PB85-240455
Methodology for Characterization of Uncertainty in Exposure Aasesraent*
ftetearch Triangle Institute
Bftaearch Triangle Park, MC
Aug 85
                    U.S. DEPARTMENT OF COMMERCE
                  National Tichnictl Infomition Sirvlci

                                  IMTIS

-------
                                           EPA/600/8-85/009
                                           August  1985
 METHODOLOGY FOR CHARACTERIZATION OF UNCERTAINTY
             IN EXPOSURE ASSESSMENTS
                       by

                 Roy W. Vhitaore
           Research Triangle lostitutc
        Research Triangle Park,  NC  27709
          ' Contract No.:   68-01-6826
          Task  Manager:  Jaaes V. Faleo
          Project  Officer:  Jobn Varren
OFFICE OF HEALTH AND EHVUOWOKIAL ASSES9fENT
     OFFICE OF RESEARCH AND DEVELOPMENT
    W.S."ENVIRONMENTAL PROTECTION AGENCY
            WASHINGTON, DC 20460

-------
EPA/600/8-85/009
Htthodolooy for Characterization of Uncertainty 1n
Exposure A
-------
                                 DISCUIHER
     Tbt information la thii report tut bctn fuadtd wholly or la part by the
Oaitod lutu toviroBMBUl Protection Aftncy mrftr Contract Bo. 66-01-6826
to tkt ItMtreh TtUaili laccltutt.  It tut Wto wb>et of tht Aftncy't ptcr
oad alaiviatntivt nvltv. *ad it •«• btco cpprevid for pvbllcttioo •• aa IPX
torant.  lantioB of trado amavs or coaMreUl product! dott aot eoaitltutt
ooJoraoMat or roeoawadatioa for ott.
                                     11

-------
                              TABLE OF CONTENTS

                                                                      Pige

IIST OF TABLES  	

LIST OF FIGURES	

FOREWORD	

ABSTRACT	      vi

ACKNOWLEDGMENTS	     vi"

1.   nCTRODUCTION AND OVERVIEW	      1

     1.1  Whit it an Exposure Aasessnent?	      }
     1.2  Characterisation of Uncertainty  	      5

2.   ASSESSMENTS BASED UPON LIMITED DATA	     12

     2.1  Liaited Data for Directly Measured Exposures 	     12
     2.2  Liaited Data for Bedel Input Variables 	     13
          2.2.1  Population Exposure Models  	     13
          2.2.2  Peraonal Exposure Models  	     14

3.   ASSESSMENTS BASED UPON ESTIMATION OF INPUT VARIABLE
     DISTRIBUTIONS 	     17

     9.1  Estimation Based on Input Variable Distributions ....     17
     3.2  Licitins Distributional Results  	     21

4.   ASSESSMENTS BASED UPON DATA FOR MODEL INPUT VARIABLES ....     25

     4.1  Interval Estimates of Exposure Percentiles:   Input
          Variable Distributions Not Known 	     25
     4.2  Interval Estiaiates of Exposure Percentiles:   Input
          Variable Distributions Known 	     26

5.   ASSESSMENTS BASED UPON EXPOSURE MEASUREMENTS	     29

6.   A06REGATING ACROSS SOURCES, PATHWAYS, AND/OR ROUTES OT
     IXPOSURE	     94

     6.1  frimMnint Estimated Expoaure Distributions	     94
     6.2  Umitiof Distributional Result*	     35

7.   COMBINING OVER DISJOINT SUBPOFULATIONS	     91

B.   SUMMARY . .  .	     «

REFERENCES	     *2

APPENDIX A	     A*J


                                     111

-------
                              LIST OF TABUS
 1       lOBury of Primary Methods for Chsracterisinf Uncertainty
         for Exposure Assessments  ................     8

A.I      Selected Percentiles for Estisuted Dove Resultinj free
         Bquilly Likely Input Variable Combinations   .......    A-8
         Selected Perccntiles for the Do»e Dittributiea
         tnm tht PretiMd Model «ad Loput Variables Biatributioai    A- 12
A.9      telectod Perccatilei for Satiaattd Dose itaultiat froa
         Alternative Bquilly likely Input Variable Coabia*tioni. .    A-15

A.4      95 Percent Confidence Interval  EatiMtes for Selected
         Pcrccntilet of the Distribution of Dose (T) Baaed on
         Model-Predicted Doses	    A-19

A.5      95 Percent Confidence Interval  EatiMtes for Selected
         Pereentilcs of the Dose Distribution (Y) Assuaiag Unifon
         and Beta Distributions for IR and ED/600, Bespectivcly. .    A-25

A.6      95 Percent Confidence Interval  EstiMtes for Selected
         Pereentiles of the Dose Distribution (T) Directly from
         Exposure Data	    A-30
                                     1v

-------
                                LIST OF FIGURES

Titure                                                               Pa££

  1       Zategrated Exposure Asaessaent	     2
           ii
  2       Baak AssessMBt/Risk Haaageaent Process	     4

  3       Hypothetical Jeiat Relative Trequeacies  for Exposure
          Duration (ED) aad lagestion Bate (XX)	    18

 A.I      Dlstributioa of Estimated Dotes Resultia* fro* Equally
          likely Zaput Variable Combination*	    A-7

 A.2      Hypothetical Subjective Estisute of tke  Distribution of
          Exposure Duration (ED)  	    A-9

 A.3      Predicted Dose Distributioa Resultiag fro* the Presuaed
          Model aad laput Variable Distributioas	    A-ll

 A.A      Distribution of Estimated Dose Resultiai froo Alternative
          Equally likely laput Varisble Combination!   	    A-U

 A.5      listofrao of the Hypothetical Staple of  Site 500 for
          Exposure Duration (ED)	    A-17

 A.6      HistofraB of the Hypothetical Staple of  Siae 500 for
          Xatestioa Rate (IR)	    A-18

 A.7      95  Percent Coafidence Interval Estiaates for Selected
          Perceatiles ef the Dose Distributioa Based en Model-
          Predicted Doses	    A-20

 A.B      Coafidence Region for the Estiaated Beta Distribution
          Paraaeters 6 tad $	    A-23

 A.9      95  Percent Confidence Interval Estiaates for Selected
          Percentiles of the Dose Distribution Assuaiag Uniform and
          Beta Diatributioas for XR aad ED/600, Respectively  . .  .    A-26

 A.10     95  Perceat Ceafideace laterval Estimates for Selected
          Percentiles ef the Dose Distribution fro* Directly Meaaured
          Doses	    A-31

-------
     The Exposure Asiessacnc Group (EAC) of tht U.S.  Eaviroaacntal Protection
Afency'i (EPA) Office of Research aad Developaeat has three a*io  fuoctiou:
1) to aoaduct exposure assessaeats; 2) to reriev aaaesaaents and  related
docaajeatai nd 3) to develop galdaliaoe for Afeney oipoeure aeeeieMBti.  The
actiTltioe vader aaeh of tbeee throe fuaetioae are aupporced by and respond to
the Mode of the various IPA profraa offices.
     Aa part of ita fvactioa to develop guidaliaae, IAC coaducte  reaearch
projects for the purpose of developing or refialof toehaiques used in exposure
Aaeeeeveat.  This docuateat presents the results of auch a project and serves as
a oupport doeuaent to EPA*a lateria Ouidelleei for Ispoaure Aaseesa«nt.
Specifically, thie report describee procedures vhich  can  be used  to cherecteritc
•acortaiaty ie ostiasted oxposure aad coataiaa aa aaaaple to doaoastrete the
atility of the awst coaplez techaiques.

-------
                                   ABSTRACT

     Virtually  all  exposure •••ciMcats  except  these  based  upon Matured
 exposure  levela for a  probability •••pic of population swabers  rely 1900 a
 •odel  to predict exposure.  The aodel aay be any aatheaatical function, aiaple
 or coaplex,  that  estimates the  population distribution of  exposure or  an
 individual  population  amber's  exposure «s  a  function of me  or e»re input
 variables.  Whenever •  Bedel that has aot been  validated is wed as the basis
 for aa  exposure  assessaent,  the aaeertaiaty  associated  with the exposure
 aMesaaent My be substantial.  The priaary characterisation of uncertainty is
 at least  partly qualitative ia  this case, i.e., it iacludes • description of
 the assuaptioni iaberent  ia the  aodel aad  their Justification.  Plausible
 alternative aodels  ahould be discussed.  Sensitivity  of the exposure assess-
acnt to  aodel foraulation  can  be investigated  by replicating the •sscssaent
 for plausible alternative aodels.
     Vhen  aa  exposure  assessaent is  based opon  directly  aeasured exposure
 levels for a  probability  aaaple  of population a*abers, uncertainty  can be
freatly  reduced and  described  quantitatively.  In  this  CMC,  the  priaary
 •ources of uncertainty are aeasureoent errors end stapling errors.  A thorough
quality assurance prograa should be designed into the  study to ensure that the
awgnitude of •easureaent  errors can  be estiaated.  The effects of all sources
of  randoo  error  ehould  be Measured  quantitatively  by  confidence interval
eatiaatet  of  paraaeters  of  interest,  e.g.,  perccatiles of  the  exposure
distribution.   Moreover,  the effect  of randoa errors  can be reduced by taking
• larger aaaple.
     Vhenever the  latter  is aot feasible, it is aoaetistes possible to obtain
at  least eeae  data for exposure ead aodcl iaput variables.  These dsts ahould
be  ased to  assess  goodness of fit of  the Model and/or presuaed distributions
 af  input variables.  This substantially reduces the amount  of quantitative
 uncertainty for  estimation of  the distribution of exposure and  is strongly
 recoMaended.  It is recognised,  however, thst it a«y aot always be feaaible to
 collect such data.

-------
                               ACKNOWLEDGMENTS

     Ac rather would like to acknovled|e that tuny KTI staff ambers, act all
of whoa) eta bt Matleotd la this  space,  ceatributed te preparatioa ef thii
report,   Several  discussion teaaieas  attended  by various XII  ataff •cabers
ceatribated freatly te the  preparatiea ef this  report.  The anther thaaki Dr.
Bartia Haaso|lia fer May helpful  euffeitioas aad for coatributioa ef Figures
1  aad 2,  which he  had  prepared  fer ether ITI Mseereh.  Appreciation is
earteaded  to  Or. Jaates  Chroay, Dr.  Douglas Pnsainnil.  aad Or. Clark Allen for
their austestioas aad constructive  criticise.   Thaaks is extended to Mr. Fred
IsBereun aad Mr. George Xircher  fer their technical assistance ia preparation
ef Appcadix  A.  The author  thanks the  following  peer  reviewers  for  their
careful review  ef  the draft wersiea ef  this doosaeat aad their suay helpful
aufttestions:   Dr.  Doaald  Porcella,  Or. Doaiaic  Ditero, Dr. John Speafler, Dr.
J. Julian Chisel*,  Mr.  John L.   Beaahaw;  aad  Dr.  Vill  M.   Ollisoa.   The
lesdersbip aad  insightful  direction  provided  by  the EPA task Master, Dr.
Jaaes Falco, is also apprecisted.   last, but act least, the author thanks his
aecretarial staff for their patient preparation ef  several  revisions of this
Mnuscript.
                                    V111

-------
                          1.  INTRODUCTION  AND  OVERVIEW

      AD exposure assessaeat  oftea progresses  through several stage* of refine-
 ment.  The  purpose  of  this  document  is  to provide  technical  guidance for
 characterisation ef  uacertaiaty  for  assessments  at various  stages of refiae-
     •
 •eat, from  assessments  based upon limited  initial  data to  those based upon
 ••tensive data.
      Tali docmneat is written for the professional who is actively involved in
 the performance ef exposure  assessments.  It ie intended to facilitate use ef
 the  latest  statistical methods  for characterising uncertainty  for  exposure
 essessaeats.
      This introductory chapter begins with a brief nectien that discusses the
 nature of an exposure assessaeat.  This section serves as a poiat of reference
 for the remainder of  this docuaent.   It is else intended to aake this docuaeat
 e. jeatially  aelf-contained.   An overview  ef the characterisation of uncertain-
 ty for exposure  aaaesaaeats at various  stages ef refinement  is  presented in
 the final  section of  this chapter, which  also serves es a guide to the regain-
 ing chapters.
 1.1  WAT  IS AN EXPOSURE ASSESSMENT?
      An  exposure  assessaeat quantitatively  estiaates  the  coatact ef  an af-
 fected populatioa with a substance under  investigation.  Typically, the aagni-
 tude,  duration,  sad/or  frequency ef coatact ere estiaated.   The U.  S.  En-
 vironmental  Protection Ageacy (EPA) has prepared a  draft  handbook  [1]  to
 provide guidance pertinent to performing  an exposure assessaeat.  la addition,
                                                                         t
 refereaces [2]  end (3] are  valuable  sources  ef information with regard to the
••ethodology  of  exposure Msessaeats.
      An  integrated exposure  sasessaeat  quantifies  the contact  ef populatioa
 •embers  with tae substance  under investigation via  nil routes  ef  exposure
 (inhalation, ingestion,  nnd dermal) nnd  all  pathways  from sources to exposed
 individuals  (e.g., from maaufacturiaf plants  to plant workers snd/or consumers
 •f manufactured  products).   As  shown  in  Figure 1,  an integrated  exposure
 Assessment also  •atimates the sine  of population  affected by each  exposure
 roui«.  nincc  an integrated exposure  assessment is  eitea needed to enable
 regulatory decisions/ characterisatioa  of  uncertainty is addressed  in this
 context.

-------
Figure 1.   Integrated Bxporare AsMmwnt

-------
      The  overall uncertainty associated with both the exposure assessment  and
 the  dose-response  relationships Bust be addressed by risk assessments.   Charac-
 terization of uncertainty for exposure assessments is addressed by this docu-
 ment; ether  comparable documents address  characterization of  uncertainty  for
 risk assessments per ee.   The primary purpose of an exposure assessment it to
 e&able riak asaessaeat aad possible aeleetioa of a regulatory option.  Regula-
 tory decisions emit be capable of withstanding close acrutiay (perhaps even in
 e court ef lav).  Therefore, the characterization of uncertainty for aa expo-
 sure assessment  ahould be very thorough aod based upon state-of-the-art quan-
 titative methods to the extent possible.
      The type of exposure (internal dose, external dose, etc.)  investigated by
 an integrated exposure assessment depends open the input needed to enable risk
 assessnent (see Figure 2).  Zn most cases an exposure assessnent should  ideal-
 ly quantify the  dose ef the substance mmder investigation that is absorbed by
 individual  body  organs  in erder  to  enable the  most reliable estimates  of
health effects.  Presently it is practical to quantify absorbed organ-specific
doses only  in certain circumstances--e.g., when the affected  organ is the akin
 (denul route) or  when exposure is due to radiation.   Moreover,  the absorbed
•hole-body  dose  often cannot be  quantified.  Instead, indicators of the  ab-
sorbed dose are  often used by necessity in exposure assessments.   These indi-
cators of  absorbed doses  may be either environmental  exposure  or body  burden
 (0)i  (*))•  Environmental  exposure  refers  to the  levels  ef the substance
onder investigation  contacted through the air, water, food,  soil, etc.  Body
burden  refers to  levels  of the substance in body  fluids or tissues,  e.g.,
blood,  Brine, breath, fatty  tissues,  hair, mails, etc.  In the  remainder of
this  document, the ten "exposure" will usually be used in a generic sense to
represent  either the absorbed  internal dose er aa indicator ef  the absorbed
dcae because both have been used aa the basis for exposure assessments.
     Aa  integrated  exposure assessment  generally  partitions  the affected
population into  aubpopulatioas  euch  that ell embers ef e  eubpopulation  are
effected by the  earn* eeurces ef exposure.  Tor  each aource, there can  be  ene
er more pathways by which the aubstaoce mader atudy travels from the aoarce to
the affected population.  lor each aource and pathway, e population •ember  can
be exposed via one er more ef three absorption routes:  inhalation, ingestion,
end  dermal.  An  integrated exposure assessment  eften begins by estimating
exposure  for  a  aubpopulatioa via  a apecific  combination ef source, pathway,
                                  -3-

-------
Figure 2.   Rink Acsesmwnt/Risk HaiMgewnt Proce**

-------
 and route.  Exposure is then aggregated across psthways  and  sources  to esti-
 mate exposure via a particular route for aubpopulation member*.  Exposure may
 also be aggregated across routes if this aggregate  is relevant to estimation
 ef risk.  Finally, the aises of the aubpopulatieas  should  also be estimated.
      An exposure assessment addresses  the  exposures experienced by  the sub-
 jects (human or aonhuman) of a specified population.   The most ceaplete charac-
 terisation ef  these exposures is  the  population distribution of  individual
 exposures.   If  the  population  distribution ef expoaure  was  known,  then  all
 summary parameters  of the distribution would alee be  known,  such as  the fol-
 lowing:
      1.   Minimum of all possible exposures.
      2.   Kaxiaua of all possible exposures.
      3.   Mean,  aedian, end ao*e ef the population exposures levels.
      4.   Standard deviation ef the expoaure distribution.
      5.   Proportion of subjects whose exposure exceeds any specified  exposure
          level.
      6.   Pcrcentiles of the exposure distribution.
Since knowledge  of  the diatributioa ef exposures implies knowledge of all the
above parameters that are ao useful for evaluation of risk,  estimation of this
distribution  is  a  primary goal ef.exposure asaeaaments.  However,  aome expo-
aure  aasessaents Bay be  limited  by the available data to  estimating summary
parameters  of the exposure  distribution,  e.g., the mean exposure  or certain
percentiles ef the exposure distribution.
 1.2   CHARACTERIZATION OF BHCERTAim
     Inference   (5]  addresses  characterisation ef ancertainty for e  source
severity index  to  be ased an an aid in regulatory decision aaking.  Interval
•stiaatios  fer  the  aesa value ef the severity index baaed  ea "error  propaga-
tion foxaulasn  is  discussed.   Assuming that the imdex ia a  function ef random
7*ii«wlcs  for which direct confidence interval  estimates  ere available,  a
first-order Taylor series  epproximation ef the  function ia  recommended  for
computation ef  as  approximate  confidence interval estimate  for the mean value
ef the  index.  A Taylor aeries approximation is  alao  recommended for deriva-
 tion  of bounds  on  the systematic  error,  or hiss, in estimation of  the Bean
eeverity  index.   Tables of  error propogstion formulas for random  errors  and
                                  -5-

-------
 systematic  errors are  provided.   Reference  (1]  recommends  these  methods  for
 characteriaiag  uncertainty for exposure  assessments.   However,  these  methods
 have limited applicability for the following  reasons:  (1) The primary purpose
 of am exposure assessment is usually to estimate the distribution of exposure,
 mot  a mean severity index,  and  (2) direct confidence  interval  estimation of
 the mean index would be preferable to using "error propagation formulas."  The
 present document will concentrate on estimation of the  distribution of expo-
 aares  aad on characterisation  of the uncertainty associsted  with this esti-
 mated distribution.
     A  review af methodology for characteriaiag  uncertainty is  provided by
 reference  |f].    This  article discusses  uncertainty for a predicted  output
 variable T when  the output  can  be  expressed as a deterministic  function of
 several random variables.  Since  it is the  uncertainty of the predicted output
 that  is  of  interest  in this  article,   the  probability distribution  of  the
 output variable aad  the variance  of this distribution are advocated as primary
 characterixations  of  the  uncertainty associated  with the predicted  output.
Vithin this context, the estimated distribution of axposures could be regarded
as a characterixation  of the uncertainty with regard to the prediction of en
individual population  member's  exposure  level, T.  However, estimation of the
population distribution of exposures is the primary goal of an exposure assess-
ment, aad it  is  the uncertainty  with regard to this estimated distribution of
exposures that must be addressed.  Thus, reference (6) is a valuable reference
with regard to  characterisation o'f uncertainty, but the present document will
 rely mainly upon somewhat different methods that more directly address charac-
terising uncertainty for an estimated distribution of exposures.
     The present document 'is not iateaded to be a general  review of method-
ology for characterising uncertainty.  lastead, this document focuses on those
aethods for characteriaiag macartaiaty that are most appropriate for exposure
assessments,  the determination of which methods are appropriate  for charac-
 terising the macerteiaty of an estimate depends apoa the underlying parameters
feeing estimated, the  type aad extent of slats available, aad the estimation
 freceduies •tilised.  Therefore,  this  document is laraely organised eccording
 to the types of 4ata aad estimation procedures that «ay reasonably form the
 basis for  en  exposure essessmeat.  As a result, e great deal of this document
                                  •6-

-------
 actually addresses estimation procedures for exposure assessments.   The aethod-
 elegy  ceaaidcrcd most appropriate for  characterizing  the  uncertainty associ-
 ated vith aa exposure aaaessaent ia tbeo diacuaaed with regard to each eatiaw-
 tiea procedure,   for example,  when the population diatribution of  exposure  is
 being  eatimated, characterization ef uncertainty eddreaaea the  poaaible  dif-
 ferencea between the eatimated diatributioa ef expoaures «ad  the true popula-
 tion diatributioa of exposures.
     An expoaure aaaeaameat ia eftea baaed ea eae er  more expoaure aceaarios
 •ad eaaociated mathematical models.  Aa expoaure accaario ideally coasiders
 •11 potential eoureca of the •ubataace that could ceme la contact vith popula-
 tiea eubjecta.   Expoaure  predictioa  models baaed opoo traaaportation aad  fate
 ef the eubataace would then be uaed to describe  the pathways from aources  to
 coatact vith population Members.  Theae aodels auy be very  coaplex, uaiag  data
 •a levels ef the eubataaee at aources aad/or ambient monitoring aites aad the
 geographic diatributioa  ef the target  populatioa to  eatiaate  the population
diatributioa ef  expoaures (|1), (7], (g](  (9)).  Alternatively, these aodels
caa be relatively aiaple  models,  auch aa amltiplicative factora, that predict
•a individual population  aeaber'a  peraoaal expoaure as a  function of aeveral
input variables  (aee, e.g., references (3] ead [10J, ead Appeadix A).
     Table  1  provides  an  overview  of  the primary methods  recoanended for
characterizing uncertainty for expoaure eaaeaaaenta.   Both  qualitative and
quantitative actbods for  characterizing uncertainty are presented  for each  of
the following five types ef data:  •
        1.   Measured  personal expoaure  levels  for a  eaaple  of  population
               obers.
        2.   Measured model  iaput variables for  a aaaple of population mem-
             bcrs, given a model that predicts  peraoaal exposure aa  a  function
             ef input variables.
        3.   Estixated Joint  diatribution ef tjodel  iaput variables,  given  a
             ajodel that predicta  peraoaal •xpoaure as • fuaction of the iaput
             variables.
        4.   limited data  for model  iaput variables, given a model that pre-
             dicta peraoaal  •xpoaure «a  a fuaction  ef the iaput  variablea.
        S.   Data en levels  ef • eubstance et fixed aites (possibly including
             eourccs) and geographic distribution ef pepulatioa meabera, given
             a model  that predicta the pepulatioa exposure  diatribution frea
             aucb input data.
                                  -7-

-------
                                     I.
                                                                                     fot iMtMtoHatM.
atlllMi
                                                             QMlltatltW Ml
                                                                                                     ItottMi
                  •la •«
                                                    I.  IMtatlaM af Ike
                                                «Mta*   f.  CMfMMM
                                                             ••  U9 •MpMVft 4l0telMB%(Wi MM/vV S€|M?V
                                                                       •f tft* ••!••!« MatHkatlaa.
                                                         I.  MtlMtlM «ff
                                                             • IM, Ml tM
                                                                                                                    •f axtorllal a»a-
             f«r •    I.  tl«ltatl«w «ff tk*
                        «M •»•«•• €• at
                                                    1.  tollflltf «f lk»
                                                                               «M!«I    I. CMflttM* latotwl Mt
                                                                                           «f tkt i^tnu «i«trltatl«i Mi/«t
                                                             If
                                                                                                               ••• •!•• available f«r

»
                                                                                        ft.  wttartlw *)ff Ik* MMltlHty «T Mr
                                                                                           MttMlM to flMBlM* •ItCIMllW t
                                                                                           tl«M •! IM MOTl.
              AMHMtlM «f M*t
           lMt |I«M • M*>l that
f«t«lcta MHIMJ i»yiliri a* • f«w-
UM af tM l*Mt var
                     I.  ««tl«ltf «ff IW
                        ran
                     t.  ll«ttatl«M «ff tM Mt> w
                        Mala for the foprt variable
                        a-litrlMtlaa*.
                                                                                        I.  Ol !<•••! «t fit *f IM MtMM«4 J*IM AUtrl-
                                                                                           MtlM «f MaVI l«Mt varlakl** If 4at« ar*
                                                                                           a»allakl« for tt» <•§•* mrlmlc* far • aaa^lr
                                                                                           «f
                                                                                        ff.  tatlMtlM 9i tftt MMltfvltf *f
                                                                                           MtlMtM *ff f>|Ma,!M« alUtMtlv*
                                                                                           tl«M *f tM MWl.
                                                                                            If «aU m avallakl* far katk I
                                                                                           tiaa
                                                                                                                             «arl>»ln
             ar a»a>l laa«t varl-
•klM. «I«M a MaVI that antflrta
Plliiaal §•!••»! aa a faaxtlan af
th» laMt varlablaa.
                     I.  ValMlty af tM ac
                            I.
                                                   1.  llaltatlaM af th» 4at*.
                                                                                        I.  BatlMtlaa) af tk* MMlllvltf af Mrawl»r
                                                                                           catlMtM to alaMlfcU rMM« af tkr l«a»t
                                                                                           varlahlc* «•!•«• aaaVar alanlkl* alt*ni»ll«r-

-------
                                               T«kl* I.
                                                       »rlMfr MKk*M far j»***rt*H*tM. Mortal****
           •»*• •till***                      «wlllrtl«t NrtkM*                 «M**U*tl** HMk*M
                           *t fl**4  I.  V*IMU* «f Ik* yoriUtlM         I.  btlMtlM «ff tto
                                                                                    to plM
              (!•*• • w4tl tint      t.  ItollatlwM »f tfce !•*•% varl-
        th* p^MtttlM !•»•••"          •*•'* <'•»••                        t-  VrilfMloB mi Uw
            tnm Mck lar^ «•!•.                                             MtlMllwi. Mi MtlMtlM •!
                                                                           •ccvrwy Mi ff«cl«l«* •*• all
                                                                           ••ck ^MMItMlw* ctonriOTlMttM •( *w«r-
                                                                           talaty My act k* Mi«lkl« «ltk tkls type »f
                                                                           at*.


*lt !• MMMi IkM Ik* VMfM* *f Ikt CMMBt* MMIMtlH I* t« MtfMt* Ik* il*tr1k*tl«i *f MfMM* Mi/**
 lk*n*l (*.t>* ••••• >M> ••* *)lk *crcc«tllr«( •••!«• pocalkle *>*««>rc( Bra*«rtl«i «f MMlMlM wllk
 ***clffl*4 lcv*l. »tc.).

-------
         	•- ••«-aea uiai tfte purpose ef the exposure asset.aent is  to
 estimate the  distribution cf exposure and/or summary parameters thereof (e.g.,
 mean,  5th  and  95th perceotiles,  Mxiaua possible  exposure, proportion  of
 population with exposure exceeding • specified level, etc.).
     The  initial exposure assessae&t  for a •ubstance  it often based upon a
 very limited  amount of data.  Under auch conditions, the primary characteriza-
 tion of  eBcsrtainty may  be  qualitative.  Exposure assessments  based  upon
 limited data  are discussed in Chapter 2.
     Chapter  3  then discusses assessments baaed  upon  the estisuted  joint
 probability distribution of model input variables, liven a model that predicts
 ptnoaal opoaure  as a  function of the iaput variables.  Depending  open the
 mtsnt of  data available to validate the  awdel as  well as the input variable
 distributions, this  type of assessment has the potential  for  reducing uncer-
 tainty relative to the Mthods discussed in Chapter 2.   It alao has the poten-
 tial for a more quantitative characterirition of the uncertainties.
     The uncertainties associated vith an txpoaure assessment  can  potentially
 be  reduced further  by basing the assesament on measured values of  awdel input
 variables  for a  sample of population embers, given  a awdel that  predicts
 personal exposure aa a  function of the input variables, as discussed in Chap-
 tar  4.   Finally, Chapter 5  discusses exposure assessments based  on  measured
 exposure levels  for a aample  of population members, which has the  potential
 for minimising the uncertainties aaaociatcd vith an assessment.
     Vhen the  initial exposure assessment for a substance is  performed,  only
 limited  data  relating  to  exposure will  usually be available.  As more re-
 sources are devoted  to  an  exposure assessment, a more  refined assessment may
 be generated.  There may be many atages of refinement for sn exposure assess-
SKnt.  The degree of refinement ultimately meeded for sn sxposure saaeaament
efcpeads upon  the accuracy meeded to enable risk management decisions, sa  well
 ma the coat of obtaining adequate data, the time available  for data  acquisi-
 tion, etc.                                   ,. ,~>-*r>, : .« v • v
     Vhen  the exposure  levels  experienced by individual  population  subjects
 sre  measured, the  measured exposures  generally rtflect the total  expoaure
 across  nil sources  snd pathways,  mud possibly  across  nil routes as well.
 Bovrrer, mmen the modeling  approach ia employed for sn exposure nsaeasment, a
 ejodel  often  predicts sn individual  population member's exposure due to  s
 specific  combination of source, pathway, snd  route.   Methods for  combining
                                  -10-

-------
estiaated exposure  distributions  over sources, pathways, and routes  are  dis-
cussed  in  Chapter 6.   Finally,  Chapter 7  discusses coabining the  estiaated
exposure distributions  over  disjoint  subpopulations to produce  sn estiaate of
the exposure distribution  for the total population.  Methods for characteriz-
ing the uncertainties associated with these estiaated  exposure distributions
•re also discussed in Chapters 6 and 7.
     A  hypothetical  czaaple  of a subpopulation exposure assessment  that  pro-
gresses through  several stages of refiaeaent is presented in Appendix A.   The
sections of the  exaaple in Appendix A correspond directly  to the sections of
the first  five  chapters of  this report end ere auabered  accordingly (e.g.,
Section 4.1 of Appendix A  corresponds directly to  Section  4.1  of Chapter 4).
                                  •11-

-------
                  2.  ASSESSMENTS BASED UPON LIMITED DATA

     The initial exposure assessment for • substance My be based upon limited
data  for directly  measured personal exposures, concentrations in the environ-
ment,  and/or  variables  that can be used to  predict personal exposure.  These
data  say be  either  extant data  or data  produced by  • Mall-scale study.
Characterization of uncertainty for  auch  an  initial  exposure  assessment de-
pends  mainly upon  whether the  data ere direct  measurements of exposure  or
measurements  of variables  that can be used  to predict exposure.   The purpose
of this chapter is to discuss characterisation of uncertainty in each of these
situation!  when data  are  limited.   Characterisation ef  uncertainty  for  an
assessment baaed upon directly measured exposures for a  small sample of indi-
viduals is discussed first.  Characterization of uncertainty is then discussed
for  estimates of  the distribution  of  exposure  based upon  limited  dsta  for
model  input variables.   Both the  population exposure  models that  directly
estimate the  population  distribution of exposure  and the  personal  exposure
models that estimate  the  exposure experienced by a  sample of population mea-
bers are addressed.
2.1  UNITED DATA FOR DIRECTLY MEASURED EXPOSURES
     Historically,  the discovery of a new environmental pollutant often begins
when  a physician  recognizes  a clustering  of cases with common clinical fea-
tures that are unique and unexpected for the community in which the clustering
has  occurred.   If  the  physician • reports  this  situation  to the  appropriate
public health authorities, an initial study on a small sample of unusual cases
may be undertaken.  Alternatively,  if the physician reports  the  situation  in
the  medical literature,  a followup  study  by public health authorities may
result.  Under  these circumstances,  the  initial assessment  is likely  to  be
based on limited extant data, i.e., data for a sample ef  the unusual cases and
standard baseline laboratory  values.   Zf a potential  health  problem  was dis-
covered, a more refined assessment bssed upon a probability sample1 of popula-
tion members  would  almost always be needed  to permit defensible risk manage-
ment decisions.
 *A  probability sample  is  a aample  for which every  populstion member has  a
 chance  (i.e.,  positive  probability)  of inclusion in the sample.  The simplest
 example  is a  simple  random sample,  in which every population  member has  the
 same  positive  probability  of inclusion in the sample.  See reference [11]  for
 an  introduction to probability sampling methods.
                                  -12-

-------
      If an assessment  is  bated upon a  Halted  database  of Measured personal
exposures,  a  priaary aspect of the uncertainty characterization is whether or
not  the data  result  froa a probability sample.  Characterization of uncertain-
ty  for an assessment based upon a valid sample of directly Measured exposures
is  discussed la  Chapter  5.  Statistics calculated  from a  small  probability
sample generally  have  relatively large standard  errors, i.e., a  relatively
hith level of uncertainty.   However, a  small probability sample provides  the
basis  for  more  rigorous  quantitative  characterization  of uncertainty  than
would be possible for a small non-probability sample.
      If the measured personal exposure levels for a small sample of population
•embers were  not so low that health risks were known to be negligible, a more
refined exposure  assessment might be pursued.  Two ways of refining the expo-
sure assessment would be as follows:
      1.   Measure  exposure levels for a large enough sample to make inferences
          with a  predetermined  level of precision using the methods discussed
          in  Chapter 5.
      2.   Develop  an  exposure  scenario and  associated exposure  prediction
          model  to be used  ss  the basis for an assessment  ss  discussed  from
          Section  2.2 through Chapter 4.
Since  the  exposure variable needed for predicting risk  is  often not amenable
to direct observation, the latter model-based approach is often pursued first,
if not exclusively.  Hence, the model-based approach to refining an assessaeot
is discussed  first in this document.
2.2  LIMITED DATA  FOR MODEL INPUT VARIABLES
2.2.1  Population  Exposu*-' Models
     An  initial  exposure  assessment is often based upon  extant data or other
easily obtainable  data regarding ambient concentrations and/or emission levels
for  known  sources of the substance  in question.  Such data may then  be  com-
bined  with  data  on the geographic distribution of population members to esti-
mate  the population distribution  of exposures  using  an  elaborate  population
exposure model  ([1], (7],  (6],  (9], (12]t [13], [U]).  Such exposure assess-
ment?  te«i t9 be subject to a high level of uncertainty.  However,  the charac-
terization of uncertainty for these assessments is often primarily  qualitative
due  to the nature  of the underlying data and the estimation methods.
      Characterisation  of uncertainty  for assessments bssed upon  population
exposure models  consists  primarily of a description of the limitations of the
population  exposure model  and limitations  of the data  for the model  input

                                  •13-

-------
variables  (see especially  reference  I 12]).   In  addition,  sensitivity  of  the
predicted exposure distribution to plausible changes in the model and/or input
variable values  can be quantified (see references I 13 J end | 14]).  Additional
quantitative characterisation  of  uncertainty My not be possible, but valida-
tion  of the exposure  model, confidence interval estimates  of  relevant para-
meters  of  the  exposure distribution, and estimation of  the accuracy end pre-
cision  of input variable measurements should all be pursued whenever possible.
2.2.2   Personal Exposure Models
     An initial exposure  assessment say  be based open limited data  (e.g.,
estimated  ranges)  for the  input  variables of e model that predicts personal
exposures.  The  personal  exposure model would be bssed  upon an exposure sce-
nario that  considers  all  potential sources of the substance that could result
in contact  with  population members.   Knowledge of transportation and  fate of
the substance would be used to model the pathways from sources to contact with
population  subjects.  These models can be very simple, such es models involv-
ing only multiplicative .factors   (see,  e.g.,  references  [3]  aod   (10],  and
Appendix A).  A  personal  exposure model is any model,  simple or complex, that
expresses en individual populstion member's  exposure  es a  function  of  one or
more input variables.
     If the available data  ere  only sufficient  to  support estimates  of  the
ranges  of the  input variables, the exposure assessment  might be limited to a
sensitivity analysis.   Tne purpose  of the sensitivity  analysis would be  to
identify influential aodel input variables and develop bounds on the distribu-
tion of exposure.   A  sensitivity anslysis would estimate the  range of expo-
sures that  would result  •• individual model input variables were varied from
their miniauia to their maximus) possible values with the other input variables
held at fixed  values,  e.g., their mid-ranges.  In addition, the overall mini-
mum end maxima possible exposures would usually be estimated.
     The uncertainty for  en exposure assessment ef this type would  be charac-
terised  by presenting  the  ranges  of -exposure  generated  by the sensitivity
analysis, by inscribing the limitations of the data used to estimate plausible
ranges  ot   model input variables,  end by discussing justification for  the
model.  Justification of  the model should include s description of the expo-
sure scenario, choice of model input variables, end the functional form of  the
model.  Sensitivity to the  model formulation should also  be  investigated by
replicating the  sensitivity analysis for plausible alternative models.
                                  -14-

-------
              	r	  w-r.._.w «.Jki.u.«.c.  «,  toe sensitivity analysis
 presented no potentially  significant  health rick, there Bight be no  aeed  to
 refine  the assessaent.  If both the aiaiaua tad aaxiaua exposures presented •
 potentially  significant  health  risk,  it would  be known  that the  exposure
 oceaario represented  a significant health problea without refining the assess-
 •cat.   van the aiaiaua does aot prtseat • potentially significant health risk
 •ad the •BSiawa does,  thea greater iaportanee is placed on choosing  a ensnary
 parameter of the exposure distribution (e.g., the Man or a apecific percent-
 lie) aa the basia  for a regulatory deciaioa.  A a*re la-depth exposure assess-
 •eat enabling estimation of the distribution of exposures ptnits  selection  of
 •ay auaaary parameter (aiaiaua,  aaxiaua,  aeaa, or a percentile, etc.) as the
 basis for a regulatory decision, and peraits the decision Baker to aeleet that
 euaaary parameter  after the  fact  (i.e., after the expoaure  distribution hss
 •MB eatiaated).
     The  aensitivity  BBSlysis can  be eahaaccd  by cocputiag the  predicted
 exposures  that  result froa all possible iaput variable coabiaatioas.  If each
 Input variable has only a finite act of possible values, the aet of all possi-
ble  coabiaatioas  of  the  Iaput  variables  caa be  foraed,  and the  predicted
exposure  caa be coaputed  for each coabiaatioa.  The Iaput variable coabiaa-
tioaa resultiag ia txposures  that coastitute  poteatial health problaas can
thereby  be  Ideatified.   These exposure  predictioas caa  be  used to fora  a
probability  distribution of exposures,  which will be referred to  as  the expo-
sure ocasitivity diatributioa, by coaputag the relative frequency  of  occurence
•f  each exposure  level or  interval  of  exposures.  The  resulting  exposure
eeasitivity  distribution  is   equivslent  to  the  population  distribution  of
exposures  only  if all coabiBstioas  of the Iaput variable values  are  equally
likely  to  occur la the target populatioa.  The exposure aeasitivity  distribu-
tion caa also be  derived for personal exposure aodels based  open coatiauous
iaput vsriables with fiaite ranges by represeatiag theae variables  by equally
•paced  points,  at the liait as the equal  epaces  beceae Infinitely  aaall and
She BBBber of points becoacs Infinitely large, the procedure for eoaputing the
•xpoaure  aensitivity  diatributioa  ia  equivalent to eatiaating the probability
diatributioa of expoaurcs that results froa atatistically independent continu-
ous Input variables with Dnifor* diatributioas on the estiaated ranges.  Thus,
the  exposure aensitivity  distribution can be derived analytically or by Hoate
Carlo eiaulatioa based upon independent Uniform laput  variable distribution!.
                                   •13-

-------
The Monte  Carlo  aethod consisit of raadoaly generating  input  variable values
•ad  using  these to  eocpute  corresponding txposure  levels,  genersting the
exposure teasitivity distribution via replication*.
     Statistics  based upon  the exposure  se&sitivity distribution should be
interpreted in tens  of all possible iaf .* variable  combinations.  For  exaa-
ple, tkt 13th perceatilc of the exposure sensitivity distribution would be the
exposure level exceeded by only 5 perceat of the exposures nsultiat  froa all
possible ceabiaatiofts of input variable values.   Although this  distribution of
exposures is aot an estimate of the population distribution (oaless  all possi-
ble coabiaations of the iaput variables actually arc equally  likely), it has
nore otility for regulatory decision makiag than a standard sensitivity analy-
ois.  For exaaple, if 99 pereeat of the computed axposures are  below the  level
that represents a significant health concern, the proportioa of the  population
at  risk is  almost  surely aaall  (albeit  possibly smaller  or larger than  1
perceat).
     Characterixation  of aacertainty  for the  aahaaced  oeasitivity  aaalysis
vould include  a  discussion  of liaitations of the data  and  justification for
the awdel as discussed above for the usual sensitivity aaalysis. Sensitivity
to ajodel fonulstioa  should also be investigated by  estimating  the distribu-
tions of  exposure that result  froa  using the  sane Uaifora input  variable
distributions with plausible alteraative  models aad comparing  the estiaated
perceatiles of the induced txposure distributions.
                                  -16-

-------
              3.  ASSESSMENTS BASED UPON ESTIMATION OF INPUT
                          VARIABLE DISTRIBUTIONS
      If • model  has been formulated that expresses an  individual'• personal
 Upoiure as • function of oac or more input variable*, the population  distri-
 bution ef aposure CM be estimated analytically or by Monte Carle simulation
 from an estimate of the Joint distribution ef the  model input variables.'  A
 dlatribution ef exposure derived  in  this  way ie not equivalent to one baaed
 •pen Matured exposures free e personal monitoring study.  The derived  distri-
 bution merely reflects the assumed personal exposure model nnd the probability
 distributions need to represent  the model input variables.  All such  assump-
 tions nhould be explicitly stated and justified.  Ideally, model input vari-
 ables should be  represented by  empirically validated  probability distribu-
 tions.   Bovever,  the joint distribution of model input variable's can be esti-
mated fresi  discussions  with  subject-cutter  experts   (e.g.,  vis  nnivariate
histograms when the input  variables may be treated ns  statistically indepen-
dent).   The  estimated population  distribution of exposures vill be equivalent
to the  exposure  sensitivity distribution discussed in Section  2.2.2 only for
the  case when all  the input variable distributions actually  arc  independent
Uniform  distributions.  Hence, the purpose of this chapter is to characterise
uncertainty  for exposure assessments  based mpon personal exposure  models that
csn be nsed to estimate the population distribution of  exposures froa informa-
tion  concerning input variable distributions.
9.1  ESTIMATION BASED OK INPUT VARIABLE DISTRIBUTIONS
     Vhen  the  data  for model  input variables are limited, discussions vitb
subject-matter  experts c»o  be nsed  to formulate  an  estimate ef the joint
distribution of the input  variables.   Such discussions  can be  used to inves-
tigate  independence  ef  the  input variables  and  determine  the  approximate
location nnd shape of the input variable distributions.  Qualitative knowledge
•f the  input variable distributions can then be represented muantitatively by
input variable probability distributions.   The  input variable distributions
should  be validated quantitatively usiai goodness-of-fit  tests whenever suf-
 ficient data nre available (see, e.g., Chapter • ef reference (13], er  Section
 4.3 nnd Chapter 6 ef reference
•Equivalently, one  can estimate the mar-final distributions, i.e., the distri-
bution  of  each input variable ignoring the others, if the input variables can
be treated as  independent.
                                  -17-

-------
     Discus.^ons with subject-Batter experts for the purpose of estinitins  the
 Joint  distribution of  aodel input variables should first detemine vhetber or
 not  the Bedel input variables can be treated as statistically independent  for
 the  population of interest.  If sjodcl input variables arc assisted to be inde-
 pendent,  the  theoretical and/or eapirical justification of  the  aesuBcd inde-
 pendence  ahould bo discussed.  If two swdel input variables cannot be treated
 •a  independent,  their  joint probability  distribution can  be  estimated  by
 etetiCBimiag the relative frequency of occurrence for aroaa in the cross prod-
 vet  apace defined by the two variables.  As a hypothetical saunple, the joint
'probability  distribution of  ingestion rate  (IF)  and exposure  duration (CD)
 •enld  bo  ostiMted for the aiodel in Appendix A by the relative frequencies of
 occurrence  shown in figure 3.   (These relative  frequencies are hypothetical
 for  the purpose of illustration only; these variables are treated as indepen-
 dent variables  in Appendix A.)  In addition, the etartinal distribution of each
 of the correlated variables,  i.e., the diatribution *f each variable ignoring
 (the  others, should be  derived fro* the joint distribution and reconciled vitb
 available  knowledge of the surginal distributions.  Of course,  if the input
 variables  are sot highly correlated or  if the swdel is aot aensitive  to  the
 assumption of independence,  it  suy be sufficient to treat  the variables as
 independent and to acknowledge  Minor departures free) the independence assump-
 tion as a aource of uncertainty.
     figure  3.  lypothetical Joint Xelative frequencies lor
                bpoaure Duration (CD) and tegestion late (IK)
                                   -18-

-------
      The input variables  of an exposure  model  that are  statistically  inde-
 pendent of all  ether input variables can be  considered  individually  by sub-
 ject-cutter experts, one  variable  at •  tisje.   The expert  that  is consulted
 with retard  to establishing  a  probability distribution  for  a factor  in  an
 exposure model aeed set be  conversant in statistics.  A probability distribu-
 tion can be estimated free  auch information as  the following   general  shape
 •f the  distribution, smallest possible value,  largest possible value,  aw it
 likely  value  (mode),  average value (swan), the median value aucb that half the
 values  of the variable are  larger and half are smaller, and/or percentilea  of
 the probability distribution (e.g.,  the 5th aad fSth pcrceatiles).  The sub-
 ject-matter expert aeed  not be able  to specify all  of these statistics  in
 order to estimate the diatribution.   la sea* cases, the subject-matter txpert
 Buy be  able to provide a histogram to be used as the probability distribution
 of a  model  input variable.
      Such subjective  probability distributions  should be confined  in  discus-
 sions with  other subject-matter experts.  The Delphi technique can  be  used  to
 obtain the most reliable consensus of opinion free a group of experts by  using
 a  series of  questionnaires interspersed  with  controlled feedback of  group
 opinion  (17).  If such discussions do not yield a consensus of opinion, sensi-
 tivity of the  model to the alternative input variable distributions suggested
 by  experts  should  be investigated.   Alternatively, direct measurements of the
 model input variables  and/or exposure  Itself  may be  used to estimate the
population distribution of exposures as discussed in Chapters 4 and  5.
     Discussions  with  subject-matter experts,  empirical  data,  and/or  the-
 oretical argumenta may auggest that a particular family of-probability  distri-
 butions  is  appropriate for  representing a model  input variable.  For  input
variables with e finite  range of possible values, distributions such as the
Uniform, Beta, Truncated Vormal, or Truncated lognormal may be supported.  The
parameters  of auch distributions can usually be estimated from s measure  of
 central  tendency  (e.g.,  mean, median,  or mode),  a measure  of variability
 (e.g.,  •tamdard deviation  or a mercantile),  mnd  an estimate  of the finite
 range of possible  values.   Reference (IB] provides  guidance  with regard  to
 estimation of parameters of continuous probability distributions.
      The  following considerations may be useful when trying to  determine  an
 appropriate  probability  distribution to  represent a  model, input  variable:
      1.   A histogram distribution may be s very useful  initial  estimate for
          the distribution of a model input variable.

                                  -19-

-------
          If • Model  input  variable has a finite range of possible values and
          if  a  smooth  probability function is  desired,  a Beta distribution
          (scaled to the dtsirtd range) My be appropriate.  The Beta probabil-
          ity  distribution  is  very  flexible;  its probability  function can
          ••SUM • variety of shapes by varying its parameters.
     S.   If «n  input  variable only assuats values grester than COM sjiaiaua
          value  (e.g., is positive-valued; then • GasjM, Veibull, or logaonal
          tittributioa aay be  appropriate.   The fiaaM distribution  is sjore
          flexible,  end  it*  probability function  caa •••use a  variety of
          •hapes by varying it* parameter*.
     4.   If aa iaput variable has aa aarestrieted range of possible vsluei, a
          Venal, Student's t,  or Logistic distribution My be an appropriate
          probability  distribution.   These  distributions  differ  sjaialy in
          their tails.  la •dditioa, traacated versions ef  these variables My
          be ased in aituatioas 2 aad 3 described above.
     The  example in  Appendix A  vacs  Hoate Carlo •iemlation  to predict the
diatribution of  exposure  from presuaed independent distribution* for exposure
duration  (ED)  aad  ingestion  rate  (IR),  the  a»dcl  input variables.   For •
population of V  sjeabers,  each person would have his ova iadividual value for
each of these variables.   The  actual Joint distribution of  the input variables
would then be the distribution of the N paired values of ED and IX experienced
by the iadividual population  embers.   If the correlatioa of these variables
was aot negligible,  the Joint  distribution of these variables would have to be
•stiMted (e.g., using jeiat  relative  frequencies such  as ahovn in Figure 3)
la order  to produce a reliable  estiMte of  the population  distribution of
exposure usiag the persoaal exposure aodel.
     then qualitative knowledge of the input variable distributions is used to
•ntiMte the population distribution of exposure, aaccrtaiaty is characterized
by discussing justification for the presoMd swdel aad iaput variable distri-
butions.  The choice of sjodcl iaput variables tad the faactioaal fora of the
•jodel ahould be  discussed. Empirical data aad/or theoretical arguments that
aqpport the model ahould  be  preseated.  Use  of a ejodel  aad  iaput variable
distributions that have been  empirically validated is recommended in order to
induce  aacertaiaty.   Alteraative models  aad/or alternative  iaput variable
attributions ahould also be discussed,  teasitivity to these alterastives cac
be  iavestitated  by  astiMtiag the distributions that result  fro* plausible
alternatives aad comparing the perceatiles af the iaduced exposure distribu-
tions.  All available data, even if the data are limited, ahould  be used to
validate the presumed input variable distributions aad the  predicted distribu-
tion of exposure.
                                  -20-

-------
 3.2  1IHITJNC DISTRIBUTIONAL RESULTS
      Given a  matheaatical model that expresses personal exposure as a function
 of several input variables, the distribution of  exposure  can be derived ana-
 lytically or  by using Monte Carlo simulation •• described above.  The distri-
 bution of  ejcposure  is  then induced by the distributions of the  model  input
 variables. lower, as the number of input variables becomes large, it can be
 •hem that the  induced  exposure distribution approaches  a  apecific limiting
 distribution  for certain classes  of swdels.  Ihe  true  exposure distribution
 for a model based open several input variables Bay then be approximated by the
 limit of a sequence of conceptual models with mere and more input variables.
 for these  classes of  swdels,  .the veana,  variances, and  covariaaeos  are the
 ssost important  characteristics  of the input variable  distributions  because
 they are  aufficient  to  determine  the limit ing  distribution.   Other  shape
 characteristics  of  the  input variable distributions then  becoae  of lesser
 importance for models with aeveral input variables.
     Limiting distributionsl results can easily be derived for models that can
 be  represented as a apecial case  of the following model  (such  as  the models
 discussed  in references [3] and [10], and Appendix A),
         n      d
     i • n x,  /  n  x, ,
               i*n+l x
where  all input  variables, Xj  through  X^,  are  independent  positive-valued
random variables (which includes positive constants as s apecisl case).   Since
all input variables  are positive, this model can aqulvalently be expressed  as
            n            d
     In Ys  Z  In X. -   2   In X, .                                  (2)
           i»l     *   i«n*l     *
It then follows  that if the independent  random variables  In Xj through In  X^
are aacb  that a central limit type  of result applies (i.e., there is  mo one
input vsrisble  (In X^) whose variance is large relative to the variance of  In
I), then the  limiting distribution  of the mxposore, Y,  is lognomal  
-------
 rately  reflects the  true  state of  nature and so long as  there  is  no single
 iaput variable whose variance dominates the total variance.  Data for exposure
 aad/or  Model iaput variables  should be  used to validate  the aodel whenever
 possible.
     Siace  the  sjeaa aad variance are sufficient statistics  for the  logaoraal
 distribution,  estimation  of  these  parameters  will  completely specify  the
 liatttiag  logaonal  distribution whoa the exposure model is  of the  fon given
 by Equation (1).  In fact, sny sjeasure of central tendency (e.f., acan, median,
 •r aede)  together with say acssure of dispersion (o.|., standard deviation or
 a mercantile) can be aaed to estimate the parameters af the limiting lognormal
 distribution.   Vhen directly avaaured  aiposuro  data are available  for indi-
 vidual  population subjects, those data say be asod to directly astiaate  the
 paraacters  of the limiting, lofnoraal distribution.   It is also possible to  use
 tike sjeaas, variances, aad covsrisaces of the aodel input variables to cstiaate
 the  parsaetera  of  the  liaitinf lognormal  distribution  of  exposure  without
 formulating  apecific probability distributions for the aodel iaput  variables,
 an shove below.
     A consistent estimate of the Bean of Y is the ratio

     E m «  n  ECX.) /  n   ECX.),
                    x            *
where E denotes an unbiased astiaate  of the swan (sn ostiaate  for  which  the
                                                                       m
expected value  over  all possible staples is the  populstion  aean)  snd EC is a
consistent estimate  of  the aean (an estiaate that converges to the population
Man as the  sample site increases).  Vhen  the  iaput vsrisbles can be treated
as  independent,  the  numerator  and denominator astiaatea are unbiased, but  the
ratio is s bissod estimate of T.  When  the iaput vsrisbles cannot be treated
as  independent, standard practice would be to continue to ase Equation  (3)
anyway, which still yields a consistent astisute af T.
     A  consistent  estimate af  the variance (?e)  af  the  exposure distribution
caa ae obtained by waiag a  first-ardor Taylor aeries approximation.  This
variance approximation  say be  expressed in general far a function  af a raadom
variable* as -fellows: •
        Tff.,....!)) •  2   I g  (S)l» V (X )
      C    1      B     i«l    Wi           *
               42  I   2    [  g  (X)J I g  (X)] Cav(Xi.X,)l           (*)
                               wi        " J              J
                                  -22-

-------
 where JB if the partial  derivative  operator; £  is  the mean of X., .... X  ; V
 •od Cov  denote  unbiased estimates of  variance and  eovariance;  and  V  ii a
 consistent estimste of variance (see, e.g., rcfcrtact |20] and Section 10.6 of
 fftftrace 121]).   Thii  mult  can bt applied directly to the model  given by
 Equation  (1) because  this vvdel it easily differentiate vita respect to  the
 input variables,  Xi-  Moreover,  whenever the model  input  variables  are  all
 independent,  es  is eften presumed to be true, the second ten in  Equation  (4)
 becomes sero.
      It ebould be noted that the Taylor  aeries variance  approximation, Equa-
 Uon (4),  is the basis for the "propagation of error formulas" found is refer-
 ences  (1] and  |5J.   Tbe application here is very different,  Befereaces  (1)
 aad (5) use the Taylor series to approximate the variance of the mean  of model
 predictions as a  characterisation of the uncertainty vith regard  to the popu-
 lation  awan.  The  present  application  is for  estimation  of the population
 variance  of the  exposure distribution when  the bssic  fen of  the  exposure
 distribution,  e.g.,  legnonal,  has already been established.  Thus,  Equation
 (4)  is  being  used to facilitate estimation of  the population distribution of
 axposurc.  The uncertainty vith regard  to this estimated distribution is  the
 uncertainty that  must be  addressed,  not that  for the estimated mean of  the
 distribution.
     Consistent estimates of the mean aad variance of T frost Equations (3)  and
 (4)  can be used  to fit the limiting lognormal distribution of Y directly.   No
 apecific  distributional  form need' be assumed for the model  input variables,
Xj.  A*  noted earlier,  any  estimate of  the  central tendency of Y,  together
vith any estimate of variability or a percentilc of the distribution of Y,  can
 also be used  to  directly estimate* the limiting distribution  of Y.  Moreover,
whenever the estimate sf risk simply involves multiplication of  the  exposure
by  one  or more positive factora, these results extend  directly to estimation
 of the probability distribution for risk.
     The  lognormal  distribution for  txposure  T is attained  for  the  exposure
sjodels  given by Equation  (1) as  the number of model input variables becomes
 infinitely large.  Za practice, the number of factora in a  model will always
 be  finite. The importance of the limiting lognomal result  ie that  the pre-
 dicted  exposure  distribution vill  be approximately lognormal when there  are
 neveral input veriables, none of which dominates the variability of exposures.
 Thus, when the model has a sufficient number of factors, the means, variences,
                                  -23-

-------
and  covariance*  of the input variable distributions  arc  typically sufficient
te  determine  the  limiting  distribution.   Other shape characteristic! of the
input variable distribution! then beCOM Itss important.
     Although  the limiting  distribution could  be estimated  directly for a
model  vith several  factors, research  into  the  rate of conveyance to '-he
limitiag  distribution would be  aeeded  for proper  interpretation.  Empirical
research  would be needed to determine when • sjodel had a  sufficient aumber of
factors to Me  the limiting distribution as aa adequate  approximation to the
true distribution.  The results vould likely depend open the type of model and
iaput variables.
     If locations (3)  and  (4)  are ued to  estimate the limitiag lognoraal
exposure  distribution  for  a model of the type given by Equation (1),  adequacy
af the approximation of the true exposure distribution by  the limiting diitri-
bution ahould be discussed.   The results ef an empirical  atudy en convergence
to  the  Halting distribution Bight be  cited.  If data en measured exposures
for a aaaplc af  population members were available, these data ahould be used
to test the adequacy of the limiting distribution.   Characterisation of uncer-
tainty vith regard to the estimated exposure distribution vould also need to
address aacertaiaty  vith regard  to the  estimated mesa and variance  given by
Equations  (3) and  (4).  If possible, confidence interval estimates ef  the aean
aad variance ahould be derived.   Sensitivity ef the predicted exposure distri-
bution to plausible  ranges of  values  for these  parameters  ahould  then be
investigated.  Justification of  the presumed exposure model  ahould  be dis-
cussed, as veil as limitations  of  the data used to estimate  means and vari-
ances er to validate the model or limiting distribution.
                                  .24.

-------
          <•.  ASSESSMENTS BASED UPON DATA FOR MODEL 2KPUT VARIABLES

      The  exposure asscssaeot bated upon » estiaate of  the  joint probability
 distribution for  tbc  input  variables  of  a  personal  exposure  aodel can  be
 nfiaed by collecting  saaplc survey data for all aodel  input variables.  The
 population diatributioa of  exposure  can then be estimated  by coaputing the
 •odel-based  exposure for each  saaple aeaber.  Tbese  exposure values can  be
 wed to directly coapute confidence interval estiaates for perceotilcs of the
 exposure  distribution.   If  specific families  of probability  distributions are
 accepted  as  rtpraseatint the joint distribution of the nwdel input variables,
 staple  survey  data can be used to coapute joint confidence interval estimates
 for  the parameters of  the input variable distributions.  These interval esti-
mates  for distribution  paraa>eters can  then be  ased  to cosjpute  confidence
interval estiaatea for percentiles of tt». exposure distribution.   Vsing either
•ethod  of  estiaation,   interval  estiaates for  percentiles  of  the  exposure
distribution are a useful quantitative characterisation of uncertainty for the
exposure assessment.
     The purpose of this chapter is to discuss characterisation of uncertainty
for exposure assessaents based upon aodel input data collected for a saaple  of
population aeabers.  Coaputation  of confidence  interval estiaates for percen-
tiles of the exposure distribution directly froa aodel-based  exposures for the
aaaple  subjects  is  considered first.   Derivation  of equivalent  confidence
intervals free confidence intervals on paraaetera of the input variable distri-
butions is then considered in Section 4.2.
4.1  INTERVAL ESTIMATES Of EXPOSURE PERCENTIIES:   IKPUT VARIABLE DISTRIBUTIONS
     DOT KNOWN
     For  suny  personal  exposure  aodels,  it suy  be  aore feasible and less
costly  to  collect data  en the aodel input variables, cather  than  en the expo-
sure itaelf.  The population aay be thought af as a aet af •  subjects, each  of
which experiences  specific  values ef the aodel input  variables end resulting
exposure.  If  the input  variable values have been recorded for a  saaple  of
population subjects  (see, e.|.,  the  hypothetical data  in Attachaent A.I  ef
appendix  A), then a. aodel  csn  be ased to ceapute the predicted  exposure for
•ach  saaple  aaaber.  Treaties  these  ceaputed exposures es if they were en-
sured exposures,  the aethods described  in Section 10.18  of  122)  end Section
12.9  ef (23] can be used to coapute confidence interval estiaates for percen-
tiles of the exposure distribution.
                                  -25-

-------
     Interval  estimates  for percentiles  of the  exposure  distribution  ire •
useful  quantitative  characterisation of uncertainty for the exposure assess-
•cat.   lovever,  interval estimates  ce«puted •• distuned in this aection are
haaed  tfm aeveral  assumptions,  primarily  the presumed  model  for personal
opoaarea.  luce, characterisation of uncertainty for  the  exposure assessment
requires  JutificatioD of  the model  and ether  assumptions.   The empirical
and/or  theoretical  haaia for the model should he diacuaaed.   Uae of a vali-
dated model always  decreases the  overall uncertainty.  Design of the aample
amrvey  that ia teed  to produce the data ahould aloe he discussed, aa veil as
accuracy aad frociaioa of  the  meaaurmmeat processes.   If a probability aample
{e.g.,  a aiso>le raadosi  aaaplc)  was not  sjaed, the haaia  for  any inferences
hoyoad  the  individuals la  the  aaaple aust he discussed.  Sensitivity to ewdel
foraulatioa caa he iaveatifated by aatiMtiag the  distribution of exposure for
plaaaible altaraetive awdela if aaaple aurrey data have heen collected for the
input variables of the alternative Models.
4.2  IITTIRVAL ESTUUTES OF EXPOSURE 1EKCEVTIUS:   HIFUT VARIABLE DISTRIBUTIONS
     WOW
     When the  Joint  probability  distribution  for the  eicposure  awdel input
variables can  he represented by  a apecific faadly  of probability distribu-
tions,  confidence interval  estinates for  perceatilea of the exposure distri-
bution  can  he  coaputcd free) a joint confidence region for the parameters of
the input veriable  distributions.   In  general,  let the vector  6 denote the
peraaeters  of  the  joint probability distribution for the model  input vari-
ables.  Let 6  denote  a 100(1-a) percent confidence region  for the aodel input
parameter •, i.e.,
     Prob I § c 6 }  • !•;.
A 100(l-«)  percent confidence  totarval  estimate of the p-th perceatile of the
ejxfoaure distribution ia than given by
     i *,(•> I**).
      T (£) daaotea the  p-th percaatile of the eopoaure diatribution imduced
by  the  penomal axposare aodal  when the values  of the Imput parameters are
f ir:f .ct .the «alue -«.   That ia, the. «et of .all .fth fereaatilea ,• Ty( of the
induced exposure  distributions generated by the  aet of tee forms a 100(1-0)
pereent confidence interval  estimate of the p-th percentile of the population
exposure distribution.  Hence, the  interval from the amalleat to the largest
                                  •26-

-------
 of these  induced p-th percentiles, i.e.,*
      [  In£  { Yp(6)  | 6e6  }, Sup  { Yp(6) | 6c6 } J,
 constitutes,** interval estimate of the p-th percentile of the exposure distri-
 bution  haviag  • confidence cetlficicnt of at least 100(1- •) percent.   IB prac-
 tice, It  it set feasible  to derive the exposure distributions corresponding to
 •11 0e6.   lower, when the personal exposure sjodel is  mathematically tract-
 able (e.g., continuous and  monotonically increasing, it may be possible to
 deduct  those values of 6c6 that will generate the IBS1lest and-largest induced
 p-th percentiles.   la  this case, only these two estimates of the p-th percen-
 tile Deed  to  be  computed ia order to geaerate the confidence interval esti-
 •ate.
     TWo  advantages of  this  procedure with respect to that  presented in  Sec-
 tion 4.1  are  that eaaller aample aiaes may be  aufficient and that the confi-
 dence interval ostiMtes  should be somewhat aarrover for  a  given sample  size
 because additional  intonation has been  osod (i.e., the  assumed joint prob-
 ability distribution  of  the  input variables).  A major  disadvantage  is  that
 this procedure is Bore difficult to implement.
     for  txaaple, suppose that an exposure model is based upon a independent
 input vsrisbles.  When  the input variables are statistically independent, the
 joint confidence level  for the iaput variable distributions is the product of
 the  individual input variable confidence  levels.  Thus,  95 percent confidence
 interval  estimates  for  perceatiles'of the exposure distribution can  be gene-
 rated froa  confidence interval estimates  of the parameters  of the a  indepen-
 dent variables,  with the  level of confidence being (0.95)1'  for the param-
 eters of each  input variable.  If in addition each iaput  variable distribution
 is  iadeied  by  1 parameters for which independent astiautors exist, the joint
 (0.55)1/k confidence region  for each iaput variable distribution is the cross
 product apace  defined by interval estimates for the 1  individual parameters
 with a  (0.9S)1'"* confidence level for each interval.4  The resulting  joint 95
•Xoehaicslly^from the infimua (Inf) to the aupremua (Sup) of the  induced p-th
percentiles corresponding to all values of 6 in 6.
tfioafideacc interval  estimates  of iaput variable parameters usually rely on  a
model,  e.g.,  approximate Normal  distribution for a aamplc mean.  Vben U ie
large,.this model may act be adequate for confidence interval estimates et the
(0.95)1'** confidence level.
                                   -27-

-------
percent  coafideace  region  for the k£ parameters of the input variable distri-
butions  generates  a set of eorreaponding probability  distributions  for  expo-
sure via the  model, baaed upoa the fc  iadcpendcnt  input variables.   Given any
particular set of input variable distributions, the corresponding distribution
of exposures caa be generated analytically or by using Moate Carlo aimulation.
For  any given pcretatilc  of  the exposure distribution, the 95 percent  lover
•ad  upper confidence  limits  arc the minimum tad maximum, rtipectively, of the
estimates of  that  percentile, baaed upon all exposure distributions  generated
by  the  Joint 95  percent confidence region  for the parameters  of the  input
variable distributions (see, e.g., Section 4.2 of Appendix A).
     These interval estimates for percentiles of the exposure distribution are
s useful quantitative characterisation of uncertainty for the exposure assess-
swat.   Bovever,  they  are  baaed upon  many assumptions.   Characterisation  of
uncertainty for  the exposurs  sssesssjcnt is not complete without s  thorough
discussion of sll assumptions.  Justification for the  model  should  be  dis-
cussed as veil as limitations of the data and design of the sample survey used
to collect  the  data.   If  s  probability sample was not  used, any  inferences
beyond the sample subjects themselves must be justified.  Any assumptions used
in computing  the  confidence  interval  estimates, such ss independence of  model
input variables  or estimators  of  parameters of the input variable  distribu-
tions, should be explicitly stated and justified.  Sensitivity to model formu-
lation  caa be  investigated  by estimating the  distribution of  exposure  for
                                   •
plausible alternative models if sample survey data  have been collected for the
input  variables  of  the alternative models.   Using validated input  variable
distributions greatly reduces  the uncertainty  resulting from essuming  par-
ticular  input variable distributions.   If exposure  measurements are  available,
the personal expoaurc model should be validated also.
                                   •28-

-------
             5.  ASSESSMENTS BASED UPON EXPOSURE MEASUREMENTS

     A major  reduction  in the uncertainty associated with an exposure assess-
ment  can be achieved by directly measuring  the exposure  for  a sufficiently
large  representative sanple  of  aeabers  of  the  affected population.   This
reduction in uncertainty is achieved by eliminating the use of a model to pre-
dict exposure.  Moreover, a siodel can be validated and model parameters can be
estimated if  data  are also collected for all potential model input variables.
     The concept of directly measuring personal exposure to a substance is not
new.   For example,  occupational  exposures to  radiation and carbon  monoxide
(CO)  have  been measured  for years using  personal exposure monitors  (PEMs).
However, PEMs which measure concentrations that are expected to be an order of
magnitude  lower than  occupational concentrations  have  not previously  been
readily  available.   In  the  past  decade, advances in  technology have enabled
manufacturers  to miniaturize  circuitry  and  components,  thus  making  direct
measurement  of personal  exposures much  more  feasible  [24].   As a  result,
several  personal exposure  monitoring  studies  have  been performed  recently
([25], [26],  [27]).   These  studies generally monitor an indicator of internal
dose,  such  as  continuously  monitored  concentrations in  the   contsct  zone,
rather than internal dose per se.
     Given a  database of measured exposure values  for  individual populatioo
subjects, such as the hypothetical sample in Attachment A.2 of Appendix A, the
population distribution of exposure may be estimated directly.   In particular,
the sample exposure  distribution  estimates the population distribution.   If a
probability  sample  design was  used  to select the sample  members,  a  weighted
estimate of  the exposure 4  ttribution should be computed  (see,  e.g.,  sample
survey inferences in [11], [21], [22], and [23]).  Use of the sampling weights
is especially  important  if  high-risk segments of the  population (e.g.,  indi-
viduals  with  potential  occupational  exposure,  or individuals  expected  to be
susceptible to health effects) have been oversampled.  Typically, the sampling
weights  account  for  unequal  probabilities of selection and weight adjustments
to account for variations in response rate.
     Tb* •••sured exposure values  may also be used to directly compute confi-
dence  interval estimates for  percentiles of the  exposure distribution (see,
e.g.,  Section  10.18  of  [22],  or  Section 12.9  of  [23]).  These confidence
intervsl  estimates  for  percentiles  of the  exposure distribution  define an
                                  -29-

-------
approximate  confidence region for the exposure  distribution.   When the expo-
sure  distribution  is  estimated from  measured exposures  for  a  probability
staple of population members, confidence interval estimates for percentiles of
the exposure distribution are the primary characterization of uncertainty for
the exposure assessment.  Design of the survey used to collect the data should
also be  discussed,  ss well as accuracy and precision of the measurement tech-
niques.  Ho  mention of an exposure prediction model  is  necessary because the
exposure levels  were  measured directly, rather  than  calculated  froo a model.
If, however,  the measured exposures are not based on a probability sample, it
should be  acknowledged that  no strictly valid  statistical  inferences can be
•ade beyond  the units  actually in the  sample.   If  inference procedures are
implemented  for  a  nonprobability sample or if the sample design is ignored in
the analysis,  the'  assumptions  upon which these inferences are  based (e.g.,
treatment of  the sample as if it was a simple random sample, or assumption of
an underlying model) should be explicitly stated and justified [28].
     When personal exposures are being monitored, the exposure distribution is
often very skewed, with most of the sample members having exposure levels that
are less than the detection limit of the monitoring equipment, especially when
monitoring exposure to  toxic  substances  ([27],  [29]).   In  this  situation,
estimation of  the  distribution  of  exposure may require a  very large sample
size.   Estimation  of the  proportion  of population members  with a  detectable
level in specified analysis  subgroups  and in  the total population may then
become an important analysis  activity.   Sometimes no sample members are found
to have  a  detectable  level.   When no sample members  have  detectable levels,
the power5  of the  hypothesis test that the proportion  of  population members
with a  detectable  level  exceeds 1 percent,  2  percent, etc., may  become the
primary  analysis (29].  It can then be  inferred that the  proportion of the
population with  a  detectable  level does mot exceed  seme fixed percentage for
which the power  of the test  is  high.   Qualitative characterization of uncer-
tainty for the  assessment would be basically unchanged,  but the power of the
hypothesis test for  specified  alternative  hypotheses  would be the primary
quantitative characterization of uncertainty.
•The  power of  the test  is  the probability  of correctly rejecting  the  null
hypothesis that  no population members have a detectable level when the alter-
native  hypothesis  (that  the  proportion of  the population with  a detectable
level exceeds 1 percent,  2 percent, etc.) is true.
                                  •30-

-------
      Whenever  personal  exposures  are  aeasured  directly,  quantification  of
measurement  errors  becoaes  an  important  aspect  of  the characterisation  of
aacertainty.   The tens  "accuracy,"  "precision," and "bias" are used to de-
acribe  specific  aspects af measurement error ((3], 130], II.6-II.8 ef  [31]).
She ten "bias"  refers to systematic errors, i.e., errors that  have a aoazero
expected value.   Vaay chemical measurement techniques are  kaown to produce
biaaed  meaauremeats.  Far example,  the aaouat af a ceataaiaant ia a specimen
•ay be  ceasisteatly greater  than the measured amouat.  The ten "precision"
refers  to the repeatability ef a measurement.  Bigh precision means there  is
little  variation between repeated measurements  en the seme ebject, i.e., the
measurements  exhibit a  low  coefficient  ef variation  (ratio  ef the standard
deviation af repeated measurements divided  by the mean).  Finally, accuracy
refers  to. the cumulative effect ef bias end precis ion.  Bigh accuracy occurs
ealy  when the  bias  ten ia eaall ead the precision ia high.  Either high bias
er low precision results in low accuracy.
     Vhen  a  survey  collecting personal exposure  data  ia properly designed,
measurement  error components can be quantified.  Moreover, estimation proce-
dures can be developed that compensate for the estimated  measurement biss.  A
thorough  quality assurance plan that eaables estiaation ef measurement error
components is presented  in reference  (32).  The design presented ia reference
(32]  enables estiaation  ef everall  precision,  bias, aad accuracy ef a param-
eter  estimate via the  relative standard deviation (USD), relative bias (RB),
aad  relative root mean  square error (BMSE), respectively.  These statistics
are defined es follows:
      1.   The BSD is calculated as  the ratio af standard  error  ef a paraaeter
          estimate divided by the parameter estimate itself.
     2.   The BB is calculated  aa  the ratio af the  estimated absolute bias
          divided by the peraaeter estimate.
     3.   USE • (BSD* + BB*A
     •toother  autistic that  is  aemetimes  ased fcy the BPA as a  measure  af
•verall  accaracy af a measurement  procedure ia the "total error" defined  aa
the  sua af  the  absolute bias plus  twice the coefficient af variation.  IFA
guidelines specify that  measurement accuracy is acceptable if the total error
is  less  than  SO percent af the mean  (33].  Vhen the measurement  error  ia
aaacceptable, improved •easurement  techniques are aeeded before say reliable
easlyses  can be performed  [33].  Vhen the measurement  error is acceptable, but

                                  -31-

-------
 the aeasureaent process has « nonzero bias,  analysis  technique* caa coapeasate
 for the  bias.   Reference  132] discusses a survey design sod estiaatioa proce-
 dure that enables estimation of the swan exposure,  corrected for the estimated
 MM  bias.   Beference (34]  suggests  th*t estimates  of mesa  levels bt multi-
 plied by Ue relative biss to produce unbiased estimates of population waas.
 finally,  reference  (35] tabulates  the true  significance level  of a two-sided
 Jqfpotheais Ust as a fuaction of the seminal significance level and bias for a
 survey design  that  dees Bet allow estimation of  the sjaaauroBcat bias.  Uacer-
 tatnty  is then described  qualitatively by eayiag that significant hypothesis
• test  results would  aot  he  affected smleae the hies was such  greater than
 expected.
      If  previous research  aad/or theoretical derivetieas  suggest  that  the
 dspoaurc  distribution should be a particular atatistical  probability distri-
 bution (e.g., Qatfora, Normal,  or  Logaoraal), the aoaiteriag data aay be used
 (to ostiBate  the parameters  of  this  distributioa.  A measure of caatrsl ten-
 dency (e.g., swan,  mediae,  or  Bode) together with  a Beasure  of variability
 Ct.g., standard deviation or a  parcentile)  is usually aufficieat to sstiaate
 »uch  a  distribution.  Vhen the  atatistical probability  distribution  has
 liBited theoretical range, auch  as the Unifon distribution, the saaple size
 and the  smallest and/or largest  data values are  usually needed ia order to
 estiaate  the extreme values  of the distribution when they  are not  known a
 priori.   leference  (IB] provides formulas for ostiBstioa of the paraaeters of
 Best cesBwaly need probability  distributions.  Vhen a theoretical probability
 distributioa has been fit  to the data, a atatistical test of goodness of fit
. should be performed to determine whether or mot  the theoretical distribution
 provides a reasonably good fit to the monitoring mate  (eee, e.g.., Chapter 8 of
 |15],  or lection 4.5 and Chapter 6 of |1«J).
     Vhen total personal exposure vis a particular route, e.g., inhalation, is
 •neitored, the  mate provide  • .direct estimate  of the attribution of total
 exposure  for members of the sampled population.  Since aggregation of ssti-
 anted exposure  distribution over aources and pathwaya ia mot required, the
 swaitoring data provide a more  defeasible vatimate of the diatributioa *l.
 «sposttr<  than  the model-based estimates  discussed  la Chapters 2,  3, and 4.
 Models that link observed  exposure to emission aonrcea may he needed, however,
 ia addition  to  Boaitoriag  data, in order to astiaate  the risk associated with
 roguletery options.   If  data were collected  for potential model input vari-
 ables as  well  as exposure  (see, e.g., the hypothetical exaaple in Attachment
                                   -32-

-------
A.2  of  Appendix A), then  the  paraaeters  of a code! relating exposure  to  the
potential input variables  can  be estiaated using statistical regression tech-
aiques.  Moreover, lack ef fit can be tested for • hypothetical  sjodel,  e.g., •
•odel bated open the physical transport and fate process.  Such  aodels  linking
sources aad exposures ere very isjportant for assessing the potential iapact ef
regulatory actions.   Using Models  validated  in this way  greatly  reduces  the
tuaatitative uncertainty associated with exposure •aeesssjents.   thfortuftstely,
the  analytical  SMasureaent difficulties «nd costs associated with collection
•f directly ewasued  exposure levels often severly lix.it or  even prevent col-
lection of eueh data.
                                   -33-

-------
     6.  AGGREGATING ACROSS SOURCES, PATHWAYS, AND/OR ROUTES OF EXPOSURE
          •

 4.1  CQHBXiniG ESTIMATED EXPOSURE DISTRIBUTIONS

 jeaerally wttMtei the  distribution  of teul expoaure via • particular route
 (inhalation,  iaaeation,  «r  dermal)  for  tat subjects of a  nubpopulatioa  ai
 diecuasad  ia Chapter 5.  Bovever, an exposure eaaeaameat baaed 1900 a e»del
 •anally estimates the diitributioo of exposure for subjects in a aubpopulatioa
 for  a particular combination mi .Aouxca, jXttOsjav, ^ad *x*ate. JnUcratioa  of
                                              tt tmaii ••Hifl -fa fjfrtet *m -«j»ti-
•ate tha tetei exposure distribution for an absorption route.
     If tha index  a,  where s«l, 2,  ..., S( indexes the unique combinations of
aources aad pathways for a particular route, r, and aubpopulatioa, t, then the
route r exposure for the i-th aubjoet in the aubpopulatioa can be expressed as
                   6
The  distribution  of T. »(i)  can be  derived  analytically or by  Monte Carlo
                       *»*
siawlatioa from the distributions ef  Yf §  t(i).  In either case, an important
consideration ia whether or not the  S  aource-psthway combinations are indepen-
dent  ef each  ether.   If not,  the  essessor ahould determine  by sensitivity
testing whether the correlated  cembinstioas must be simultaneously modeled so
ns to incorporate the  correlation.' For example, if two pathways are strongly
correlated because they originate from the  same source, the two pstbways would

value for ene  pathway  would then always be equal to the aource value for the
ether pathway for each iteration.
     The total rente r exposure for  the i-th subject in aubpopulatioa t can be
computed from  Equation (7) by  animation over the independent source-pathway
   »inatioes indexed by  s.  The distribution  of  route  r exposure may then be

         xpeemre distributions.  Thus,  the Monte Carle method could be Impla-
   nted  by aimultaaeously ^taeratiAf-values for the independent source-pathway
combinations end their num.  The result would be estimated exposure distribu-
tions for the independent aource-pathway combinations end theirsum, the total
      «4iai

-------
      In tone uses, the risk due to exposure to • subataace My be  related  to
 total exposure over  all routea  (e.g.,  vbea there ia  BO risk for expoaures
 below sea* threshold level).  Eatiaaticn of total exposure  ia  alao accessary
 to judge whether regulation of • particular route of exposure  ia an effieieat
 approach to eoatrolling total population exposure. The total exposure experi-
 enced by the 1-th subject ia aubpopulatioa t ia then  given by
              3             3   S
      Tt(i)  • 2  T    (i) -2   2   Y      (i) ,                     (8)
       *      r»l  *fl      r»l ••!   »•••*
•here r iadexes the three txpoaure routea (Inhalation, ingestion, and dermal).
      Xa  order to eatiaute  the  distribution of the total exposure, Tt(i),  it
would be important to  detcniae  whether or  not  the exposures via the three
absorption  routes  vere  independent randoa variablea.  If act,  the  correlated
routea would have to he treated aiawltaaeoualy ia  a Moate Carlo simulation,  as
discussed  above  for correlated source-pathway combinations.  When the suana-
tioa  ia  over iadependent routes,  the Monte  Carlo aianilation  values for the
independent  routea  auy he  added  to produce  a simulated value for the total
•xpoaure.  The aianilatcd Monte  Carlo diatributioa of total cspoaure can then
be generated by a large nunber of repetitioas of this tiaulatioa.
     When the aubpopulatioa diatribution of exposure  (either total exposure  or
route  r  expoaure) has been  estiauted aa described above, uncertainty concern-
log the  individual route and/or souree-pathvay expoaure  distributions should
he  characterised  aa described  ia 'Chapters 3-S.   Ia  edditiea, these methods
should  be applied,  to  the  exteat possible,  to the  aggregated subpopulation
exposure  diatribution.   For example, if  peraoaal •xpoaure has been measured
directly  for  varioua  routea sod/or sourcc-pathvaya for a sample of members of
• suhpopulatioB, then the aggregate exposure has  also been measured directly,
aad the sjethoda  described  la  Chapter S  csa be  nsed to compute confidence
interval  eatixutes  for fercentiles of  the aggregated exposure diatribution.
6.2  UOTING DISTRIBUTIONAL ttSULTS
     When the total exposure given by Equation (7) or (•) is the ausnation of
•evere! tsrss, central liaiit theory provides iaiportant distributional results.
In  either case,  the'total expoaure is written  ea  the  sm of the positive
rsndos) variables T  . »(i).  fence,  if s central  liaiit type of rcault appliea
                  —i—i—
to the variablea Tr   t, then the liaiting expoaure distribution ia the Vernal
distribution  aa  the* nuaber of  tena  in the auamation in Equation (7)  or (g)
                                  -33-

-------
becomes large.   ••••,  when the  total  exposure is the SUB ever several  sourcei,
psthwsys, and/or routes,  and no one ten has a variance that ii large  coapired
to the  variance of the  total, then the  for.  of the distribution for total
exposure iar lubjtets IB subpopulation t will be  approximately Nonal (see,
e.g., toction 14-1.1 of (19]).
     The MM and variance  for the total  ejrpoiurc, Tr t,  can  easily  be esti-
eated fro*  the  Man,  variance, and  covariance for individual source-pathway
combinations as  follows:

     •l»rit(i)J  • 2   I[Y    (i)|, and                               (9)
        *•*       *      *•••*
          *»*
                  a*l
                     8
             (i)] •  X
                      8-1   8
                    2  2    2   Covtt     (i), Y   ,   (i)J,           (10)
where  E,  Var, Cov  repreaent  the population mean,  variance,  and covariance,
respectively.  Of  course, when  the  aource-pathvay  combinations  iadexed  by s
are  assumed to  be independent,  the second ten  ia Equation (10)  is  aero.
Substituting unbiased or  consistent  cstinatei directly into Equations (9) and
(10) provides corresponding unbiased or  consistent  estimates  ef  the  Man and
variance ef the total exposure, Y  t, for route  r for members af subpopulation
t.  These estimates may be used to .fit the limiting  Normal distribution direct-
ly.  In fact, any  measure of central tendency (e.g., mean, median, or mode),
together with any measure of variability  (e.g.,  etaadard deviation or a percen-
tile),  can be  used to estimate the limiting Venal distribution for  total
exposure.
     She  limiting  Bonal  elistribmtiea far total exposure ia  attained as the
amnmer ef tens in the  Summation ia Equation  (?)  ar (8)  becomes infinitely
large.  In practice,  the  number ef tens mill always me finite.  Xhe primary
importance  ef the limiting Venal  distribution is  that  the distribution for
total  exyotare  will be approximately Venal eftea there  ere eeveral  terms ia
the  emanation, none ef which dominates the variability ef the total exposure.
.Ia this case the Mans, variances, and cevariaacea  ef the tens in the aamma-
tiea an  Mst important because they are aufficieat to determine the limiting
distribution.  Other  shape characteristics of the exposure distributions for
the  individual tens are then less important.
                                  -36-

-------
     Although  tbe Uniting distribution could be estimated directly when  the
 total  exposure is tbe BUB of several terms, research into the rate  of  conver-
 gence  to the liaiting distribution would be Deeded for proper interpretation.
 Empirical  vesearch would  be Deeded  to detenine when the Dumber ef tens in
 the  summation  was sufficiently large to use the limiting Venal distribution.
 The  results would likely depend upon tbe sources and pathways  being  summed  and
 their  relative importance.
     Xf Equations (9) and  (10) arc used to estimate the liaiting Venal distri-
 bution for total exposure, adequacy ef the approxiaation ef the true expoaure
 distribution by the limiting distribution ahould be discussed.  An empirical
 Dtudy  en convergence to tbe limiting distribution slight be Deeded.  If  total
 exposure was measured for a sample ef population members, these data should be
 used to teat  the adequacy ef tbe limiting  distribution.   Characterisation ef
 uncertainty with regard to the distribution ef total exposure would also Deed
 to address uncertainty vith regard to the estiaated mean and variance given by
Equations  (9)  and  (10).   If possible, confidence  interval estimates  of  tbe
•can and variance should  be derived.  Sensitivity ef  the predicted exposure
distribution to plsusible rsnges of values for these parameters should  slso be
investigated.  In addition, limitations ef the data used to estimate aeans  and
variances  or  to  validate  tbe  distribution of  total exposure  should  be dis-
cussed.
                                  -37-

-------
                7.  COMBINING OVER DISJOINT SUBPOPULATIONS

     Afltr the  teul  aubpopulatioo t distribution of axposure has been esti-
•ated  for trery  aubpopulatien, the  «ixei of  the  aubpepulstions  should be
astimated.   The  muBber  of  oubpopuUtion  Mmbers experiencing  • particular
level of exposure eao be estimated fro* ao estimate of the site of the subpop-
ttlation tad  aa  estimate  of the properties of subpopulation members upcrieac-
ia| that  exposure level.  Moreover,  the proportion! of the total population
•sobers IB the  individual  aobpopulations can be used to combine  the estimated
•UstiibotiOM for the oubpopulationj  to produce an estimate of the distribu-
Uoa of exposure for the total population.
     Bnccrtainty  concerning  the oices of  the oubpopulations sbould  be ad-
dressed by diacuasini limitations of tbe data and methods  need to estimate the
aobpopulation sixes.   Confidence interval estimates of the subpopulation sizes
•bonld olso be produced whenever possible.
     If the  total affected  population of  V subjects  is partitioned  into T
disjoint anbpopulatiena  that each  contain >t aubjects, the estimated cumula-
tive probability  distribution function for  total route r  axposure is given by
     Fr(y) « Frob (Yr < y)
              T    .
           •  I   (IT  / ») Prob (Yr    < y)
             •el   *              r»l
           •  *  Ptrr t <*>•                                         (11)
             •si  •  • • •
      •      *
where N  and N are the estimated sizes of the t-th subpopulation and the total
population, respectively. Pt is  the astimated proportion of the population in
aubpopulation t, and Ff t is the eatimatod cumulative probability distribution
function for route r axposure in aubpopulation t.   Xf estimates of the route r
tmtpopnlatlen distribution  functions,  T>|t,  are replaced by estimates of the
tistribution  functions,  ?t,  for the  total  over  all  routes, Equation (11)
produces  the astimated  distribution  function far  total exposure over all
mates far the-total population.
     Vhen  the distribution  of axposure mas been aggregated across subpopula-
tion* as  described  above, uncertainty concerning the individual aubpopulation
axposure  distributions is  characterized as  described  in  Chapters  3-6.  In
                                  •38-

-------
 addition,  uncertainty concerning  the  total population exposure  distribution
 (cither  for route r exposure or total  txposure)  can  be  addressed in terns of
 percentilea of this distribution.  The p-th percentile of the route r exposure
 distribution is that txposure level, y  . such that
      p»Prtb 
-------
can be iitiaated is follows
      -  •'      (Pe .) 0-PC .)
     V«(pC(t) « -^	Soi-  .                                    (17)

IB ftDcral, the variance expreaiien |iven by Equation (17) veuld be aultiplied
ty  • 4tsi|o  tffeet  (or variaaee inflation factor)  to  reflect unequal  aaaple
•fcleetioa  probabilities  and/or  differential  adjuataents  to  coatpeaaate  for
•rcrrey  aoartapoaae (ace taoplc  aurvey aaalyaia  in  (11), (21],  (22),  (29)).
                                   •40-

-------
                                8.  SUMMARY

     A floury of  the  priaary aethods  recoaaended  for  characterising uncer-
 tainty for  exposure  •••••uenti vas  presented in  Table  1.   Virtually all
 exposure  asaessaents  tictpt thoie  baaad open Matured exposure levels for a
 probability aaaple of population aeabers  rely opoa  a aodel to predict expo-
 sure.   The aodel aay be any aatheaatlcal  function, staple or coaplex, that
 astiaates  the population distribution  of exposure  or an  individual population
 •saber's  exposure as a  function  of one or aore input variables.  Whenever a
 eiodel  that has aot been validated ia used aa the basis for an exposure aasess-
 aent,  the  uncertainty associated  vitb  the exposure assesaaent stay be eubstan-
 tial.  The priaary characterisation of uncertainty  ia at  least partly qualita-
 tive in this case, i.e., it includes a description  of the assumptions inherent
 in the aodel and their justification.   Plausible alternative aodels ahould be
 discussed.  Sensitivity of the exposure asseaaaeat  to aodel femulation can be
 iaveatigated  by  replicating the assessaent for plausible  alternative aodels.
     When  an exposure  aasessaeat  ia  based upon  directly aeasured exposure
 levels  for a  probability aaaple  of population aeabers,  uncertainty  can be
 greatly  reduced   and  described  quantitatively.  In this  case,  the priaary
 aources of uncertainty are aeasureaeat errors and aaapling errors.  A thorough
 quality assurance prograa ahould be designed into the atudy to ensure that the
Mgnitude  of aeasureaent  errors  can be estiaated.  The effects  of all sources
 of  randoa  error ahould  be aeasured  quantitatively  by confidence interval
 eatiaates of paraaetera of interest, e.g., perceatiles of the exposure distri-
 bution.  Moreover,  t>« effect  of raadoa errors can be  reduced by taking a
 larger aaaple.
     Whenever the latter  ia  not feasible, it ia aeaetiaes possible to obtain
 at least eoae data  for exposure and aodel input variables. These data ahould
 be uaed te assess goodness of fit ef the aodel aad/or areauaed distributions
 of  input  variables.  This substantially reduces the eaount  ef quantitative
 uncertainty for  estiaation of  the  distribution ef  exposure end ia strongly
 recoaaended.  It ia recognised, however, that it aay  not  always  be feasible to
 collect such data. •
                                  •41-

-------
                                  REFERENCES
 ID Callahan, H. A.,  Johnson, R.  H.,  KcGinnity, J.  L.,  HcUughlin,  T. C.,
     •warn,  D.  J.,  and  6lisak,  P.  V.    (library,  1983).   Draft Handbook
     for Perforaiat  Exposure Aaaeaaaenta.  Prcp«rtd by  the U. 8. Environmental
     Prottctloa Agency, Office of Health and Environmental Assesscent.

 |2) Executive Office  of  the  President,  Office of Science  and Technology
     Policy (Hay, 1984 Draft).  Chcaical Carcinogens;  Review  of the Science
     aad Ita  Application*.   Federal ietitter. Vol. 49, Me.  100, 21594-21637.

 (3) Rational  Heaearch Council,  CotBittee en  Prototype Explicit Analyse> for
     Pesticides (1980). Retulatina  Pesticides.  Vatienal Acadesiy of Sciences,
     VaahiBftoa,  B.C.

 (4) Calabreae, Cdvard  J.   (1982).   The Role  of Exposure  Oats  in Standard
     Settint.   Toxic Substances Journal. Vol. A, 12-22.

 (5) 0.  8.  Environmental Protection  Agency (1978).  Source Aaaesaaent:  Analy-
     ais of Uncertainty-Principles and Applications.        EPA»600/2-76-004u.
     FMO-131485.

 [6] Cos, D. C.t  and laybutt, P.  (1981).  Methods for Uncertainty Analysis:  A
     Comparative  Survey.  Risk Analysis. Vol. 1, No. A, 251-258.

 (7] Hiller,   C.  (1978).    Exposure Aaaeaaaent Hodelint:  A State-of-the Art
     Review.   EPA-600/3-78-065.
 (8] Drake, R. L. and Bar rafer, 8. H.   (1979).  iiathesMtical Models for Ataos-
    pberic Pollutants.  Electric Power Research Institute.

 |9] Hoachandreas,  D.  J. and Stark, J. V. C.  (1978).  The 6EOMET Indoor-Out-
    door Air Pollution  Hodel.  EPA-600/7-78-106, P
(10)  EPA Enpoaure Aaaeaastent Group (Hay 1984).  itochaatic Processes Applied
     to Risk Analysis  of TCDD Contaainated Soil;  A Caae Study.

Ill)  Ealton, 6.  (1983).  Introduction to Survey Saaolini.   lage Publications,
     •overly Hills.
112) Baamiaeu"i  I*t Jones, A., Steifcrvald, B.( aad ThAapson, V. (Scptaaber
     1984 Draft).  The Magnitude aad Hature of the Air Tonics Problen in the
              ites.   OSEPA  Office  ol  Air and  Radiation,  Office of  Policy,
                  Evaluation.

113) Johnaoo,  R. Jl.« Jr., iemhaxdt, ». .E.,-Haleon, *.-§.,..and Galley, H. «.,
     Jr.  (1973).   Aaaeaastent  cf Potential Radielotical Health Effects fro»
     Radon la  •aturaVGaa.   EPA-520/1-73-OQ*.

114] U.  8. Environawntal  Protection Agency (1983).  Huaan Exposure to Ataes-
     Pberic Concentrations of  Selected  Cheaieals.    Office  of   Air  Quality
     Plannini  and  Standards.


                                 -42-

-------
                            REFERENCES (coat.)


 US] Daaiel,  Wayne  W.  (1978).   Applied RoaaarMetric Statistics.   Routhton
     Hifflia, Boston.                                            '

 (16] Cooover,  W. J.  (1971).   Practical Roaparaaetric Statistics.  John Wiley
     and SOBS, Rev York.

 (17] liaatoae, Rarold A., tad luroff, Hurray, eda.  (1975).  The Delphi Method:
     Techniques  and Applications.   Addiaon-Wesley,  leading,  Massachusetts.

 [18] Johnson, R.  I.,  and Rotx, S. (1970).   Distributions in Statistical  Con-
     tiauoua Caivariate Distributions. •oughton Miff HA, Boston.

 (19] Burrill,   Claude   W.    (1972).   Measure, latetration. and Probability.
     McGrav-Jlill, Rev York.	

 (20] Woodruff, Ralph S.  (1971).   Simple Method for Approximating the Variance
     of •  Complicated Estimate.   Journal  of the American Statistical Associ-
     ation. Vol. 66, 411-414.       	

(21) Kendall,  Maurice  6.,  and   Stuart,  A.  (1968).  The Advanced Theory  of
     Statistics. Voluae 3.  Rafner, Rev York.

(22] Banaen, H.  R.,  Burvitx,  W.  R.,  and Madow, W. C. (1953).  Sample Survey
     Methods and Theory.  Voluae 1.  Wiley,  Rev York.

(23] Kish,  I. (1965).   Survey  Sampling.  Wiley, Rev York.

(24] Wallace, L. A., and Ott, W.  R. (1682).  Peraoaal Monitors:  A State-of-
     the-Art  Survey,   Journal of the  Air Pollution Control Association.  Vol.
     32, 602-610.                 I

125] Speagler, J. D.,  and Soczek,  M. 1. (1984).   Evidence for Improved Ambient
     Air Quality  and  the Reed for Personal Exposure Research, Environmental
     Science and Technology. Vol.  18,  Ro. 9,  268A-280A.

(26] lartvell, T. D., Claytoa, C. A., Michie, R. M., Wbitmore, R. W., Xeloa,
     I. S.( Joaea, S.  M., aad  Wbitchurst, D.  A.  (1984).  Study of Carbon Monox-
     ide Expoaure of Resideats of Washiattoa.  DC. aad Denver. Colorado.
     IPA-600/S4-84-031, PBB4-183516.

(27] Pellissari, E. D.,  Bartvell, T. D.,  Sparacino, C. M.,  Sheldon,  L.  S.,
     tbltsjort, R. W.,  lAialnier, C., aond  Zeloa, I.  S.  (1984).  Total EXPO-
     ture Aasessaeat Hethodoletv  (TEAM) Study;   for** Beasea-Kortbern Rev
     	       1392/03-35.   Research Triaajle Institute, Research Triangle
     PorkTRC.
                    *
(28] Banaen, H. H., Madow,  W. C., tad Teppint, R.  J.  (1983).  An Evaluation
     •f Model-Dependent and  Probability Sampling Inferences IB Sample Surveys.
     Journal of the American Statistical Association.  Vol.  78, Ro.  384,  776-
     F9T
                                  -43-

-------
                           REFERENCES  (coot.)


(29] MasoB,  X. 1.,  HcFadden,  ».  D., Ituaccaiooe, V. C., and HcGrath, C.  S.
     (1981).     Survey cf DBCP Distribution in Ground Water Supplies and  Sur-
     ftct Hatar Fondi.  KTI/W4/05-08J1. Xcaaarch Triaailc laititute. Kcaearcb
     Triable Park,  J»C.

(90] littahtrt, C.  (1968).   Expression of the ItocerUiotits of Final Btsults.
     '   ice, Vol. 160. 1201-1204.
131] ftdcrcr,   V.  T.  (1973).   lUtiiticB and locittv:  Data Cellection and
     lattrprttatioB.  Hareel Dckkcr, Ktw York.

|S2] local, X. H.( Pitraea,  S. A..  Byera,  ». 1.,  ud  Bandy, K. V.  (1981).
     Batioaal  iuaun Adipose Tiaaue Surrey Quality Aaauraocc Project Plan.
     RI1/1I64/21-11.  Ktacarch Iriao«lc laatitutt, Xaaaarcb Trianflc Park, DC.
(33] Picrioa,  S.f  ud  lucaa,  X.  (1983).   Bitpanic HAKES Pilot Study;   Hei-
     i
                                               pai_	
            t£f Volatile aad 8««iT8latil« Ortanic Ccapounds la Blood and
       inc gpeciacaa.  £PA 5»0/5-83-001.
(34J  Bestafcld,  J.  M.. lueas, X. H., tt al. (1983).  Uad and Cadniua Levels
     IB Hoed  of Baltimore. Maryland, tity Public School Teachers:  World
     health OrtaBiiatioB/llBited MatioBS EBviroaaeatal Protraat.        Midwest
     Xaaaarcb  li
      sacarcb lavtitute, Kaaaas City, Hiaaouri.

(35] Lncas, X. M.t  Xaaaacchiene, V. 6., aad Helroy, D. X. (1982).  Pelychlori-
      ited Bipbecvls  in HUMP Adipose Tiaaue and Mother*a Milk.   XTI/1864/50-02F.
      eaearcb Triangle Institute, Research Triangle Park, NC.
     •at
     Kei

-------
      APPENDIX A
A HYPOTHETICAL EXAMPLE
        A-l

-------
                    APPENDIX A:  A HYPOTHETICAL EXAMPLE

A.I  INTRODUCTION
     The purpose of  this  appendix is to provide • tipple hypothetical example
of •  aofuencc  of exposure assessments with a charrctcriaetion of uncertainty
for each.   The sequence  begins with ao aaaeaament based  1900 limited initial
data and ends  vith aa assessment based oa auffieiaat monitoring data to vali-
date and/or empirically estimate  a personal exposure model.  This example is
founded  to clarify concepts  presented in the  text of  this report.   It is
oecofnittd  that many  expoaure aaaeaaMnts aay mot Nlatc  directly  to this
OBaapIe.  mvvertheless, it is expected that this example  vill provide a useful
Quide to characterization of uncertainty for exposure assessments snd clarifi-
cation of concepts discussed in this report.
     The example is based open a recant case study by the EPA Exposure Assess-
ment Group.1  The  cese study was concerned vith an  assessment of human expo-
aure to TCDD (specifically 2,3,7,8-TCDD) resulting free) TCDD contaminated aoil
located in  an insecure disposal  aite  auch as  a  aanitary  landfill.   It was
assumed that the aitc  vas not located en residentisl or  agricultural property
and that access was sufficiently restricted ao that aignificsnt human exposure
occurred only  offsite.   Two routes  of exposure,  dermal  and infection, were
considered  for  exposure  resulting from deposition of  dust or eroded aoil en
rosidential  areas.   For  the purposes  of this  illustration,  only  the soil
                                   •
ingestion route vill be considered.   The EPA case atudy  focuied en the popu-
lation of individuals exposed  to  one part per billion  of TCDD in the residen-
tisl aoil.  A  more refined assessment vould have accounted for variability in
the residential aoil contamination levels.
A.2  ASSESSBEKT BASED UPON LIMITED ZVITIAL DATA
     The initial exposure assessment for a substance may  be based upon limited
data for  directly measured  personal exposures  and/or variables that  can be
wed to  predict personal  exposure.   Assessments based upon limited data for
directly measured exposure levels vill be considered first.   Assessments based
vpeo limited data for the input  varieties ef a personal exposure model vill
•hen be discussed in lection A.2.2.
SEPA  Exposure  Assessment ' Group  (May  1984).   Stochastic Processes Applied
to Riik Analviit af TCDD Conteminsted Soil:  A Case Study.
                                  A-2

-------
 A.2.1  Halted Data for Directly Measured Exposures
     Aa  aaaessaent My  begin vitb  directly Matured exposure levels  for  a
 Mall  •••pit of population •cabers.   The axpoaure data My  be  extant data,
 a.|.,  data froa Mdieal  records  or routine Mnitoring data.   A aaaple turvey
 aight  be deaifncd to coapile extant exposure data for aaaple aubjects.  Selec-
 Uen of  the iaitial aaaple aight be purposive (e.g.,  o iiaple of individual*
 with opecific health tffccta aad  potential axpoaure  to the  iiibaunce aader
 investigation).  If a potential  health problem was discovered, • sore refined
 aaeesaMnt  baaed apon a  probability aaaple of populatioa ambers would alaost
 always be Moded to enable risk SUM gown t decisions.
     Uhea  aa assessaeat  ia  baaed upon a  limited database of awaaured peraoaal
 exposures,  a priaary aspect of  the cbaracterisatioa of uncertainty is whether
 or aot the data roault  frost a probability aaaple.  Characterisation of uncer-
 tainty for  aa aaaesaaeat baaed upon a probability aaaple of directly aeaaured
 exposures  ia discussed  in Section  5. However,  atatiatica  calculated from  a
 •Mil  probability  aaaple  |cnerally  have  relatively  large ataadard  errors,
 i.e.,  a relatively high level  of  uncertainty.
     If the Maaured peraonal  cxpoaure levels  for a aaall saaple of population
MBbers were  not ao  lew  that  health risks were known to be negligible, a more
 refiaed exposure assessaent aight  be  pursued.  Two waya of refining the expo-
 sure assesaaent would be  as follows:
     1.   Heaaure exposure levels for a large  eaougb saaple to Bake inferences
          with a predetermined level  of  precision using the aethods discussed
          ia Section 5.
     2.   Develop  an er?osure  sceaario  and associated exposure  prediction
          aodel  to  be used as  the basis  for aa assessaent as  discussed  in
          Sections 2.2 through 4.2.
 Since  the  exposure variable needed for predicting  riak ia  oftea aot aaeaable
 to direct observstion, the latter aodel-based approach ia oftea puraued firat,
 if aot axcluaively.  trace, the  awdel-baaod approach  to refining the assess-
awnt ia preaeaud firat.
 A.2.2  LiiiUd Data for Model  Input Variablea
     If liaitad initial data on exposure  indicate that a auk stance My preaeat
 a significant health concern,  the exposure assessment My be refined by devel-
 oping  a peraoaal axposurc aodel beaed upon aa exposure scenario that coaaidera
 all  potential aources  of the  aubstaace  that  could  reault  in  contact with
 population  Mabera.  The aodel could be  very aiaple,  possibly involving only

                                  A-3

-------
 multiplicative  factors,  or be  very cooplex.   ID the present context, a per-

 eonal  exposure model  is  toy  equation  that  expresses  a population Member's

 expoeure as • function of one or aero input variables.

      for example, suppose  that the follevinf model was  adopted  to predict the

 dose of TCDD absorbed ia a lifetime * y the i-th populatioa member for the EPA
 case study citad above:8

      Ki) •  (ftC(i)] lED(i)) IlJl(i)] (AF(i)J (c(i)],                  (l)

 where T • Dose in Baits of aaaograms of TCDD absorbed in a lifetime,

      1C • TCDD concentration in infested aoil
         • 1 pat par billion (by definition in the IPA ease study),

      ED •  Exposure  duration ia oaits  of days of  contaminated  soil ingestion
           par lifetime,
                               \
      ZX •  foil imgestioa  rata  for that portion of the lifetime  during which
           aoil is ingested in anits of grans of aoil ingested per  day,

      AT •  Absorption fraction  ia units of grans of TCDD absorbed  per gran
           of TCDD ingested, and

       e » A positive randon arror ton with an oxpected  value of one.'

 For on  initial  exposure assessment,  only limited  data would typically be

 available concerning the populatioa distribution of model input  variables. For

 •sample, the data may aopport estimation of mo more than approximate ranges of

 values for the input variables.  For the model (1), suppose that the following

 estimates of  the ranges of  the input  variables are based open  very liaited
 dats and/or aubjective considerations:4
 •This model is adapted  from that meed in the IPA  case atudy.  The case atudy
 oodel included a potency  factor ao that risk was  estimated  rather than expo-
 Bare.  The  model alao  otaadardiaed  the exposure  by  dividing by the exposed
 fforaoa'a body weight and  lifetime.   The body weight and lifetime.factors are
 not presently  considered  for the  aake of  simplicity.  Alao, lifetime, body
••weight, and exposure duration stay well -be correlated, especially  Cor short
 lifetimes, making the Monte Carlo aimulatiea much more complex.

 *Ime raadoa error ton  represents  the cumulative multiplicative effect of all
.tsnobserved or latent variables not explicitly included in the model.
                   *                                           •
 4Xn practice, it My be relatively oaay to collect  data with  which to estimate
 the population distribution of some model input variables,  e.g., SC and AT for
 oodel  (1).   The diatribution of other input variables, e.g., ID and IK, or
 (their product—the total grama of aoil ingested per lifetime, may  be much more
 difficult to estimate.  One may not be sble to improve upon  expert subjective
 estimates of the ranges of such variables.

                                   A-4

-------
      CD:   0 to 900  days  of  cootaaiaatcd toil iaieition per lifetime.
      IR:   0 to 11 fract  of  soil ingested per day of ingest ion.
      AT:   0.10 to 1.00 |rao> of TCDD absorbed per gram ingested.
 Such data  would typically verve  ai the  basis  for performing a  sensitivity
 analysis  for the Model.   Tbe purpose of the  sensitivity  analysis  would be  to
 identify  influential model  input variables and develop bounds en the  distribu-
 tion of  lifetime  absorbed dose.   Tbe  aaasitivity analysis  Bight  begin  by
 ••tinting the  maximua  possible  dose.   In  the  worst case  (ED»900,  UNll.
 AF«1.0),  the ejcpected lifetime dose would be 9.900 aanograms of TCDD.  If this
 doae was  not  considered to be a potentially aignificant health concern, the
 exposure  assessment might be  teninated.   If one were sot  confident that the
 prtdieted  dose truly represented the worst case,  further investigation Bight
 be dcsirsble.
      If the worst  case  did present a aignificant health concern, the sensi-
 tivity  analysis Bight  be pursued  further.   For example,  one Bight  use the
Bodel to estimate the range of dose that rtsults as oach input variable varies
over  ita  range  of  possible values, with  the other variables  fixed at their
Bidrange, i.e.,
     ED:  Dose  varies  from 0  to  4455 nanograms  as CD vsries from  0  to 900,
          with IR • 5.5 and AF • 0.90.
      IB:  Dose varies froo 0 to 4455 nanograas as  IR varies  from 0  to 11, with
          CD * 450 and AF • 0.90.
     AF:  Dose  varies  from 1980 to  2475 nanograas as AF  varies from  0.80  to
          1.00, with CD « 450 and IR • 5.5.
Since the predicted dose  is  Best sensitive  to changes in  exposure duration
 (ED)  and  aeil  ingestion  rate  (IR), collection of additional data  for these
variables,  or  their product—the  lifetime  aeil iagestion—would be warranted.
•incc the plausible variation  in  the absorption  fraction has lass  tffect  on
the  predicted  dose, leas effort Bight he directed toward collecting  data  on
this  factor;  it Bight  even he reasonable to  treat this factor  aa a  known
consist fcr aubsequent modeling.
     The acnsltivity analysis can he enhanced by computing the predicted doaes
that  result from  all poasible  input variable combinations.   If each  input
variable  had only  a  finite set of  possible valves, the  aet  of all possible
combinations of the input  variables could be formed, and the predicted dose
could be  computed for each combination.  These predicted doses could be used
                                  A-5

-------
  to fora • probability distribution of dose by computing the relative  frequency
  of occurrence of each dose or interval of doses.  The resulting dose  sensitiv-
  ity distribution it  equivalent  to the population distribution of dose only if
  all combinations of  the  input variable values arc equally  likely to occur in
  the target population.   The  doae sensitivity distribution can also be derived
  for personal exposure models based upon continuous input variables with finite
  ranges  by  representing  these  variables by  equally-spaced points.  At the
  liait, na  the equal spaces  hecosie  infinitely avail  and the number of points
  becomes  infinitely large, the  procedure for  computing the doae  sensitivity
  distribution is equivalent to estimating the probability  distribution of dose
  that results  fro*  statistically  independent  continuous input variables vith
  Oviform distributions on the estimated  ranges.  For  the present example, the
  dose sensitivity distribution  was generated  by the Monte  Carlo method vitb
  9,000 iterations using independent Uniform distributions for ED,  IR,  and AT on
  the ranges 0 to  900,  0 to 11, and 0.80  to 1.00,  respectively.1  The  resulting
  distribution of  dose  is  shown  in figure A.I.  Estimates  of selected percen-
  tiles of this distribution arc  shown in Tnblc A.I.   Statistics bssed upon the
  •lose sensitivity distribution are  interpreted in tens* of all possible input
  variable  combinations, rstber  than the  population distributon of dose.  For
  example,  the 95tb  percentile of  the dose sensitivity  distribution  (approxi-
  mately 6400 nanograms  of  TCDD) ia an estimate  of the  lifetime  dose  exceeded by
  only 5 percent of  tbe  doses  resulting from all combinations of input variable
  values.
       The  uncertainty  for this  initisl  tsposure assessment would be charac-
  terised by describing the limitations of the data used to estimate plausible
  ranges  of model input variables  and by describing the basis  for the exposure
  model.  A discussion  of the limitations  of tme model  and justification for the
  •jodel would  me  o  major  component of  the character ins t ion  of uncertainty.
  flnusible alternative  models should  at least he  discussed.   Sensitivity to
        ^^                               •  '  »•«»•*  »««•<  '.   * »  +  l—.  .
  model formulation should be  sddressed by repeating the  Honte Carle simulation
  for plausible eltemative models.  Tbe  choice of a specific number iterations
"•Jbnte Carlo  simulations were  first ferformed  vith 100, 1,000,  3,000, mnd
  10,000 iterations.  All  simulations presented in  this  appendix ere based on
  3,000 iterations because there was  little difference in estimates  based on
  3,000 and 10,000 iterations.  The  additional cost of  10,000 iterations did not
  appear to be justified.
                                    A-6

-------
p
E
R  20-
C
E
N
T

   10-
      0
      4000

PREDICTED DOSE  CNG/LIFETIME3
   Flgvre A.I.  Diftritratien of Ettfmted Ooies ResiiltinR fro* Equally Likely tnpat VaHable Combinations

-------
Table A.I.  SELECTED PERCENTILES FOR ESTIMATED DOSE  RESULTING FROM
                  EQUALLY LIKELY INPUT VARIABLE COMBINATIONS
Pcrctatilc
•
O.01
0.02
0.03
O.O4
0.05
0.06
0.07
0.06
0.09
0.10
0. 15
0.20
0.25
0.30
0.35
0.40
0.45
0.50
O.55
0.60
0.65
0.70
'0.75
0.60
0.85
0.90
0.91
.92
.93
.94
.95
.96
.97
.96
.99
Point
EstiMte
11.57
26.15
39.64
59.55
76.44
103.72
126.05
151. SI
172.36
194.40
314.61
467. 12
616.11
790.69
995.34
1202.6S
1406.98
1624.66
1909.63
2214.34
2571.43
2951.19
3S55.48
9924.13
4552.30
5316.73
5473. 16
5651.56
5865.06
4116.02
4379.43
4654.90
-->4939.50
7297.60
7828.86
                                                Itproauctd from
                                                Mii»v«iUbU copy.
                              A-8

-------
 for  the Monte  Carle simulation should be justified, noting that more  itera-

 tieas  would improve  the  precision of tstiMtcs at  somewhat greater  expense.

 A.3  ASSESSMENT BASED UPON ESTIMATION OF IKPUT VARIABLE DISTRIBUTIONS

     Discussions  with subject-sutler experts  (e.g., medical doctors)  can be

 used to fonulate  an estimate of  the  joint distribution of the model input

 variables,   for  example,  auppose that discussions  suggested  that the  input
 variables  for  model (1) were all atatistically independent and  that  the fol-

 lowing input variable distributions wore supported:
     ID:  Discussions auggested that the distribution of exposure durstion was
          a  symmetric continuous distribution with  the  following properties:

          1.   Minimum duration • 0 days.
          2.   Maximum duration • 600 days.
          3.   Mcsn duration • 300 days.
          4.    10 percent of exposure durations are ISO days or less.
          5.    10 percent of exposure durations are 450 days or more.

          This information auggeats that the distribution of aspesure  durstion
          is as shown in  Figure A.2.  This distribution  can be  parameterized
          by fitting a Beta probability distribution for ED/600.4  In  particu-
          lar,  the  Beta distribution with o « B • 3.1 is a good approximation
          to the distribution of ED/600 shown in Figure A.2.

     XR:  Discussions auggested that the ingestion rate varied  readonly from
          0 to 5.0 grans per day over the population menbers, i.e.,  a
          Unifen (0, 5.0) probability distribution was aupported for  IR.

     AT:  Discussions suggested that tbe absorption frsction Bust be very
          close to one, if not identically one.
Probability
Density
     Figure A.2.  Hypothetical Subjective Estimate of the Distribution of
                               Ixposure Duration (ED)
•See Johnson, *. 1., and 8. Koti (1970).  Distributions in Statistics;
mou» Univariate Diatributions. Chapter 24 fo
distribution.
	   Contio-
or the parameterization of the Beta
                                  A-9

-------
 The  estimated probability  distributions for  the  model input  variables  were
 used  to  generate  an  estimated distribution  of lifeline dose,  assuming  tbe
 •bore  Joint distribution for the input  variables  for  model  (1) using a Monte
 Carlo  aiaulation baaed upon 3,000 iterations.   The  estimated distribution of
 4ose ia ••own in Figure A.3.   Selected  perceatiles  free this distribution as
 •hewn  in Table A.2.
      She ptlBiary characterisation  of oacertaiaty for this exposure assessaent
 would  be a  discussion of the Justification fer the presumed model  and input
 variable distributions.  Any  available data,would be  used to validate  the
 input  variable distributions  aad predicted distribution of dose.  Discussions
 vith subject-cutter experts  are not aufficient to establish confidence inter-
 val estimates for perceatiles of the estimated distribution of dose.   However,
 differences ia the expert opinions ceaceraiai the input variable distributions
 ahould  be discussed.   Moreover, sensitivity of  the  estimated  distribution of
 dose to plausible alternative siodels ahould be discussed.
      Coapariaon of Figure A.3  to Figure A.I illustrates the effect of discus-
 aioaa with  aubjeet-Mtter experts for this hypothetical aaposure asaessaent.
 Several changes occurred  ia  the hypothetical aaaaple.   The distribution of ED
 was changed froa Unifora on the range (0,900) to a scaled Beta distribution on
 (0,600).  The range of the Uaifora IR distribution was  modified froa (0,11) to
 (0,5).   Finally, the Uaifora  diatribution  of AF oa  the  range  (O.8., 1.0)  was
 replaced with the  fixed value unity.  Such changes in  the  ranges of the input
 variables Bight occur  if  conservative estimates of input variable ranges vere
 aaed for the initial axpoaure aaaeaaaent.  The aet effect of all these changes
 an the  worst caae acaaario ia  rather dramatic.  The worst case  aceaario  now
 predicts  the  Maximum lifetime  dose  to  be  3,000 aanograms of  absorbed TCDD,
 iaatead of 9,900 aanograms.
      Vhcn comparing Figures A.I aad A.3, another important  change ia  that only
 figure A.3 can be iaterpreted aa an aatimate af the population distribution of
 dose.   It can be  ao  iaterpreted becauae  it  ia based  upon estimates  of  the
 distributions af the  aodel  input  variables.  Figure  A.I was generated  by
 treating all values within  conservative estimates  af the raages of  the aodel
••impat variables as ooually likely;  it represents the dose  aaaaitivity distri-
 bution, aot an estimate af the population distribution  of dose.
      In order to iaolate  the affect of replacing the Vnifora distribution  for
 ED with a acaled  Beta distribution, an additional  simulation was  performed
                                   A-10

-------
  40H
P
E
R  20-
C
E
N
T
       o
2000       •    4000

          PREDICTED  DOSE CNG/LIFETIfO
   Figure A.3.  Predicted Dose Distribution Resulting from the Pretwwd Nedet and Inpat Variable Diatrihutlons

-------
Ttble A. 2.  SELECTED FERCENTILES FOR THE DOSE DISTRIBUTION RESULTING FROM
                THE FRESUKED MODEL AND INPUT VARIABLES DISTRIBUTIONS
Fercentile
0.01
0.02
0.03
0.04
0.05
0.06
0.07
0.08
0.09
0. 10
0.15
0.20
0.25
0.30
0.35
0.40
0.45
0.50
0.55
0.60
0.65
0.70
' 0.75
o.eo
o.es
0.90
0.91
0.92
0.93
0.94
0.95
0.96
C.97
0.98
0.99
Joint
SitiMtc
12.57
23.63
35.49
48.77
61.09
76.02
67.27
101.28
114.63
125.63
197.32
259.01
321.68
385.25
452.75
519.19
890. 15
668.44
751.23
628.63
919.94
1019.62
1116.64
1244.86
1383.62
1547.47
1592.77
1626.67
1664.63
1721.11
1767.70
1642.66
1921.01
2O34.71
8210.60
                                  A-12

-------
 using the aane  input distributions that  produced  Figure A.3, except  that  a
 Uniform (0,600)  distribution was used for the ED distribution.   Tbus,  Figure
 A.4 is based upon a  Monte Carlo simulation with 3,000 iterations  and indepen-
 dent input variables  with  the following  distributions:   (1)  Uniform (0,600)
 for Er; (2) Uniform  (0,5.0) for IK; and (3)  AF • 1.00.  Selected percaatiles
 from the resulting distribution ef dose are presented in Table A.3 for compari-
 aoa to Table A.2.  Comparison of these tables reveals that the exposure distri-
 bution baaed upon a  Uniform distribution for  ED (Table A.3) is  flstter  and
 •ore disperse than that based upon the Beta distribution (Table A.2).  This is
 a result of the fact  that the Uniform distribution is itself flatter and more
 dispersed than the Beta distribution.
 A. 4 ASSESSMENT BASED  UPON DATA FOR MODEL IKPUT VARIABLES
     The  exposure assessment based upon an estimate of  the  joint probability
 distribution  for model input  variables  can be  refined by  collecting  sample
 surrey  data for  all  personal exposure model input varisbles for a probability
 •ample  of population members.   The population  distribution of  exposure  can
 then be  estimated by  computing the predicted exposure  for  each aample member
based  upon  tbe  model.  These  predicted  exposures  can  be used   to directly
 compute confidence interval  estimates  for percentiles of the exposure distri-
bution.  If the  joint distribution of the model input varisbles is hoove,  the
aample survey data can be used to compute joint confidence interval estimates
 for the parameters of  tbe input vsrisble distributions, which can  then be used
to  compute  confidence interval  estimates for  percentiles of  the  exposure
 distribution.    Using  either method of  computation,  interval estimates  for
percentiles of  tbe exposure distribution are  a useful  quantitative  charac-
terixation of uncertainty for an exposure assessment.
     Development  of  confidence  interval  estimates  for percentiles  of  the
exposure  distribution would ordinarily require  that sample  survey data be
 collected for exposure  and model  input variables.  That  is, a  probability
•ample  of population  members  would be •elected, and data would  be collected
 for  these individuals for  the exposure end model  input variables.  For  the
example aader consideration, the exposure of interest is the lifetime absorbed
dose  of TCDD via the ingestion  rente and the deposition pathway.   The model
 input variables  are the number of days of ingestion of contaminated aoil in a
 lifetime, ED; the rate of soil ingestion, IB,  in grams per day; and the absorp-
 tion  fraction,  AF.   Unfortunately,  none of these  variables are  amensble  to
                                  A-13

-------
p
e
R
C
E
N
T
                                                                                      • • '
                               PREDICTEO DOSE  CNG/LIFCTIMEJ
    A.4.  Dittritmtlofi e>f tstimted Doae Resulting tnm Alternative Cifially Likely Input Variable Co«blnation

-------
Table A.3.  SELECTED PEXCENTILES FOR ESTIMATED DOSE RESULTING FROM
            AITEXNATIVE EQUALLY LIKELY INPUT VARIABLE COMBINATIONS
Ptrccotile
0.01
0.02
0.03
O.04
0.05
0.06
O.07
o.oe
0.09
0.10
0.15
0.20
0.25
0.30
0.35
0.40
0.45
0.50
'0.55
0.60
0.65
0.70
0.75
0.60
O.65
0.90
0.91
0.92
0.93
0.94
0.95
0.96
0.97
0.96
0.99
Point
IstiMte
2.98
7.64
S3.57
19.66
27.51
as.se
43.09
49.62
86.27
44.57
101.46
154.39
206.61
266.19
331.07
401.69
464.14
852. 1&
645.oe
726.24
649.74
979.36
1134.76
1314.16
1503.40
1770.46
1642.22
1915.40
1960.76
2079.25
2156.24
2256.32
2328.70
2457.85
2626.13
I Itprorfuctd .from
I bttt
                                                         cop
                             A-15

-------
direct  observation.   For  the take of illustration only, this  example  will be
coati&ued as  if the absorption fraction were known  to  be  unity and the  other
variables  could be  measured dirtetly  for a tuple  of individuals from the
target  population.
  .4.1   Xmterval_Estim«tes ef Iicposure Percentiles:   Input Variable  Distribu-
               >t Inovn
     For many personal  exposure models, it My be  such  easier  end  less costly
to  collect data en the model  input variables, rather  than on the exposure
itself.  Beace, given model (1), tappoae that data  have  been collected  for the
exposure duration, ED, and the iagestion rate, ZR,  for a probability Maple of
subjects from the  target  population.  For illustration, hypothetical outcomes
for a simple  random sample of aise  500 were  generated  using a symmetric Beta
(e * B * 3-D  distribution for  ED/600 sad  an independent Uniform (0, 5.0)
distribution  for  XR.  The 300  pairs of values generated  for  ED  and IR are
listed  in Attachment A.I.  Histograms of the ED and  XR  distributions for this
hypothetical sample are shown in figures A.5 and A.6.
     The predicted dose was computed for each of the  500 sample members as the
product of  the ED  and  XR values  shown in Attachment A.I  in  accordance with
model (1), assuming  that  AT  is unity.  Treating these  500  predicted doses as
if they were  measured  doses,  equal tail area 95 percent confidence interval
estimates were  computed for  percentiles of the distribution of dose.9   These
confidence interval estimates are listed for selected percentiles in Table A.4
ond presented graphically in Figure*A.7.
     Computation of the interval estimates for perccatiles of the distribution
of dose shown  in  Table A.4 and  Figure A.7 is based upon:  (1) the exposure
model (1); (2) sample survey .data on the input variables (Attachment A.I); and
(3) the assumption that the  absorption fraction (AT) is unity.  The interval
estimates of p«rcontilcs of the distribution are a  mseful fuantitative  charac-
terisation of uncertainty for the exposure assessment.  Bovever, chsracterisa-
tion of mneertainty for the exposure  assessment  is mot •complete without a
thorough discussion  of  all mssumptioas, especially Justification of the model
matd U compute the predicted doses.  The design of  the sample survey  used to
collect the -data should also-fee  discussed.  'Boreovcr, •tasitivity  to the
Y8ee  Bansen, M.  H., V.  H. Burwits,  and  V. C.  lUdov (1953), Sample Survey
Methods  and  Theory. Section  10.18;  or Kish, Leslie  (1965), Survey  Saapling.
Section  12.9.
                                  A-16

-------
I5.O-
12.5-
P
E
R
C
E
N
T
 7.5-
 S.O-
 2.5-
              50    100
          Pigare A.S.
                      150  200  250   300  350  400   450  500  55O   60O

                           EXPOSURE DURATIONCDAYS)

                     M of the Rypethettcal  Savple of Size 500 for Exposure Duration (CD)

-------
       12.
       10.
    P
    E
    R
    C
    E
7.5-
S   T   5.0-
        o.o-
                   0.5    1.0    1.5   2.0   2.5   9.0    3.5    4.O    4.5    5.8

                                 INGESTION RATTCGRAHS/OAY)
              Figure A.6.  RistoRrM or the Hypothetical Sanple of Size 500 for Inge*lion Rate (IP)

-------
Table A.4.  95 PERCENT CONFIDENCE INTERVAL ESTIMATES FOR SELECTED PERCENTILES
              OF TEE DISTRIBUTION OF DOSE (Y) BASED ON MODEL-PREDICTED DOSES
j^^^te^

^EfTt^' ff'i* • * *^S 0>w01
|P0|jfeh$ f?»~...&~j 8). 02
tey^^/j?'.. •' J j| Q,03
*-^^^ ' *^*7~""O.04
1:"!'i 0.05
"'- .""* O.06
.07
.06
.09
.10
.15
.20
.25
.30
.35
' .40
0.45
0.50
0.85
0.60
0.65
0.70
O.75
0.60
O.85
0.90
0.91
O.92
O.93
0.94
0.95
* O.96
.-.*..< 0.97
'• — 	 _
W ' • ' •.*»
lover
i Limit
• ^^t
•80 •
4). 80
t.43
i3.26
23.79
88.36
48.86
86.38
47.01
78.97
186.50
196.79
257.81
285.10
•39.16
400.67
462.55
859.40
424.04
494.09
776.75
•85.71
981.62
1086.97
1243.56
1416.62
1443.15
1829.68
1899. 17
1448.46
1749.26
1859.73
1097.96
3)007.93
9143.18
Poiat
UtiMte
7^^
••V
10.49
93.12
88.46
49.83
41.23
74.49
•1.32
114.38
137.07
106.11
935.09
276.87
•32.24
991.30
452.37
848.56
408.42
482.09
781.23
•73.85
955.25
1085.00
1173.99
1331.26
1850.54
1430.07
1468.86
1747.71
1847.36
1880.79
1941.24
9027.12
9142.16
8980.78
Upper
limit
• A 4K
• V.29
95.28
44.49
87.01
72.48
•2.15
124.46
141.70
149.59
162.53
212.77
266.90
311.68
373.27
445.83
834.44
404.82
476.24
748.56
•61.61
949. 18
1020.47
1135.79
1291.32
1428.95
1725.20
1779.50
1666.79
1885.84
1944.57
2018.62
2127.76
2171.92
9295.75
9428.30
                                 A-19

-------
i.e-l
    c
    u
    N
    U
    L
    A
    T
    X
    V
    E

    P
    C
    R
    C
    E
    N
    T
0.4-
0.2-
        0.
                                                               **          ••
0
                       500
 1000        1500

EXPOSURE CMS/LIFETIME*
                                                               2500
Pigvre A.?. 95 Perccat Coafldmce Interval Catinte* for Selected Pereentllea ef the BMC DUtrlbatim
                                    on Model-Predicted Doses

-------
essuaed  model  should  be  investigated  by  computing  the predicted doses  and
confidence  iaterval estimates for plausible alternative models.   If any direct-
ly measured exposure data are available (data on absorbed doses for the pret-
ant hypothetical example), these data should be ased to validate the predicted
distribution of dose.
A.4.2 Interval Estimates of Exposure Percentilet;  Input Vsrisble Distribu-
       tions Known
      Coafideace iaterval estimates vere derived for perceatiles of the  expo-
sure  distribution  ia Section A.4.1 primarily by computing the doses predicted
by model  (1), given dsta for ED sad I*.  When the joint probability distribu-
tion  for  the personsl exposure model input vsriablcs is presumed to be haovn,
confidence  iaterval estimates for perceatiles of the exposure distribution can
be  generated from a joint  confidence  region for the parameters of  the joint
distribution of all model input variablea.  The level of coafideace essocisted
vith intervsl estimates of exposure distribution perceatiles is  then the level
of  confidence associsted vith the joint coafideace region  for  the psrameters
•f  the model  Input variables.   Vhca  the  input variables  are  statistically
independent,  the  joint  confidence level  is  the product  of the  confidence
levels for  the  parameters of the iadividusl input variables. For example, if
the exposure model has oaly two input variables that are independent,  then 95
percent confidence iaterval estimates of the exposure distribution percentilet
arc generated from the iadividual 97.5 percent (V0.9S) confidence regions for
the parameters of the two iaput variables.
     Thus, for model (1)» 95 perceat confidence iaterval estimates for  percen-
tiles  of  the distribution  of lifetime abaorbed  dose vill  be  geaersted fron
97.S perceat confidence  regions  for the parameters of the ED end XR distribu-
tions, assuming that  the variables are atatistically independent and that the
absorption  fraction,  AT, is  fixed at one.  For  the purpose of illustration,
assume that expert opinion  supports  the Bets distribution as the  true  prob-
ability distribution  for ID/600,  and  that the Vmifon  (0,6) distribution is
•apperted as the population distribution for IB.  Ae data ia  Attachment A.I
vill  then me msed  to determine 97.5 perceat coafideace regions  for the param-
eters  of  the ED aad 1R  distributions  ia accordance vith these assumptions.
First, a  97.5  percent confidence iaterval estimate  vill be derived  for the
parameter 6 of  the IR distribution.  Second,  the dsts  for ED vill be  used to
                                  A-21

-------
determine  an approximate 97.5 percent joint confidence  region for  the  param-
•ttrs  of the Beta  distribution for ID/600.  The  95  percent  joint  confidence
teflon  for  the  input variable parameters  will then be  used to generate  95
percent  confidence  interval estimates for percentiles of the exposure distri-
bution.
     Since  the  presumed probability distribution for IR is Uniform  (0,6), the
Unique  Minimum Variance Unbiased  Estimator (UMVUE)  for  the  parameter  6 is8
     t • (a+1) 1*^  / n,
•here fl^a> !• the largest iagestion rate observed in  • simple  randosj  sample
•f aiae  a.   Using this estimator, the UMVUE for the pa***eter 6 based upon the
daU for IF in Attachment A.I is  5.005.  Moreover, the standard trror  (i.e.,
the estimated standard deviation) of this estimate can be shown to be
     •E (6) » 21^ / VaTaTn • 0.010.
Since  the  maximum of  a simple random  sample of site 500 vill have  approxi-
nately a formal probability distribution, an approximate equal tail area  97.5
percent (V0.95) confidence interval estimate of 0 is from 4.982 to 5.027.
([Rote that the true value 6, 6 » 5.0, which would mot be known with real  data,
Lies within this interval.)
     The data for  exposure duration (ED) in Attachment A.I were  used to com-
pute maximum likelihood  estimates ef  the parameters  a and  0 for the  Beta
probability distribution for CD/600.9  The maximum likelihood  estimates of
these parameters were:  fi • 2.97 and 9 • 2.93.  A joint 97.5 percent (V0.95)
confidence region for these parameters was then approximated by independent
                                                              *
OS. 7 percent (4Vo795) -confidence interval estimate* for 6 end  B.to  Using
«*ee ratel, Jagdlsb t. (1973).  Complete Sufficient Statistics aad HVU Isti-
oaters.  Cemmunicstiona in Statistics. Vol. 2, Vo. 4, pp. 327-336.
°*ee I.  1. Johnson  aad S.  loti (1970), Pistributioni in Statistics;  Contin-
nous Vaivarlste Distributions, Section 24.4.
fi°The  joint confidence level  it 97.5 percent  (VoTw)  assuming independence,
Diace  tke  joint confidence level for independent estimates is  the  product of
the individual coafidence levels.
                                  A-22

-------
 the  Halting Normal diitributieo for  these  maximum  likehihood  estimates,  the
 equal  uil  area  98.7 percent confidence interval estimates vere computed to be
 from 2.54 to 3.40 for 6  and  2.51 to 3.35 for 0.  (Vote that the true values,
 0 «  8  « 3.1ft, which would Dot be known with rtal data, lie^ within these inter-
 vals.)  If the  maximum  likelihood estimators for ft and  0  were statistically
 imdepeadent, the croas product space defined by parameter values within the
 above  ranges would be • joint 97.5 percent (VOJ) confidence region for these
 parameters,  lastoad,  the estimators ars positively correlated (correlation ~
 0.15), and the joint confidence ration is shaped something like the ellipsoid
 •haded in figure A.8.  Since analytic specification of the 97.5 percent confi-
 dence  region shown  in this figure  is very  complex,  the joint 97.5  percent
 confidence  interval  estimate  for the Beta distribution parameters o and 6 was
 approximated by  the  cross product apace defined by  the Interval estimates of
 the  individual  parameters,  i.e., the rectangle  in Figure  A.8.   This is  a
 conservative estimate since the probability content of the rectangle is neces-
 sarily greater than that of the ellipsoid.
   2.51
              2.54     3.40
     Figure A.8.  Confidence region for the estimated Beta distribution
                                 parameters 6 and 0

     The cross product space defined by the 97.5 percent (V0.95) confidence
iaterval estimate ef the parameter t of the IX distribution and the joist 97.5
perceat  confidence region for  the  parameters a and 0 of the  ZD distribution
femeratea a  corresponding act ef dose distributions via model (1).  fince the
joint level of confidence associated with this cross product space is approxi-
mately 95 percent, an  approximate  lower (upper) 95 percent confidence  limit
for any  mercantile of the distribution of dose is given by the minimum (maxi-
mum) ef that percentile for ell dose distributions generated by input variable
parameters within this cross product space.

                                  A-23

-------
     Since  the  predicted  dote it computed as tae product of EC tod  IK for the
 model  under consideration, the  lover (upper) confidence limits for the dote
 diitribution ere  generated by the ID and U  distributions  that give the most
 probability to  nail  (large)  values of ED and IB.   Hence,  the lover limit of
 the parameter estimate  for the IR distribution, 6 • 4.982,  was used to gene-
 rate lower  confidence  limits  on the dose distribution,  and  the upper limit, 6
 • 5.027,  MS  uod to generate upper confidence limits.   Since  the mean of the
 Beta distribution varies  directly with o, the lower  liait « • 2.54, was used
 to generate the  lower  confidence limits for dose,  cad the upper limit,  o *
 3.40, MI Mtd to generate upper limits.  Since the parameter 0 affects euioly
 the variance of the Beta distribution for ID/600, it  MS accessary to consider
 both the  upper  and lover limits  on  0 for generating each oet of dose confi-
 dence limits, both  lover  and  upper,  fence, the following procedure was used
 to compute  the  95 percent confidence limits on  the  distribution of dose.  To
 geaerate  the  lower confidence limits,  two Monte Carlo aimulations were per-
 formed as follows:
     1.   The first  simulation consisted of  3,000  iterations  coaputed using
          model (1) with  the  following inputs:   AT • 1.0; IR distributed as a
          uniform variate with 0 • 4.982; ED/600 distributed as a Beta variate
          with o • 2.54 and 6  » 2.51.
     2.   The second  simulation  consisted of 3,000  iterations  computed using
          model  (1) with  the  following inputs:   AT • 1.0; IR distributed as a
          Uniform variate with 6 • 4.982; ED/600 diatributed as a Beta variate
          with a • 2.54 and 0  » 3-35.
                                   •
Fercentiles were estimated for each of the dose distributions generated by the
Hoate Carlo method.   The  lover  95 percent confidence limit estimate was the
•mailer of  the  two estimates produced  in this way  for «ach percentile.  The
same procedure was used to generate the upper 95  percent confidence  limits lor
etosc, except that 6  • 9.027 and 0 • 9.40 was Mad for the Monte Carlo simula-
tions,  and  the  larger of  the  two  estimates MS used for each perccotile.
   .. Aeae  95 percent comfidamca interval estimates are listed for selected
mercentilas ef the distribution ef dese in Table  A.9  and praacnted graphically
in Figure A.9.    The  point estimates of percentilea nhown  in Table A.5 and
figure  A.f  were computed directly from  the  900 doses predicted by model (1)
Iron the  data in Atuchment A.I.  Thus, the point eatimatea in Table A.5 are
 identical to the  point  estimates ahown in Table A.4. The confidence interval
 estimates ere  similar, but mot identical,  because  they are computed using
                                  A-24

-------
Table A.5.  95 PERCENT CONFIDENCE INTERVAL ESTIMATES FOR SELECTED PERCEKTILES
            OF THE DOSE DISTRIBUTION (Y) ASSUMING UNIFORM AKD BETA DISTRI-
                        BUTIONS FOR XR AND ED/600, RESPECTIVELY
Perceatile
0.01
0.02
.03
.04
.05
.06
.07
.06
.09
0.10
0.15
0.20
0.25
0.30
O.35
0.40
0.45
0.50
0.55
0.60
0.65
0.70
0.75
o.6o
0.85
0.90
0.91 .
0.92
.93
.94
.95
.96
.f7
.98
.99
lever
liflit
4.53
12.72
21.22
91.21
42.48
53.09
40.77
70.27
90.10
90.51
136.23
175.91
222.60
278.46
330.77
390.97
451.73
515.17
563.62
655.95
741.83
022. 19
931.30
1045.89
1160.18
1329.36
1371.03
1399.65
1444.76
1513.55
1945.16
1439.32
1794.95
1624.22
1993.04
.Point
Eitiaite
7.96
10.49
23.12
95.66
49.53
41.23
74.49
91.32
114.38
137.07
166.11
235.69
276.87
332.24
991.90
452.37
846.96
406.42
462.09
751.23
973.95
955.25
1055.00
1173.99
1331.26
1550.54
1430.07
1446.66
1747.71
1847.38
1980.79
1941.24
2027.12
2142.16
2250.78
Upper
LiBit
20.39
40.14
93.15
72.43
90.36
101.55
117.74
135.16
150.55
167.92
255.08
326.73
997.25
476.95
554.75
439.71
726.46
612.43
900.41
994.06
1098.52
1204.34
1309.51
1445.75
1590.58
1782.67
1616.17
1663.04
1914.19
1964.64
2036.44
2122.65
2224.31
2348.62
2486.36
                                  A-25

-------
        t.l
   c
   u
   M
   U
   L
   A
   T
   I
   V
   E
0,
        0,
£   P
^M   ^»
    R
    C
    E   0.2-
    N
    T
        e.0-
                                   »   •
                                t    »
                        i  »       '
                      •  »      «
                    i  »     i
                        400
                              60O
120O
I60O
                                        EXPOSURE  CMC/LIFETIMES
      Figure A.9.  93 Percent Cmfldmce Interval CstlMte* for Selected Percent! let of the fto«e Diatributien
                          A«*u»!nft Unifora and Beta Diatribntlons for 1R and ED/600, Reopectively

-------
 different procedures,  and the interval estimates shown in  Table  A.5  sake use
 of the  additional  assumptions that • Uaifera (0,6) distribution is accepted as
 the probability distribution for IK and an independent Beta  distribution is
 accepted for D/600.  The interval estimates shown  in Table A.5  should theo-
 retically to aarrower than those ahovn in  Table A.4 due to the  use  of these
 additional assumptions,   However, the  opposite is  true.   This auy be  due to
 the conservative  nature  of  the procedures  used to estimate the  joint  confi-
 dence region for the parameters of the ED and TR input variable distributions.
 Moreover, the  >ormal  distribution say  BOt^me • food  approximation of  the
 distribution! of the parameter  estimators (0,  a, and 0) vhen such high  levels
 of confidence,  97.5 and 98.7, are necessary.
      Computation of  the  interval  estimate for  percentiles  of the  distribution
 of dose  shown  in  Table  A.5  and  Figure A.9 is based open:  (1)  the  exposure
model  (1);  (2) sample survey data  oa the  input variables (Attachment  A.I);
 (3) the assumption that the absorption fraction (AT)  is unity;  (4) the assump-
tion that the exposure duration (ED) aad aoil ingestion rate (IX)  are  indepen-
dent  random  variables;  (5) the  assumption that the distribution  of ingestion
rate  is Uniform (0,6);  and  (6) the assumption  that the distribution  of  the
scaled  exposure duration  variable  ED/600  can be characterised  by the  Beta
probability distribution.  The resulting interval estimates for percentiles of
the dose distribution are a useful quantitative characterization of uncertain*
ty for  the exposure  assessment.   However, characterisation of  uncertainty for
the exposure assessment  is not  complete without a thorough discussion  of the
justification  for  all  assumptions used in deriving these  interval estimates.
The assumption that  a particular family of probability distributions  can be
used to  represent  a  model input variable, e.g., the Beta  family  of distribu-
tions for ED, can  be tested by standard goodness of  fit tests.11  limitations
of the  data  aad design of the aample survey that produced the data should be
discussed.   limitations  of the  procedures  used  to  compute joint confidence
interval  estimates  of the  parameters of  the  input  variable distributions
should  also  to  discussed.  The number  of iterations used in the Bonte  Carlo
slr^lstions  should to  justified, aclBovledgiag that sjort iterations  would
produce  mere precise estimates at  higher cost,  finally, sensitivity to  the
"See Daniel,  Wayne V. (1978), Applied Henparameteric Statistics. Chapter 8;
or Conover,  V.  J.  (1971), Practical Honparametric Statistics. Section 4.3 and
Chapter 6.

                                  A-27

-------
 assuaed  model  should be  investigated  by confuting the predicted  doses  and
 confidence Interval estimates  for plausible alternative models.
 a.S. aJSESSttXT BASED UPON EXPOSURE MEASUREMENTS
      Zhe iiaal  atage of refinement for an exposure •••tsucnt  ii  to collect
 Maple eurvey data en exposure for a probability aaaiple of population members.
 Such a database can  greatly docraase  the  uncertainty eaaociated vith estima-
 tion ef the population distribution of exposure if a sufficiently large sample
 la selected.  Ibis reduction  in  uncertainty is achieved mainly by eliminating
 the  mac  ef  a  sjodal  to predict  exposure.  Moreover, a  model,  which My  be
 meeded to relate exposures to  sources, can be validated and its parameters can
 be estimated, if data arc  alao collected for ell model input variables.
      Suppose that  BO alternatives to model (1) were under  consideration,  end
 data were collected for absorbed  dose (Y), exposure duration (ED), and iages-
 tion rate  (XX), aesumiag  that the  absorption fraction  (AT)  was unity,   for
 illustration, hypothetical outceaes  for a eimplc raadom  sample  of  site11  500
 wera generated  using  a symmetric Beta  (a • B «3.1) distribution for ED/600 and
 ea independent Uaifon  (0,  5.0)  distribution  for  IK.  Bather  than assuming
 that sjodel  (1)  predicts  an individual's  dose exactly, the dose, Y(i), absorbed
 by the  i-th population member  was  generated fro* the model
      In (Y(i)]  • 1.10 In (ED(i)] -» 0.90  In (IF(i)] + ln(t(i)),        (2)
 •here la denotes the aatural  legarithsi, and  the  error ten, ln(c), follows a
 Bonal  probsbility distribution vith a Man of xero and  a  atandard deviation
                                   •
 ef one-third.1* touivalently, the dose absorbed by the i-th population swaber
 vas generated as
      I(i) • lED(i))1'10  UR(i))0-90  C(i)  ,                           (3)
 •here c is a random  error-ten with e legaenal distribution having a ewdian
 •f uity.  The raadosj error tana in  tjodels  (2)  aad (3) can be thought ef as
 the collective effect of  siiaor explanatory variables act explicitly included
 "In practice, the sample aise aaeded  to achieve acceptable precision within
 •pacified cent constraints awst be derived.  Sample aiaes much larger than 500
. may often be meeded.
 "It ia recognized that data generated  by this model ere much eimpler and more
 veil-behaved than  any real exposure data.   Lacking real data  with which to
 illustrate the proposed procedures, a aimple  hypothetical  database has been
 generated.
                                   A-28

-------
in  the  model.   The data generated for i hypothetical  simple  random aaople of
500 Mabers of the target population are presented in Attachment A.2.
     Given  •  database of  Matured exposure values  such  as the hypothetical
doses ahowc in Attachment A.2, the population distribution of exposure may be
estimated directly.   The  cample exposure  distribution is  a direct estimate of
the population distribution  of exposure.   The Measured  exposure  values My
also be oaed to directly confute confidence interval  tstiMtes for percentiles
of  the  exposure distribution ea ehovn in Table A.6  end Figure A. 10."  These
confidence  interval   estimates  for ptrcontilos of  the exposure distribution
define  an approximate confidence region fer the expoaure  distribution.  When
the catiMte ef the  diatribution of expoaure ia baaed upon Masured exposures
for a probability aample  of population members, confidence interval estimates
for percentiles ef the exposure distribution are  the priaury  characterization
ef  oaccrtainty for  the expoaure  essesse«nt.   liaiitations of  the  data and
design  of  the aurvey naed to  collect  the  data ehould alao be diacussed.  No
mention ef  e  model   ia necessary  because  the  exposure  levels  were Masured
directly rather than calculated from e ewdel.
     If a  database such  as that  shown  in Attachment A.2 is available, the
parameters ef e Model relating peraonal expoaure  to  explanatory variables for
which data were collected  My be estimated using  etatistical  regression  tech-
niques.   Moreover, lack ef fit  can  be tested for a hypothetical model  based
upon a  postulated exposure scenario, auch  as model  (1).   Both of these  tech-
niques will be illustrated below.  '
     Suppose that model (1),  with an absorption fraction, AT,  ef unity, is
expected to predict an individual'a lifetime abaorbed dose. This model csn be
expreased equivalently aa
      la T(i) • In ED(i) * la XX(i) + la «(i),                        (4)
where T(i)  ia the dose experienced  by the i-th population member.  It would
then be reaaonable to ase the data to eatimatc the  coefficients e^ end  o^ ef
the regression equation
     la T(i) • Oj In DCi) * fl2 In IR(i) *  I*,                        (5)
"See Bans en, H. B., Burvitc, V. »., end Madow, V.  6.  (1953), Basn>le 8
Methods and Theory. Section 1.18; or Xiah, 1. (1965),  Survey Sampling.
  tction 12.9.

                                  A-29

-------
Table A.6.  95 PERCENT CONFIDENCE INTERVAL ESTIMATES FOR SELECTED
          KRCENTIiES OF THE DOSE DISTRIBUTION (Y) DIRECTLY FROM
                        EXPOSURE DATA
                         lover        Peiat       Upper
         Ptrctttile       Uait      IitiMte      limit
            0.01          8.46       2O.30      31.70
            0.02          19.74       84.18      78.86
            0.03          23.96       78.18      97.05
            0.04          88.31       90.82     128.52
              05          78.39      105.47     137.30
             .06          90.06      129.99     151.25
             .07         100.43      137.32     191.73
             .08         127.78      149.44     206.15
             .09         134.82      181.19     228.47
             .10         143.55      204.65     256.13
             .15         239.87      319.20     971.06
            0.20         954.23      994.30     480.99
            0.25         414.71 .     820.98     898.39
            0.30         834.03      422.84     476.32
            0.35         436.54      703.46     804.16
            0.40         720.13      830.11     912.92
            0.45         846.09      932.71    1066.57
            0.50         948.75     1099.23    1176.51
            0.55        1101.77     1181.76    1353.50
            0.60        1212.82     1363.13    1477.58
            O.65        1377.45     1489.36    1633.91
            0.70        1830.40     1490.40    1600.59
            0.75        1714.53     1853.71    1982.21
            0.60        1893.70     2075.07    2312.22
            O.85        2202.19     2402.35    2667.51
            0.90        2626.97     2966.36    8233.86
            0.91        2727.32     9084.35    9305.52
            0.92        2916.67     9169.80    9408.31
            O.93        9081.00     9297.61    9592.00
            0.94        9185.46     9937.49    9831.06
            O.95        9296.06     9464.11    9889.16
            0.96        9)397.85     8809.79    4033.09
            •.97     ' "8)472.42   -^894.85 •"tf*253.92
            0.98        8881.38     4092.06    4692.71
            0.99        4106.45     4447.16    4466.43
I
                                       fttproduced from
                                       ••»» tvaiUbU copy.
                        A-30

-------
    0.8-
c   o.o-
u
N
U
L
A
T
I
V
E
    0.4-
P
e
R
C
e   0.2-
N
T
    0,
         0
                 IOO8     2O0O     3000
                            EXPOSURE CNG/LIFETIME5
A.W.  95 P«rceat Confidence Interval Cttioates for Selected Perceetile* «f the I>Me Distribution
                                fron Directly Heaaured Do*e»

-------
where  C* is  •  Morally distributed random error tern with Dean  zero,   Using
the data la Attachment A.2, the estimated regression equation is
     la T(i)  •  1.10 la £D(i) * 0.91 la XR(i).                         (6)
The  analysis of variance  Uble for this regression equation is  presented as
ftiiplsy  A,i.   the  multiple coefficient  of  deteniaation,  Ra,  is  a useful
measure  of goodness  of fit of the model.11  The value of R2  is  very  large for
•the  present example  (R*«0.998)  because  the  hypothetical data  were  geaerated
from  a model of the  type  bciag fitted,   la practice, ouch nailer values of
1*,  e.g.,  1* > 0.60, may bo  regarded as  Indicating  a reasonably good fit.
     laving  fitted the regression model  (6), adequacy of the model (4) can be
tooted by  toatiag  the hypothesis that the regression coefficients, 0j and Oj,
arc both uaity.  Za  this case, the P-value of  the  hypothesis test is aeen in
Dioplay A.I  to  bo  essentially aero.  That  ia,  there  is  almost  no chance that
the estimated values of C. aad o. can differ froa unity by the observed aaount
or sjore  when the  true values of o. and  o2 are both unity.  Thus,  the  hypo-
thesis  that model  (1)  generated  the data can be  rejected  based  upon this
hypothetical aaaple of SCO measured exposures when the true state of  nature is
described by awdel (3).  Of course, the  true atate of nature  is never  known
for a real exposure assessment.
 »»The multiple .coefficient of  determination, R1, «ay  he interpreted os  the
mwportion of variability  ia  the observed exposures taplaimed by  the model.
 It ia  the square of the multiple correlation coefficient,  which  mcssures the
partial  correlation between the  dependent variable and the independent  vari-
 ables.
                                  A-32

-------
DISPLAY A.i: ANALYSIS OF VALANCE TAME f
             REGRESSION EQUATION
DEF VARIABLE! LOG. DOSE

                   sun or         MEAN
SOURCE    BF      SQUARES       6QUAF.E      F VALUE       FRO*  F

MODEL      2    23484.986    11742.493   112743.629       O.OOC1
ERROR    498    81.867778     0.104152
U TOTAL  900    23536.854

    WOT HSE     0.322726     ft-SOUARE       0.9978
    SCP MEAN     4*774368     ADJ R-SQ       0*9978
    C.V.          4.76393

NOTE: NO INTERCEPT TERM xs USED. R-SOUARE xs REDEFXNED.
                PARAMETER     STANDARD   T FOR HO:

VARIABLE  BF     CSTXHATE        CRROR  PARAHETER-0   PROK > IT!  TYPE  XX 6S


LOG.XR     1     0.911218     0.014461       43.010       0.0001     413.514

LOG.ED     1     1.102879  0.003020861      365.088       0.0001    13882.336


TEST:            NUMERATOR:    ?o,803e  BF:    2   r VALUE:  679.eno
                 BENONINATOR: 0.104152  IT:  498   PROII  >F :   o.oooi
                                       A-33

-------
ATTACHMENT A.It  EXPOSURE  DURATION (ED) AND 1NGESTJON RATE (IB)  FOR
                A HYPOTHETICAL SIMPLE lANOOH IAHPIC OP SCO PEOPLE
                     THE  TAI6ET POPULATION
 It
II
ID
XI
ID
ID
                                IR









10
11
12
IS
14
IS
14
17
18
19
20
11
12
15
14
25
24
17
to
19
10
11
12
13
14
IS
14
17
38
2)9
40
Ml
42
41
44
48
44
47
48
49
00
890.43
814.40
408.44
447.18
ISO. 84
188.89
182.29
420.17
459.12
450.48
155.22
100.14
418.41
194.13
228.30
114.87
842.12
129.49
149.24
148.81
74.99
404.88
273.09
130.12
19J.45
252.09
407.47
114.19
197.10
140.30
243.93
158.40
420.39
174.70
145.24
141.91
203.41
148.00
J04.ll
488.84
480.21
444.08
•80.99
94.11
244.08
111.11
198.21
417.01
82.99 <
115.47 1
.92
.25
.43
.43
.99
.19
.78
.44
.99
.43
.41
.19
.42
.70
.74
.48
.49
.11
.73
.11
.99
.02
.89
.42
.19
.14
.40
.91
.04
.42
.72
.48
.49
.40
.84
.21
.43
.19
.11
.44
.01
.17
.90
.17
.tl
.tl
.18
.90
>.84
t.49
tl
82
83
84
15
t4
17
18
89
80
41
82
85
84
45
84
87
88
89
70
71
72
73
74
75
74
77
78
79
80
81
82
83
84
85
•4
87
88
•9
90
91
92
91
94
91
94
97
98
99
100
171.17
447.14
170.43
188.41
144.11
190.10
187.03
87.41
841.90
144.93
192.84
190.85
182.49
111.24
139.34
442.77
124.05
108.89
278.44
102.18
142.25
185.15
242.74
179.92
111.84
88.94
118.18
1 20. :14
123.07
105.42
197.14
279.84
85.84
470.18
211.74
804.74
177.40
180.88
188.10
140.47
407.41
427.99
147.81
171.84
144.71
199.91
140.08
124.41
H. 48
.49
.71
.15
.55
.94
.88
.19
.99
.03
.24
.12
.43
.44
.74
.24
.83
.27
.80
.75
.44
.71
.21
.05
.73
.40
.14
.80
.48
.25
.75
.41
.51
.13
.45
.84
.43
.17
.77
.97
.71
.17
.98
.84
.89
.44
.91
.44
.72
.81
"?S

101
102
103
104
108
104
107
108
109
110
111
112
111
114
115
114
117
110
119
120
121
122
123
124
12S
124
127
128
129
130
111
132
133
134
135
114
117
118
119
140
141
142
141
144
148
144
147
148
149
ISO
118.54
142.14
179.78
119.49
73.44
147.19
120.48
171.22
75.28
148.33
282.11
413.10
114.39
121.84
362.24
184.95
170.34
855.35
272.88
144.87
415.49
149.05
197.44
144.39
185.87
489.48
133.19
802.49
145.89
144.77
73.24
407.04
174.11
133.77
112.18
83.97
189.57
177.97
t20.ll
101.79
179.42
til. 94
til. 04
110.04
198.25
142.18
172.41
123.74
147.22 j
141.94 j
.84
.40
.77
.75
.34
.24
.43
.54
.07
.93
.03
.04
.18
.77
.93
.79
.39
.90
.25
.78
.14
.24
.15
.59
.43
.47
.01
.41
.39
.90
.92
.93
.01
.52
.40
.44
.28
.29
.44
.17
.89
.39
.81
.15
.47
.13
.89
.79
1.89
1.44
                                  A-34

-------
ATTACHMENT A.It IXPOSURE DUIATIDN (ID) AND IN6E5TION IATE  (It)  FOR
                A HYPOTHETICAL SIMPLE lANDON SAMPLE Of 500 PEOPLE
                MOM THE TARGET POPULATION
 ID
ID
II
ID
ID
XI
ID
ID
It
151
152
185
184
ISS
ISt
157
158
159
ItO
Itl
142
its
144
its
Itt
U7
ItO
It*
170
171
172
175
174
175
174
177
178
179
180
181
182
185
184
185
184
187
188
189
1*0
191
1*1
Ifl
1*4
1*5
1*8
1*7
1*8
J99
00
190.71
189.82
148.1*
140.80
814.24
95.49
174.07
119.01
ItS. 22
129.18
855.89
125.87
454.11
150.00
•18.90
419. *7
112.51
487.82
45S.14
418.85
70.77
115.04
454.19
172.28
244. t9
429.04
140.99
177.84
440.08
495. Ot
lit. 97
•08.29
It?. IS
117.45
151.75
111.02
•15.42
•18.45
197. SI
185.18
144.45
188. *8
141. IS
184.00
418.10
182.88
17*. IS
155.84
142.78 !
72.40 <
.45
.82
.1?
.05
.14
.75
.25
.17
.48
.02
.47
.22
.85
.84
.44
.84
.78
.85
.88
.7*
.42
.74
.84
.20
.10
.88
.52
.93
.80
.88
.77
.84
.48
.10
.97
.85
.59
.17
.84
.48
.19
.44
.45
.•8
.40
.91
.82
.78
!:»?
101
202
105
104
105
104
10?
108
109
110
111
112
115
114
111
Sit
117
118
119
120
121
222
223
124
125,
124
127
128
229
ISO
151
252
ISS
1S4
155
ISt
IS?
til
159
•48
•41
•42
•41
•44
•45
244
•47
•48
{49
50
102. tO
500.53
843.97
404.74
417.91
442.45
ItS. 25
144.84
77.55
•25.80
185.14
172.14
178.15
241.54
482.47
•20.72
218.50
404.58
•20. Ot
175.50
244.42
42*. 92
810.91
157. tl
120.54
194.92
253.80
445. tt
255. t9
•58. It
212.74
420.2?
415.19
419.05
119.04
410.10
1*8.42
•12.41
lit. 12
•72.80
274.48
409.85
•11.10
115.15
425.52
179.17
511.55
.52
.53
.20
.52
.55
.00
.74
.05
.92
.85
.82
.46
.85
.1?
.89
.59
.18
.79
.84
.75
.51
.It
.04
.34
.85
.21
.58
.78
.82
.80
.75
.45
.71
.54
.12
.IS
.4*
.09
.54
.14
.11
.18
.40
.11
.58
.SI
.17
111.74 .»
ill:!! I'M
tsi
152
111
II*
tss
IS*
157
ase
tst
1(0
s«i
it*
its
•44
its
It*
It?
tta
It*
170
I?l
172
I7S
174
175
I7t
177
178
271
ISO
181
182
•SS
184
185
ISt
117
188
18*
•to
III
Itl
its
1*4
t»S
1*8
1*7
IK
loo
101. Ot
SOS. 82
178.10
434.31
**.S8
ItS. 57
147.10
457.74
181. tl
145.17
154.54
414. S*
81.05
ISt.ll
121.41
190.25
457.18
•St. 72
100.71
ISS.2*
148.02
S10.59
.450. IS
IS1.7S
422.33
ISt. 88
252.25
217.10
247.fl
108.27
140.13
535.47
144.51
115.77
182.70
120.81
107.27
129.08
KS.Ot
409.41
151.94
814.78
I4S.4S
MS. 87
14*. SO
ISS. 18
140.74 .
152.71
188.48 <
185.47 0
.11
.09
.49
.tl
.1?
.44
.41
.92
.75
.57
.80
.9}
.02
.57
.IS
.42
.81
.IS
.95
.SS
.07
.44
.82
.7)
.91
.18
.90
.55
.24
.75
.47
.94
.77
.77
.15
.SB
.24
.11
.20
.51
.27
.48
.95
.IS
.84
.02
.78
.7*
.94
.42
                              A-35

-------
ATTACHMENT A.It  IXPOSUU OUftATlON CIO) AND 1NOESTION IATE  (II)
                A HYPOTHETICAL SIMPLE IANDOM SAMPLE OP  500  PIOPlE
                MOM TNC TAI6ET POPULATION
 ID    10
II
10
10
tl
10
10
II
IS1
102
101
•04
JOS
104
107
108
lOf
110
111
111
us
114
111
114
117
IIS
119
120
121
122
12 S
124
12 S
124
127
12B
129
ISO
1S1
1*2
1SS
114
1SS
ISI
1S7
ISO
•19
140.
141
•42
•41
r,t.
•41
144
•49
•48
149
SO

S9S.4S
144.82
112.91
189.78
•80.82
194.89
114.22
2*2.17
127. SS
171. 78
417.80
144.05
90.1*
180.47
140.14
194. 91
144.48
18*. 90
144.81
1*7.49
4*1.71
•41.98
145.85
147.18
•19.42
492.89
144.89
118.92
•0.74
til. 79
441.88
140.11
120.99
114.12
114.07
124.19
477.19
191.11
117. SO
470.8*
178.41
421. 4S
•••.17
.19
.19
.11
.11
.92
.89
.99
.04
.19
.OS
.SO
.17
.81
.12
.44
.12
.42
.14
.40
.91
.01
.87
.21
.44
.04
.15
.8*
.41
.00
.OS
.21
.9.
.14
.99
.OS
.41
.28
.17
.12
.72
.17
.IS
.88
19S.41 1.81
418. 8) 1-44
•44.74 t.10
177.44 4.99
•12. SS 1.12
1SS.S8 4.04
4*5. 86 1.42

SSI
152
US
•14
SSS
154
117
158
189
140
141
142
143
144
I4S
84 4
147
148
149
170
171
172
171
174
17S
174
177
178
179
180
181
182
•85
••4
•85
•84
•87
•88
••9
•90
•91
•91
89 >
•94
*9S
•94
197
198
299
00

•21.42
115.00
170.04
•91.11
100.10
108.90
119.49
99.45
197.01
197.05
•15.05
411.10
78.44
149.01
178.78
US. 21
•04.72
•58.84
401.45
408.44
12S.77
244.44
455.21
199.11
414.78
•81.40
12*4.25
124.18
415.04
•08.01
484.95
170.74
•14.94
•11.14
418.94
427.27
148.89
8*2.81
•14.8*
488.11
114.81
••4. SO
•98.81
45*. 49
•88.11
441.41
171.44
128.47
.09
.04
.85
.•4
.49
.40
.11
.44
.24
.95
.SB
.18
.89
.84
.14
.00
.09
.05
.44
.SI
.49
.88
.19
.•1
.98
.70
.44
.00
.72
.45
.15
.74
.11
.11
.07
.19
.78
.09
.99
.81
.70
.48
.•9
.08
.10
.14
.14
.44
127.20 1.15
14.14 1.99
A-36
401
402
401
404
408
404
407
408
409
410
411
412
415
414
418
414
417
410
419
420
421
422
425
424
42S
424
427
428
429
450
451
412
415
484
411
414
417
418
419
440
441
441
441
444
-445
444
447
440
449
450
107.24
100.44
472.08
100.11
•41.48
149.52
197.84
257.77
•11.45
•44.95
•51.72
119.59
155.74
101.77
•44.18
488.20
•99.44
104.04
•44.18
114.15
192.40
195.7*
112.18
•12.28
184.02
184.09
149.22
•12.82
•41.22
408.24
110.92
•12.13
•75.43
411.08
440.18
178.93
•74.17
424.88
•07.01
•19.82
•09. 45
480.02
71.85
•84.12
94.49
•02.44
•99.70
94.19
103.70
141.10 ,
.05
.95
.90
.44
.43
.04
.03
.99
.39
.44
.27
.41
.20
.51
.74
.49
.40
.41
.03
.39
.13
.SS
.73
.54
.11
.00
.73
.00
.33
.84
.55
.90
.15
.•5
.54
.54
.43
.11
.01
.44
.09
.45
.04
.47
.19
.80
.78
.94
l:ii

-------
ATTACHMENT A.li IXPOIUtE DUIATION (ED) AND IN6E5TION BATE fit)  FOt
                A HYPOTHETICAL SIMPLE BANDON SAMPLE OF 500 PEOPLE
                MOM THE TAI6ET POPULATION

 ID    ID     II
4SI
412
411
414
4S5
4S4
417
456
419
4«0
4*1
442
4ft)
444
4ft 5
4««
447
448
4ft9
470
471
472
473
474
475
474
477
478
479
410
461
462
46J
484
485
414
487
488
489
4*0
491
492
49 >
4*4
4*5
4*4
4*7
4*8
4*9
180
143.40
1)1.44
448.71
442.81
151.98
106.11
1)0.8)
147.5)
409.58
157.48
169.04
4*6. JO
•74.41
404.49
140.25
256.04
18). 19
48). 11
454.15
152.49
10*. 58
144.74
•41.))
85). 89
116.21
170.24
111.00
145.42
146.70
422.17
24). 22
140.45
4)7.45
101.48
144.43
140.**
171.78
220.78
177.90
443.75
1)9.90
4J2.94
144.82
48.84
12*. 15
1)7.94
.30
.24
.81
.85
.8)
.))
.85
.48
.99
.17
.97
.62
.78
.14
.22
.25
.27
.45
.21
.44
.99
.29
.87
.22
.57
.40
.78
.52
.44
.74
.12
.8)
.4)
.18
.87
.27
.28
.24
.21
.47
.61
.1*
.12
.84
.72
.18
814.77 1.21
4)4.08 4.47
illl', III
                                A-37

-------
ATTACHMENT A.It  DOSE (V),  EXPOSURE DUIATION ICO). AND INCEST10N
                •ATI (II)  FOR A HYPOTHETICAL S1HPIC IANDOH
                OP seo PEOPLE rion THE TARCET POPULATION
 IB    ID
1R
ID
XR









10
11
12
19
14
IS
It
17
18
1*
20
21
22
13
24
25
2*
27
26
29
10
11
12
IS
14
IS
It
17
SB
19
40
41
4»
43
44
4S
44
47
48
i!
12C.II
441. 44
•4S.18
411. IS
219.71
US. 12
149.04
181. IS
105.4*
12*. 10
•44.94
4S4.19
246.44
189.97
197.93
470.92
211.12
4S4.74
170.77
124.59
2S7.SO
418.19
217.44
198.72
112.49
104. 9S
194. 49
141.83
141.84
288. OS
284.45
141.74
411.44
14.22
141. 10
124.04
170.87
970. f 8
810.09
447.41
••9.81
•S4.I7
141.44
•18.81
144.47
485.18
104.71
258.42
.48
.17
.47
.79
.91
.42
.41
.45
.70
.70
.09
.04
.97
.85
.45
.S9
.08
.14
.08
.SO
.42
.40
.24
.85
.90
.97
.44
.98
.11
.52
.11
.04
.29
.24
.12
.87
.16
.10
.62
.16
.11
.81
.22
.11
.S4
.64
.11
.15
123.82 |. 81
487.17 S.84
2410.44
194.44
1894.55
2548.79
2912. S4
948.04
113.20
S792.94
61.74
1714.82
1474.04
14.18
465.50
29SS.47
4S9.32
444.81
1411.28
1944.45
620.63
841.49
1442.03
1847.24
1145.74
683.99
1388.27
1349.73
1874.05
454.27
1181.76
681.44
1123.49
24.91
164.03
• .44
1711.20
144.64
1724.16
1169.80
1235.82
1124.42
•616.07
1982.29
•41.17 •
215.78
1298.17
4011.81
1710.27
1101.41
1915.77
717.14
81
•2
S3
14
S3
84
87
86
89
40
41
42
43
64
45
44
47
46
• 9
70
71
72
73
74
75
74
77
76
79
60
61
82
• 3
•4
• 5
• 4
•7
•8
•9
to
tl
92
$3
94
95
94
97
98
100
114.01
•41.64
433.42
171.17
192.13
404.05
824.02
128.47
270.65
271.41
228.39
157.70
174.44
227.77
422.48
448.77
133.88
192.01
146. S3
145.54 •
209.23
147.06
209.99
123.47
68 .-84
239.64
292.97
112.22
284.70
150.41
424.11
118.80
419.42
176.92
417.76
119.18
440.80
•24.12
•7.28
•41.14
•42.49
212.68
471.13
110.14
484.61
474.11
121.16
404.13
.82 .
.01
.87
.42
.87
.97
.05
.11
.37
.74
.37
.65
.64
.84
.85
.01
.02
.16
.70
.23
.22
.14
.42
.61
.79
.45
.76
.27
.01
.26
.64
.20
.11
.27
••2
.40
.44
.94
.12
.18
.26
.19
.89
.11
.12
.04
.64
.SI
157.80 4.84
178.42 1.95
1029.58
1189.81
1230.45
1509.21
454.77
2426.0*
1727.41
2198.37
280.19
1119.49
94.13
1396.74
1123.25
1182.99
1227.11
20.33
•77.82
1099.40
748.43
97.43
1144.43
117.24
481.88
8641.90
68. 9*
257.77
2211.77
1889.22
134.45
154.51
932.71
1824.03
•291. 45
T02.49
1261.00
241.06
•619.47
1241.80
119.20
144.14
1444.80
41.14
1730.96
119.11
117.12
74.91
447.97
440.42
1470.48
1179.37
                              A-3B

-------
ATTACHMENT A.2i DOSE IV), IXPOSURE DURATION CED),  AND INGCST10N
                •ATE (ID FOR A HYPOTHETICAL SIMPLE RANDOM SAMPLE
                OF SOO PEOPLE FROM THE TARCET POPULATION
 ID
ID
IR
ID
ID
IR
101
101
103
104
10S
10*
107
108
109
110
111
lit
111
114
US
11*
117
118
119
120
111
122
123
124
its
12*
127
120
129
ISO
1S1
1S2
113
114
135
11*
117
ISO
1S9
140
1*1
142
14)
144
148
• • •
*•»•
147
140
149
ISO
191.10
l«*.17
127. 42
117.70
100.04
414.07
US. 91
10S.28
270.1*
144.02
471.19
192. 98
1*9.41
2*2.44
119.14
24*. 12
tfl.tO
ft. 97
140.09
4*2.41
107. «0
ISO. 09
101.05
49S.70
111.19
199.81
lit. 10
114.82
192.1*
120.91
S1J.SS
180.19
202.97
444.00
229.2*
197.17
00.40
401.8*
104.72
142.12
400.10
411.11
IOS.20
•42.82
141. 1*
8)1.77
111.10
424.10
.29
.10
.7*
.ts
.08
.8*
.17
.00
.71
.43
.«7
.03
.00
.00
.22
.21
.1*
.70
.02
.82
.11
.11
.72
.19
.24
.7*
.40
.40
.S3
.88
.0*
.87
.St
.23
.88
.14
.9*
.10
.27
.74
.04
.01
.4S
.00
.18
..JS
.10
.*s
!!!:»! US
1470.4*
•71.82
708.03
1841.93
tfO*. 12
til. OS
irs.08
127. «1
1*4.04
1002.71
819.77
It. 40
129. *4
1101.7*
1247.1*
141.49
140*. 70
•39.93
1140.40
2489.98
104.21
1009.10
21*7.90
839.04
1*91.01
1409.3*
101.17
90.02
•«0. 09
1282.8*
4053.12
1242.44
070.81
211.13
930.17
SO*. 00
400.1*
•92.01
111*. 07
•071.2*
1411.08
•297.01
1)78.77
•78.33
1044.41
049.42
1054.70
•179.19
i$i:ti
181
182
183
184
118
18*
187
180
1S9
1*0
1*1
1*2
1*1
1*4
1*8
l««
1*7
1*0
1*9
170
171
172
171
174
17S
17*
177
170-
179
100
101
182
103
104
US
10*
117
100
109
100
191
192
101
194
108
19*
197
190
199
00
100.14
•81. tO
1*9.94
374.75
149.11
414.42
t*0. 88
108. 98
179.84
402.82
4**. 00
77.2*
•10.11
•20.49
422.00
170.13
•72.01
8*4. 2S
144.42
•44.74
•19.79
147.19
•94.81
••4.71
204. -42.
•34.07
07.03
•89.10
•21.41
183.22
•92.80
•74.12
•41.18
144.23
•72.01
•44.14
•4*. It
00*. 04
•74.08
92.19
•19.78
104.82
•09.00
•09.01
1*0.44
410.18
124.07
•8«.tl
.01
.10
.17
.80
.11
.01
.to
.09
.11
.00
.20
.02
.1*
.10
.02
.1*
.04
.02
.87
.1*
.01
.11
.81
.17
.21
.•7
.01
.•2
.13
.09
.01
.4*
.29
.•4
.00
.10
.94
.08
.*7
.09
.10
.t4
.78
.09
.•1
.•9
.09
.08
U9.91 {.74
1.02 1.44
1905.**
1744.85
1008.72
1809.79
1121.10
1227.70
1*04.09
4092.04
• US. 47
702.43
1979.11
405.46
179S.I4
1404.14
071.49
219.80
1100.74
1174.80
011.79
•92.70
1804.82
408.4*
14J1.S8
. 794.14
1027.42
4447.10
1 129.9*
11*3.04
2*10.90
2402.15
400.12
180.9*
077.04
897.84
1944. 95
•209.2*
•221.72
•70. 45
1119.80
94.82
100*.**
192.47
400.01
•100.17
442.47
•It.lO
1101.49
1149.19
K0.74
7.22
                               A-39

-------
ATTACHMENT A.2t DOSE (V). EXPOSURE DURATION (ED).  AND
                •ATE (II) FOR A HYPOTHETICAL SIMPLE  RANDOM  SAMPLE
                OP 100 PEOPLE PROM THE TAR6ET POPULATION
 IB
ID
II
CD
IR
101
•02
10)
104
tos
•04
107
toe
109
tio
III
112
til
214
21S
Il«
117
210
lit
220
221
222
223
224
22S
224
227
220
229
210
2S1
212
211
214
2SS
214
217
110
219
240
241
242
141
244
243
24*
247
240
J«9
SO
1*2.44
277. 45
91.10
442.02
110.27
104.09
211.00
111.11
40S.20
•20. 84
209.17
144.49
227.78
401.44
44.19
S97.4S
199.70
400.10
491.40
244.10
117.18
277.11
407.47
07.71
S9S.42
477.10
125.41
S72.SO
022.04
170.08
171.01
170.02
208.47
111.10
120.14
402.04
114.47
170. SI
2)4.41
140.08
411.74
•24.14
144. IS
140.47
248.24
404.12
71.41
IIS. 71
140.90 :
400.48 1
.11
.44
.00
.47
.10
.14
.14
.17
.10
.14
.29
.22
.12
.24
.40
.10
.99
.70
.41
.95
.02
.42
.48
.21
.44
.24
.44
.14
.44
.88
.14
.04
.44
.21
.77
.49
.01
.70
.24
.79
.to
.44
.12
.11
.'« 1 -
.01
.74
.19
1.07
.00
1104.71
1404.04
414.19
1090.77
1101.10
448.04
1427.40
411.49
108.79
144.90
020.47
728. SO
489.24
1181.95
S04.0B
1801.49
089.10
2007. SI
S00.24
1021.07
274.90
047.11
080.07
I9S.29
1102.14
I40S.01
1894.11
2412.4$
181.17
701.44
1002.00
10.10
.989.09
8447.12
1410.00
194.04
1023.19
1149.70
721.43
494.02
1188.97
1209.17
114.71
194.10
•40.04
1109.41
211.00
1499.04
008.07
00.19
281
282
281
114
288
284
287
280
289
240
241
242
243
244
248
244
247
240
249
270
271
272
271
274
278
274
277
270
279
200
201
202
201
104
208
•04
207
100
209
•90
291
•92
•93
•94
198.-
•94
297
290
III
S90.74
438.27
194.00
•91.10
133.20
494.73
280.01
417.81
442.78
78.90
281.78
449.49
97.12
240.94
210.94
•82.74
247.38
412.39
•38.33
210.44
•24.34
•41.31
348.3*
107.04
•on 44
100.10
194.49
•92.27
•40.71
210.91
140.74
110.92
210.40
194.00
•11.10
•40.71
•44.10
•47.29
•29.01
•10. 10
412.21
•41.47
414.94
141.14
129.09
•40.40
114.91
•12.17
255:15 <
.28
.08
.42
.14
.12
.19
.89
.04
.40
.74
.41
.20
.74
.70
.11
.98
.05
.01
.91
.01
.24
.17
.97
.00
.87
.40
.82
.94
.94
.71
.74
.71
.01
.11
.40
.11
.07
.78
.04
.47
.11
.00
.04
.41
.07
.17
.48
.44
1.94
.04
711.04
428.48
138.41
1084.04
3401.92
1040.05
1402.27
1702.38
3411.15
149.44
1224.78
4105.41
240.21
1145.52
1184.71
1177.04
814.11
3949.05
440.11
748.41
1444.91
22.92
1745.54
20.10
191.04
091.20
2014.14
908.87
2414.12
2401.81
1739.48
474.19
898.40
142.72
101.19
1282.01
1099.21
1412.49
1712.29
117.07
2411.30
172.21
1009.79
981.92
2182.31
' 1177.11
421.28
704.74
JS50.11
1490.40
                                A-40

-------
ATTACHMENT A.2t  DOSE (V).  IXPOSUIE DURATION (CD).  AN8  1N6ESTION
                •ATE (IR)  FOR * HYPOTHETICAL SIMPLE  RANDOM  SAMPLE
                OP 100 PEOPLE PIOH THE  TAISET POPULATION
 20
ID
IR
                                   ID
ID
                                         IR
101
102
10)
10*
SOS
101
107
108
109
110
111
112
IIS
114
SIS
11*
117
lie
11*
120
121
122
12S
*24
12S
524
127
128
129
ISO
111
112
1SS
SI*
IIS
SS4
117
•11
•If
1*0
•41
1*2
Ml
144
•41
•44
•47
}"
ISO
S72.12
24*. 97
141. SS
218.9*
1«S. IB
411.49
IIS. 48
1S7.S8
454.84
1*8.75
11.17
tis.ts
1«7.«0
84. 41
110. U
128.89
181.22
SOS. 24
1*8.09
112.40
S94.*7
445. «S *
1*7.41
187.41
1SS.4S
287. «S
1*9.87
2)1.39
1*1.41
28). 82
114.22
117. 7S
241. «4
17*. 79
1*1.28
170.1*
428.*!
491.79
412.18
2)2.01
415. SI
111.00
ass. is
.4)
.5)
.71
.08
.45
.17
.24
.08
.9)
.5)
.5)
.45
.17
.74
.OS
.93
.48
.00
.52
.17
.7*
.18
.IS
.««
.10
.40
.89
.74
.94
.S4
.OS
.50
.47
.OS
.27
.19
.47
.41
.8*
.17
.*«
.ts
.11
4SS.lt 1.4}
•S9.fl 0.1S
toi.st r.7*
•74.11 l.tl
171.09 l.SS
«i:» t:!J
1159. S*
209.72
•10. 2i
1*2.11
911.9*
1111.77
2144.92
10.25
1417.**
S130.ll
84.32
1083.00
1*34.0*
420.17
14.29
S105.25
8154.47
1978. S)
105.91
•55.72
.1117. *9
•S27.77
1125.09
204.45
S49.ll
119.51
520 ..96
10*1.80
S02.S8
1799.85
420.4*
8*9.48
872. *S
780.04
tsss.s*
•sst.ss
14)4. *7
1954.07
•ISO. 88
1475.41
4204.94
1157. 1)
1175.75
1772.88
141.91
1101.57
4*47.5)
1122.10
1441.90
850.11
151
152
55)
154
155
15*
157
158
159
1*0
1*1
1*2
1«S
1*4
1*5
S**
1*7
1*8
1*9
170
171
172
17)
174
175
17*
177
178
179
180
181
182
18 S
SS4
•S5
>•*
•87
•88
189
190
•91
•92
19)
•94
•95
•9*
197
•98
HI
105.52
599.24
101.34
152. SI
121. 85
18S.S5
179.44
420.29
SS4.19
190. 43
224.80
111.91
15*. 81
282.8*
S2S.S8
15*. 07
475.01
194.84
18*. 18
171.97
•14. 89
189.57
147.24
144.12
102.48
2S7.S4
444.45
229.14
154.47
419.15
144.95
144.70
S47.S4
480.74
•10.10
SS9.S9
•55.41
SS4.SS
145.94
212.11
425. SS
177.48
.5*
.55
.1)
.18
.45
.08
.20
• 20
.52
.04
.49
.19
.92
.17
.79
.25
.94
.84
.5*
.18
.18
.44
.04
.72
.5*
.44
.9*
.17
.79
.71
.97
.55
.10
.2*
.74
.48
.97
.25
.80
.18
.40
.89
1S7.71 0.42
119.01 1.47
197.41 1.11
458.45 a.tl
191.17 4.11
•42.79 S.I8
«S:H Ml
1SSS.BB
4048.5*
105.47
87). C6
910.51
1084.55
1154.70
1851.95
2S27.10
10)2.80
1575.7*
1259.1*
19*9.2*
483.41
188.7)
1189.54
2577.04
1554.59
I42S.10
1714.55
S1S.47
757.21
1215.70
2825.99
191.54
7*S.04
2075.07
421.29
1011.45
. 1451.50
449.5*
499.92
78.18
1497.48
1948.10
•77.42
948.49
•015.70
444.29
122.80
1945.44
149.11
111.40
441.44
154.92
~S)18.)S
•518.14
1550.98
1401.40
400.00
                             A-41

-------
ATTACHMENT A.2i DOSE (V). EXPOSURE DURATION (ED). AND IN6ESTION
                •ATI (II) FQft A HYPOTHETICAL SIMPLE IANDOH SAMPLE
                OP 100 PEOPLE MOM THE TAI6ET POPULATION
 If
ID
II
ID
ID
II
401
402
40)
404
405
40*
407
400
409
410
411
412
411
414
415
41*
417
410
419
420
421
422
423
424
425
42*
427
420
429
410
411
412
411
414
415
41*
417
410
4>9
440
441
441
44)
444
440
44*
447
440
449
400
110.05
4*1.00
1*1.04
1*2.00
•11.01
•90.0*
•50.04
1*1.**
•10.44
1*0.40
141.10
1*7.4$
•10.92
4*1.91
19*. 10
114.41
190.17
114.90
•09.19
140.79
411.11
144.14
490.79
124. *5
470.10
129.00
151.0)
107.9*
450.47
114.9)
114.01
402.1*
421.07
411.11
491. *4
•40.04
144.00
101.01
110.11
•10.97
107. *9
•4.91
115.02
491.04
•71.0*
•40.09 ;
•54.00
415.04

.1*
.12
.OS
.79
.12
.21
.0*
.17
.00
.10
.11
.«)
.91
«*1
.92
.07
.92
.41
.19
.•2
.0)
.04
.04
.02
.99
.47
.12
.1)
.)7
.74
.01
.10
.44
.4*
.0*
.90
.42
.0)
.79
.05
.•5
.71
.14
.25
.02
.42
.•9
• *0
.40
.75
179.25
1*72. *2
*4*.13
1102.09
740.42
•4*4.11
0711.1*
795.11
11*7.71
544.11
78.45
514.21
•22.10
1190.9)
«040.18
•04.1)
010.10
10*. 10
120.72
794.17
•109.07
904.09
4120.40
87*. 08
•111.51
•90.11
*!.!«
•10.05
211.11
105). 71
744.70
1085.25
1947.19
•911.10
499.02
•102.81
710.44
417.12
•47.41
1021.90
792.91
•12.17
•81.40
1115.21
1*8.97
10*0.00
•010.09
•00.07

451
452
45)
454
455
454
407
450
459
4*0
4*1
4*2
4*1
4*4
4*5
444
4*7
4*0
4*9
470
471
472
471
474
475
47*
477
470
479
400
401
402
40)
404
405
40*
407
400
409
490
404.48
122.02
10S.90
451.45
490.01
•10.12
•10.10
194.00
•01.44
110.09
•41.11
145.71
450.11
417.10
402.41
151.15
124.19
407.89
111.02
188.5*
117.49
171.11
177.40
79.94
149*72
114.02
477.41
157.01
1)0.1*
41*. 17
192.00
•19.9*
1*1.29
125.7*
171.41
•09.00
149.01
1*4.77
•07.05
•21.11
"491^448.2* '1
492
491
494
490
494
497
490
too
•90.04
194.00
199.01
.24
.52
.41
.27
.00
.12
.01
.00
.07
.*!
.70
.*1
.12
.1*
.to
.9*
.10
.17
.47
.0*
.«*
.21
.74
.11
.07
.25
.04
.05
.27
.10
.19
.25
.05
.42
.07
.05
.14
.11
.94
.11
.11
.11
.00
.•9
•01.01 4.07
1*2.74-8.40
112.0) 1.-94
172.19 4.00
ill. 20 4.90
14.95 4.94
1*7.21
422.8*
958. ))
1494.14
•002.01
451.19
0)8.8*
1114.80
011.77
1202.4*
1044.41
1)43.13
08). 91
17)1.13
125.08
71*. 41
1551.7)
112.47
1)74.**
4114.09
15). 42
191.48
1522.42
171.08
924.12
1094.14
1151.10
1121.83
711.5*
1842.41
1210.12
11*2.79
17). 01
1421.20
127*. 43
814.05
044.27
17)1.97
002.42
120.2)
' -1797.12
1952. OS
140*. OS
'•50.7*
•01.00
241.11
449.09
10*4.**
K87.70
«*.!«
                             JU42

-------