very precise esti-
     mate for the emission factor, since precision in source
     severity is limited by the precision in available
     measurement technologies.

The last conclusion should not be construed as implying that
sampling and analytical procedures are unimportant in the Source
Assessment Program.  In the first place, bias in sampling and
analysis does have a significant effect on the total uncertainty
as shown in Table 4.  Secondly, in many cases the effect of
random uncertainty in emission factor will be considerably
greater than indicated in Table 4.  For sources having negligible
plume rise, for example, the total uncertainty will be reduced by
69% to 74% over that used in Table 4, and the contribution from
the random uncertainty in emission factor will be correspondingly
greater.  Thirdly, reliable emissions data are of value in their
own right, above and beyond their utility in estimating source
severity.
                                46

-------
                            SECTION 6

              SOURCE ASSESSMENT:  EMISSIONS BURDEN
INTRODUCTION

For criteria pollutants, the source assessment screening process
makes use of a second decision index, the national emissions
burden, in addition to source severity.  For a given source type
and given criteria pollutant, the national emissions burden, NB/
is defined as follows:

                                  M
                             NB = gB                          (65)
                                   n

where  M  = annual mass emissions of given criteria pollutant
        "   from the given source type, kg
       M  = annual mass emissions of given pollutant from all
        n
            stationary sources nationwide, kg
Equation 65 defines the "true" value of N .  In practice, it is
calculated as follows:

                               CAP   E.F )
                                       R;                     (66)
where   CAP^, = total production capacity of source type, kg/yr
         EF  = representative emission factor for source type,
               g/kg
       M^    = estimate of Mn obtained from 1972 National
               Emissions Report (11) ,a kg
 The National Emissions Report contains data on emissions from
 both stationary and mobile sources.  Only the data for station-
 ary sources are used in estimating Mn.
(11)  1972 National Emissions Report; National Emissions Data
     System (NEDS) of the Aerometric and Emissions Reporting
     System (AEROS).   EPA-450/2-74-012, U.S. Environmental
     Protection Agency, Research Triangle Park, North Carolina,
     June 1974.  422  pp.
                                47

-------
Note that the emission factor in Equation 53 is not the same as
that used to calculate source severity.  The latter emission
factor is for a single emission point within a representative
plant, while the emission factor in Equation 53 is the total for
all emission points within a representative plant.

As noted in Section 5, the national emissions burden is presently
used for screening a source type as follows:  the source type is
considered to be a potential candidate for control technology
development only if ND > 0.001.
                     D ~~
The objective of this section is to perform an analysis for NB
similar to that given in the previous section for source severity,

UNCERTAINTY IN ND
                B
The uncertainty in N_ arises from the following components:
                    D

   • There is a random uncertainty in production capacity
     as discussed in the previous section.

   • There is uncertainty in the emission factor which has
     both random and systematic components.

   • There is uncertainty in nationwide annual emissions,
     due to the use of the NEDS data base, which has both
     random and systematic components.

For convenience, let A, B, and D denote the "measured" values of
the variables as follows:
                            A = CAP
                                   T
                            B = EF
                                  R
                            D = M
                                 n
Let ar, br, and dr be the relative random uncertainties associ-
ated with A, B, and D.  Using the error propagation formulas from
Table B-l, the random uncertainty in N_, is found to be:
                                      _,
                                      a
er =
                                  V - V
(67)
Considering next the systematic uncertainty, let



                         -b „      -d „
                               48

-------
 be systematic error bounds on EF^ and Mn,  and let bsu,  bsฃ,  etc.,
 denote the corresponding relative error bounds.  Using the formu-
 las from Table B-8 yields the following systematic error bounds
 on N :
                               b   + d
                             = -  -                         (68)
                          su
                              b  „ + d
                               sฃ  ^ su                        (69)
                               1 + d
                                    su
The total uncertainty in N  is the sum of the random and  system-
atic components:
               /N     //\     /N     /\  ,_    on    Q V
               ... - ปK2 + br2 + dr2 +  ฐ"   ,                 (70)
                                       /\     /\
                                       b_n + d.
                                             su
Thus, we have the following uncertainty interval for the true
value of N_.:
          c

             
-------
1 December 1976).  There are approximately  100  categories  which
contribute to the national  totals.  Hence,  the  variability in  the
totals is estimated  (assuming all categories  contribute  equally
to the total) to be:
                       Vioo(ฐ.Q5)2 =  fo.oos
                            100

The  95% confidence  limits are approximately  twice  this  value,  or
ฑ0.01.

The  systematic error bounds on Mn were obtained  by assuming  that
the  NEDS totals are systematically low due to  incomplete  inven-
tories.  Assuming the emissions inventories  are  at least  70%
complete gives ds& = 0 and dsu = 0.3.

With the above values, Equations 70 and  71 become:
                         ^\  *\i /\
                         eu = br + 0.5

                         /v  f\j /v
                         e, = br + 0.3


The  uncertainty interval for N_, is then:
                              c

           (NB)C (0.7 - br) <  (NB)T <  (NB)C (br + 1.5)         (74)


STATISTICAL TEST OF HYPOTHESIS

Inequality 74 can be used as in the previous section to formally
set  up a statistical test of hypothesis.  The null  hypothesis  in
this case is:

                       H0:  (NB)T > 0.001


This hypothesis will be accepted if the upper limit in  Inequality
61 exceeds 0.001.   Therefore,  the critical value for the  test,
(N*)c,  is given by:

                              =   o.ooi                       {75)

                                br + 1.5


Inverting Inequality 74 yields:

                    (NB)T             (N )
                  ~	5-2- <  (NB)C < 	^-V                 (76)
                  br+1.5     " ^   0.7-br
                               50

-------
The  upper end point of this  interval becomes infinite  for br
=  0.7, which places an upper bound on thง acceptable uncertainty
in emission factor.  With this value of br, Equation 62 becomes
 (N*)c 3*  0.0005.


As br  ranges between  zero and  one,  (Ng)c  varies from 0.00067  to
 0.00042.  Sin^e the level of the test,  a,  enters the calculation
only through br,  the  critical  test value  is  again nearly  indepen-
dent of  the test  level.3
SUMMARY

The above analysis leads to the following guidelines for use of
the second source assessment decision index:

1.  The critical value of national emissions burden should be
    set at 0.0005 = 0.05%.

2.  The random uncertainty in emission factor should be less
    than ฑ70%.

The emission factor referred to here is the total emission factor
for the representative plant.  However, the above restriction can
be met if the uncertainty in the emission factor for each emis-
sion point in the representative plant is required to be less
than ฑ70%.

The analysis of the first source assessment decision index result-
ed in an upper bound on the random uncertainty in emission factor
for criteria pollutants of ฑ23%.   Hence, the present value of
ฑ70% theoretically places no additional restriction on sampling
and analytical procedures.  In practice, however, just the oppo-
site may be true.  The value ฑ23% may be impractical and, there-
fore,  may be ignored entirely, while the value ฑ70% may be
reasonable, and, therefore, represent a real restriction.
 ar has been neglected compared to br.
 Criteria pollutants are covered by Case A.
                               51

-------
                            SECTION 7

          OFFSET CALCULATIONS:  PLANT EXPANSION PROBLEM
INTRODUCTION

As part of EPA's overall program to achieve and maintain National
Ambient Air Quality Standards, the individual states are required
(through State Implementation Plans) to determine the amount of
emissions reduction necessary to offset the probable impact of
increased population, industrial activity, motor vehicle traffic,
and other growth factors.  Of particular concern in this section
is the policy regarding increased emissions due to increased in-
dustrial activity through expansions of existing plants.

One way to offset emissions from a plant expansion is to require
the company to make a corresponding reduction in emissions from
its existing unit.  Such a policy entails the associated problem
of determining a posteriori whether the company is complying with
the policy.  Such a determination is not entirely straightforward
because the emission rates are not known exactly.  This uncer-
tainty should be taken into account in the decision-analysis
process .

The objective of this section is to analyze the uncertainty
associated with the above problem and construct an approximate
statistical test as the basis for making a decision regarding
compliance with the offset policy.

WORKING EQUATIONS

The difference in emission rates of a given pollutant before and
after plant expansion can be expressed as follows:

                      Q2 = Qi - AQi + QN                     (77 )
                      Q2 - Ql = QN -

where   Q2 = emission rate after expansion, g/s
        QI = original emission rate, g/s
        QN = emission rate from new unit, g/s
       AQj = reduction in emission rate from original unit, g/s
                               5-2

-------
 Both QN and AQi  are composed of terms corresponding to the indi-
 vidual emission  points within the plant.   Thus,

               Q   = ฃQNi      (sum over all  emission           (73)
                    i         points  in new unit)
 and
     AQj =  I(AQi).       (sum over  all  emission  points  in
           j              original  unit which  have  different
                         emissions after  expansion)
        =  SQjj -  ZQ1(j                                        (79)


In these equations Q^ and Qj'j denote  emission  rates  from the
jth. emission point in the original unit before and  after  expan-
sion, respectively.

Equations  77, 78, and 79 are the basic  relationships governing
the problem.  Note that the difference  (Q2  - Qj) depends  only on
emissions  from the new unit and those emissions  from the  original
unit which are changed in the expansion.

UNCERTAINTY IN (Q2 - Ql)

Uncertainty in (Q2 - QI) is due to uncertainties in the individ-
ual emission rates QNI/ QI-J' an<^ QI'-J*  Once the random and sys-
tematic uncertainties in these values have  been  determined,  the
uncertainty in (Q2 - QI) can be obtained using error propagation
formulas for addition and subtraction as outlined below.

Consider first the random uncertainty.  Let BN and  B^  denote the
measured values of QN and AQlf  and let brn  and brA  be  the  corre-
sponding random uncertainties.  Then the random uncertainty in
(Q2 - QI)  is, according to the subtraction  formula  in  Table B-l,
Appendix B:

                       e  =

where

                     b2
                     D rN   "" rNi

and
                     rlj  ' - ri'j -  \" rij
                               53

-------
 Therefore,  we have:
               er '   b     +          + b2'                   (80)
where brNi/ ^riv  anc^  ^ri'j  are t'ie ran<3om uncertainties  asso-
 were  rNi/   riv anc   ri'j   e   ie  ran   OU
                                             V "XT   " A /
                   IN  ,         a  ,             N    A
                       JIN           ฃA                   Sfc

be  systematic error bounds for Q ,  AQlf  and (Q2 - QI).  Then,
using  the  addition and subtraction  formulas from Table B-8, we
obtain:

            esu = buN + bฃA = Z'D--"-  +  ^ib"'-! + b.--i-;i           (81)
                = bฃN + buA =       +  Eb     + b'             (82)
where bujj-^,  b  ^,  bu i ^ ,  etc., are systematic  bounds correspond-
ing to QNi,  Q*   and
The total  uncertainty in (Qa - QI) is the  sum of the random and
systematic components.   Combining Equations  80,  81,  and 82, we
obtain:
    e  =     M-       T      ,
    u   L  rNi    \  rl]     rl
                             - H/2 + 2
                             D/       i



      =  Zt>2 XT- + Z^b2  , • + fa2 , . • M /2 + ฃ*
       I .   rNi   . y rl]    rl'j/l      .
                                       .  •       - .   bnil .\       (84)
                                       ฃNi    .1  ul]    ฃ1 '
The uncertainty interval for  (Q2 - QI) is  given by:
             BN -  BA - 6ฃ - (Q2 - Ql} ~ BN  -  BA  + 6U           (85)

Alternatively,  Inequality 72 may be written  as:

          (Q2  -  Qi)c - eฃ < (Q2 - Q!)T <  (Q2  - Qi)c + eu       (86)

where  (Q2 -  Qi)ip  denotes the true value of the  emission rate dif-
ference and  (Q2 - Qi)c = BN - BA is the calculated difference.

                                54

-------
TEST OF HYPOTHESIS

Inequality 86 forms the basis for a formal test of hypothesis.
The hypothesis to be tested is:
               H0:   (Qa ~ Ql)T > ฐ' i.e., Q2 > Q!

This is tested against the alternative hypothesis:

               HA:   (Q2 - Qi)T < 0; i.e., Q2 < Qj


The following simple procedure is proposed for carrying out the
test:

1.  Specify level of test, a  (a = probability of Type I error)

2.  Determine (Q2 - QI),, = B  - B
                       \^f    IN    ฃA

3.  Calculate e  according to Equation 83.   [Note:  the b  's
    should be (1 - 2a) x 100% confidence limits, where a is the
    specified level of the test.]

4.  If  (Q2 - Q!)c + eu is ^ 0, reject H0; otherwise, accept H0.
    (Note:  If H0 is accepted, the plant is considered out of
    compliance.)

Notice that compliance requires that the emissions from the orig-
inal plant be reduced by the amount of emissions from the new
capacity plus an additional amount, eu, which compensates for the
uncertainty in the data.

The two kinds of incorrect decisions that can be made using the
above procedure are:

Type I Error;  Reject H0 when it is, in fact, true; i.e., decide
               that the plant is in compliance when it is, in
               fact, in violation.

Type II Error;  Accept H0 when it is, in fact, false; i.e.,
                decide the plant is in violation when it is, in
                fact, in compliance.

The probabilities of making the two types of errors in decision
making are as follows:

1.  Probability of Type I error is 1 a.

2.  Probability of Type II error is a function of a,  (Q2 - QI)Q,
    eu, and e^.   It must be determined individually for each
    specific problem.
                                55

-------
Theoretically, the risk of a Type II error can be controlled by
specifying the sample size to be used in obtaining emissions data,
However, the effect of a Type II error is to require the plant to
make an additional reduction in emissions.  In practice, there-
fore, specification of the sample size involves an economic trade-
off between the cost of reducing uncertainty in the data and the
potential cost of reducing emissions.

NUMERICAL EXAMPLE

In order to illustrate the test procedure, numerical values have
been assigned to the variables involved in the test.  These val-
ues are for illustration only, and do not correspond to actual
field data.
1.   Take a = 0.05 [Note:
    confidence limits.]
                          the br's should be  (1 - 2a) x 100 = 90%
    Let
    brN = 6 g/s
    brA = 8 g/s
                         e  =
                                 rN
                                           = 10
3.   Let
    buN = 4
        = 2


        = 3
     uA
    bฃA = 6 g/s
                        6su = buN + bฃA = 10 9/S
                                    buA = 5
    Then
    e  = e  + e   =20 g/s
     u    r    su      ^
       = e
              esฃ = 15 g/S
    Let
    BN = 50 g/s
    BA = 65 g/s
                      (Q2 - QI)C =
                                           = -15 g/s
                               56

-------
6.  Uncertainty interval for  (Q2 - QI)T is then
+e
                                  +20
    Since  (Q2 - QI)Q + eu = +5 9/s' the nul1 hypothesis is
    accepted; i.e., it is concluded that  (Q2 - QI)T > 0 and the
    plant is out of compliance.  Notice that if the uncertainty
    in the data were not taken into account, the opposite con-
    clusion would be reached since the nominal value of  (Q2 - QI)
    is -15 g/s .

    An "effective" operating characteristic curve  for the  test  is
    illustrated in Figure 9.  This curve  is obtained by  treating
    the uncertainty interval  (86) as an exact 90%  confidence  in-
    terval.  This approximate treatment indicates  a relatively
    large risk of a Type II error in this particular example

    (PType II = ฐ'79)-
            1.0
            0.8
          V 0.6
          j_>
          o*

          | 0.4
          O_


            0.2
             0-4
             ~
                  -15 .
     -5    0    5

        'VVr
10
                         15
20
        Figure  9.   "Effective"  operating characteristic
                    curve  for  example  problem.
SUMMARY
The data  required  to  quantify the uncertainty in the emission
rate difference  before  and  after plant expansion consist of ran-
dom and systematic error  bounds on the following quantities:

    • Emission  rates from  each emission point in the new unit.
                              57

-------
   • Emission rates before and after expansion from each emis-
     sion point in the original unit which has different
     emissions after expansion.

The above data can be used to perform a simple, approximate sta-
tistical test to determine whether the plant is in compliance
with offset policy.  The essence of this procedure is the follow-
ing:  In order to ensure that air quality is not degraded, emis-
sions from the original plant must be reduced by an amount
sufficient to compensate not only for the emissions from the
added capacity, but also for the uncertainty in emissions data.
                              58

-------
                            SECTION  8

               COMPARISON OF ALTERNATIVE CONTROLS:
             REASONABLY AVAILABLE CONTROL TECHNOLOGY
 INTRODUCTION

 The problem to be addressed in this section involves comparison
 of the control efficiencies of two alternative control technolo-
 gies.  It is assumed that a plant currently employs a given con-
 trol technique with a nominal  (measured) control efficiency which
 is less than that of reasonably available control technology
 (RACT) for the given application.  In order that the plant be
 required to adopt RACT, it should be established that RACT is
 indeed superior to the installed control technology.  Since
 neither of the two efficiencies can be known with certainty, this
 determination should not be based on the nominal efficiencies
 alone.

 The purpose of this section is to formulate a statistical test to
 serve as the basis for deciding whether RACT is superior to a
 given installed control technique.  Only the simplest case, a
 single pollutant emitted from a single source, is considered.

 GOVERNING EQUATIONS

 The control efficiency, ej, of the installed device is given by

                           Qi - Q.2       Q2

                                                              (87)
where Qj and Q2 are the uncontrolled and controlled emission
rates, respectively.  Letting e2 denote the control efficiency
of RACT, we have

                                     Q2
                      e2 - ej = e2 + QY - 1                   (88)


For the purpose of this example, it is assumed that EJ is meas-
ured by measuring QI and Q2 .  The value of e2 is assumed to have
been determined previously on a similar but different installa-
tion.  Therefore, the measured value of e2, together with its
                                59

-------
           associated random and systematic uncertainties, is assumed to be
           given.

           We now wish to develop equations for the total uncertainty in the
           difference (ฃ2 - EI).  Let BI and B2 be the measured values of
           Ql and Q2 with associated random uncertainties brj and br .  Let
           (e2)c ke the given nominal value of e2 with associated random
           uncertainty 3r-  Then the nominal value (e2 - EI)C of tne control
           efficiency difference is given b
                   (E2 - EI)C =  (e2)c + 3 -- 1
                                                                         (89)
           The associated random uncertainty, e , is obtained using error
           propagation formulas for addition and division:
where
V
                               and
                                                                         (90)
           Considering next the systematic uncertainty, let
+b
                    +b
          B
                           ' B2
                                  „
                                                             +e
           be systematic error bounds for Qj, Q2 , e2 an^  (E2 ~ EI)ซ  Using
           addition and division formulas for propagation of systematic
           errors yields:
                              esu = 3u +  B -

where
                 "/s"
                     again denotes relative uncertainty.
           The total uncertainty in (e2 - ei) is the sum of the random and
           systematic components.   From Equations 79, 80, and 81 we obtain
                                          60

-------
                        ,
                       ri
                               r2
                                                              (93)
      V  2   /B
     -™r  + hr
                       ri
                            + b
(94)
The uncertainty interval for (e2 ~

         (e2 - ฃi)  - e  < (e2 - E
                                       is given by:

                                      <  (e2 - ฃi)
(95)
where  (e2 - EI)T represents the true value of the control effi-
ciency difference,  (e2 - ฃi)c is given by Equation 89, and e  and
e  are given by Equations 93 and 94.
 J6

TEST OF HYPOTHESIS

Inequality 95 forms the basis for a statistical test of hypothe-
sis.  The idea behind the test is the following:  if the lower
uncertainty limit  (e2 - ฃi)c - e  exceeds zero, then we can con-
clude that (e2 - EI)T is greater than zero, i.e., that RACT has
a higher efficiency than the installed device.  If the lower
uncertainty limit is less than or equal to zero, then we cannot
conclude that RACT is better.

The null hypothesis for this case is that RACT is no better than
the installed control device, i.e.,

                      H0:   (E2 - EI)T < 0


The alternative hypothesis  is that RACT is better; i.e.,

                      HA:   (ฃ2 - EI)T > 0


The test is formally carried out as follows:

1.   Specify level of test,  a (a = probability of Type I error).

2.   Calculate (e2 - ฃi)c from Equation 89.

3.   Calculate e  from Equation 94.  [Note that the random uncer-
    tanties should be (1 - 2a) x 100% confidence limits.]

    If (e2 - e1)c - e  > 0, reject H0 (conclude RACT i;
    otherwise accept H0 (conclude RACT is not better).
                               61

-------
The two types of incorrect decisions that can be made using  the
above test are:

Type I Error;  Conclude RACT is better when  in  fact  it  is  not
               better.

Type II Error;  Conclude RACT is not better  when,  in fact,  it
                is better.
NUMERICAL EXAMPLE

In order to illustrate the test procedure, numerical  values  have
been assigned to the variables involved in the  test.   These  val-
ues are for illustration only, and do not represent actual  field
data.

1.  Take a = 0.05  [Note:  the random uncertainties should be
     (1 - 2a) x 100 = 90% confidence limits.]
2.  Let
4.
              = 0.10; i.e., the nominal efficiency of  the
                            c
    installed device is  (ฃ1)^, = 0.90.
3.  Let the nominal efficiency of RACT be  (ฃ2),-, =  0.95.   Then
    (ฃ2 - El)  = 0.05.                        c
              = 0.20
     ui
    er = o.oi
              = b
                 u2
Then,
      V  2
     =>3T.  +
B,
 l

B~
                      ri
+  b.
                               T2
                                       - •
     = 0.048

5.   (ฃ2 - EI)
                    = 0.05 - 0.048 = 0.002  >  0
Therefore, we reject the null hypothesis and  conclude  that  RACT
has a higher control efficiency than the installed  device.
                                62

-------
SUMMARY

Before a plant is required to install RACT in preference to a
previously installed alternative control method, the superiority
of RACT should be demonstrated in light of uncertainties in emis-
sions data.  A simple approximate statistical test can be employ-
ed for this purpose.  The essence of the test is the following:
in order to demonstrate superiority, the RACT control efficiency
is required to be sufficiently high that the difference in
control efficiencies is statistically significant for the given
level of uncertainty in the data.

The information required to perform the test consists of the ran-
dom and systematic components of uncertainty in the two control
efficiencies to be compared.  Usually, this information will be
derived from corresponding uncertainties in controlled and uncon-
trolled emission rates.
                                63

-------
                            SECTION 9

                 CONCLUSIONS AND RECOMMENDATIONS
The principles of decision analysis can be used to form a sound
basis for evaluating the uncertainties in environmental data as
they relate to decision making.  These principles have been
applied to two numerical indices (source severity and national
emissions burden) to develop guideposts for use as an aid in
environmental decision making.  However, these principles cannot
be used in a rigorous fashion due to the complexity of the prob-
lems to be considered and due to the subjective nature of environ-
mental decision making.  Seldom, if ever, will an environmental
decision be made only on the basis of a given numerical index
(e.g., source severity).  In general, other experimental data;
social, political, and economic considerations; and human judg-
ment all enter into the decisionmaking process.  The index is
only one piece of information which the decision maker may util-
ize in arriving at a decision.  The index provides only technical
information; value judgments must be made by the decision maker
(in this case, IERL).  This fact does not alter the basic
approach used to analyze the uncertainty associated with the
index.  Therefore, in this report,  it is convenient to think of
the index alone as determining the outcome of the decision-making
process.  In practice, the index may not be used as the sole
basis for decision making; IERL may consider other factors.

CONCLUSIONS

The following conclusions pertain to the Source Assessment
Program:

   • The stochastic decision approach is more than the
     deterministic approach, to uncertainties in emission
     factors, hence, it is the preferred approach.

   • The critical value of the calculated source severity
     should be set at 0.05 or 0.08.

   • The critical value for national emissions burden should
     be set at 0.05%.

   • The following guidelines for planning sampling and
     analytical procedures should be observed:
                                64

-------
   Criteria pollutants—The random uncertainty in emission
   factor  should be  less than  ฑ23%.   If this restriction
   requires an  impractically large number of samples, suffi-
   cient samples should be collected  to maintain the random
   uncertainty  below ฑ70%.

   Noncriteria  pollutants—The random uncertainty in emission
   factor  should be  less than  ฑ23%.   If this restriction
   requires an  impractically large number of samples,
   the guideline for Cases B!  and B2  should be followed.

•  In general,  there will be a large  risk of a Type II error
   (see Section 6) in the screening procedure no matter how
   large the value of the calculated  source severity.  This
   situation results from the dominant effect of systematic
   uncertainties associated with sampling and analytical
   procedures, dispersion modeling, and health effects
   information.  This situation can be ameliorated only by
   reducing these uncertainties (e.g., by using more
   detailed modeling techniques or through a program of
   fundamental health effects research) or by devising a
   decision index other than source severity which will
   circumvent these  uncertainties.

•  Eliminating the uncertainty associated with dispersion
  modeling may be possible by appropriately defining the
   true (deterministic)  source severity, ST.  Experimental
  data presented in Appendix I indicate that the simple
  Gaussian dispersion model correctly predicts the ensemble
  median ground level concentration under Class C stability
  conditions.  Thus, by using this ensemble median value in
  the definition of ST, the simple Gaussian equation could
  be used with correction factor (y)  equal to unity.   In
  effect,  the condition y = 1.0 would be specified by the
  definition of ST.

• Results  of the source severity simulation presented in
  Appendix H indicate that the definition suggested in the
  preceeding conclusion for the deterministic source sever-
  ity would be consistent with the stochastic approach to
  source severity.   In particular,  results indicate that the
  deterministic severity so defined would yield a worst
  case value in the  sense that higher severities would be
  expected to occur  only infrequently,  on the order of
  once per year.

•  Based on experimental data cited in Appendix I,  the dis-
  tribution of ground level concentration under Class C
  stability and fixed source conditions is approximately
  log-normally distributed,  with  98%  of the values falling
  within a factor of three  of  the median value.   As noted
  above, the  data indicate  that the median concentration
  is correctly predicted  by the simple  Gaussian dispersion
  model.
                             65

-------
RECOMMENDATIONS

The following recommendations are made for conducting the Source
Assessment Program:

   • The critical value of source severity for criteria pollut-
     ants should be set at 0.05.

   • For noncriteria pollutants, the most consistent and
     logically defensible approach would be to estimate
     "acceptable" concentration using the mean TLV conversion
     factor for criteria pollutants (i.e., take F = TLV/21.4)
     and set the critical value of source severity at 0.01.
     The net effect would be to reduce the source severity
     (as originally specified for the Source Assessment Pro-
     gram) by a factor of about 15, and to reduce the
     original lower cut point by a factor of 10 from 0.1 to
     0.01.

   • For criteria pollutants, sampling and analytical strate-
     gies should be planned so as to maintain the random
     uncertainty in emission factor below ฑ23% if practicable;
     otherwise, below ฑ70%.

   • For noncriteria pollutants, the minimum number  (generally
     three)  of samples required to obtain a valid estimate of
     the random uncertainty in emission factor should be col-
     lected and analyzed.  This procedure is consistent with
     the second recommendation above,  concerning interpreta-
     tion and estimation of "acceptable" concentration.

The following recommendations are made for additional related
work:

   • The present study has served to underscore the need for
     additional information in the following areas:

        - Dispersion modeling—a simple method is needed
          to accurately predict the dispersion of airborne
          pollutants.

        - A vastly expanded data base is needed relating
          pollutant exposure to human health effects (extrap-
          olation of animal data to humans).

        - Much more information is needed concerning the
          possible bias associated with individual sampling
          and analytical techniques used to measure pollut-
          ant emissions.

It is therefore recommended that additional research be under-
taken in each of the above areas.
                               66

-------
The present study, including the above recommendations,
should be considered as a first step toward establishing
a coherent procedure for environmental decision making.
It is recommended that a followup study be undertaken to
investigate the applicability of other principles of
decision theory, such as the principle of maximum expect-
ed utility, to the types of problems considered in this
report.  The present study is not the ultimate answer
to analysis of the uncertainties in environmental data.
                           67

-------
                           REFERENCES
1. Savage, L. J.  The Foundations of Statistics.   John Wiley &
   Sons, Inc., New York,  New York, 1954.

2. Lindley, D. V.  Making Decisions.  John Wiley  & Sons,  Ltd.,
   London, 1971.

3. Freund, J. E.  Mathematical Statistics.  Prentice-Hall, Inc.,
   Englewood Cliffs, New Jersey,' 1962.  390 pp.

4. TLVsฎ Threshold Limit Values for Chemical Substances and
   Physical Agents in the Workroom Environment with Intended
   Changes for 1976.  American Conference of Governmental
   Industrial Hygienists, Cincinnati, Ohio, 1976.  94 pp.

5. Turner, D. B.  Workbook of Atmospheric Dispersion Estimates.
   Public Health Service Publication No.  999-AP-26, U.S.
   Department of Health,  Education, and Welfare,  Cincinnati,
   Ohio, 1969.  84 pp.

6. Pasquill, F.  Atmospheric Diffusion, Second Edition.  John
   Wiley & Sons, Inc. (Halsted Press), New York,  New York,
   1974.  429 pp.

7. Cheremisinoff, P. N.,  and A. C. Morresi.  Predicting Trans-
   port and Dispersion of Air Pollutants from Stacks.  Pollu-
   tion Engineering, 9 (3):3, 26, 1977.

8. Hughes, T. W., R. B.  Reznik, R. W. Serth, and  Z. S. Khan.
   Source Assessment Methodology.  Contract 68-02-1874, U.S.
   Environmental Protection Agency, Research Triangle Park,
   North Carolina, January 7, 1976.  33 pp.

9. Serth, R. W., and T.  W. Hughes.  Source Assessment:  Carbon
   Black Manufacture.  Contract 68-02-1874, U.S.  Environmental
   Protection Agency, Research Triangle Park, North Carolina.
   (Preliminary document submitted to the EPA by  Monsanto
   Research Corporation,  December 1975.)   145 pp.
                               68

-------
10.  Eimutis, E. C., B. J. Holmes, and L. B. Mote.  Source
     Assessment:  Severity of Stationary Air Pollution Sources—
     A Simulation Approach.  EPA-600/2-76-032e, U.S. Environ-
     mental Protection Agency, Research Triangle Park, North
     Carolina, July 1976.  119 pp.

11.  1972 National Emissions Report; National Emissions Data
     System (NEDS) of. the Aerometric and Emissions Reporting
     System (AEROS).  EPA-450/2-74-012, U.S. Environmental
     Protection Agency, Research Triangle Park, North Carolina,
     June 1974.  422 pp.

12.  Eisenhart, C.  Expression of the Uncertainties of Final
     Results.  Science, 160 (3833):1201-1204, 1968.

13.  Braddick, H. J. J.  The Physics of Experimental Method.
     John Wiley & Sons, Inc., New York, New York, 1954.  404 pp.

14.  Volk, W.  Applied Statistics for Engineers.  McGraw-Hill
     Book Company, New York,  New York, 1958.  354 pp.

15.  Beers, Y.  Introduction to the Theory of Error.  Second
     Edition.  Addison-Wesley Publishing Co., Inc., Reading,
     Massachusetts,  1957.

16.  Pearson, E. S., and H.O. Hartley.  Biometrika Tables for
     Statisticians,  Third Edition, Volume 1.  Cambridge Univer-
     sity Press, New York, New York, 1966.  264 pp.

17.  Welch, B. L.  The Generalization of "Student's" Problem When
     Several Different Population Variances Are Involved.  Bio-
     metrika, 34:28-35, 1947.

18.  Natrella, M. G.  Experimental Statistics.  National Bureau
     of Standards Handbook 91, U.S. Department of Commerce,
     Bureau of Standards, Washington, D.C., 1963.  504 pp.

19.  Fisher, R. A.,  and F. Yates.  Statistical Tables for Bio-
     logical, Agricultural, and Medical Research.  Oliver and
     Boyd, London, 1963.  356 pp.

20.  Jenson, V. G.,  and G. V. Jeffreys.  Mathematical Methods in
     Chemical Engineering.  Academic Press, New York, New York,
     1963.  pp. 356-360.

21.  Eimutis, E. C., and M. G. Konicek.  Derivations of Continu-
     ous Functions for the Lateral and Vertical Atmospheric Dis-
     persion Coefficients.  Atmospheric Environment, 6(11):859-
     863, 1972.
                              69

-------
22.  Tadmor, J.,  and Y. Gur.  Analytical Expressions for the
     Vertical and Lateral Dispersion Coefficients in Atmospheric
     Diffusion.  Atmospheric Environment, 3:688-689, 1969.

23.  Martin, D. 0., and J. A. Tikvart.  A General Atmospheric
     Diffusion Model for Estimating the Effects on Air Quality of
     One or More  Sources.  Presented at the 61st Annual Meeting
     of the Air Pollution Control Association, St. Paul, Minne-
     sota, June 23-27, 1968.  18 pp.

24.  Pasquill, F.  Atmospheric Dispersion Parameters in Gaussian
     Plume Modelling, Part II, Possible Requirements for Change
     in the Turner Workbook Values.  EPA-600/4-76-030b  (PB 258
     036), U.S. Environmental Protection Agency, Research Tri-
     angle Park,  North Carolina, June 1976.  43 pp.

25.  Code of Federal Regulations, Title 42—Public Health,
     Chapter IV--Environmental Portection Agency, Part 410—
     National Primary and Secondary Ambient Air Quality Stand-
     ards, April  28, 1971.  16 pp.

26.  Briggs, G. A.  Plume Rise.  AEC Critical Review Series.
     U.S. Atomic  Energy Commission, Division of Technical Infor-
     mation Extension, Oak Ridge, Tennessee, 1969.  81 pp.

27.  Moses, H., and M. R. Kraimer.  Plume Rise Determination--A
     New Technique Without Equations.  Journal of the Air Pollu-
     tion Control Association, 22 (8):621, 1972.

28.  Thomas, F. W., S. B. Carpenter, and W. C. Colbaugh.  Plume
     Rise Estimates for Electric Generating Stations.  Journal
     of the Air Pollution Control Association, 20(3):170, 1970.

29.  McGuire, T., and K.  C. Knoll.  Relationship Between Concen-
     trations of  Atmospheric Pollutants and Averaging Times.
     Atmospheric  Environment, 5(5):291, 1971.

30.  Handy, R., and A. Schindler.  Estimation of Permissible
     Concentrations of Pollutants for Continuous Exposure.  EPA-
     600/2-76-155 (PB 253 959), U.S. Environmental Protection
     Agency, Washington,  D.C., June 1976.  136 pp.

31.  Cornfield, J.  Carcinogenic Risk Assessment.  Science, 198
     (4318):693-699, 1977.

32.  Montgomery,  T. L., and J. H. Coleman.  Empirical Relation-
     ships Between Time-Averaged SOa Concentrations.  Environ-
     mental Science and Technology, 9 (10):953-957, 1975.
                              70

-------
33.  Hamil, H. F., and R. E. Thomas.  Collaborative Study of
     Method for Determination of Stack Gas Velocity and Volu-
     metric Flow Rate in Conjunction with EPA Method 5.  EPA-
     650/4-74-033 (PB 240 342), U.S. Environmental Protection
     Agency, Research Triangle Park, North Carolina, 1974.
     40 pp.

34.  Hamil, H. F., and R. E. Thomas.  Collaborative Study of
     Method for Stack Gas Analysis and Determination of Mois-
     ture Fraction with Use of Method 5.  EPA-650/4-73-026  (PB
     236 929), U.S.  Environmental Protection Agency, Research
     Triangle Park,  North Carolina, 1974.

35.  Smith, F., and J. Buchanan.  IERL-RTP Data Quality Manual.
     EPA-600/2-76-159, U.S. Environmental Protection Agency,
     Research Triangle Park, North Carolina, 1976.

36.  Hamil, H. F., and D. E. Camann.  Collaborative Study of
     Method for the Determination of Particulate Matter Emissions
     from Stationary Sources (Portland Cement Plants).   EPA-650/
     4-74-029  (PB 237 346), U.  S. Environmental Protection Agency,
     Research Triangle Park, North Carolina, 1974.  54  pp.

37.  Hamil, H. F., and R. E. Thomas.  Collaborative Study of
     Method for the Determination of Particulate Matter Emissions
     from Stationary Sources (Fossil Fuel-Fired Steam Generat-
     ors) .   EPA-650/4-74-021 (PB 234 150), U.S. Environmental
     Protection Agency, Research Triangle Park, North Carolina,
     1974.   36 pp.

38.  Hamil, H. F., and R. E. Thomas.  Collaborative Study of
     Method for the Determination of Particulate Matter Emissions
     from Stationary Srouces (Municipal Incinerators).   EPA-650/
     4-74-022  (PB 234 151), U.S. Environmental Protection Agency,
     Research Triangle Park, North Carolina, 1974.  37  pp.

39.  Hamil, H. F., and D. E. Camann.  Collaborative Study of
     Method for the Determination of Sulfur Dioxide Emissions
     from Stationary Sources (Fossil Fuel-Fired Steam Genera-
     tors) .  EPA-650/4-74-024  (PB 238 293), U.S. Environmental
     Protection Agency, Research Triangle Park, North Carolina,
     1973.   64 pp.

40.  Hamil, H. F., and D. E. Camann.  Collaborative Study of
     Method for the Determination of Nitrogen Oxide Emissions
     from Stationary sources (Fossil Fuel-Fired Steam Genera-
     tors) .  EPA-650/4-74-025  (PB 238 555), U.S. Environmental
     Protection Agency, Research Triangle Park, North Carolina,
     1973.   102 pp.
                               71

-------
41.  Hamil, H. F., and R. E. Thomas.  Collaborative Study of
     Method for the Determination of Nitrogen Oxide Emissions
     from Stationary Sources (Nitric Acid Plants).   EPA-650/4-74-
     028 (PB 236 930), U.S. Environmental Protection Agency,
     Research Triangle Park, North Carolina, 1974.   41 pp.

42.  Constant, P. C., G. Scheil, and M. C. Sharp.   Collaborative
     Study of Method 10—Reference Method for Determination of
     Carbon Monoxide Emissions from Stationary Sources—Report of
     Testing.  EPA-650/4-75-001 (PB 241 284), U.S.  Environmental
     Protection Agency, Research Triangle Park, North Carolina,
     1975.

43.  Constant, P. C., and M. C. Sharp.  Collaborative Study of
     Method 104—Reference Method for Determination of Beryllium
     Emissions from Stationary Sources.  EPA-650/4-74-023 (PB 245
     Oil),  U.S. Environmental Protection Agency, Research Tri-
     angle Park, North Carolina, 1974.  94 pp.

44.  Title 40—Protection of Environment.  Chapter 1—Environ-
     mental Protection Agency—Part 60—Standards of Performance
     for New Stationary Sources.  Subchapter C--Air Programs.
     Method 10—Determination of Carbon Monoxide Emissions from
     Stationary Sources.  Federal Register, 39 (47):9319-9321,
     1974.

45.  Guzewich, D. C., and W. J. B. Pringle.  Validation of the
     EPA-PTMTP Short-Term Gaussian Dispersion Model.  Journal
     of the Air Pollution Control Association, 27 (6) : 540-542,
     1977.

46.  Scriven, R. A.  Variability and Upper Bounds for Maximum
     Ground Level Concentrations.   Philosophical Transactions of
     the Royal Society of London,  Series A, 265:209-220, 1969.

47.  Weber, A. H.  Atmospheric Dispersion Parameters in Gaussian
     Plume Modelling, Part I, Review of Current Systems and Pos-
     sible Future Developments.  EPA-600/4-76-030a  (PB 257 893),
     U.S. Environmental Protection Agency, Research Triangle
     Park,  North Carolina, July 1976.  58 pp.

48.  Shiruaikar, V. V., and P.  R.  Patel.  Long Term Statistics of
     Peak/Mean Concentrations From A Point Sources.  Atmospheric
     Environment, 11 (4):387-389, 1977.

49.  Serth, R. W., and T. W. Hughes.   Source Assessment:  Carbon
     Black Manufacture.  EPA-600/2-77-107K, U.S. Environmental
     Protection Agency, Research Triangle Park, North Carolina,
     October 1977.  244 pp.
                              72

-------
50.   Standard for Metric Practice.  ANSI/ASTM Designation E
     380-76ฃ, IEEE Std 268-1976, American Society for Testing
     and Materials, Philadelphia, Pennsylvania, February 1976,
     37 pp.
                              73

-------
                           APPENDIX A

                ACCURACY,  ERROR,  AND UNCERTAINTY


In the analysis of errors  in  experimental data,  it is important
to distinguish the following  concepts:   1)  accuracy of the meas-
urement process, 2) error  in  the  measured value, and 3)  uncertain-
ty in our knowledge of  the true value which is measured.  The
relationship between these concepts is  indicated schematically in
Figure A-l.

  MEASUREMENT PROCESS       MEASURED VALUE        KNOWLEDGE OF TRUE VALUE

  ACCURACY	^-  (TOTAL) ERROR	^-  (TOTAL) UNCERTAINTY
  PRECISION	*-  RANDOM  ERROR	*-  RANDOM UNCERTAINTY
  BIAS	^-  SYSTEMATIC ERROR	^-  SYSTEMATIC UNCERTAINTY

      Figure A-l.  Schematic  representative of relationship
                   between accuracy, error, and uncertainty
                   and  between the  components  of each term.

The terms accuracy, precision, and  bias refer  to the measurement
process itself  (12).  Accuracy refers to  the closeness between
true and measured values which is characteristic of a given mea-
surement process.  Accuracy  (strictly speaking,  inaccuracy) is
composed of two parts:  precision (strictly speaking, imprecision)
and bias.  Precision is a  measure of the  closeness together, or
lack of scatter, in successive independent measurements when all
controllable variables  are held fixed.   Scatter in the measure-
ments is due to fluctuations  in values  of variables which are not
controlled in the measurement process.   Bias is the magnitude and
direction of the tendency  of  the  measurement process to measure
something other than what  is  intended.   Failure to maintain
isokinetic conditions during  particulate  sampling is a typical
example of bias in a measurement  process.

The inaccuracy of the measurement process results in errors in
the measured values.  The  total error in  a measurement is the
magnitude and sign of the  deviation of  the measured value from
the true value.
 (12) Eisenhart,  C.  Expression of the Uncertainties of Final
     Results.   Science,  160(3833):1201-1204 , 1968.


                                 74

-------
The  total  error  can  be decomposed into a random component and a
systematic component.   The random error in the measured value
results  from the imprecision of the measurement process, while
the  systematic error results from the bias in the measurement
process.   The decomposition of total error into random and
systematic components  is  illustrated in Figure A-2.  When the
mean of  the population of all possible measured values coincides
with the true value  in Figure A-2,  the measurement process is
said to  be unbiased.
       on
       OL
       O
       O
       O
       U-
       o
       >-
      s
      on
            TRUE
            VALUE
  MEAN OF ALL POSSIBLE
   MEASURED VALUES
                     -SYSTEMATIC ERROR-
  VALUE OF A SINGLE
    MEASUREMENT
 RANDOM
" ERROR "
                            -TOTAL ERROR
                            -MEASURED VALUES-
         Figure A-2.
Decomposition of total error into
random and systematic components.
In practice, the magnitude  of  the error in the measured value is
not known.  If  it were  known,  the measured value could simply be
corrected by the amount of  the error to obtain the true value.
The best that can usually be done is to place reasonable bounds
on the possible error.   In  the case  of random errors, the bounds
can be estimated statistically,  and  the results expressed in the
form of a confidence  interval  about  the measured value.  In the
case of systematic errors,  bounds must be estimated based on a
knowledge of the measurement process.   For example, in field
sampling work,  the measurement method  can be tested on standard
samples to determine  the bias  under  varying conditions.  The
largest value of the  bias so obtained  may then be used to esti-
mate a bound on the systematic error under field conditions.
Again, these results  can be expressed  in the form of an interval
about the measured value.

Thus, the estimation  of bounds for the error in a measured value
results in an interval  about the measured value within which the
true value can  reasonably be expected  to lie; i.e., an interval
of uncertainty.  Uncertainty,  then,  refers to our lack of know-
ledge of the true value of  a quantity  due to imperfect data;
                                 75

-------
i.e., due to error in the measured value.  The total uncertainty
can be decomposed into a random component, which is due to the
random error in the measured value, and a systematic component,
which is due to the systematic error in the measured value.  The
uncertainty is quantified by estimating bounds on the random and
systematic errors.


When measured values are used in a calculation, the corresponding
uncertainties result in an uncertainty of the calculated quantity
in relation to the true value.  The manner in which uncertainties
propagate through a calculation depends upon the functional rela-
tionship involved.  This topic is discussed in Appendix B and is
generally referred to as error propagation.  Strictly speaking,
however, it should be called propagation of uncertainty.  In this
report, then, we are concerned with uncertainty and its effect on
decision making.
                                76

-------
                           APPENDIX B

      USE AND INTERPRETATION OF ERROR PROPAGATION FORMULAS
Any experimentally determined quantity has associated with it
some degree of error for which bounds must be determined in order
for the measurement to be meaningful.  Indeed, a measured value
for which no error bounds can be determined is essentially use-
less.  When the measured value is subsequently used in a calcula-
tion, its associated error results in an error in the computed
value.  One is then faced with the problem of determining error
bounds on the computed value in terms of errors associated with
input values.

The manner in which errors (strictly speaking, uncertainties9)
propagate in a calculation depends upon the functional relation-
ship between variables and upon the type (random or systematic9)
of error involved.  In this appendix, the basic formulas for
propagating the two types of errors are reviewed, and a number
of example problems are presented which illustrate their use and
limitations.  Some brief derivations are included to point out
underlying assumptions and to indicate the extension to arbitrary
functional relationships.

As pointed out in Appendix A, the subject of the present section
should more properly be termed "propagation of uncertainty"
rather than "propagation of error."  However, in order to adhere
to convention as well as for conciseness of expression, the term
"error" is used in this section to denote both exact error (which
is useful in theoretical discussions) and uncertainty (which is
useed in practice).  The intended meaning will usually be clear
from the context.  When it is not clear, the meaning will be
indicated parenthetically.

RANDOM ERRORS

Error propagation formulas for random errors are listed in Table
B-l. for the four basic arithmetic operations.  These formulas are
special cases of the general formula also yiven in Table B-l.
The formulas give the confidence interval for the appropriate
function of two independent variables in terms of the confidence
intervals for the two independent variables.  The extension to
 See discussion in Appendix A.
                               77

-------
cases  involving more than two independent variables should  be
evident  from the form of the general  equation listed in the table.
These  formulas  as well as others are  given in various forms in
the literature  (see, for example, References 13, 14, and  15).   All
of the formulas are based on the following equation for the
variance of  a linear function of n  statistically independent
random variables (3):

         02y  = (cJVx! + (c2)202x2 +  .  .  .  + (cJ2o2xn      (B-l)


where y  = CiXi  + C?xo + . .  . + C x ,  and  the C.  are constants.
      j     i  i     *. *.            nn            i

If the X-L are not statistically independent,  then the equation
for o2   contains the covariances of the  x  variables and hence the
equations in  Table  B-l are not valid.  A derivation of the  error
propagation  formula for the  sum of two variables is first given,
and then  its  extension to nonlinear relationships is indicated.

    TABLE B-l.   ERROR PROPAGATION FORMULAS FOR RANDOM ERRORS

          Operation                  Error  propagation formula
    Addition, xi + x2        A + B ฑ \a2 + b2
     Subtraction, KI  - x2     A - B ฑ \a2 + b2
    Multiplication, XiX2     AB ฑ \B2a2 + A2b2
    Division, X!/x2           (A/B) ฑ y g^ + fir fa2
    General case:   f(x1,x2)   f(A,B)  ฑ^pg;B)] 2 a2 + [8f^B)]2 b2


    NOTE. — A ฑ a and B ฑ b are confidence intervals  for xl and x2 .
          The formulas give confidence intervals for  the various
          mathematical operations performed with Xi and x2 .  The
          formulas are valid only when A and B are  statistically
          independent.
(13) Braddick, H.  J.  J.   The Physics of Experimental Method.
     John Wiley  &  Sons,  Inc., New York, New  York,  1954.  404 pp.
(14) Volk, W.  Applied Statistics for Engineers.   McGraw-Hill
     Book Company,  New York, New York, 1958.   354  pp.

(15) Beers, Y.   Introduction to the Theory of  Error.  Second Edi-
     tion.  Addison-Wesley Publishing Co., Inc.,  Reading, Massa-
     chusetts, 1957.
                                78

-------
Derivation of Addition Formula

Let y = xj + x2 , where xฑ and x2  are  independent  variables  whose
measurements are subject to  random  errors.   The  (exact)  errors
are assumed to be normally distributed with  zero  means  and  vari-
ances o2Xi an^ 02X  .  Let A and B denote  the  estimated  means  of
xi and X2 determined  from n^ measurements  of  Xj  and  ng  measure-
ments of X2 .  Then  (i_ - a) x 100% confidence  intervals  for  the
true average values x^ and x2 are given by

                                            a
                                            xi
          xi  = A ฑ Z    .  a, = A ฑ  Z    ,   , - = A ฑ  a       (B-2)
           1        l-a/2  A        l-a/2
                                            a

             = B +~ Zi-a/2 aB = B  +-  Zi-a/2
where Zj_a/2 denotes the  (1 - a/2) percentage point  of  the  stand
ard normal distribution, where a2^ = a2x ,/n^ and  a2g =  o2x  /ng
are the variances of the two sample means A and B, and  where
                         b = Z    ,  -j=                     (B-5)
                               l-a/2  n
A  (i_- a) x 1(K)% confidence  interval  for  y  is  required.   Since
y = xj +_x2, y is normally distributed  (3),  and  the  sample esti-
mate of y is C = A + B.  Using Equation B-l,
                                             ฃ3
                                                             (B_6)
Thus, the  (i - a) x 100% confidence  interval  for  y  is
  = C * Zl-a/2 ฐc = A + B ฑ Z.
= A
                                          + B  ฑ Va2  +  b2      (B-7)
When the variances of a2Xl and o2X2 are not  known,  they  must  be
estimated by the sample variances
                                79

-------
          ,    nA                        .    nB
  2   —         A^   /       A \ 2    2   —    ^-     \T^  /       r> \ 2

                                      k

The confidence intervals  for xj and x2  are  then

                                          s.

                     A ฑ  t,   ,„  „
and

                                    '
                     1Q + 4-
where tl_GL/f  v denotes the  (1 - a/2) percentage point of  the
Student t distribution with v degrees of  freedom.   In this  case,
the above derivation is no longer valid because although

                              s2     s2
                                xi     X2
                        s2  = - + -                   (B-9)
                          C    nA     nB

is an unbiased estimate of o2Q, it is not a_sum-of-square esti-
mate.  Therefore, the random variable  (C  - y) /SQ will not follow
the t distribution.  (The special case when a2Xi =  a2x2 an<^
nA = nB is an exception.  See Example 3 below.)  Thus in  this
case, the error propagation formulas are  approximations which  are
accurate for large sample sizes n& and ng, for which the  esti-
mates s2x, and s*X2 approach the true variances cr2x and  02X2-
However, it might be expected that the formulas will not  be
greatly in error except for sample sizes  smaller than about 5,
for which the percentage points of the t  distribution deviate
markedly from those of the normal distribution.  This point is
examined further in the examples given below.

Nonlinear Relations

When y = f(xi,x2) and f is nonlinear, the above derivation  does
not apply because Equation B-l no longer  holds.  However, when
errors in measurements are small (or more precisely, when stand-
ard deviations aXl and aX2 are small compared with  mean values
Xi and x2), the problem can be linearized.  Thus, neglecting
second-order terms ,
  C E f(A,B) % f(Xi,x2) + 	—	(A - Xl) + 	^	(B  - x2)
3f(x1/x2)      _     3f(x1,x2)
                             (B -

                                 (B-10)


      eo

-------
 Since  this  expression  is  linear  in A  and  B  and  since  to  this
 degre_e of approximation the partial derivatives evaluated  at
 (xi,x2)  and (A,B)  are  equal, Equation B-l yields

                 2  _  pf(A,B)12  ,     pf(A,B)-|2 ,
               ฐ C ~  L 9*1  J ฐ A +  L  9x2   J  ฐ B          (E
For example, when y = f(x!,x2) = X}X2, then  9f(A,B)/9x1 = B and
9f(A,B)/9x2 = A, so that
The confidence interval for y is then
               a_ = AB ฑ Z,   ,  4/B2a2  + A2o2
                C          l-a/2 HA       B
                                   = AB ฑ N/B2a2 + A2b2      (B-13)

which is the multiplication formula given in Table B-l.   In this
manner, error propagation formulas can be derived for any  func-
tional relationship.  However, the formulas will be accurate  for
only small errors.

Non-Normally Distributed Errors

When error distributions deviate significantly from normality,
Chebyshev's theorem (3)  can be used to estimate confidence inter-
vals.  This theorem states that for a random variable, x,  which
has a mean, y, and standard deviation, a, the probability  that
|x - y| > ka is less than 1/k2; i.e.,


                      P(|x - y| > ka) < —                  (B-14)
                                        k2

Thus, a (i - a) x 100% confidence interval is given by

                             y ฑ ka


where  ka = ->—
            •a

Actually, the confidence level will be greater than  (i -  a) x 100%
since the above probability is strictly less than 1/k2.   The
important point is that this confidence interval is independent
of the distribution of the random variable, x.
If Chebyshev's theorem is used to compute all confidence  inter-
vals, the derivations of error propagation formulas given above
remain valid if ka is substituted for z    .  .  The formulas will
be only approximately correct if sample variances are used to
                                81

-------
 estimate  error variances or  if normal statistics are used to com-
 pute  one  confidence  interval and Chebyshev's theorem is used to
 compute the other.

 For a 95% confidence  interval,

                           l  - a = 0.95

                              a = 0.05

                             k  = 4.47
                              a

 This  value of ka is more than twice the corresponding value of
 z,   / =1.96 for a normally distributed variate.  For this
 reason, Chebyshev's theorem  is seldom used to compute confidence
 intervals.  It is recommended that it be used only when error
 distributions are known to be highly non-normal.

 Examples;  Random Errors

 Example 1—
 The following data on atmospheric emission factors for ammonia
 emitted from ammoniation and granulation plants were obtained
 from  various published sources.

   TABLE B-2.   AMMONIA EMISSIONS  FROM AMMONIUM NITRATE PLANTS
                                      Ammoniator-
                                       granulator
Dryer and
 cooler
 Number of measurements,  n                      7               16
 Mean emission factor,  g/kg                 0.503            0.316
 s,  g/kg                                    0.564            0.262
 s//n                                       0.213            0.065
 to.975/ n - 1                              2,447            2.131
 95% Confidence interval  for mean   ฑ0.521 (ฑ104%)    ฑ0.139 (ฑ44%)
The emission factor for the entire plant is the sum of the values
for the ammoniator-granulator and the dryer and cooler.  We wish
to compute the confidence interval for the emission factor for
the entire plant and to compare the result obtained using the
error propagation formula with that obtained by rigorous statis-
tical methods.

   • Using the error propagation formula,

           A = 0.503, a = 0.521, B = 0.316, b = 0.139

Hence,
                               82

-------
                        C = A + B =  0.819
              c = \a2 + b2  = V0.271  +  0.019  =  0.539

     The 95% confidence interval  is  thus  0.819  ฑ  0.539, or
     0.819 ฑ 66%.

   • The confidence interval can  be  approximated  using
     the method of Welch (16, 17).   Actually,  an  approxi-
     mation to Welch's method is  used  which  is  valid when
     sample sizes are not too small  (18).  The  following
     quantities are computed:
                   VA E
                          =  0.2132  =  0.0454
                    B ~  n
                          B
                             =  0.0652  =  0.0042
f =
         V_, = 0.2228 = "effective"  standard  deviation  of mean
          13
           (V
          V 2       V 2
         -X-+    B
                 "
                  B
                         -  2  = "effective"  degrees  of  freedom
   f = 8.37 - 2 = 6.37

     The confidence interval  is  computed  using  the  t distribu-
     tion with f '  degrees  of  freedom,  f  being  the  nearest
     integer to f,  or f = 6  in  this case.  Thus,
        = t0. 975, 6
                               =  2.447(0.2228)  =  0.545
(16)  Pearson,  E.  S.,  and  H.  0.  Hartley.  Biometrika Tables  for
     Statisticians,  Third Edition, Volume  1.  Cambridge Univer-
     sity Press,  New York, New  York,  1966.   264 pp.

(17)  Welch,  B.  L.   The  Generalization of "Student's" Problem When
     Several Different  Population Variances  Are Involved.   Bio-
     metrika,  34:28-35, 1947.

(18)  Natrella,  M^  G.  Experimental Statistics.  National Bureau
     of  Standards  Handbook 91,  U.S. Department of Commerce,
     Bureau  of  Standards,  Washington, D.C.,  1963.  504 pp.
                               83

-------
     The  95%  confidence  interval  is  then 0.819 ฑ 0.545 or
     0.819  ฑ  67%.   The agreement  with the previous calcula-
     tion is  excellent.

     Suppose  that  it  is  known  a priori that the variances
     o2Xl and o2x?  of the  two  error  distributions are equal,
      (Actually, using the  F-test,  the difference in the two
     sample variances is not significant at  the  0.05 level,
     but  is significant  at the 0.10  level.)   In  this case,
     the  confidence interval can  be  computed rigorously
     using  the t distribution.  The  method is essentially
     the  same as that employed for comparing two means (21),
     The  pooled estimate of the unknown  variance is com-
     puted  as follows:
(
                   nA -
                                            x2
                                   -  2
                                               -  0.1399
     The estimated standard deviation  of  the  mean  is then
                nA + l/nfi = 0.3741J1/7  +  1/16  =  0.1695


     The confidence limit is given by
10 . 9 7 5 / nA + n~ ~ 2
  B
                                     1/nB  =  t 0 . 9 7 5 , 2 1 (0 . 1695)


                                          =  2.080(0.1695)

                                          =  0.353
     Hence, the confidence interval  is  0.819  ฑ  0.353  or
     0.819 ฑ 43%.  It is seen that the  additional  information
     about the error variances can be used to obtain  a
     smaller confidence interval.

Example 2 —
The data are similar to those of Example  1, but the emission
factors are for particulate matter rather than  ammonia.

 TABLE B-3.  PARTICULATE EMISSIONS FROM AMMONIUM NITRATE PLANTS
                                   Aminoniator-   Dryer and
                                    granulator    cooler
Number of measurement, n
Mean emission factor, g/kg
s, g/kg
s//n
^0. 975' n ~ 1
95% Confidence limit for mean
2
0.175
0.070
0.049
12.706
0.623(356%)
12
0.230
0.173
0.050
2.201
0.110(48%)
                               84

-------
This example represents an extreme case since only two measure-
ments are available for the ammoniator-granulator.

   • Using the error propagation formula,

           A = 0.175, a = 0.623, B = 0.230, b = 0.110

     Hence,

                        C = A + B = 0.405
                      c = Va2 + b2 = 0.632

     The 95% confidence interval is thus 0.405 ฑ 0.632 or
     0,405 ฑ 156%.

     The approximation to Welch's method used in Example 1 is
     not valid in this case due to the small sample size
     involved (nA = 2).  On the other hand, the exact version
     of the method cannot be used because the percentage
     points of the appropriate distribution function have not
     been tabulated for degrees of freedom less than 8 (19).
     The approximate calculation is included here for
     completeness.
                       Va = —— = 0.00245
                        **    n—
VB =

                                 - 0.00250
           f =
                          'A + VB = 0.0704
                  (VA + VB J2
                  V
               LnA + 1 + nfi -f 1
     Therefore,
         - 2 = 9.88 - 2 = 7.88
                             f •  = 8
            to.975,8\/VA + VB = 2.306(0.0704)  = 0.162
                               85

-------
     Hence, the confidence interval is 0.405 ฑ 0.162 or
     0.405 ฑ 40%.  The agreement here is quite poor, but no
     conclusion can be made since the accuracy of the present
     calculation is unknown.

     It is assumed that a2x  = o2X2.  (Using the F test, the
     difference in the sample variances is not significant
     even at the 0.25 significance level.)  Then

                     - ls2   +  n  - l\s2
}s2   + (n^ - l\
)   xj    (  B    )
                                 - 2
                                          x2

                                             = ฐ-0278
                             + l/nB = 0.0162


       to.975' nA + nB - 2 sp^l/nA + l/nfi  = 2.179(0.0162)

                                           = 0.277

     Hence, the confidence interval is 0.405 ฑ 0.277 or
     0.405 ฑ 68%.  Since the difference in the sample vari-
     ances is not significant at the 0.25 level, the present
     calculation is probably the most accurate of the three
     methods in this case.  Thus, it appears that error prop-
     agation formulas can yield very conservative error
     estimates when small sample sizes are involved.

Example 3—
It is desirable to investigate further the behavior of the
formulas for small sample sizes since repeating an experiment
more than once or twice is often impractical.  As previously
noted, when y = xj + x2, the random variable

                              C - y
                                SC

follows the t distribution with 2(n - 1)  degrees of freedom in
the special case when o2x  = a2X2 an^ n^ = nB = n.  Hence, the
confidence limits for y are given by


                       * tl-a/2, 2(n-i)  SC

On the other hand, the addition formula yields the confidence
limits

                         * tl-a/2, n-i SC

The percent difference is thus
                               86

-------
          Vl  SC  -
               t2(n-l)  sc                t2(n-l)

where the confidence level subscripts have been dropped for con-
venience.  This  factor  is tabulated below for the 95% and 99%
 (two-sided) confidence  levels  (i - a/2 = 0.975 and 0.995).

     TABLE B-4.  PERCENT DIFFERENCES IN SIZES OF APPROXIMATE
                 AND EXACT CONFIDENCE INTERVALS




n
2
3
4
5
10
t , - t_ /.
n-l 2 0
t2(n-i)
95% Level
191
55
30
25
8

n~ ( too )

99% Level
541
116
58
37
13

It is clear from these values that error propagation formulas can
yield very conservative results for smaller sample sizes.  How-
ever, at the 95% level — the one most often used in practice — the
difference does not exceed 25% for n = 5.  A conservative error
of this magnitude in the confidence limits should be acceptable
for most purposes.  In effect, the calculated confidence interval
simply corresponds to a slightly higher significance level than
the nominal (95%) level.

When a2X! * a?-X2, the accuracy of the error propagation formulas
can be estimated by means of the Behrens-Fisher confidence inter-
val, dn_, sc (19).  Aside from philosophical considerations, the
Behrens-Fisher method may be disputed on the grounds that the
actual confidence level is not exactly equal to the nominal level,
but varies with the ratio  ฐx\/ax2  (16).  However, the method
should give a good indication of the accuracy of the error formu-
las.  The percent error in the addition formula is given by


                        Vl " dn-i
                            n-i

where dn_ l is tabulated in Table VI of Reference 19.  Since
depends on SA/SB = tan 0, the extreme value of dn-1 for
(19) Fisher, R. A., and F. Yates.  Statistical Tables for
     Biological, Agricultural, and Medical Research.  Oliver and
     Boyd, London, 1963.  356 pp.

                                87

-------
0ฐ < 0 < 90ฐ is used to estimate the maximum error.   The  values
are~listed below for the 95% and 99% levels.

       TABLE B-5.  MAXIMUM PERCENT DIFFERENCES  IN  SIZES OF
                   APPROXIMATE AND EXACT CONFIDENCE
                   INTERVALS WHEN ERROR VARIANCES  ARE UNEQUAL





n
2
4
6
t , - d
n-i r
d
n-i
95% Level
-29
-1.8
+0.35

1—1 / -I nn\
\ J.UU )

99% Level
-29
+4.3
+5.8

Significant errors occur for only the smallest  sample  sizes;
however, the errors need not be conservative.

Example 4 —
Derive error propagation formulas for the exponential  and  loga-
rithmic operations.  For the exponential operation,


                          y = f(x) = xn                     (B-15)


                          f (x) = nxn-1                     (B-16)

                          C = f (A) = An                     (B-17)

and

                            x = A ฑ a                       (B-18)

where  a = z,    ,  a.
            l-a/2  A

Following the linearization procedure described above  under
"Nonlinear Relations,"

                       ac2 = [f'(A)]2(aA2)                   (B-19)


                          ac = f ' (A)aA                      (B-20)
                             ฑ Zl-a/2(ac)


                               88

-------
                     Y = A" ฑ  Z!-a/2  ***  ^A)               (B-23)
                     y  =  An  ฑ  nAn~l(a)                       (B-24)

 Symbolically,  this  relationship  may  be  written  as

                      (A  ฑ a)n =  An ฑ nhn~l(a)               (B-25)

 For  the  logarithmic operation,

                         y = f (x)  = ln(x)                    (B-26)

                            f ' (x)  = 1/x                      (B-27)

                         C = f(A)  = In (A)                    (B-28)

 Linearizing as before,

                       ac2 =  [f (A)]2(aA2)                  (B-29)


                       ac = f (A)aA = io                    (B-30)
Therefore,


                     ^ =C  +-  Zi-a/2(ac)                     (B


                       = in (A) ฑ Zi_a/2 1  (aA)              (B-32)


                     y = In (A) ฑ |                          (B-33)

Symbolically, this result may be written


                      In (A  ฑ a) = ln(AJฑ |                 (B-34)

The above formulas may also be obtained as special cases  of  the
general error propagation formula given in Table B-l.

SYSTEMATIC ERRORS

Error propagation formulas  for systematic  errors are  given  in
Table B-6 for the four basic arithmetic operations.   The  first
two formulas are valid in general, while the  latter two are
restricted by the conditions a < |A| and b <  |B|.  In all cases,
the two variables as well as the corresponding errors are assumed
to be functionally independent.  The first two formulas together
with the linearized versions of the multiplication and division

                                89

-------
 formulas have been reported in the literature  (20).   A derivation
 of  the multiplication  formula is given here  to  demonstrate the
 method.

  TABLE  B-6.  ERROR PROPAGATION FORMULAS FOR SYSTEMATIC ERRORS


  Operation	Lower bound	Upper bound	

 Addition       A + B - (a + b)               A + B +  (a + b)

 Subtraction    A - B - (a + b)               A - B +  (a + b)

 Multiplication AB + sgn(AB)ab -  (a|B| +b|A|)  AB+ sgn(AB)ab +  (aJB| + b|A|)

  ...         A     atsl  + b|A|             A    alel + bJAl
 Division	'—'	'—'	           — +
              B   _o .     ,._,, i_ i           B
B2 + sgn(AB)b|B|          B   B2  - sgn(AB)b|B|
NOTE.—A ฑ a and Bib are error bounds for x\ and x2 •  The formulas give
upper  and lower error bounds for the four basic mathematical operations
performed with xi and Xฃ•  The formulas are valid only when xj and X2 are
functionally independent variables.

Let y =  xiX2  where KI and Xฃ are  independent variables; let A
and B represent the measured values of KI and X2;  let  e^ and eg
signify  the (exact)  errors associated with A and  B;  and let a
and b stand for upper bounds on |e,|  and  |eBl-  Then

                    y = Xlx2 =  (A  + EA)(B + eB)              (B-35)

                     y = AB + eAB  + EBA + EAEB               (B-36)


Letting  y  = AB + c,u, the  (exact)  error in the product enn is
  •    i             f\D                                      f\D
given by


                      ฃAB = ฃAB +  ฃBA + ฃAฃB                 (B-37)

Since in this  case there is no statistical basis  for a partial
cancellation  of errors, a worst-case analysis is  the only
recourse.   Hence, the maximum  and minimum values  of  eaR for
-a <  e,  <  a and -b < ฃn < b must  be determined.
   —  A  •*           —  D —

Assuming a <  |A| and b < |B|,  the results are readily  found to be
as illustrated in Table A-7.

The four cases can be combined into the single formula
sgn(AB)ab  ฑ (a|B| + b|A|), where  sgn(AB)  denotes  the algebraic
sign  of  the product AB.
(20) Jenson,  V.  G.,  and G. V.  Jeffreys.  Mathematical Methods in
     Chemical Engineering.  Academic Press, New York, New York,
     1963,  pp.  356-360.

                                90

-------
     TABLE  B-7.   MAXIMUM AND MINIMUM ERRORS IN THE PRODUCT AB


               I        I    ฃ,._  Maximum     ฃ,._. Minimum
           sgn  A    sgn  B	AB    	AB	

             +        +      ab +  aB + bA    ab - aB - bA
                           ab -  aB - bA    ab + aB + bA
             +        -      -ab - aB + bA   -ab + aB - bA
                     +      -ab + aB - bA   -ab - aB + bA
When a and  b  are  small  relative to |A|  and |B|, the product ab
can be neglected  and  the  formula reduced to

                      |eAB|max = a|B|  +  b|A|                 (B-38)

Dividing by |AB|  gives
                         AB
                                                            (B-39,
which states  that  the  relative  absolute errors are additive.
This is the form given in  Reference 20.  It can be derived more
easily by  linearizing  the  problem at the outset.  Thus, to terms
of first order,
  y = f (Xl, x2)    f(Af  B)  +     tr-(xl  - A)  +     x^ * B>

                                                            (B-40)


        y - f (A/ B) -  8f(B)(x1  -  A)  + ^(x2 - B)   (B-41)
                 e   =  9f(A, B)        3f(A,  B)
                 EAB       3xi    A        9x2   ฃB          (B-42)
Taking f(xi, x2) =
                          ฃAB =  BฃA  +  AฃB
Thus,

                      (ฃAB)max =  lBla  +  !A'b                (B-44)


                      (ฃAB)min =  -'Bla -  'Alb               (B-45)

or

                              =  alBl  + b'Al                 (B-46)

                                91

-------
From Equation B-42, the general formula for computing linearized
systematic error bounds is
              f(A, B)
          ฑ[
                          9f(A, B)
    • 1
        a +
             3f(A, B)
                                           9x2
For operations more complex than those given in Table B-6, error
bounds are best determined by the above linearization procedure
if errors are small or by direct substitution if errors are large.
In the latter case, however, the bounds on y cannot generally be
obtained by substituting the corresponding bounds on the x's into
the function f.  In fact, determining the bounds on y rigor-
ously requires solving two nonlinear programming problems, namely,
            Max



and

            Min

       eA' V  '

subject to
                   N
                         Y =
                                       B
                               eB,...,N
                                (B-47)
                   N
= f (
-a < e  < a, -b <
Examples;  Systematic Errors
                   A + eA, B + eB,...,N +
                                 < b,...,-n < e  < n
                                                            (B-48)
                                (B-49)
Example 5—
For comparison, Example 1 is reworked assuming that the errors
are systematic rather than random.  Taking A = 0.503, B = 0.316,
a = 0.521, and b = 0.139, the bounds on the sum are given by

                 A + B ฑ (a + b) = 0.819 ฑ 0.660

                                 = 0.819 ฑ 81%

Of course, the systematic error bounds will always be larger than
those for random errors because
           Va2 + b2 < a + b
                                                            (B-50)
according to the triangle inequality.  A similar relationship can
be demonstrated for the (linearized) multiplication and division
formulas.
                               92

-------
Example  6—
The  specific reaction rate of a first order chemical reaction  is
3.0  x  10~7 ฑ 50% sec"1, and the initial reactant concentration is
0.1  ฑ  20% kg-mole/m3.   It is desired to compute error bounds on
the  initial reaction rate YO = ^co assuming that the errors are
systematic.  For this problem, A = 3 x 10~', a = 1.5 x 10~7,
B =  0.1, and b = 0.02.

Using  the multiplication formulas in Table B-6,

Lower  bound—

(3 x 10~8) + (3 x 10~9) - (1.5 x 10~8 + 6 x 10~9)

                                    =  (3.3 x 10~8) -  (2.1 x 10~8)

                                    = 1.2 x 10~8

Upper  bound—

             (3.3 x 10~8) + (2.1 x 10~8) = 5.4 x 10~8

Thus,

            1.2 x 10~8  < (Y  = 3 x 10~8) < 5.4 x 10~8


or

             YQ = 3 x 10-8 ft 2-4 x 10Z8 kg_mole/m3/s


Using  the linearized version of the formulas yields

            0.9 x 10~8  < (Y  = 3 x 10~8) < 5.1 x 10~8


or

             YQ = 3 x 10~8 ฑ 2.1 x 10~8 kg-mole/m3/s

This represents an error of 25% in the lower bound and 6% in the
upper  bound.   However,  the estimate for the lower bound is con-
servative, while that for the upper bound is not.

Unsymmetrical Systematic Errors

The formulas given in Table B-6 assume that the upper and lower
error  bounds are symmetrical about the nominal values of the
variables.  This is often an unrealistic assumption for system-
atic errors.   For example,  the algebraic sign of a systematic
error may be known even though the magnitude can only be esti-
mated.   Unsymmetrical error bounds can also arise from physical


                               93

-------
restrictions on the variables;  e.g., the variables  may be
restricted to non-negative values.   In addition,  use  of the
multiplication or division formulas in Table B-6  will result in
unsymmetrical bounds.  Thus,  if a sequence of calculations is to
be performed, formulas for variables having unsymmetrical error
bounds  are required.

Error propagation formulas are  given in Table B-8 for the case  in
which
                         ~ ai 1 xi  <  A
                                          a2
                                                              (B-51)
where

and



where
            a2 > 0
                       B -
                               <  x2  < B + b2
                                                              (B-52)
       bj, b2  >  0

           TABLE B-8.
                        ERROR  PROPAGATION FORMULAS  FOR
                        UNSYMMETRICAL SYSTEMATIC  ERRORS
Operation
Addition
Subtraction
Multiplication




A
A
AB


B
Lower bound
+ B - (ai + bj)
- B - (aj + b2)
+ sgn(AB)a]Bi - (aiJB| + BI 1

-------
Examples;  Unsymmetrical  Systematic  Errors

Example  7 —
In  Example 5, A =  0.503,  a =  0.521,  B =  0.316,  and b =  0.139.
Thus, the lower bound on  KI was  0.503 -  0.521 = -0.018.   However,
xj  and X2 represent real  emission  factors and hence cannot be
negative.  Thus, it is more realistic to take HI = 0.503,
a2  = 0.521,  and bi = b2 = 0.139.   Using  the addition formula
from Table B-8 yields

Lower bound —

                A  + B -  (a! + b!)  =  0.819 - 0.642

                                   =  0.177

Upper bound —

                A  + B +  (a2 + b2)  =  0.819 + 0.660

                                   =  1.479

In  Example 5, a lower bound of 0.159 was obtained for the sum.
However, the present value of 0.177  is still more conservative
than the value of  0.280 obtained in  Example 1 using the  formulas
for random errors.

Example 8 —
Obtain an error propagation formula  for  unsymmetrical systematic
errors for a functional relationship of  the form.
                       y = f(Xl, x2) = -                  (B-53)
where  Xj, x2 > 0
Using the linearization procedure described previously, start
with Equation B-42,
                     _ 3f(A, B) ,  .     3f(Af B)
                          R     'e-^  +     s     \ BI


                                                            (B-55)
ea/ป2 = "^'V^  (ฃA)  + "XV:."/ (ej         (B-54)

               1 / ' v    2 A
                                       B-
where -aj < e  < a2 and -bi < e  < b2 .  It is assumed that  the
measured values, A and B, are both positive.  Hence, it  is
readily seen that
                     (ฃA/B2)max =  7+   T(bi>               (B-56)
                                 95

-------
                                  B2     B3

These equations can be expressed  symbolically as  follows:
                                                            (B-57)
                                   a2
                                        2  A

                                        B3
                            (bi)
                     b2
             B2
                                                            (B-58)
                 B
                                  al   o A
                                  — + — (b2)
                                  B2    B3
Since both variables are positive,  it  is  a  simple  matter to
obtain the exact error bounds  (as opposed to  the linearized
bounds) by direct substitution.  Thus,  the  upper bound is
obtained by substituting the largest value  for  the numerator and
the smallest value for the denominator.
Upper bound--
A +
(B - 1
Lower
ai A(l + a2/A) f 1 + a2/A
A i L n
Dl)2 B2
A
B2
A
B2
bound —
(1 - bi/B)2 B2f (1 - bi/B)2 "
(1 + a2/A) - (1 - 2 bi/B + bi2/B2)
-\ ฑ
X T
a2/A + 2 bi/B - bi2/B2"
T a 	

 A -
             A(l - ai/A)     A

(B + b2)2    B2(l + b2/B)2   B2
                       - ai/A
                                                -  1
            A

            B2


            A

            B2
1 +
1 -
       - ai/A) -
                                      + b2/B)2


                                     2 b2/B  + b22/B2)
                              (1 + b2/B)2


                           2 b2/B + b22/B2)

                            + b2/B)2
                                                            (B-59)
                                                            (B-60)
                                                            (B-61)
                                             (B-62)
                                                            (B-63)
                                             (B-64)
Thus,
                                96

-------
              "A/B2 max    2
              :A/B2 min
                             a2/A + 2
                            ~ai/A + 2 b2/B + b22/B
                                2/n2
                                       b2/B)
                                                           (B-65)
                                            (B-66)
These results reduce to those obtained using the linearization
procedure when the error in the denominator is small; i.e., when
 L/B
     ซ 1 and b
2/B
    ซ 1.
CONCLUSIONS
The assumptions underlying the error propagation formulas for
random errors can be summarized as follows:

   • The errors are normally distributed.  (This assumption
     is required if normal statistics are used to compute
     confidence intervals.)

   • The errors are statistically independent.

   • When the error variances are estimated from the measured
     data, the formulas are valid for large sample sizes.  In
     practice, a sample size of five should be adequate for
     most purposes.

   • For nonlinear relations, the formulas are valid for
     small errors only; i.e., a ซ  |A| and b ซ  |B|.

For systematic errors, the only assumptions are that the vari-
ables are functionally independent and, in the case of the multi-
plication and division formulas, a <  |A| and b <  |B|.  The
linearized versions of these equations are accurate for small
errors; i.e., a ซ |A| and b ซ |B|.

When both types of errors are important in a calculation, they
should be treated separately.  The total error in the calculated
value is then the sum of the random and systematic errors  (15).
When the type of error is not known, as is sometimes the case
with data obtained from the literature, the formulas for system-
atic errors should be used since they yield more conservative
error bounds.  The systematic error formulas can also be used as
a conservative approximation for propagation of  random errors
when the assumptions underlying the formulas for random errors
are invalid.  For the addition and  subtraction of quantities sub-
ject to random errors, sharper bounds can be obtained by using
the appropriate statistical method  when  the error variances are
known to be equal.
                                97

-------
                           APPENDIX C

             DERIVATION OF SOURCE SEVERITY EQUATIONS
SUMMARY OF EQUATIONS

The source severity of pollutants can be calculated using the
mass emission rate, Q, the effective height of the emissions,
H, and the primary ambient air quality standard, PAAQS, or the
threshold limit value, TLV.  The severity equations shown in
Table C-l are derived in this appendix.9

   TABLE C-l.  SOURCE SEVERITY EQUATIONS FOR ELEVATED SOURCES
                 Pollutant _ Severity equation

             Particulate matter      S = - —
                                          H2
             SO                      S =
               x
             NO
             Hydrocarbons            S = 162 Q
                                           H2
             CO                      S =
                                           H
             Others                  S =  5'5 Q
                                         TLV(H2)
DERIVATION OF EQUATION FOR x =
                            luclX

The most widely accepted formula for predicting time-averaged
ground level concentrations downwind from a point source is the
generalized Gaussian dispersion equation  (5, 6).
 For convenience, the subscript "C" for "calculated" severity is
 omitted in this appendix.

                                98

-------
              x = ™-FTT ฃXPh ^'"i  ISXP|- Tf^l  I         
-------
                           2	— = 0                       (C-7)
                               V
or
                          Oy . Oz . _s_


Substituting this result back into Equation C-2 yields the
desired result,

The validity of the assumption ay = az can be ascertained by
referring to Tables C-2 and C-3 121-24) .   These tables give the
functional dependence of a  and oz on downwind distance.  It can
be seen that for stability^Class C, Equation C-5 is approximately
satisfied.

The averaging (i.e., sampling) time associated with Equation C-9
is approximately 3 min.a  For averaging times between 3 min and
24 hr, a semiempirical correction factor given by Turner (5) is
applied to Equation c-9.b
 The two sets of values in Tables C-2 and C-3 are not entirely
 consistent with respect to averaging time.  The values for ay in
 Table C-2 correspond to an averaging time of 3 min.  The values
 for oz in Table C-3 correspond to averaging times in excess of
 some limiting value, rm, which is approximately proportional to
 emission height for emission heights up to 100 m, at which
 height and above rm is approximately 10 min (24).  Thus, Turner
 (5)  states that concentrations calculated using these values for
 Oy and az correspond to an averaging time between 3 min and
 ID min.

 The correction factor is applicable to situations in which the
 mean wind direction remains constant during the period of
 interest.  That is, as the averaging time increases, the width
 of the plume increases, but the position of the plume centerline
 does not change.  Since the mean wind direction does not, in
 general, remain constant over an extended period of time (e.g.,
 24 hr), this correction factor corresponds to  a worst case
 situation with respect to pollutant concentration.
(21)  Eimutis,  E.  C.,  and M.  G.  Kcnicek.   Derivations of Continu-
     ous Functions for the Lateral and Vertical Atmospheric
     Dispersion Coefficients.   Atmospheric Environment, 6(11):
     859-863,  1972.

(22)  Tadmor, J.,  and  Y.  Gur.  Analytical Expressions for the

                                                     (continued)

                               100

-------
TABLE C-2.   VALUES OF a FOR THE  COMPUTATION OF  a  a  (21)

Stability class
A
B
C
D
E
F
a
0.3658
0.2751
0.2089
0.1471
0.1046
0.0722

                  For the equation
                 where  x = downwind distance
                       b = 0.9031  (from Reference 40)
    TABLE C-3.
VALUES OF  THE CONSTANTS USED TO
ESTIMATE VERTICAL DISPERSION3 (23)

Usable range, Stability
m class
>1,000 A
B
C
D
E
F

100 to 1,000 A
B
C
D
E
F
<100 A
B
C
D
E
F
0.
0.
0.
1.
6.
18.

0.
0.
0.
0.
0.
0.
0.
0.
0.
0.
0.
0.
Coefficient
00024
055
113
26
73
05
C2
0015
028
113
222
211
086
192
156
116
079
063
053
2
1
0
0
0
0

1
1
0
0
0
0
0
0
0
0
0
0
.094
.098
.911
.516
.305
.18
d2
.941
.149
.911
.725
.678
.74
.936
.922
.905
.881
.871
.814
-9
2
0
-13
-34
-48
f
9
3
0
-1
-1
-0
0
0
0
0
0
0
.6
.0
.0


.6
2
.27
.3
.0
.7
.3
.35
3






For the equation
a

= cxd +

f








                           101

-------
                       -          /t  \ฐ•17
                       y    _ X   I  O I
                        max    max\t  /

or
                                       ,0.17
                                                            (C-ll)
where    t  = "short-term" averaging time; i.e.,  3 min

          t = averaging time of interest

       X    = maximum ground level concentration  corresponding to
              averaging time, t

DEVELOPMENT OF SOURCE SEVERITY EQUATIONS

Source severity, S, has been defined as follows:


                            S = ^|K                        (C-12)


where  x    is given by Equation C-ll with u = 4.5 m/s
        ITlclX

and where  F = PAAQS for criteria pollutants
             = TLV/300 for noncriteria pollutants

The averaging time associated with )^max is 24 hr  for noncriteria
pollutants;  for criteria pollutants, it is the value specified in
the corresponding PAAQS, as shown in Table C-4 (25).
(continued)
     Vertical and Lateral Dispersion Coefficients in Atmospheric
     Diffusion.  Atmospheric Environment, 3:688-689, 1969.

(23) Martin, D. 0., and J. A. Tikvart.  A General Atmospheric
     Diffusion Model for Estimating the Effects on Air Quality of
     One or More Sources.  Presented at the 61st Annual Meeting
     of the Air Pollution Control Association, St. Paul, Minne-
     sota, June 23-27, 1968.  18 pp.
(24) Pasquill, F.  Atmospheric Dispersion Parameters in Gaussian
     Plume Modelling, Part II, Possible Requirements for Change
     in the Turner Workbook Values.  EPA-600/4-76-030b  (PB 258
     036), U.S. Environmental Protection Agency, Research Tri-
     angle Park, North Carolina, June 1976.  43 pp.

(25) Code of Federal Regulations, Title 42—Public Health,
     Chapter IV—Environmental Protection Agency, Part 410—
     National Primary and Secondary Ambient Air Quality Stand-
     ards, April 28, 1971.  16 pp.

                               102

-------
TABLE C-4.   SUMMARY OF  NATIONAL  AMBIENT AIR QUALITY STANDARDS (25)

Pollutant
Particulate


SO
X



CO

Nitrogen dioxide


Photochemical oxidants
d
Hydrocarbons (nonmethane)

Averaging time,
hr
Annual
(geometric mean)
24b
Annual
(arithmetic mean)

24
3
8b
1
Annual
(arithmetic mean)
h
1ฐ

3
(6 a.m. to 9 a.m.)
Primary
standards

75

260
80
24

365

10,000
40,000
100


160

160

yg/m3



(0.03)


(0.14)
None.
(9)
(35)
(0.05)


(0.08)

(0.24)

Secondary
standards
(ppm)
60a

150
60
365
c
260
1,300

40,000
100


160

160





(0.02)
(0.14)

(0.1)
(0.5)
None.
(35)
(0.05)


(0.08)

(0.24)


a                                    0
 The secondary annual  standard (60 yg/m3) is a guide for assessing implementa-
 tion plans to achieve the 24-hr secondary standard.
 Not to be exceeded more than once per year.
C                                     3
 The secondary annual  standard (260 yg/m3) is a guide for assessing implemen-
 tation plans to achieve the annual standard.
d
 Recommended guideline for meeting the primary ambient air quality standard
 for photochemical oxidants.

CO  Severity

The  primary standard for CO  is  reported  for a 1-hr  averaging time.
Therefore,

                              t  =  60 min
                             t   =  3  min
x
 max
                           xmax\60/
                        =   2 Q
                           ireuH2
              (-)0'1'
              \60/
                                    2  Q
                           (3.14) (2.72) (4.5)H2
                                                -(0.6)
(C-13)


(C-14)


(C-15)
                                  103

-------

                           -  (3'12 *210'2)Q


Setting F equal to the primary standard for CO, i.e., 0.04 g/m3 ,
yields


                    c _ Xmax _  (3.12 x 10"2)Q               ,_  1Q>
                    S = — -- -               (C-18)
                         *         0.04 H2

or
                                 H2

Hydrocarbon Severity

For nonmethane hydrocarbons, a 3-hr averaging time  is used.

                           t = 180 min


                      -    _     / 3  \ฐ- 17
                      xmax ~ Xmaxll86/                      (C-20)
                           =  (0.5) (0.052)Q                  (C-22)

                                   H2
                      ^max       2


For nonmethane hydrocarbons, the concentration of 1.6 x  10~k  g/m3
has been established as a guide for achieving oxidant standards.
Therefore,


                       S =    ฐ'ฐ26 Q	                   (C-24)
                            (1.6 x 10~U)H2

or
                           s        _                       (c_25)
                                  H2
                                104

-------
Particulate Severity

The primary standard for particulate matter  is reported  for  a
24-hr averaging time.  Therefore,

                      -    _      (3   \ฐ-17
                      xmax " xmax\l,440/                    (C-26)

                           =  (0.052)Q(0.35)                 (C-27)
                                    H2
The primary standard for particulate matter  is  2.6 x  10  **  g/m3 .
Thus,


                       S =    ฐ-ฐ182 Q                      (C-29)
                            (2.6 x 10~1+)H2
or
                            S     -                         (C-30)
                                 H2
SO  Severity
  X
The primary standard for SOX is 3.65 x 10"4 g/m3  for an averaging
time of 24 hr.  Thus, proceeding as before,
                          max
                          max       2


                       Xmax      0.0182 Q
                                                            (c_32)
                               (3.65 x
or
                            S                               (c_33)
                                 H2
N02 Severity
Since NO 2 has a primary standard with a 1-yr averaging time,
Equation C-10 cannot be used to calculate Xmax-  Hence, the
                               105

-------
following equation is used which gives the annual mean ground
level concentration  (2):
                                                            (C-34)
To obtain the equation for Xmax' Equation C-34 is differentiated
with respect to x, and the derivative is substituted into Equa-
tion C-3.  These operations result in the following equation:
                                                           (c-35)
From Table C-3, for stability Class C, a  has the form
            = ax
                                                           (C-36)
Substituting Equation C-36 into C-35 and solving for x yields
                                                           (C-37)
Using the values a = 0.113 and b = 0.911 from Table C-3 yields

                                H1 -098
       max
                                0.137
                                                           (C-38)
and
K)
                       max
                                     = ฐ'691 H
Substituting these values for x and o  back into Equation C-34
and setting u = 4.5 m/s gives        z
           =  ฐ-ฐ314  Q
               H2.098
                                                           (C-40)
 Equation C-34 is based on a 16-point wind rose; i.e., the 360ฐ
 of the compass are divided into 16 sectors of 22.5ฐ each.  Equa-
 tion C-34 represents a worst case situation in that it assumes
 that the wind always blows from the same sector during the
 entire year.   Thus, for example, if the wind directions were
 distributed equally over the 16 sectors during the year, the
 corresponding concentration would be one-sixteenth of the value
 given by Equation C-34.
                               106

-------
Since  the N02  standard  is  1.0  x  10"1*  g/m3 ,  the  N02  severity
equation is

                     S  = - ฐ
                          (1.0 x

or
                             S                               (c_42)
                                H2.1
Noncriteria Emissions
For noncriteria pollutant concentrations,  the  averaging  time  is
24 hr.  Thus,

                                 /   3   \Q.l7
                     Xmax ~ Xmax\l,440/                     (C-43)

                                    (0>35)                   (c_44)
                               H2
Since F = TLV/300, the equation for source  severity  is

                                  0.0182 Q
                                 (TLV/300) H2

or
                           s s
                               TLV(H2)

AFFECTED POPULATION CALCULATION

The affected population is calculated using Equation C-34  to  pre-
dict average concentration, )(, as a function of downwind distance.
The value of az corresponding to stability Class C  is used in
Equation C-34; i.e.,

                        a  = 0.113 x0-911                   (C-48)
                         z
                                107

-------
The downwind distances at which x/F = 1-0 are determined by an
iterative technique;9 i.e., the roots of the following equation
are determined:
                  a uxF
                         exp - ฑ MM  - 1.0 = 0
                                                           (C-49)
where a  is given by Equation C-48.

A typical plot of "x/F as a function of downwind distance appears
as follows:
where
          and x2 are the roots of Equation C-49
As previously noted, Equation C-34 gives the concentration for a
worst-case situation in which the wind blows continuously from
one sector of width 22.5ฐ (one-sixteenth of a circle).  Hence,
the area in that sector over which x/F exceeds 1.0 is given by
(see following sketch) .
                       A' =
                            TT (x2
                                2 _
:)
(C-50)
The population density, Dp, in the vicinity of the representative
source is obtained by averaging the county population densities
for each plant in the source type.  The worst case arises when
most of the population is concentrated in the sector downwind
from the source; e.g., in a metropolitan area.b  This situation
 This calculation is not entirely consistent in that different
 averaging times are used for x and F. _Equation c-34 gives an
 annual or long-term average value for x-  The averaging times
 for F vary from 1 hr to 1 yr.   This discrepancy is considered
 acceptable in view of the very crude overall approach used to
 estimate the exposed population.
 The worst case arises when the wind blows from one sector most
 of the time and most of the population is contained in the down-
 wind sector.  The present approximation can be considered as a
 limiting case in which the wind always blows from the same sector
 and the entire population is contained in the downwind sector.

                               108

-------
is approximated by assuming that the entire population in the
vicinity of the representative source is distributed uniformly
in the downwind sector.

The population, P, in the area over which x/F>l-0 is then given
by

                       P = Dp(x22 - Xl2)                     (C-51)

or

                       P = DA (persons)                     (C-52)

where

                       A = TT(x22 - Xi2)                      (C-53)

is the area contained in an annular region surrounding the source.
The quantity, P, is designated the effected population.
                               109

-------
                           APPENDIX D

                           PLUME RISE
The problem of determining plume rise from an elevated point
source is discussed in this appendix.  Equations are recommended
for calculating plume rise in the Source Assessment Program, and
error bounds for the calculated plume rise are estimated.  These
error bounds are employed in Section 5 and Appendix I.

RECOMMENDED PROCEDURE

According to Briggs (26), as of 1969 there were over 30 plume
rise formulas in the literature with new ones appearing at a rate
of about two per year.  This multitude of formulas can be clas-
sified into two broad categories:  1) those which give the plume
rise as a function of downwind distance and 2) those which yield
a single value for the plume rise.  Although the former equations
show better agreement with observed plume rise, they are incom-
patible with the simple Gaussian dispersion equation used in
Source Assessment.  Of the equations in the second category, a
modified Holland equation appears (on the basis of extensive data
presented in References 26 and 27) to be about as good as any.
The original Holland equation is  (27)


                    AH =(- 1.5 V D + 0.04 QTT)                (D-l)
where  AH = plume rise, m

        u = wind speed, m/s

       V  = stack gas exit velocity, m/s

        D = stack diameter, m

       Qu = heat emission rate, kcal/s
        n
(26)  Briggs,  G.  A.   Plume Rise.   AEC Critical Review Series.
     U.S.  Atomic Energy Commission, Division of Technical Infor-
     mation Extension,  Oak Ridge, Tennessee, 1969.  81 pp.

(27)  Moses, H.,  and M.  R. Kraimer.   Plume Rise Determination—A
     New Technique  Without Equations.  Journal of the Air Pollu-
     tion Control Association,  22(8):621, 1972.


                               110

-------
 The heat emission rate is calculated as  follows  (30) :
                     H


 where   P = atmospheric  pressure,  dyne/m2

         R = gas  constant =  8.314 x 10 5  dyne-m/gmole-ฐK

         M = molecular weight  of effluent,  g/gmole

        Cp = heat capacity at  constant pressure  of  effluent,
             kcal/g-ฐK

        T  = stack gas temperature,  ฐK
         s
        T  = ambient  air  temperature,  ฐK
         a,

 The values calculated from  Equation D-l are generally too  low.
 For example, values  of AHcaic/ ^observed for  22  stacks given  in
 Reference  26 range from  a low of 0.04 to a high  of  1.18.   Exclud-
 ing the  lowest and highest  values,  the  range  is  0.18 to  0.66.
 The mean is 0.45.  The stack  heights range from  60  m to  180 m.

 Extensive  data are also  summarized  in Reference  30.  For small
 stacks  (30 m to  40 m) , the  average  value of AHcaic/AHo^g is 0.17.
 For medium-size  stacks (60  m  to 120 m) ,  the average value  is  0.43,
 For the  largest  stacks (150 m to 180 m) , the  average is  1.0.  For
 another  set of data  (Bringfelt data from Sweden) covering  36
 stacks ranging from  20 m to 150 m  in height,  the average value of
 AHcale/AHobs is  0.5.   The average  for all  four sets of data in
 Reference  27 is  0.33.

 Based on  the value of  0.33  for AHcaic/AH0kS,  the plume rise calcu-
 lated from Holland's  equation should be  multiplied  by a  factor of
 3.  This value is in  good agreement with a value of 2.92 recom-
mended by  Stumke  (26)  based on independent observations.   A cor-
 rection  factor of 2.2  results from  the  data given by Briggs (29)
 (2.2 = 1/0.45) .

 For electric generating  stations, an equation is available which
 has been optimized for best fit to  existing data (28) .

                      4,000 QH\ฐ.'""'          QO-^"
            AH =  0.414| - 2-       =  (8.892)— -       (D-3)
(28)  Thomas, F. W.,  S. B. Carpenter, and W. C. Colbaugh.  Plume
     Rise estimates for electric Generating Stations.  Journal
     of the Air Pollution Control Association, 20(3):170, 1970.


                                Ill

-------
The following plume rise equations  are  recommended for source

assessment:



For general sources,




                    AH =(- 1.5 V D  +  0.04  Q__)                (D-4)
                         \U      S          rl/
AH = (8.892) — -                   (D-3)
For electric utility combustion  sources,



                                    Q



                                    U0.694




ESTIMATED ERROR BOUNDS FOR PLUME RISE



Based on data  in  References  26  and 27,  error bounds for plume

rise when using Equation  D-4  are estimated to be



                              AH  ,
                       A  ,-     calc  0  _
                       0.5 <  77j - < 3.0
                           —  An .    —
                               true



                              A H

                       0.33 < Atjtrue <  2.0

                            ' AHcalc -




                  ฐ'33 AHcalc  ฑ AHtrue ฑ  2  AHcalc
         AHcalc(1-ฐ - ฐ-67) 1  AHtrue  ^  AHcalc(1'0
Therefore,
                    csu     ,  n    ฐsa
                          = 1.0; -r^	 =  0.67
                   AHcalc       '  AHcalc



For the error analysis, values  of c   and  c    are  required where
                                   SX1      S jL




                         - ฐsu  -     ฐSU                       -
and
                                 112

-------
 Since


                     A        Csu /  AHcalc   \
                     CSU = AiTT-lH  + AH  .  I                (ฐ-7)
                             calc \ s     calc/

 values of the  term  in parenthesis  are needed.   The following data
 were  obtained  from  Reference 27:


             StaCk                       AH    /H
               size        Plant             true7  s
             Small    Argonne  I       0.17)      -  n  •>
             Small    Argonne  II      0.25)    g  ~
             Medium   Harwell         0.87J
             Medium   Gernsheim       0.38>  Avg  =  0.5
             Medium   Duisberg       0.31*
             Large    Lakeview       1.34
             Large    Widows Creek    1.14       -10
             Large    Gallatin       1.49   Avg  ~  L"L
             Large    Paradise       0.89
For small stacks,

       AHtrue = ฐ'2 Hs and  Ali     =  ฐ-5 -  AHcalc  =  Q'1  Hs
                             true
Therefore,

                      AH
                        calc    _  0.1
                    H  + AH   .     1.1    w'w'
                     s     calc

For large stacks,

                           AHcalc
       AH,.    = 1.2 H  and .„      =3.0  ->-  AH   ,  =3.6H
         true        s     AH                calc        s
Therefore,

                      AH
                        calc      3.6
                    Hs + AHcalc
Note that csu is a result of errors for small stacks  and  cSฃ  for
large stacks.  Thus, error bounds for plume rise calculated from
Equation D-4 are
                         = 1.0(0.09) = 0.1                   (D-8)
                               113

-------
                     c  .  =  0.67(0.78)  =  0.5                  (D-9)
                      S A/



Error bounds for plume rise calculated from  Equation D-3 are

estimated from data given in Reference 28  as follows:



                0.6 AH  ,   
-------
Thus, error bounds for plume rise calculated from Equation D-3
are
                           csu
                               = ฐ'23
                               115

-------
                           APPENDIX E

         ALTERNATIVE METHODS FOR ESTIMATING "ACCEPTABLE"
            CONCENTRATION FOR NONCRITERIA POLLUTANTS
METHOD ORIGINALLY ADOPTED FOR SOURCE ASSESSMENT PROGRAM

For noncriteria pollutants, the "acceptable" concentration, F, is
specified as follows:
The conversion factor, G = 1/300, converts the TLV to an  "equiva-
lent" primary ambient air quality standard.  The factor 8/24
adjusts the TLV from an 8-hr work day to continuous  (24-hr) expo-
sure, and the factor of 1/100 is designed to account for  the fact
that the general population constitutes a higher risk group than
healthy workers*  This latter factor was chosen after consulta-
tion with EPA health effects experts.

METHOD BASED ON CRITERIA POLLUTANTS

An alternative method for estimating "acceptable" concentration
for noncriteria pollutants is to determine the relationship
between TLV's and primary ambient air quality standards (PAAQS)
based on the criteria pollutants.  This relationship is shown in
Table E-l.  The (geometric) mean conversion factor (i.e.,
PAAQS2it/TLV) for the four pollutants listed in the table  is
0.0467.  Thus,

                      F = 0.0467 TLV = ~-^                 (E-2)


The "acceptable" concentration based on data for criteria pollut-
ants is therefore higher than that adopted for the Source Assess-
ment Program by a factor of about 15.

Assuming that criteria pollutants constitute a random sample of
size 4 from the population of all pollutants, and further assum-
ing that the conversion factors are log-normally distributed over
this population, a 95% confidence interval for the geometric mean
conversion factor is found to be (0.0123, 0.179).  The correspond-
ing confidence interval for the inverse of the mean conversion
factor is (5.6, 81.3).  Thus, the value of 300 obtained by

                                116

-------
        TABLE E-l.   RELATIONSHIP BETWEEN TLV'S AND PRIMARY
                    AMBIENT AIR QUALITY STANDARDS

Criteria pollutant
Particulate matter
Sulfur dioxide
Carbon monoxide
Nitrogen dioxide
PAAQS,
pg/m3
260
365
10,000
100
Averaging
time,
(fcavg) ป
hr
24
24
8
8,760
Estimated
24-hr
PAAQS2k,
pg/m3
260
365
8,670
370
TLV,
yg/m3
10,000
13,000
55,000
9,000
Conversion
factor,
PAAQS ?u /TLV
0.0260
0.0281
0.158
0.0411

   •Calculated using the method given in Reference 29.  The equation used was

                      PAAQS24 = PAAQS(tavg/24)b


   where b =  0.13 for CO and 0.22 for N0ฃ.   These exponents are averages  of
 .  the values given in Table 2 of Reference 32.

Method 1  lies  well  beyond the upper  end point, 81.3, of  this
interval.

METHOD BASED ON LD50  DATA

The problem of estimating permissible concentrations of  pollut-
ants from available health effects data has been studied by Handy
and Schindler  (30).   Their results yield the following two esti-
mates of  "acceptable" concentration  in terms of animal LD50
values:9
                      F  =  1.07  x 10-1* (LD50)

                      F  =  8.1 x 10~5
(E-3)
Equation E-3 is based  on a statistical correlation between  TLV's
and animal LDso values for some 240 different compounds  (30).
 LD5o is the dosage which  results in mortality to 50% of exposed
 population.

(29)  McGuire, T., and  K. C.  Knoll.   Relationship Between Concen-
     trations of Atmospheric Pollutants and Averaging Times.
     Atmospheric Environment,  5(5):291, 1971.
(30)  Handy, R., and A.  Schindler.  Estimation of Permissible
     Concentrations of  Pollutants for Continuous Exposure.  EPA-
     600/2-76-155 (PB  253  959),  U.S.  Environmental Protection
     Agency, Washington, D.C.,  June 1976.   136 pp.
                               117

-------
Equation E-4  is based on  an  assumed permissible  body  concentra-
tion of pollutant  together with  an assumed  biological half-life
of  30 days  (30).   Thus, nearly identical  results are  obtained  by
two entirely  different approaches.

In  order to compare  Equations E-3 and  E-4 with Equations  E-l and
E-2, the least  squares fit to the relationship between TLV and
LD50 data obtained in Reference  30 is  utilized:

                          LD50 =  34.5  (TLV)                   (E-5)

With this relationship, Equations E-3  and E-4 become

                          F = 0.0037  (TLV)                    (E-6)

                          F = 0.0028  (TLV)                    (E-7)

For comparison, Equation  E-l is  rewritten as follows:

                          F = 0.0033  (TLV)                    (E-8)

Clearly, Equations E-l, E-3, and E-4 are  essentially  equivalent
definitions for F.   In fact, the coefficient 0.0033 in  Equation
E-8 is the mean of the two coefficients in  Equations  E-6  and E-7.
It  should be noted,  however, that for  a particular compound, the
value of F based on  TLV may differ by  more  than  an order  of magni-
tude from the value  based on LDsg.  This  variability  is due to
the large deviations of individual data points from the least
squares fit to the TLV versus LD50 data  (30).

METHOD BASED ON DOSE-RESPONSE RELATIONSHIPS

The three previous methods have dealt  with  the estimation of an
"acceptable" concentration using currently  available  health
effects data.   The present subsection  considers  the hypothetical
situation in which a dose-response relationship  for the human
population has been established for a  particular pollutant.  Such
a relationship is illustrated in Figure E-l.

In this figure, probability,  P,  of a lifetime response to pollu-
tant exposure is given as a function of the concentration, C, to
which exposure occurs.   The solid line represents the curve
fitted to experimental data obtained at high concentrations.
The dotted line represents a statistical  extrapolation  (31) of
 (31) Cornfield, J.  Carcinogenic Risk Assessment.  Science,
     198(4318):693-699, 1977.


                               118

-------
     Figure E-l.
Schematic representation of a dose-response
relationship,  with a positive no-response
level.
the experimental curve to the no-response level  (P = 0).  The
dashed lines represent 95% confidence limits for the experimental
curve.

The estimated value, D, of "acceptable" concentration in this
case would be the concentration at which the extrapolated experi-
mental curve intersects the C-axis.  The associated uncertain-
ties, d^ and du, would be obtained from the 95% confidence limits
as indicated in Figure E-l.

The positive no-response level exhibited by the dose-response
relationship shown in Figure E-l is not the only type of behavior
that can occur at low concentrations.  The extrapolated experi-
mental curve may pass through the origin as illustrated in Figure
E-2, or it may intersect the P-axis as illustrated in Figure E-3.
In order to determine an "acceptable concentration in the latter
two cases, an "acceptable" probability, P*, of a response to
pollutant exposure would have to be arbitrarily chosen.  The
estimated value, D, of "acceptable" concentration and the asso-
iated uncertainties, dฃ and du, would then be obtained as shown
in Figures E-2 and E-3.
                               119

-------
                            APPENDIX F




         CONSTRUCTION OF OPERATING CHARACTERISTIC  CURVES
This appendix  explains the construction of  the  operating charac

teristic  (OC)  curves  presented in Section 5.  Attention is

focused specifically  on construction of the curves presented in

Figure 4.  For convenience,  this figure is  reproduced here as

Figure F-l, to which  the plotted points have  been added.
            i.o



            0.9



            0.8



            0.7



            0.6



            0.5



            0.4



            0.3



            0.2



            0.1



             0
     NO SYSTEMATIC

      ERROR
SYSTEMATIC

ERROR = e,..
             0.01
                       0.1
                                  1.0
                                                       100
           Figure F-l.   Operating characteristic curves
                        for Case A with  b   =0.10.
GOVERNING  RELATIONSHIPS


                                       s\

From Equation  38  (Section 5)  (taking  br  =  0.10)
                       e  = e  + e   =17.14
                        u    r    su
                            e  + e  „ =  0.89
                             r    sฃ
                                        (F-l)





                                        (F-2)
                         -V
     b 2 + 0.01  =  0.14
                             e   =17
                              su
(F-3)




(F-4)
                                120

-------
                           esฃ = ฐ'75
                                        (F-5)
The uncertainty interval for source severity is therefore

                  ScC1 - ^) ^ ST < Sc(l + eu)              (F-6)
or
 0.11
                                    18.14
(F-7)
OC CURVE FOR SYSTEMATIC ERROR EQUAL TO -e
                                         sฃ
The OC curve gives the (fiducial) probability that ST < 1.0 as a
function of Sc•   The fiducial probability distribution of Sm is
shown schematically in Figure F-2 for the case in which the
systematic error is -esฃ = -S^ esฃ.  The distribution is centered
at the point Sc - eSฃ = 3^(1 - esฃ).   Since er represents a 95%
confidence limit in this case, the 2.5% and 97.5% points of the
distribution are located at a distance of er = SQ er on either
side of the center point.
 Figure F-2.  Schematic representation of  fiducial  distribution
              of S   for systematic error equal  to ~esn-

From Figure F-2, when the central point of the distribution  is at
1.0, P(S  < 1.0} = 0.5.  This situation occurs when

                        c/'i.— o^ 	 in                    ( "F — R ^
sc =
                         1 - e
                                   0.25
                                        = 4.0
(F-9:
The point SG = 4.0, P = 0.5 is shown plotted in Figure F-l.
                               121

-------
Similarly, the 0.975 probability point is found from the equation
                     SC = 0.25 + 0.14 = 2'56                (

The point SG = 2.56, P = 0.975 is also plotted in Figure F-l

The 0.025% point is found in a similar manner.
                     S  =
                      C   0.25 - 0.14    '

The point S  = 9.1, P = 0.025 is shown plotted in Figure F-l.

Additional points on the OC curve are obtained by assuming that
the probability distribution in Figure F-2 is a normal distribu-
tion.9  For illustration, take S- = 7.0.  The mean of the distri-
bution of S  is then

                 Sc(l - egฃ) = 7.0(0.25) = 1.75            (F-14)


The 97.5% point is

               Sc(l - egฃ + er) = 7.0(0.39) = 2.73       .  (F-15)

The 2.5% point is

               Sc(l - egฃ - er) = 7.0(0.11) = 0.77         (F-16)


The distribution is shown schematically in Figure F-3.  The prob-
ability that ST < 1.0 is equal to the area under the normal curve
to the left of 1.0.  In order to compute this area, the standard
deviation of the distribution is needed.  Since the 97.5% point
of a normal distribution occurs at 1.96 standard deviations from
the mean,

                      1.75 + 1.96 a = 2.73                 (F-17)

or

                            a = 0.50                       (F-18)
 Note that the three points previously determined do not depend
 on this assumption.  Thus, only the precise shape of the OC
 curve between these points is affected by this assumption.  The
 precise shape of the curve is of little interest in the present
 study.

                                 122

-------
              t/1
              s
              CO
              o
              Qi
              Q.
Figure F-3.  Schematic  representation  of  the  distribution of
             S   for  systematic error equal  to -e  .  and SG = 7.0.

Next, the standard normal deviate,  z,  corresponding to S~ = 1.0
is calculated.
                               •sm  -  y
                            z =
                       _ 1.0 ~  1.75 _   ,  ,n
                      z	--	1.50
                                                            (F-19)


                                                            (F-20)
The area, F(z), under the standard normal  curve  to  the  left  of z
is given in Table F-l (17).  Because of the  symmetry  of the
normal distribution,

                         F(-z) = 1 - F(z)                   (F-21)

Thus, entering Table F-l with  z = 1.50,

                        F(1.50) = 0.9332                    (F-22)

                 F(-1.50) = 1  - F(1.50) =  0.0668            (F-23)

The point Sc = 7.0, P = 0.0668 is shown plotted  in  Figure  F-l.
The complete OC curve is constructed by repeating the above  calcu-
lation for other values of S .
OC CURVE FOR SYSTEMATIC ERROR EQUAL TO e
                                        su
The fiducial distribution of S-p is shown  schematically  in
ure F-4 for the case in which the systematic  error  is e
The 0.5 probability point is obtained  from

                                su
                                   )
SC   1+17
                                  = 0.0556
                                                               ^
                                                             S,,e
                                                            (F-24)


                                                            (F-25)
                              123

-------
       TABLE F-l.  AREA UNDER THE NORMAL CURVE, F(z)  (14)
                       "•>
            ,,=/
— oo

z
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1.0
1.1
1.2
1.3
1.4
1.5
1.6
1.7
1.8
1.9
2.0
2.1
2.2
2.3
2.4
2.5
2.6
2.7
2.8
2.9
3.0
3.1
3.2
3.3
3.4
0.00
0.5000
0.5398
0.5793
0.6179
0.6554
0.6915
0.7257
0.7580
0.7881
0.8159
0.8413
0.8643
0.8849
0.9032
0.9192
0.9332
0.9452
0.9554
0.9641
0.9713
0.9772
0.9821
0.9861
0.9893
0.9918
0.9938
0.9953
0.9965
0.9974
0.9981
0.9987
0.9990
0.9993
0.9995
0.9997
0.01
0.5040
0.5438
0.5832
0.6217
0.6591
0.6950
0.7291
0.7611
0.7910
0.8186
0.8438
0.8665
0.8869
0.9049
0.9207
0.9345
0.9463
0.9564
0.9649
0.9719
0.9778
0.9826
0.9864
0.9896
0.9920
0.9940
0.9955
0.9966
0.9975
0.9982
0.9987
0.9991
0.9993
0.9995
0.9997
0.02
0.5080
0.5478.
0.5871
0.6255
0.6628
0.6985
0.7324
0.7642
0.7939
0.8212
0.8461
0.8686
0.8888
0.9066
0.9222
0.9356
0.9474
0.9573
0.9656
0.9726
0.9783
0.9830
0.9868
0.9898
0.9922
0.9941
0.9956
0.9967
0.9976
0.9982
0.9987
0.9991
0. 9994
0.9995
0.9997
0.03
0.5120
0.5517
0.5910
0.6293
0.6664
0.7019
0.7357
0.7673
0.7967
0.8238
0.8485
0.8708
0.8907
0.9082
0.9236
0.9370
0.9484
0.9582
0.9664
0.9732
0.9788
0.9834
0.9871
0.9901
0.9925
0.9943
0.9957
0.9968
0.9977
0.9983
0.9988
0.9991
0.9994
0.9996
0.9997
0.04
0.5160
0.5557
0.5948
0.6331
0.6700
0.7054
0.7389
0.7704
0.7995
0.8264
0.8508
0.8729
0.8925
0.9099
0.9251
0.9382
0.9495
0.9591
0.9671
0.9738
0.9793
0.9838
0.9875
0.9904
0.9927
0.9945
0.9959
0.9969
0.9977
0.9984
0.9988
0.9992
0.9994
0.9996
0.9997
0.05
0.5199
0.5596
0.5987
0.6368
0.6736
0.7088
0.7422
0.7734
0.8023
0.8289
0.8531
0.8749
0.8944
0.9115
0.9265
0.9394
0.9505
0.9599
0.9678
0.9744
0.9798
0.9842
0.9878
0.9906
0.9929
0.9946
0.9960
0.9970
0.9978
0.9984
0.9989
0.9992
0.9994
0.9996
0.9997
0.06
0.5239
0.5636
0.6026
0.6406
0.6772
0.7123
0.7454
0.7764
0.8051
0.8315
0.8554
0.8770
0.8962
0.9131
0.9279
0.9406
0.9515
0.9608
0.9686
0.9750
0.9803
0.9846
0.9881
0.9909
0.9931
0.9948
0.9961
0.9971
0.9979
0.9985
0.9989
0.9992
0.9994
0.9996
0.9997
0.07
0.5279
0.5675
0.6064
0.6443
0.6808
0.7157
0.7486
0.7794
0.8078
0.8340
0.8577
0.8790
0.8980
0.9147
0.9292
0.9418
0.9525
0.9616
0.9693
0.9756
0.9808
0.9850
0.9884
0.9911
0.9932
0.9949
0.9962
0.9972
0.9979
0.9985
0.9989
0.9992
0.9995
0.9996
0.9997
0.08
0.5319
0.5714
0.6103
0.6480
0.6844
0.7190
0.7517
0.7823
0.8106
0.8365
0.8599
0.8810
0.8997
0.9162
0.9306
0.9429
0.9535
0.9625
0.9699
0.9761
0.9812
0.9854
0.9887
0.9913
0.9934
0.9951
0.9963
0.9973
0.9980
0.9986
0.9990
0.9993
0.9995
0.9996
0.9997
0.09
0.5359
0.5753
0.6141
0.6517
0.6879
0.7224
0.7549
0.7852
0.8133
0.8389
0.8621
0.8830
0.9015
0.9177
0.9319
0.9441
0.9545
0.9633
0.9706
0.9767
0.9817
0.9857
0.9890
0.9916
0.9936
0.9952
0.9964
0.9974
0.9981
0.9986
0.9990
0.9993
0.9995
0.9997
0.9998

Figure F-4 .   Schematic representation of fiducial distribution
of
                   for systematic error equal to e
                            124

-------
The  0.975 probability point is given by
                   SC = 1 + ll   0.14 = ฐ-0551

Similarly, the 0.025 probability point is given by
                   SC = 1 + 17- 0.14 = ฐ-0560              

These three points are shown plotted in Figure F-l .  Since  these
points are nearly colinear on this plot, the remainder of the OC
curve is approximated by drawing a straight line between the
points.  Of course, the OC curve must eventually bend over  at the
ends and approach the horizontal lines P = 1.0 and P = 0 asymp-
totically.  However, these portions of the curve are of little
interest here.

"EFFECTIVE" OC CURVES

The "effective" operating characteristic curve shown in Figure 5
(Section 5) can be constructed using the appropriate branches of
the curves in Figure F-l.   The upper branch of the curve for
systematic error equal to esu and the lower branch of the curve
for systematic error equal to -esฃ are connected by a horizontal
line .

Alternatively, Figure 5 can be constructed by working directly
with an "effective" probability distribution.  The "effective"
fiducial distribution of ST is obtained by treating Interval F-6
as an exact 95% confidence interval for ST-  The resulting  dis-
tribution is shown schematically in Figure F-5.  The area under
each branch of the density function is 0.5.  The 2.5% point
of the distribution is Sr(l - e_n - er) , and the 97.5% point is
       "•     * \         ^ป      Sx,    •*-
scd + esu + erj '

By following the procedure described in the previous subsections,
it can be verified that the density function shown in Figure F-5
generates the "effective" OC curve of Figure 5.
                               125

-------
l/l
s
Q
CD

-------
                           APPENDIX G



 RELATIONSHIP OF SAMPLING AND ANALYSIS BIAS TO SYSTEMATIC ERRORS





In this appendix, the relationship of sampling and analysis bias

to systematic errors is developed.



Let 6  = upper bound on positive bias in sampling and analysis

    $  = upper bound on negative bias in sampling and analysis



where  3 , B~ > 0



Then, letting Bmeas and B^rue represent the measured and true

values of the emission factor,
     B     - B.
a+ .   meas    true
p  >  - -
                                         B.    - B
                                           true    meas
                                              -
B
true
(1 -J- Q 1 *•ป "R
X i D J ^ D
M / - meas
B
meas
true - 1 + 6+
B \
meas ^
P 	 5 	
true
B > B (l - B
meas - true ^
B
„ ^ meas
Btrue S , -
/
^ meas
By definition of b   and b
                  SV1      S X- /
  B
   meas
Therefore,
                     B
                     1-6"
                            ^ Btrue ^ Bmeas
                 = B    (l + b  V
                    meas \     su/
                                                             (G-l)
and
                                127

-------
and
meas
                      1  +  6

or

                          1
                          T=Bmeas(1 ' O                 (G'2)
                                =  1  + b                       (G-3)
                         1-6

or

                           1
                         1  +  6n

Therefore,
                                =  1  - b                       (G-4)
                          b    =  _                           (G_5)
                           su    1-6
                                 1  +  6
                                128

-------
                           APPENDIX H

SOURCE SEVERITY SIMULATION AND PROBABILISTIC SENSITIVITY ANALYSIS


BACKGROUND

In the beginning of the Source Assessment Program, a decision
index, "source severity," was developed.  For noncriteria pol-
lutants, i.e., pollutants other than particulate matter, sulfur
dioxide, nitrogen dioxide, hydrocarbons, and carbon monoxide, sta-
tionary source severity was defined as


                           S = -S^I-Q-                       (H-l)
                               TLV(h2)

where    S = stationary source severity
         Q = air pollutant emission rate, g/s
       TLV = threshold limit value, g/m3
         h = physical emission height, m (typically, chimney or
             stack height)

This simple and usable expression was derived after a number of
conservative, simplifying assumptions had been made.  For screen-
ing and evaluative purposes, this form of the decision index was
deemed appropriate at the time.

During the course of the Source Assessment Program, questions
arose regarding the effects of uncertainty in various parameters.
Addressed here in one combined analysis are a more complete  formu-
lation of the severity model, physical distributions of param-
eters, and uncertainties in several parameters.  In this simula-
tion study are included effects of varying atmospheric stability;
variations in wind speed; uncertainties in the ratio of peak con-
centrations, Xpeafc, to 24-hr maxima, X2i+; uncertainties in the
emission factor; uncertainties in lateral dispersion coefficients,
Oy; uncertainties in vertical dispersion coefficients, az; varia-
tions in temperature; and resulting variations in plume rise.
D. B. Turner's well-documented air dispersion model (5) has  been
used.  The analytical effort to rigorously calculate the effects
of all uncertainties is formidable if not impossible.   A Monte
Carlo simulation was used to investigate effects of uncertainties
in input parameters.
                              129

-------
DESCRIPTION OF REPRESENTATIVE PLANT

The simulated representative plant was a coal-fired, steam elec-
tric utility assumed to be emitting beryllium from one virtual
point or stack.  Simulated plant constants and fuel parameters
are shown in Table H-l.  Meteorological data for atmospheric sta-
bility and wind speed correspond to central Alabama and were used
because they were readily available in a form suitable for this
simulation.  The meteorological data were the joint frequency of
occurrence of atmospheric stability and wind speed.  Frequency-
distributed variables and their parameters are shown in Table H-2.
Since there were only six wind speed ranges and six stability
classes, these parameters were not distributed but were used by
means of table lookup procedures.

Equations that were used in the severity simulation, ordered in
calculational sequence, may be found in Example 1 at the end of
this appendix.  The source severity calculation for this particu-
lar source, emitting beryllium, is shown in Example 2.  This
severity, subsequently referred to as deterministic severity, was
calculated to be 18.6.  If plume rise is included in the calcula-
tion, the resulting deterministic severity is 3.3.  The latter
calculation is given in Example 3.

SIMULATION RESULTS

The ground level concentration of a pollutant emitted by an
elevated point source is a function of several variables, one of
which is the downwind distance.  For this initial analysis, a
distance was used where the mean of the ground level concentra-
tion distribution is a maximum.  This distance could not be calcu-
lated directly and was approximated by a binary search.  It was
found to lie between 2 km and 4 km.  A distance of 3 km was thus
used for all simulations.  Results of the baseline simulation,
where all variables were distributed, and of probabilistic sensi-
tivity analyses are presented in Table H-3.

GENERAL COMMENTS ON RESULTS

In a Monte Carlo simulation, since distributions are used as
inputs and not single deterministic values, the result is also a
distribution of values—in this case, a distribution of values
for severity.  When a probabilistic sensitivity analysis was per-
formed, the variables were fixed one at a time at either a high
level or low level.  However, since the remaining variables are
still distributions of values, the resulting severity will again
be a distribution of values.

The sensitivity analyses were performed by fixing each of the
variables that were normally distributed at the +2o level and the
-2a level.  Thus, in the second row of Table G-3, the notation
SpO.05 designates that the value of P (the ratio of peak to 24-hr

                              .-1-30

-------
    TABLE H-l.  PLANT AND FUEL PARAMETERS USED IN SIMULATION
         	Parameter	Value

         Coal consumption rate, g/s             34,532,
         Stack height, m                           100
         Barometric pressure, mb                 1,013,
         Stack gas temperature, ฐK                 ^08,
         Fraction of beryllium in coal, ppm        ^.3,
         Stack I.D., m                            4>23h
         Stack gas velocity, m/s                  14.1
         TLV of beryllium, g/m3               2 x 10~6
          1,092,000 metric tons/yr (nominal value for a
          500-MW powerplant).   (1 metric ton equals 106
          grams; conversion factors and metric system
          prefixes are presented at the end of this
          report).

          Nominal value for a 500-MW powerplant.
         c
          Reference 4.
          TABLE H-2.   ASSUMED DISTRIBUTION PARAMETERS

Variable
xpeak _ p
x24-hr
Emission factor
Ambient temperature, ฐ
az
ay
Atmospheric stability
Wind speed
Assumed
distribution
Normal.
• Normal .
K Normal.
Normal.
Normal .
None (table lookup)
None (table lookup)
Mean
6.2a
0.85
286.1
1.0
1.0
•
Standard
deviation
0.93a
0.347
4.9
0.255
0.051

Lower
limit
4.3
0.17
276.5
0.5
0.9

Upper
limit
8.1
1.53
295.7
1.5
1.1


NOTE. — Blanks indicate
numerical value is not
applicable.



(32)  Montgomery,  T.  L.,  and J.  H.  Coleman.   Empirical Relation-
     ships  Between Time-Averaged S0? Concentrations.  Environ-
     mental Science  and  Technology,  9 (10) :953-957,  1975.
                               131

-------
TABLE H-3.   SIMULATION RESULTS OF PROBABILISTIC
            SENSITIVITY ANALYSES

Baseline
So
SP0.05
Po.05 = 4.3
SP0.95
PO . 95 = 6-1
SEF0.05
EFo.05 = 0.17
SEF0.95
EFfl.95 = 1-53
Tao.05
Tao.os = 276.5
STa0.95
Ta0.95 = 295.7
ZO . 05
ฐz = ฐz - ฐ'5 ฐz
'Sa
Z0.95
az = az + 0.5 az
Sa
yo.os
ฐy = ฐy - O'1 ฐy
ayo.9s
ฐy = ฐy + O'1 ฐy
Stability Class A
Stability Class F
.Stability D
'Uo.05 = 2.05
Stability D
U0.95 = 6.99
SP0.5
Pfl.5 = 6.2
SEF0.5
EF0<5 = 0.85
Mean
Sx
0.0828
0.122
0.065
0.0163
0.148
0.076
0.086
0.027
0.139
0.097
0.08
0.0242
9.3 x 10~8
0.0153
0.112
0.085
0.812
Standard
deviation
0.
1.
0.
0.
1.
0.
0.
0.
1.
0.
0.-
0.
4.75
0.
0.
0.
0.
788
11
592
153
38
749
832
367
06
842
689
168
x 10~5
154
72
773
776
S
1
2
1
0
3
1
2
0
2
2
1
0
<0
0
1
1
1
0.95
.85
.70
.44
.372
.41
.89
.05
.97
.28
.05
.70
.34
.01
.65
.48
.62
.91
P(S
0
0
0
0
0
0
0
0
0
0
0
0
1
0
0
0
0
< 0.1)
.585
.561
.597
.688
.541
.586
.568
.806
.451
.576
.586
.513
.0
.764
.147
.581
.565
P(S
0
0
0
1
0
0
0
0
0
0
0
1
1
0
0
0
0
< 1.0)
.841
.771
.887
.00
.703
.853
.827
.954
.675
.8?j
.859
.00
.0
.999
.832
.838
.824
P(S
0.
0.
0.
0.
0.
0.
0.
0.
0.
0.
0.
0.
0.
0.
0.
0.
0.
> 1.0)
159
229
113
0
297
147
173
046
325
177
141
0
0
001
168
162
176
                      132

-------
maximum concentrations) was fixed at the -2a level where P is
equal to 4.3.  The third row shows that for this sensitivity
analysis, P was fixed at the +2a level  (or 8.1).  Since results
of the probabilistic simulation are also distributions, several
parameters are available for making comparisons.  One of the obvi-
ous parameters is the mean of the resulting distributions; other
parameters are listed in Table H-3 as column headings and include
the standard deviation, the severity at the 95% cumulative level,
and the probability that the severity is <0.1, <1.0, and >1.

All of the resulting distributions are nonnormal.  Looking at the
equations in Example 1, the mean severity values shift in the
appropriate direction.  Thus, when the emission factor is low,
the mean severity is low; and when the emission factor is high,
the mean severity is high with respect to the baseline case.

When the ambient temperature is low, the buoyancy factor in the
plume rise equation (Equation H-4) is high; therefore, the plume
rise is higher and thus the severity is less than the baseline
case.  The opposite is true when the ambient temperature is high.

The effect of oz is more complex, since az is both in the exponen-
tial term as well as in the denominator of the premultiplying
factor in the Gaussian plume equation (Equation H-7).  The effect
of the exponential term overshadows the premultiplier term, as
can be seen from the lower value of az yielding a lower value of
severity when compared to the baseline case.  The opposite is
true for higher values of az.

In the case of ay, the effect is straightforward since ay is used
in only the denominator of the premultiplying factor and thus the
low values of ay give higher values of severity compared to the
baseline case.  The opposite is true for higher values of ay.

For stability Class A, a lower value of mean severity is obtained
since values of ay and az are large in this stability class.  The
exponential term thus becomes unity, and only the premultiplying
term becomes significant.  Since ay and az are in the denomina-
tor, a lower value of mean severity is obtained compared to the
baseline case.  For stability Class F, values of ay and az are
very small, smaller than the plume rise by a significant factor.
Thus, for example, if H = 300 m and a  = 30 m,
                                     Z

                  exp [-0.5(!ฃ?_)2] = 1.9 x 10-22              {H-2)


and the severity is much lower than the baseline case.

Stability D was fixed to evaluate the effect of fixed wind speed.
Using a low value of wind speed results in a higher plume rise;
this in turn gives a lower value of severity.  The wind speed
term in the premultiplier in the Gaussian plume equation is

                               133

-------
overriden by the plume rise term.  The converse is true when the
wind speed is higher.

The last two rows of Table H-3 show results when mean values of
the parameters were used.  When P was set equal to its mean value
of 6.2, a mean severity approximately equal to the baseline case
was obtained.  The same was true when a mean emission factor of
0.85 was used.  No subsequent simulations using mean values of
the parameters were run.

SPECIFIC RESULTS

One indicator of sensitivity is the value of S at the 95% cumula-
tive level.  These results are in Column 3 of Table H-3.  If the
true emission factor is higher than the value being used, the
highest severity at the 95% cumulative level is the value of 3.4.
Low values of P and high values of az are also sensitive vari-
ables.  The uncertainties in az and ay are the main reason for
saying that the dispersion model is no better than a "factor of
2 or 3."  However, in this simulation, separation of the individ-
ual effects and the overall result should be a good approximation
of the "true severity" for this source.

The column labeled "P(S > 1.0)" is another indicator of sensitiv-
ity and is a way of looking at the cumulative severity distribu-
tion from a different perspective.  In one case, if concern
relates to the magnitude of the severity at the 95% probability
level, Column 3 would be used for sensitivity.  If, on the other
hand, interest involved the magnitude of_ the probability that
S > 1.0, Column 6 would be used.  In either case, P, a , and the
emission factor are the most sensitive variables.

LARGE SYSTEMATIC ERROR IN EMISSION FACTOR

One final question that was addressed involves the effect of a
large systematic error in the emission factor.  First, plant and
fuel parameters are assumed to be known exactly.  While this is
not strictly true (e.g., stack gas flow rate and temperature may
vary), uncertainties are assumed to be small.  The emission
factor in this analysis is expressed as a fraction of beryllium
that is emitted based on the amount of beryllium in the coal.  To
preserve the uncertainty distribution on the emission factor and
to facilitate computation, the fraction of beryllium is set in
coal (2.3 ppm)  at 5 times the baseline value (11.5 ppm).  Results
of this simulation are shown in Table H-4. ~

CONCLUSIONS DERIVED FROM THE SIMULATION ANALYSIS

If this simulation approach is assumed to yield a "truer" picture
of the actual severity at this representative plant (the true
severity distribution being obtainable only by long-term sampling
for 24-hr beryllium maxima in a comprehensive sampling network),
the following observations can be made:


                             134

-------
          TABLE  H-4.  SIMULATION OF A LARGE SYSTEMATIC
                      ERROR IN THE EMISSION FACTOR
            SD
                    0.95
                                0.1)    P(S < 1.0)    P(S >  1.0)
   0.414   3.94
    9.5
        0.523
0.62
0.38
   • The deterministic severity is conservative by a factor
     of 225 when compared to the mean simulated severity for
     the baseline case (18.6/0.0828).

   • The largest value of S at the 95% probability is 3.41
     for a high emission factor when looking only at emission
     factor precision.  This represents a conservative factor
     of 5.5 compared to the deterministic severity.

   • Even if the systematic error in emissions is a factor of
     5, the deterministic severity is conservative by a
     factor of 2, (18.6/9.5).

Thus, one may conclude that the first decision index, the
deterministic source severity, is indeed a worst case estimate
and can be used as one of several tools to aid the EPA decision
maker in the screening process.
Example 1.  Equations Used in the Severity Simulation

                       Q = (CONS)(EF)(%Be)

where     Q = emission rate, g/s
       CONS = coal consumption, g/s
         EF = emission factor for beryllium, g/kg
        %Be = fraction of beryllium in coal
                  'V
           Ah =
                   U
  1.5 + (2.68 x 10~3)p
                                              (H-3)
                                              (H-4)
where
Ah

V
 s
 d
                u =
               T  =
plume rise, m
stack gas exit velocity, m/s
inside stack diameter, m
wind speed, m/s
barometric pressure, mb
stack gas temperature, ฐK
 Ordered in calculational sequence,
                               135

-------
               T  = air temperature,  ฐK
                cl
       2.8 x 10~3 = constant with units of  (mb)"1 m"1
                          a  = Ax0-9031
                                                             (H-5)
where  a  = lateral dispersion coefficient, m
        x = downwind distance, m
        A = function of stability class as  shown  below
                    Stability class
A
B
C
D
E
F
0.3658
0.2751
0.2089
0.1471
0.1046
0.0722
                            = A!X
                                                             (H-6)
where                 a  = horizontal dispersion coefficient,  m
       AI, BI, and Cj
       (for x > 1,000 m) = functions of stability  as  shown  below
where
           Stability class
                                Ai
BI
Ci







X =
H =
h =

A
B
C
D
E
F

0.00024
0.055
0.113
1.26
6.73
18.05
Q f ^ r
\f — ™ ^^W*"t 1 lit
x ™yฐzu exp[ ฐ-
2.094
1.098
0.911
0.516
0.305
0.18
-/ H\21
* l ll
u
\ z/ J
-9.
2.
0.
-13
-34
-48.

6
0
0


6

short-term peak concentration, g/m3
h + Ah =
physical

effective emission
stack height, m
Y = X-
*21t p
height,


m





                                                             (H-7)
                                                             (H-8)
where
       X2i+
         P
             maximum 24-hr concentration, g/m3
             empirically derived peak to 24-hr maximum  ratio
                     S =
                          (TLV) (8/24) (1/100)
                                                             (H-9)
                               136

-------
where    S =  source  severity
       TLV =  threshold  limit value, g/m3

Example 2.  Source Assessment Deterministic  Severity  Calculation

                      .  Q =  (CONS) (EF) (%Be)                  (H-10)

where     Q = emission  rate, g/s

       CONS = coal consumption rate,  g/s
            = 3.4532 x  1(T  g/s

         EF = emission  factor for beryllium
            = 0.85

        %Be = fraction  of beryllium in coal
            = 0.23 x 10~5 g Be/g coal

          Q = (3.4532 x 104) (0.85) (0.23 x  10~5)
            = 6.75 x 10~2 g/s

          S =  5'5 Q                                        (H-ll)
              TLV(h2)
              (5.5) (6.75 x  10"2)

               (2 x  10~6) (100) 2
            = 18.6

Example 3.  Deterministic Severity Using Plume Rise Correction

The plume rise is calculated using Equation  D-3.

                                   Q  0 .kkk
                        Ah = (8.892) — -                  (H-3)
where  QH ^ 67 x 10" 3 YS      b                             (H.12)

        h = 100 m

        u = 4 . 5 m/s

       T  = 407.95 ฐK
        S
       T  = 286.1 ฐK
        3.
       V  =14.1 m/s
        o
        d = 4.23 m
 Equation D-l  (uncorrected Holland equation) gives  a  plume  rise
 of 65 m and a source severity of 6.8.  Equation D-4  (corrected
 Holland equation) gives a plume rise of 195 m and  a  source
 severity of 2.1.

                               137

-------
Thus,
         QH =  (67 x  ID'3) (14.1) (4.23)2(1,013) (4084~8286j
         QR = 5,137  cal/s
         ..   8.892(5,137)ฐ-l+1+lt
         An = 	
                  (4.5)0.69^
         Ah = 139 m
          H = h +  Ah  =  239 m
          S =  (5.5) (6.75  x IP"2)
                (2  x 10~6) (239)2
          S S 3.3
                              138

-------
                            APPENDIX I

         QUANTIFICATION OF UNCERTAINTY IN SOURCE SEVERITY


 Inequality 32,  derived in Section 5,  can be used directly in the
 decision-making process by computing values of eฃ and eu corre-
 sponding to each individual calculated severity, Sc.   Alternative-
 ly,  Inequality  32 can be used to derive general guidelines for
 decision making in the^Source Assessment Program by computing
 generalized values of e& and eu.  The latter approach is followed
 in this report.  In this appendix,  generalized values are derived
 for  each of the individual uncertainties required in the calcula-
 tion of e^ and  eu.

 RANDOM UNCERTAINTIES IN AVERAGE  CAPACITY AND EMISSION HEIGHT
 (ar  AND cr)

 For  organic and inorganic chemicals source  types having  small
 numbers of plants, ar and cr are usually zero since data for all
 plants are usually available.  For  source types having large num-
 bers of sources,  survey data are generally  obtained from a suffi-
 cient number of plants to make these  terms  small.   For example,
 in a source assessment of asphalt hot-mix plants,  600 plants were
 surveyed from a population of approximately 4,000  plants to yield
 95%  confidence  limits on mean capacity and  mean stack height of
 ฑ4%  and ฑ3%,  respectively,   Thus,  for most  source  types, ฑ5%
 appears to be a conservative value  for these terms.   Then

                        ar2  + 4 cr2  =  0.01                   (1-1)

 SYSTEMATIC UNCERTAINTY IN EMISSION  FACTOR (b   AND b   )
                                             S Xf      o U.

 Estimates  of  systematic  and random  errors involved in sampling
 and  analytical  procedures have been determined for a  number of
 standard EPA  methods  via collaborative testing.   The  available
.results  are  summarized in Table  1-1 (33-43).   This table shows
 that average  biases  for  standard EPA  methods range rom -35% to +10%,
 (33) Hamil, H. F., and R. E. Thomas.  Collaborative  Study  of
     Method for Determination of  Stack Gas Velocity  and Volumet-
     ric Flow Rate in Conjunction with EPA Method  5.   EPA-650/4-
     74-033  (PB 240  342), U.S. Environmental  Protection Agency,
     Research Triangle Park, North Carolina,  1974.   40 pp.
                                                        (continued)

                                139

-------
     TABLE  1-1.   COLLABORATIVE  TEST  RESULTS  FOR EPA  SAMPLING AND  ANALYTICAL  METHODS
EPA










High volume.
(Ambient air.


Method
2
3
S
6
7
8

9
10
104

b
)


Chemiluminescent
(NO ), ambient air.)
X
Chemiluminescent (photochemical
oxidants, ambient air) .




NDIR (ambient




air) .
Bias,
* of standard
concentration



0
0
-2
(Analysis only.)
+1.4% Opacity.
-2 to +10
-20

0


0
-35 to -15
(From 0.05 ppm
to 0.50 ppm.)



+2.5
Precision (standard deviation) ,
% of mean concentration
Within lab
5.5
10 to 30
10 to 30
4
7

60
2% Opacity.
2 to 7
44
3
2
(At 250 pg/m3)
1.3
(At 1,000 pg/rn3)
7 to 8
(At 100 yg/m3)
9.2
(At 0.05 ppm)
5.9
(At 0. 1 ppm)
3.2
(At 0. 5 ppm)
20
(At 3 mg/m3)
Between lab
5.6
15 to 35
20 to 40
5.5
10

65
2.5% Opacity.
4 to 13
58
3.7
4
(At 250 pg/m3)
2.6
(At 1,000 ug/m3)






30
(At 3 mg/m3)
95% (20) confidence
limits, % of mean
Within lab
11
20 to 60
20 to 60
8
14

120
4% Opacity.
4 to 14
88
6
4
(At 250 pg/m3)
2.6
(At 1,000 pg/m3)
14 to 16
(At 100 ug/m3)
18
(At 0.1 ppm)
12
(At 0.1 ppm)
6.5
(At 0.5 ppm)
40
(At 3 mg/m3)
Rptwppn 1 ab
11.2
30 to 70
40 to 80
11
20

130
5% Opacity.
8 to 26
116
7.4
8
5.2







60
(At 3 mg/m3)
Minimum
detectable
limit Reference
36
37
38-41
3 ppm 38,42
38,43,44

38
38
20 ppm 38,45
38,46
3 mg 38
25 pg/m3 38


10 pg/m3 38
0.3 mg/m3 ^8





0.3 mg/m3 38
NOTE.—Blanks indicate no information available.
 Values listed are for a single measurement.  For n measurements, the values shouJa be divided by >^n.

-------
 Biases were determined by measuring  a  gas  of  known  composition
 (standard)  under controlled conditions.  Values  reported  are  aver-
 ages of a number of runs  by various  collaborators at  different
 levels of pollutant concentration.   Since  bias can  vary greatly
 among collaborators and with pollutant concentration,  the range
 of the individual measurements would be more  appropriate  for
 determining systematic error bounds.   For  example,  in  the measure-
 ment of carbon  monoxide by EPA Method  10  (44), bias in individual
 tests ranged from -12% to +18%,  and  there  was a  definite  trend
 from positive to negative bias with  increasing CO concentration.
 Hence,  more realistic  error bounds would be obtained by using the
 limits -12% and +18% for  bias rather than  the average  values of
 -2%  to +10%.  For the  purpose of the present  analysis, however,
 the  average values will be used.
  The  value of  -35%  for negative bias is probably a conservative
  estimate for  source  assessment anyway, since  it corresponds to a
  method  for  sampling  ambient air rather than a stack sampling
  method.
(continued)
(34)  Hamil,  H.  F.,  and R.  E.  Thomas.   Collaborative Study of
     Method  for Stack Gas  Analysis and Determination of Moisture
     Fraction with  Use of  Method 5.   EPA-650/4-73-026 (PB 236
     929), U.S. Environmental Protection Agency,  Research Tri-
     angle Park, North Carolnia, 1974.

(35)  Smith,  F., and J. Buchanan.  IERL-RTP Data Quality Manual.
     EPA-600/2-76-159, U.S.  Environmental Protection Agency,
     Research Triangle Park,  North Carolina,  1976.

(36)  Hamil,  H.  F.,  and D.  E.  Camann.   Collaborative Study of
     Method  for the Determination of  Particulate  Matter Emissions
     from Stationry Sources  (Portland Cement  Plants).  EPA-650/
     4-74-029 (PB 237 346),  U.S. Environmental Protection Agency,
     Research Triangle Park,  North Carolina,  1974.   54  pp.

(37)  Hamil,  H.  F.,  and R.  E.  Thomas.   Collaborative Study of
     Method  for the Determination of  Particulate  Matter Emissions
     from Stationary Sources  (Fossil  Fuel-Fired Steam Generators).
     EPA-650/4-74-021 (PB  234 150), U.S. Environmental  Protection
     Agency,  Research Triangle Park,  North Carolina,  1974.   36 pp.

(38)  Hamil,  H.  F.,  and R.  E.  Thomas.   Collaborative Study of
     Method  for the Determination of  Particulate  Matter Emissions
     from Stationary Sources  (Municipal Incinerators).   EPA-650/
     4-74-022 (PB 234 151),  U.S. Environmental Protection Agency,
     Research Triangle Park,  North Carolina,  1974.   37  pp.

(39)  Hamil,  H.  F.,  and D.  E.  Camann.   Collaborative Study of
     Method  for the Determination of  Sulfur Dioxide Emissions
     from Stationary Sources  (Fossil  Fuel-Fired Steam Generators),
     EPA-650/4-74-024 (PB  238 293), U.S. Environmental  Protection
     Agency,  Research Triangle Park,  North Carolina,  1973.   64 pp.
                                                      (continued)

                               141

-------
From Appendix F,

                          b
                           su
                           su   1-3
                           s         -
                           SJt   1 + 3

where  $~ and  3  are bounds on_the negative and positive bias,
respectively.  Substituting 3  =0.35 and 3+ = 0.10 yields

                            b   = 0.5
                             su

                            t>   = 0-1
                             sฃ
                                       A
, RANDOM UNCERTAINTY IN EMISSION FACTOR  (b )

The random uncertainty in emission factor will be treated as an
independent variable since it can (theoretically) be controlled
by specifying the number of samples to be collected and analyzed
 (continued)
 (40) Hamil, H. F., and D. E. Camann.  Collaborative Study of
     Method for the Determination of Nitrogen Oxide Emissions
     from Stationary Sources  (Fossil Fuel-Fired Steam Generators).
     EPA-650/4-74-025  (PB 238 555), U.S. Environmental Protection
     Agency, Research Triangle Park, North Carolina, 1973.  102 pp.

 (41) Hamil, H. F., and R. E. Thomas.  Collaborative Study of
     Method for the Determination of Nitrogen Oxide Emissions
     from Stationary Sources  (Nitric Acid Plants).  EPA-650/4-74-
     028  (PB 236 930), U.S. Environmental Protection Agency,
     Research Triangle Park, North Carolina, 1974.  41 pp.

 (42) Constant, P. C., G. Scheil, and M. C. Sharp.  Collaborative
     Study of Method 10—Reference Method for Determination of
     Carbon Monoxide Emissions from Stationary Sources—Report of
     Testing.  EPA-650/4-75-001  (PB 241 284), U.S. Environmental
     Protection Agency, Research Triangle Park, North Carolina,
     1975.
 (43) Constant, P. C., and M. C. Sharp.  Collaborative Study of
     Method 104—Reference Method for Determination of Beryllium
     Emissions from Stationary Sources.  EPA-650/4-74-023  (PB 245
     Oil), U.S. Environmental Protection Agency,  Research Tri-
     angle Park, North Carolina, 1974.  94 pp.

 (44) Title 40—Protection of Environment.  Chapter 1—Environ-
     mental Protection Agency—Part 60—Standards of Performance
     for New Stationary Sources.  Subchapter C—Air Programs.
     Method 10—Determination of Carbon Monoxide  Emissions from
     Stationary Sources.  Federal Register, 39:9319-9321, 1974.

                               142

-------
The  range over which b  can be expected to vary can be determined
by reference  to Table I-l, which lists the precision of standard
EPA  methods.

The  precision of each measurement method was determined by an
analysis of variance  (ANOVA) of collaborative test results
obtained under field conditions.  The "within-laboratory" stand-
ard  deviation is obtained from the residual, or error, variance
in the standard ANOVA procedure.  It is an estimate of the impre-
cision inherent in the sampling and analytical procedures and is
the  imprecision which would be obtained if repeat measurements
were performed under identical conditions by the same personnel
using the same equipment.  The "between-laboratory" standard
deviation is  obtained from the total variability in the test
data, and is  a measure of the total imprecision due to 1) the
inherent imprecision in the method, 2) differences in personnel
and  equipment among laboratories, and 3) any process variations
which occurred during the tests.  In the absence of process
variation, the difference of the between-laboratory and within-
laboratory variances is a measure of the variability due to
differences in factors such as personnel or equipment and is
designated "laboratory bias."  (Note that laboratory bias is a
component of  imprecision rather than bias.)  In this case,

        a2             = a2           + a2                   (1-4)
          between labs     within lab     laboratory bias

Assuming that errors are normally distributed and that estimates
of standard deviations are based on a large number of degrees of
freedom (The  latter assumption is not justified in some cases due
either to poor experimental design or to a limited number of
usable data points.), the 95% confidence limits are approximately
equal to twice the standard deviation.  These values are listed
in Table I-l  for both the within-laboratory and between-laboratory
components.

Usually, at least three samples are collected and analyzed.  Thus,
dividing the  95% confidence limits by /3,  the largest random
uncertainty in emission factor should be about 75%.  Larger
values could result from large process variations, however.
Measurement of trace elements and polycyclic organic material
(POM) may also result in larger uncertainties due to the rela-
tively large imprecision of analytical methods used for these
species.   For the purpose of this analysis, it will be assumed
that the random uncertainty in emission factor varies between 0
and  100%;  i.e.,
                              /\
                          0 < br < 1.0                       (1-5)


SYSTEMATIC UNCERTAINTY IN EFFECTIVE EMISSION HEIGHT (c   AND c  )
                                                      S Xf      S U

Bounds on the systematic error in effective emission height due
to use of semiempirical equations to estimate plume rise are

                               143

-------
derived in Appendix C.  The bounds corresponding to the modified
Holland equation  (Equation C-4) are used in the present analysis:
                            c   =05
                            ฐsa   ฐ'b

UNCERTAINTY IN CORRECTION FACTOR FOR DISPERSION EQUATION  (m)

There are no experimental data which can be compared with disper-
sion calculations under the conditions of interest in this report;
i.e., fixed-source parameters, wind speed of 4 . 5 m/s, and Class C
stability.  However, recent measurements by Guzewich and Pringle
(45) permit such a comparison for (approximately) fixed-source
parameters and Class C stability.  A total of 33 measurements was
made under Class C stability conditions with wind speeds ranging
from 1.6 m/s to 13.2 m/s.  Emission rates varied from 8.4 g/s to
9.8 g/s;a other source parameters remained constant during the
measurements.

The distribution of the ratio of measured to calculated concentra-
tion is shown in Figure I-l.b  Calculated values were obtained
using simple Gaussian dispersion theory together with Briggs1
plume rise formula  (26) .  Data are approximately log-normally
distributed with a mean of 1.0 and a variance of 0.3136.  The 1%
and 99% points are approximately equal to 1/3 and 3, respectively.

Data variability can be assigned to variability in the following
factors :

   • Measured values
   • Wind speed
   • Emission rate
   • Estimated plume rise
   • Other atmospheric parameters (other than wind speed and
     stability class)

Only the last of these contributions is pertinent to the present
analysis.   That is, error bounds for the correction factor, y,
based on Figure 1-1, should be on the conservative side.
 Emission rates were accurately known since the "pollutant" was a
 tracer substance injected into the stack gas stream.

 Raw data were supplied by R. L. Hanson, Office of the Project
 Manager for Chemical Demilitarization and Installation Restora-
 tion, U.S. Department of the Army, Aberdeen Proving Ground,
 Maryland.
(45) Guzewich, D. C., and W. J. B. Pringle.  Validation of the
     EPA-PTMTP Short-Term Gaussian Dispersion Model.  Journal of
     the Air Pollution Control Association, 27 (6) :540-542, 1977.

                               144

-------
      is 1.0 -
       S 0.8 -

      ^ 0.6 -

         0.4 -
         0.2 -
         0.010.050.10.20.51 2   5  10 20 3040506070
90 95  98 99 99.8 99.9 99.99
        Figure 1-1.  Comparison of measured and predicted
                     values of ground level concentration.

Therefore, the value m = 3 will be used in the analysis  as  a
conservative estimate.

The above "factor of three" is consistent with the  best  case
estimate given by Turner (5).  This  factor corresponds  to  the
estimation of the pollutant concentration at  a given  point  down-
wind of the source rather than estimation of  the  maximum concen-
tration, Xmax-  Theoretical and empirical analyses  indicate that
a "factor of two" may be more appropriate for estimation of
Xmax (46-47).  However, the "factor  of three" will  be used  here
in order to take into account uncertainty in  the  averaging  time
correction (see following subsection).

Data in Figure 1-1 indicate that the uncertainty  in the  disper-
sion equation could be eliminated by appropriately  modifying  the
(46) Scriven, R. A.  Variability and Upper  Bounds  for  Maximum
     Ground Level Concentrations.  Philosophical Transactions of
     the Royal Society of London, Series A,  265:209-220,  1969.

(47) Weber, A. H. Atmospheric Dispersion Parameters  in Gaussian
     Plume Modling, Part I, Review of Current  Systems  and Pos-
     sible Future Developments.  EPA-600/4-76-030a (PB 257 893),
     U.S. Environmental Protection Agency,  Research  Triangle
     Park, North Carolina, July 1976.   58 pp.
                              145

-------
definition of source severity.  The value of Xmax was defined as
the value which would result from the source emitting into an
atmosphere with a wind speed of 4.5 m/s and Class C stability.
If a  large number of tests were made holding source parameters,
wind  speed, and stability class fixed, but allowing otheฃ atmos-
pheric parameters to vary,3 a distribution of values of Xmax
rather than a single value would be obtained.  Figure 1-1 indi-
cates that the Gaussian dispersion model correctly predicts the
median of this distribution (median correction factor equals
l.Q).  Therefore, by specifying the value of Xmax to be used in
the definition of source severity as the above ensemble median,
inherent uncertainty associated with the dispersion model would
be eliminated.

Of course, in interpreting the above modified version of source
severity, it would have to be kept in mind that actual ground
level concentrations under conditions of Class C stability and
wind speed of 4.5 m/s could be "a factor of three" higher or
lower than the median value.  Hence, "actual" severities could be
higher or lower than the nominal value by a "factor of three"
under these atmospheric conditions.

From the standpoint of decision making, the modified version of
source severity would result in a decision based on the median
value of Xmax fฐr Class C stability and wind speed of 4.5 m/s
rather than on the upper 97.5% value of Xmax-

UNCERTAINTY IN AVERAGING TIME CORRECTION FACTOR (tQ/t)0-17

The averaging time correction factor given by Turner (5),
(to/t)0*1 , is a semiempirical relationship derived from data on
lateral and vertical diffusion coefficients in steady winds.  It
applies to situations in which the mean wind direction remains
constant over the period of interest.  Thus, it corresponds to a
worst case situation with respect to pollutant concentration.

A number of studies have been conducted to determine the relation-
ship between measured concentration and averaging time for ambi-
ent air monitors located at fixed sites (29, 32, 48).  Results of
these studies tend to predict considerably lower mean concentra-
tions than does the above correction factor, which is to be
expected since the mean wijd direction does not, in general,
remain constant over an extended period of time.  The diversity

 For example,  stability Class C defines a range of atmospheric
 stability conditions rather than a single unique condition.
 Previously,  it was assumed that other atmospheric parameters
 were held fixed so that a unique atmospheric state (and a cor-
 responding unique value of x^   )  was obtained.
(48)  Shiruaikar,  V.  V., and P. R.  Patel.  Long Term Statistics
     of Peak/Mean Concentrations From A Point Source.  Atmos-
     pheric Environment,  11 (4):387-389, 1977.

                              146

-------
of  results obtained  in the three studies cited above also empha-
sizes the fact that  the relationship between concentration and
sampling time is strongly dependent upon local meteorological
conditions; e.g., wind direction statistics.

For the purpose of the present analysis, it is assumed that the
mean wind direction  remains constant during the period of inter-
est.  This assumption is consistent with the concept of a fixed-
state receiving atmosphere assumed previously.  The uncertainty
associated with the  averaging time correction factor under these
conditions is then required.  In lieu of other information, this
uncertainty is assumed to be included in the uncertainty associ-
ated with the dispersion equation, and the latter uncertainty is
conservatively taken to be a "factor of three" (see preceding
subsection) .

The effect on source severity of relaxing the assumption of con-
stant mean wind direction is considered in Appendix G.
                                         ^       /\
UNCERTAINTY IN ACCEPTABLE CONCENTRATION (d   AND d  )
                                          SV1      S J6
It  is assumed that for each noncriteria pollutant, i, the "true"
value, FT, and calculated value, FC, of the "acceptable" concen-
tration are given by

                              = (TLV)i Gi                   (1-6)
                         (Fc)i =  (TLV)i
where   (TLV).  = TLV of pollutant i as listed in Reference  4
           G.  = conversion factor which converts  (TLV). to an
                equivalent PAAQS                       1
            G = an "average" conversion factor

Since (TLV). is taken as the nominal value listed in Reference  4,
there is no1uncertainty associated with this value in  the context
of the present analysis.  All variability due to the methodology
used in establishing TLV's is associated with the conversion
factors, G..

From Equations 1-6 and 1-7 are obtained

                           (FT)i   fi
                           Wh~ G

                        F ) .
                    In -7—j-i-  = In GL - In G                 (1-8)
                               147

-------
It is assumed that the In Gi are normally distributed  (i.e.,  the
G^ are log-normally distributed) over the population of pollut-
ants with mean y and variance a2.  It then  follows  from Equa-
tion I-8_ that ln(FT/Fc) is normally distributed w^th mean
y - In G and variance a2.  If it is assumed that: G  is  the  geo-
metric mean of the distribution of Gj_, then In G =  y,  and  the
distribution of ln(FT/Fc) nas a mean of zero.

Now it is assumed that criteria pollutants constitute  a random
sample of size four from the population of all pollutants.  From
Table E-l, the values shown in Table 1-2 are obtained.

     TABLE 1-2.  CONVERSION FACTORS FOR CRITERIA POLLUTANTS

Pollutant
Particulate matter
Sulfur oxides (SOX)
Carbon monoxide (CO)
Nitrogen oxides (NO )
J\.
Gi
0.0260
0.0281
0.158
0.0411
In G.
-3.650
-3.573
-1.845
-3.191

Using these values, the estimated mean and standard deviation of
the In G.  are

                           x = -3.064

                           s = 0.8395

Thus, the estimate for G is


                      G = ey = ex = 0.0467                   (1-9)

The estimated 95% confidence interval for ln(FT/Fc) is  ฑ1.96 s;
i.e. ,


                                /M
                     -1.645 < ln(=i)< 1.645
                                \ C/


                                FT
                        0.193 < =rฑ. < 5.18
                              "  C ~

                     0.193 FC < FT < 5.18 FC                (1-10)

                 /\       /\
By definition of d   and d „,
                  su      sx,


                             ^ FT <

-------
From Equations 1-10 and 1-11,
                            d   = 4.2
                             su
In the Source Assessment Program, G = 0.0033, so that In G = -5.70
In this case, the estimated 95% confidence interval for F^/FQ  is
given by

                    [-3.064 -  (-5.70) ] ฑ 1.96 s

That is,

                                FT
              2.64 - 1.645 <  In =ฑ-  < 2.64 + 1.645
                                FC  "

                               FT
                        2.70  < =ฑ- < 72.6
                              ~ FC ~

                     2.70 FC  < FT < 72.6 FC                 (1-12)

Comparing Equations 1-11 and  1-12 yields
                           /\

                           dsu = 71'6
Thus, there are three sets of values for the uncertainty in
"acceptable" concentration.

Case A.  No uncertainty in acceptable concentration.
                          s\     s\
                          d „ = d   =0
                           si    su

Case B! .  Acceptable concentration uncertain; TLV conversion
          factor equal to geometric mean value.
Case 62 .  Acceptable concentration uncertain; TLV conversion
          factor equal to 1/300.
3Note that d ฃ > 0 by definition.

                              149

-------
dsu = 71'6
    150

-------
                           APPENDIX J

           EXAMPLE CALCULATIONS FOR SOURCE ASSESSMENT:
                    CARBON BLACK MANUFACTURE
In this appendix, the principles developed in the body of the
report are applied to a source assessment study of carbon black
manufacturing operations .(49).  In this study, emission  factors
for the main process vent were measured at a typical carbon black
plant.  Mean emission factors and 95% confidence limits  for the
mean values  (computed from the experimental data) are listed in
Table J-l.  These values were taken directly from Reference 52.
Also listed in Table J-l are the random and systematic uncertain-
ties associated with the measured emission factors.  The values
of br are simply the 95% confidence limits for the mean  emission
factors.  The systematic uncertainties for nitrogen oxides and
carbon monoxide were obtained by applying Equations G-5  and G-6
to the biases listed in Table 1-1 for EPA Methods 7 and  10,
respectively.  The systematic uncertainties for the other species
in Table J-l were equated with the generalized values given in
Table 3  (Section 5), since no information is available concerning
biases in the methods used to measure emissions of these
compounds.

TREATMENT OF ST AS A FIXED PARAMETER

The random uncertainties^in average plant capacity and average
emission height (ar and cr) were set equal to zero since data
were available for all plants in the industry.  Values given in
Table 3 (Section 5) were used for the uncertainties in plume
rise, dispersion equation, and "acceptable" concentration.

For each pollutant listed in Table J-l, the following quantities
were computed:

    • Estimated severity,  S  (including plume rise).
    • Lower and upper bounds for the true severity,  S .
    • The critical test value,  S*.
                                \^
(49) Serth, R. W,, and T. W. Hughes.  Source Assessment:  Carbon
     Black Manufacture.  EPA-600/2-77-107k, U. S. Environmental
     Protection Agency, Research Triangle Park, North Carolina,
     October 1977.244 pp.
                               151

-------
Results for criteria pollutants are presented in Table J-2.  Cri-
teria pollutants correspond to Case A of Section 5; i.e., no
uncertainty in "acceptable" concentration.  The critical test
value derived in Section 5 by the generalized analysis was  0.05.
The values for nitrogen oxides and carbon monoxide are signifi-
cantly higher than the generalized value because of the smaller
systematic uncertainties in emission factors for these two  species
    TABLE J-l.
    EMISSION FACTORS AND ASSOCIATED UNCERTAINTIES
    FOR CARBON BLACK PROCESS VENT

Material
emitted
Particulate matter
Nitrogen oxides
Hydrocarbons
Carbon monoxide
Hydrogen sulfide
Carbon disulfide
Carbonyl sulfide
Isobutane
n-Butane
POM (total)
Measured emission
factor , g/kg
0.11 ฑ 70%
0.28 ฑ 15%
50 ฑ 48%
1,400 ฑ 19%
30 ฑ 82%
30 ฑ 76%
10 ฑ 99%
0.1 ฑ 80%
0.27 ฑ 57%
0.002 ฑ 52%
b
r
0.70
0.15
0.48
0.19
0.82
0.76
0.99
0.80
0.57
0.52
*sฃ
0.1
0
0.1
0.09
0.1
0.1
0.1
0.1
0.1
0.1
b
su
0.5
0
0.5
0.02
0.5
0.5
0.5
0.5
0.5
0.5

TABLE J-2.
EXAMPLE CALCULATIONS FOR CRITERIA POLLUTANTS
(CASE A:  NO UNCERTAINTY IN ACCEPTABLE CONCENTRATION)

Material
emitted
Particulate matter

Nitrogen oxides
Hydrocarbons
Carbon monoxide

Calculated
severity, S
0.004

0.032
4.2
0.36

Uncertainty interval
S — e < S  S^- for these species), but not
with regard to nitrogen oxides and particulate emissions  (since
SG < S* for these species).
                              152

-------
 For noncriteria pollutants,  calculations were made for both Case
 BI  (TLV conversion factor of 0.0467)  and Case B2 (TLV conversion
 factor of 0.0033).  Results  of these  calculations are given in
 Tables J-3 and J-4,  respectively.   The critical test values
 obtained by the generalized  analysis  in Section 5 were 0.01 for
 Case BI and 0.05 for Case B2.   The critical test values listed in
 Tables J-3 and J-4 are essentially the same as the generalized
 results since the same uncertainty values were used in both
 cases.
   TABLE J-3.
EXAMPLE CALCULATION FOR NONCRITERIA POLLUTANTS
(CASE BI:  TLV CONVERSION FACTOR = 0.0467)

Material
emitted
Hydrogen sulfide
Carbon disulfide
Carbonyl sulfide
Isobutane

n-Butane
POM (total)
Calculated
severity, S
ซu>
0.28
0.11
0.030
0.000014

0.000042
0.030
Uncertainty interval
0
0
0
0

0
0
<_ ST 1
- ST 1

-------
                           GLOSSARY
accuracy:  Closeness between true and measured values which is
     characteristic of a given measurement process.

atmospheric stability class:  Class used to designate degree of
     turbulent mixing in the atmosphere.

bias:  Magnitude and direction of the tendency of a given mea-
     surement process to measure a quantity other than the
     nominal or intended quantity.

criteria pollutant:  Emission species for which an ambient air
     quality standard has been established.

decision index:  Numerical index used as an aid in decision
     making.

emission factor:  Weight of material emitted to the atmosphere
     per unit weight of product produced.

expected value:  First moment about the origin (i.e., the mean)
     of the probability distribution of a random variable.

Monte Carlo simulation:  Simulation in which variables are
     represented by probability distributions.

national emissions burden:  Decision index defined as the ratio
     of emissions of a criteria pollutant from a given source
     type to emissions of the same pollutant from all stationary
     sources nationwide.

Neyman-Pearson theory:  Most widely accepted theory of statis-
     tical inference.in which population parameters are re-
     garded as fixed.

noncriteria pollutant:  Emission species for which no ambient
     air quality standard has been established.

operating characteristic curve:   Curve which gives the (fiducial)
     probability that the true severity is less than one as a
     function of the calculated severity.

plume rise:   Distance above point of emission where plume rises
     due to its momentum and buoyancy.


                               154

-------
precision:  Closeness together, or lack of scatter, in replicate
     measurements which is characteristic of a given measure-
     ment process.

random error:  Component of the total error which results from
     imprecision of the measurement process.

random uncertainty:  Component of the total uncertainty which
     results from random error in the measured value.

reasonably available control technology (RACT):   Available,
     economically feasible control technology required for
     attainment of national ambient air quality standards (as
     specified in Code of Federal Regulations, Title 40,
     Chapter I, Part 51.  Requirements for Preparation, Adopt-
     ion, and Submittal of Implementation Plans).

representative source:  Hypothetical source having typical values
     of parameters which is used to characterize an industry
     or source type.

source severity:  Decision index defined as the ratio of maxi-
     mum time-averaged ground level pollutant concentration to
     an acceptable pollutant concentration.

systematic error:  Component of the total error which results
     from bias in the measurement process.

systematic uncertainty:  Component of the total uncertainty
     which results from the systematic error in the measured
     value.

Type I error:  Error in decision making which results from
     rejecting the null hypothesis when it is true.

Type II error:  Error in decision making which results from
     accepting the null hypothesis when it is false.

uncertainty:  Measure of the lack of knowledge of the true
     value of a quantity due to imperfect data;  i.e., due to
     errors in the measured values of the quantity.
                              155

-------
             CONVERSION FACTORS AND METRIC  PREFIXES
                                                      (50)
                          CONVERSION  FACTORS
 To convert  from

Degree Celsius  (ฐC)
Degree Kelvin  (ฐK)
Gram/second  (g/s)
Joules (J)
Kilogram (kg)
Kilometer (km)
Meter (m)
                                     To
Meter'1
           (m2)
Meter3 (m3)
Meter/second
Metric ton
Metric ton
Metric ton
Newtons (n)
Pascal (Pa)
Pascal (Pa)
Watts (W)
                 (m/s)
Degree Fahrenheit
Degree Celsius
Pound/hr
Calories
Pound-mass  (avoirdupois)
Mile
Foot
Foot2
Foot3
Mile/hr
Pound-mass
Kilogram
Ton (short, 2,000 pound mass)
Dyne
Millibars
               r\
Pound-force/inch'' (psi)
Calories/min
                                Multiply by
                                                       ^
                                                       t*
            = 1.8 tฐ + 32
            = tฐ - 273.15
                    7.937
             2.388 x 10"1
                    2.205
             6.214 x 10"1
                    3.281
              1.076 x 101
              3.531 x 101
              2.237 x 103
              2.205 x 103
              1.000 x 103
                    1.102
                                                           1.000  x  10
                                                           1.450  x  I0~k
                                                            1.434 x 101
                            METRIC PREFIXES
  Prefix  Symbol   Multiplication factor
       Kilo
       Mega
       Micro
            k
            M
            P
      10
        3

        -6
                                                     Example
1 kPa =  1 x 103 pascals
 1 MJ   '     ""
 1 pg
= i x lu pascal'
= 1 x 166 joules
  1 x 10~6 gram
(50) Standard  for Metric Practice.  ANSI/ASTM Designation
     E  380-76  ,  IEEE Std 268-1976, American  Society for Testing
     and  Materials, Philadelphia, Pennsylvania, February 1976.
     37 pp.
                                   156

-------
                                 TECHNICAL REPORT DATA
                          (Please read Instructions on the reverse before completing)
 1. REPORT NO.
  EPA~600/2~78~004u
                            2.
                                                       3. RECIPIENT'S ACCESSION NO.
 4. TITLE AND SUBTITLE
                SOURCE ASSESSMENT: Analysis of
                                                      5. REPORT DATE
                                                       August  1978
                                                       6. PERFORMING ORGANIZATION CODE
 and E.C.Eimutis
                    ,  T. W. Hughes, R. E. Opferkuch,
                                                      8. PERFORMING ORGANIZATION REPORT NO.

                                                        MRC-DA-632
9. PERFORMING ORGANIZATION NAME AND ADDRESS
Monsanto Research Corporation
1515 Nicholas Road
Dayton,  Ohio 45407
                                                       10. PROGRAM ELEMENT NO.
                                                       1AB015; ROAP 21AXM-071
                                                       11. CONTRACT/GRANT NO.
                                                       68-02-1874
 12. SPONSORING AGENCY NAME AND ADDRESS
 EPA, Office of Research and Development
 Industrial Environmental Research Laboratory
 Research Triangle Park, NC  27711
                                                       13. TYPE OF REPORT AND PERIOD COVERED
                                                       Task Final; 11/76-3/78
                                                       14. SPONSORING AGENCY CODE
                                                        EPA/600/13
 is. SUPPLEMENTARY NOTES IERL-RTP project officer is Dale A.. Benny, Mail Drop 62, 919/
 541=2547.  Similar previous reports are in the EPA-600/2~76~u32 and -77-107 series.
 is. ABSTRACT r^ rgp0rj. gjves results of: an analysis of the uncertainties involved in cal-
 culating the decision parameters used in the Source Assessment Program; and a
 determination of the effect of these uncertainties on the decision-making procedure.
 A general procedure for performing an  analysis of uncertainty is developed, based on
 the principles of error propagation and statistical inference.  It is shown that this sim-
 ple and straightforward method represents and approximation to standard statistical
 techniques. The approximate method is illustrated by application to four problems  in
 the area of environmental control.  The general procedure Is used to establish guide-
 lines for conducting air emissions studies in the  Source Assessment program.  In
 particular, guidelines are established for precision in field sampling and analytical '
 work, and for setting critical values of decision parameters.
17.
                              KEY WORDS AND DOCUMENT ANALYSIS
                 DESCRIPTORS
                                           b.lDENTIFIERS/OPEN ENDED TERMS
                                                                      COSATI Field/Group
 Pollution
 Uncertainty Principle
 Statistical Inference
 Statistical Distributions
                        Sampling
Pollution Control
Stationary Sources
Source Assessment
Error Propagation
13B
20J
12A

05A
Field Tests
13. DISTRIBUTION STATEMENT
 Unlimited
                                           19. SECURITY CLASS (This Report)
                                           Unclassified
                                                                    21. NO. OF PAGES

                                                                       180
                                          20. SECURITY CLASS (This page)
                                           Unclassified
                                                                   22. PRICE
EPA Form 2320-t (S-73)
                                      157

-------