CALCULATION OF THE FINAL ACUTE VALUE  FOR

 WATER QUALITY CRITERIA FOR AQUATIC  ORGANISMS
              RusselL J.  Erickson

Center for Lake Superior  Environmental Studies

       University of Wisconsin-Superior

          Superior,  Wisconsin  54880
              Charles E.  Stephan

     U.  S,  Environmental  Protection Agency

   Environmental Research Laboratory-Duluth

            6201 Congdon  Boulevard

           DuluCh, Minnesota  55804
   ENVIRONMENTAL RESEARCH LABORATORY-DULUTH
      OFFICE OF RESEARCH AND DEVELOPMENT
     U.S.  ENVIRONMENTAL PROTECTION AGENCY
           DULUTH,  MINNESOTA  55804'

-------
                                 DISCLAIMER




     This report has been reviewed by Che Environmental Research




Laboratory-Duluth, U.S. Environmental Protection Agency, and approved for




publication.  Mention of trade names or commercial products does not




constitute endorsement or recommendation for use.
                             AVAILABILITY NOTICE




     This document is available to the public through Che National Technical




Information Service (NTIS), Springfield, Virginia  22161,

-------
                                  ABSTRACT
      The  Final  Acute Value  (FAV)  for a material, which is an integral part  of
 ehe  procedure for  deriving  water  quality criteria  for aquatic organisms,  LS
 an estimate  of  th« fifth  percentile of a statistical population represented
 by the  set of Mean Acute  Values (MAV) available for the material, a MAV being
 the  concentration  of the  material that causes a specified level of acute
 toxicity  to  aquatic organisms  in  some taxonotnic group.  A new procedure for
 calculating  FAVs has been developed under the assumption that sets of MAVs
 are  random samples of  such  populations.  Based on  examination of available
 sets of MAVs,  it was inferred  that FAV estination  would be best served by
 assuming  that  the  populations  have a log triangular distribution.  Also,
 because this  or any other assumption will"likely not completely hold over  the
 entire  range of data in all  sets, it was judged that FAV estimation  should be
 based on  a subset  of the  data  near the fifth percentile.  Based on
 simulations,  it was determined that a FAV for a set of MAVs  would be best
 calculated by (a)  assigning  each  MAV a cumulative  probability P^*R/(N+1)
 (R»rank,  N»nuraber  of MAVs in the  set), (b)  fitting a line to Ln(MAV) versus
/Ffl  using the  four  points with P^ nearest 0,05 and using the geometric
mean functional relationship to estimate slope, and (c) calculating  the FAV
 as the concentration corresponding to Pg*0,Q5 on this  line.  Major
modifications of this  new procedure were found to  result either  in only minor
 changes in FAVs or  in  FAVs  at  variance with  the data.  The  old  procedure for
calculation of FAVs was judged to have some  theoretical and  practical
 shortcomings  that  make it less desirable than the  new  procedure,  but  FAVs  by
 the  two procedures were generally similar.   A procedure based on  extreme
deviation from random  sampling generally did not produce greatly  different
FAVs.
                                     i ii

-------
                                  CONTENTS









Abstract	•	...... 111




Tables   	   v




Figures  	   v




Acknowledgments  , »	vi








Introduction	    1




Conclusions	13




Statement of Problem	   16




Description of Percentile Estimation Methods   	  19




Selection of Distribution 	   26




Selection of Percentile Estimation Method  and  Subset  Size  .  	  .   33




Application of Rscomnended Procedure and Alternatives to Data Sets   .  .   39




Discussion	   50









References	   53




Appendix I	   54
                                     IV

-------
                                     TABLES
Number                                                                   Pa|e



  1.   Example Sets of Species Mean Acute Values	   4




  2.   Example Sets of Family Mean Acute Values	9




  3.   Average Goodness-of-Fits for In(SMAV) Data Sets	32




  4.   Average Goodness-of-Fits for In(FMAV) Data Sets	32




  5.   Means and Standard Deviations of Estimates of Fifth Percentile




         (K<^~) by Various Methods, for 10,000 Samples from a Standard




         Triangular Distribution  	  35




  6.   Mean True Cumulative Probabilities (£5) of Estimates of Fifth




         Percentile (PCx^)) by Various Methods, for 10,000 Samples




         from a Standard Triangular Distribution  ,	36




  7.   FAVs Calculated from SMAV Data Sets by Old Procedure, Recommended




         New Procedure, and Various Modifications of Recommended New




         Procedure	40




  8.   FAVs Calculated from FMAV Data Sets by Old Procedure, Recommended  .




         New Procedure, and Various Modifications of Recommended New




         Procedure	42
                                     FIGURES









  1,    Standard Probability Density Plots for Rectangular,




         Triangular,  Normal, and Biexponential Distributions   ,,..,.  30

-------
                               ACKNOWLEDGMENTS




     Charles Norwood helped design Chis project  and  Robert W.  Andrew




performed some of the calculations.
                                     VI

-------
                                  INTRODUCTION




      On  November 28, L980, the U.S. Environmental Protection Agency published




"Guidelines  for Deriving Water Quality Criteria for the Protection of Aquatic




Life  and  Its Uses" as Appendix B of an announcement of the availability of water




quality  criteria documents (1).  Calculation of the Final Acute Value (FAV) is




an important part of the process described in these Guidelines.  A FAV is a




concentration of a material derived from an appropriate set of Mean Acute Values




(MAVs),  a MAV being the concentration of the material that causes a specified




level of  acute toxieity to an aquatic taxon in laboratory tests.  The FAV is




defined  to be lower than all except a small fraction of the MAVs that are




available for the material.  The fraction was set at 0.05 (i.e., the FAV  lies at




the fifth percentile of the MAVs) because other fractions resulted in FAVs that




were  deemed  too high or too low  in comparison with the sets of MAVs  from  which




they  were obtained.  However, if the set contains a MAV for an important  species




that  is  lower than the calculated FAV, the FAV is set equal to that  MAV.




      In order to be useful, the  procedure for obtaining a FAV  from a set  of MAVs




must  be  objective so that different parties will obtain the same FAV from  a set




of MAVs.  The development of a reasonable mathematical framework for FAV




calculation was therefore necessary.  In addition, it is desirable that  the




rationale for Ch« calculation procedure be relatively easy  to  understand  and




that  the computations' be as simple as possible.  Section IV.1-0 of the




Guidelines described a procedure for calculating a FAV from a  suitable  set of




Species Mean Acute Values (SMAVs).  Because of criticism of this  procedure,  this




project was initiated to define  the general problem of calculating a FAV,  to




evaluate alternative procedures, and to recommend the most  appropriate
                                       1

-------
 procedure-   This  project was not  intended to evaluate che definition of the FAV




 or the procedures for  obtaining MAVs.




      Development  of  an  appropriate  procedure for calculating FAVs requires  che




 availability of typical sets of SMAVs.   Some of the water quality criteria




 documents  (1)  contain  such  sets in  Table 3 of the section on Aquacic Life




 Toxicology.   Twenty data sets  for freshwater species and seventeen for saltwater




 species  were considered to  be  acceptable for the purposes of this project




 because  they contained  SMAVs from at least eight families in a variety of




 taxonowic  and  functional group*.  Th«§« data s«tt (TabIt I) contain froo 8  co 45




 SMAVs  for  a  variety of  organic and  inorganic materials.   Because ail  acceptable




 sets  of  SMAVs  (that w*re available  at  the completion of  this  project  in  May,




 1982)  were used and because th*y  include a diversity of  species  and materials,




 this  group of  37  data  sets  should be repr«s«nt*tive of the data  sees  from which




 FAVs will be calculated.




      There is  son* concern  that FAVs would be more appropriately bated on  a




 taxonomic level higher  than species  (e.g., family).  Statistical analysis of




 data  sets similar to those  in Table  1 has shown that differences between




 families are usually greater, often by an order of magnitude or  more,  than




 average  differences within  families  (2).  Therefore, if  a set of SMAVs has a




disproportionate number of  species  from a sensitive or insensitive  family, che




 FAV might bs undesirably affected.   For example, of che  29 SMAVs  for  zinc  in




 fresh water,  six  are from Salmonidae and are all among the twelve  lowest SMAVs.




 Resolution of which taxonomic  level  is most appropriate  is not  of  concern  here,




but because  the definition of tht FAV might be so modified, Family  Mean  Acute




Values (FMAVs), the geometric mean  of  all the SMAVs available  for  a family, were




computed for  all data sets  in Table  1  and are reported in Table  2.  Subsequent



                                       2

-------
analysis will consider how the use of these two different taxonoraic levels might




affect  recommendations about the procedure for calculating a FAV.  This does




riot, however, constitute an endorsement of either species or family as che most




appropriate taxonoraic level.




     This report will first define the problem of FAV calculation and then




discuss the general methods available for estimating percentiles.  Next, the




example data sets  in Tables 1 and 2 will be examined to determine an  appropriate




statistical distribution to use in the FAV calculation procedure.  Simulated




samples from the selected statistical distribution will then be  used  to




determine the procedure most appropriate for calculation of the  FAV.   Finally,




the procedure selected will be applied to the example data sets  and the




significance of deviations from various assumption! of the procedure  will be




evaluated,

-------
TABLE I.  EXAMPLE  SETS OF  SPECIES  MEAN  ACUTE  VALUES.'
COPPER
(FRESHWATER)
Rank SMAV
45
44
43
42
4!
40
39
38
37
36
35
34
33
32
31
30
29
28
27
26
25
24
23
22
21
20
19
18
17
16
15
14
13
12
11
10
9
8
7
6
5
4
3
2
1
260.
150.
148.
145.
117.
91.8
47.9
46.5
35.2
23.1
22.9
21 .8
20.1
18.9
14,4
10.1
8.4!
5.81
5.37
5.00
4.95
3.97
3.29
2.80
2.28
2.20
2.20
2.13
2.13
2.12
1 .99
1.83
1.68
1.42
1.34
1.23
1 .07
1.02
0.91
0.91
0.76
0.55
0.43
0.28
0.23
DOT
(FRESHWATER)
Rank SMAV
42
41
40
39
38
37
36
35
34
33
32
31
30
29
28
27
26
25
24
23
22
2!
20
19
18
17
16
15
14
13
12
11
10
9
8
7
6
5
4
3
2
!



1230.
362.
192.
175.
68.
67.
54.
48.
48.
40.
33.
25.
18.
17.
14.
12.
10.
9.3
8.5
8.0
7.3
7.8
7.3
5.0
4.9
4.3
4.0
3.9
3.5
3.2
3.0
3.0
2.6
2.4
.9
.9
.7
.7
.6
.4
1.1
0.36



CADMIUM
(SALTWATER)
Rank SMAV
31
30
29
28
27
26
25
24
23
22
21
20
19
18
17
16
15
14
13
12
11
10
9
8
7
6
5
4
3
2
1














50600.
50000.
21200.
21000.
19200.
12200.
10100.
6600.
5290.
4100.
3940.
3800.
3500.
3440.
2930.
2590.
2410.
1800.
1710.
1670.
1480.
1220.
1080.
760.
645.
320.
169.
144.
135.
78.
41.3














CADMIUM
(FRESHWATER)
Rank SMAV
29
28
27
26
25
24
23
22
21
20
19
18
17
16
15
14
13
12
11
10
9
8
7
6
5
4
3
2
1
















138.
135.
134.
133.
125.
91.4
86.7
80.7
55.9
54.7
47.0
38.2
35.9
30.3
28.0
22.3
19.7
12.2
7.01
3.57
2.87
1.67
1 .1?
0,87
0.29
0.09
0.04
0.03
0.02
















TOXAPHENE
(FRESHWATER)
Rank SMA¥
29
28
27
26
25
24
23
22
21
20
19
18
17
16
15
14
13
12
11
10
9
8
7
6
5
4
- 3
2
1
















180.
28,
26.
24.
20.
15.
14.
14.
14.
13.
13.
12.
11.
10.
9.8
9.2
8.7
6.3
6.
4.2
4.1
4.
3.
3.
3.
2.5
2.3
2.
1.3

















-------
\BL£ 1.  Continysd
ZINC
(FRESHWATER)
Rank SMAV
29
28
27
26
23
24
23
22
2!
20
19
18
17
16
15
14
13
12
1!
10
9
a
7
6
5
4
3
2
1
2260.
1019.
732.
716.
708.
699.
331.
524.
413.
567.
315.
293.
285.
255.
172.
169.
92.8
82.6
81.4
64.9
57.9
37.6
49.3
42.0
26.2
23.1
21.2
9,09
9.89
ENORIN
(FRESHWATER)
Rank SMAV
28
27
26
23
24
23
22
21
20
19
18
17
16
15
14
13
12
11
10
9
8
7
6
5
4
3
2
!

352.
64.
60.
34.
32.
5.9
4.7
3.1
2.1
1.8
1.5
1.3
1.2
1.1
1.0
0.85
0.78
0.76
0.75
0.69
0.54
0.47
0.46
0.44
0.41
0.33
0.32
0.15

MERCURY
(SALTWATER)
Rank SMAV
26
25
24
23
22
21
20
19
18
17
16
15
14
13
12
11
10
9
8
7
6
5
4
3
2
1



1680.
1260.
400.
315.
230.
223.
158.
116.
98.
98.
89.
84.
79.
70.
60.
50.
17.
14.
14.
14.
10.
7.6
6.6
5.6
4.8
3,5



ZINC
(SALTWATER)
Rank SMAV
24
23
22
21
20
19
18
17
16
15
14
13
12
11
10
9
8
7
6
5
4
3
2
1





70600.
50000.
39000,
24600.
9460.
8100.
6330.
4090.
3640.
3380.
2440.
2160.
1780.
1450.
1270.
1000.
950.
591.
498.
400.
321.
310.
290.
166.





UNDANE
(FRESHWATER)
Ran*. SMAV
22
21
20
19
18
17
16
15
14
13
12
It
10
9
8
7
6
5
4
3
2
1







676.
485.
460.
207.
141 .!
138.
90,
83.
68.
67.1
64.
55.6
49.
45.
44.
44.
40.
32.
32.
10.5
10.
2.








-------
TABLE 1.   Contlnuad
COPPER
( SALTWATER )
Rank
22
21
20
19
18
17
16
15
14
13
12
1 1
10
9
8
7
6
5
4
3
2
1
SMAV
600.
560.
526.
487.
412.
364.
330.
181.
141.
138.
136.
129.
128.
124.
120.
86.
69.
52.
50.
39.
31.
28.
HEPTACHLOR
(SALTWATER)
Rank
19
18
17
io
15
14
13
12
11
10
9
a
7
6
5
4
3
2
1
SMAV
194.
138.
112.
is m
J J t
50.
32.
14.5
10.
8.
6.22
3.77
3.4
3.
3.
1.5
1.06
0.36
0.8
0.057
NICKEL
(FRESHWATER)
Rank
22
21
20
19
13
17
16
15
14
13
12
11
10
9
8
7
6
5
4
3
2
1
SMAV
2230.
2030.
1540.
1080.
1010.
730.
720.
665.
659.
627.
509.
507.
457.
440.
440.
401.
388.
302.
234.
208.
78.5
54.0
0 1 ELDS 1 N
(FRESHWATER)
Rank
19
18
17
16
15
14
13
12
11
10
9
8
7
6
5
4
3
2
\
SMAV
740.
620.
567.
250,
213.
130.
41.
39.
24. •
22»
20.
15.
10.8
8.1
8.
6.1
5.0
4.5
2.5
0 1 ELOR I N
(SALTWATER)
Rank
21
20
19
18
17
16
15
14
13
12
11
10
9
a
7
6
5
4
3
2
1

SMAV
50.0
34.0
31.2
23.0
19.7
18.0
14.2
10.8
10.0
8.9
S.6
7.0
6.0
5.0
5.0
4.5
3.5
2.3
1.5
0.9
0.7

LINQANE
(SALTWATER)
Rank
19
ia
17
16
15
14
13
12
11
10
9
a
7
6
5
4
3
2
1
SMAV
3680.
450.
103.9
66.0
60.0
56.0
47.
35.0
30.6
28.0
14.0
10.0
9,0
7.3
6.28
5.0
5.0
4.44
0.17
ALDRIN
(FRESHWATER!
RanK
21
20
19
18
17
16
15
14
13
12
It
10
9
8
7
6
5
4
3
2
1

SMAV
19000.
4900.
180.
143.
50.
45.9
42.
34.
32.
28.
27.
27.
21.
16.
10.
9.
8.
7.4
6.1
4.5
4.

CHROMIUM(VI)
(SALTWATER)
Rank
19
18
17
16
15
14
13
12
11
10
9
8
7
6
5
4
3
2
1
SMAV
105000.
93000.
91000.
57000.
32000.
30500.
22000.
17200.
15000.
10000.
7500.
6600.
6300.
4400,
4300.
3650.
3100.
2000.
2000.
ENDS IN
(SALTWATER)
Rank
2!
20
19
13
17
16
15
14
13
12
1 1
10
9
a
7
6
5
4
3
2
1

SMAV
14.2
12.
3.1
1.8
1 .7
1.2
1 .1
0.95
0.65
0.63
0.6
0.36
0.31
0.3
0.3
0.28
0.1
0.094^
0.05
0.048
0.037

CHROMIUM* M 1)
(FRESHWATER)
Rank
18
17
16
15
14
13
12
1!
10
9
3
7
6
5
4
3
2
1

SHAV
1075.
728.
633,
233.
224,
224.
224.
191 .
191 .
139.
161 .
138.
136.
132.
123.
113.
47.
33.4


-------
3LE 1.  Continued
HEPTACHLOR
(FRESHWATER)
Rank
13
17
16
15
14
13
12
t!
10
9
8
1
6
5
4
3
2
1
SM*V
320.
148.
101.
81 .9
78.
61 .3
47.3
42.
29.
26.
24.
23.6
13.1
7.8
2.3
1.8
1 ,1
0.9
TOXAPHENE
(SALTWATER)
Rank
14
13
12
11
10
9
a
/
6
5
4
3
2
1
SMAV
1120.
324.
43.8
21.
16.
8.2
4.5
* ^
4.4
1.4
1.1
1.1
0.5
0.11
NICKEL
(SALTWATER)
Rank
17
16
15
14
13
12
11
10
9
8
7
6
5
4
3
2
1

SMAV
350000.
320000 .
150000.
49000.
47000.
25000.
17000.
9670.
7960.
6360.
2030.
1180. -
634.
600.
508.
310.
152.

CHROMlUM(Vi)
(FRESHWATER)
Rank
!4
13
12
11
10
9
8
7
6
5
4
3
2
1
SM*¥
195000.
134000.
120000.
69000.
59900.
59000.
45100.
304 QQ.
30000.
25000.
6800.
6400.
3100.
'67.
DOT
ALORIN
(SALTWATER)
Rank
17
16
15
14
13
12
11
10
9
8
7
6
5
4
3
2
1

SMAV
89.
7,9
7,0
6.0
4.0
3.9
2.0
1.8
1.6
1.1
1.0
0.68
0.6
0.53
0.4
0.38
0.14

CHLORDANE
(FRESHWATER)
Rank
14
13
12
11
10
9
3
7
6
5
4
3
2
1
SMAV
190.
82.
59.
58.
57,
56.
45.
40.
37.
26.
25.
15.
6.3
3.
(SALTWATER)
Rank
16
15
14
13
12
11
10
9
8
7
6
5
4
3
2
1


SMAV
100.0
36.0
33.0
33.0
25.0
17.0
13.0
12.0
9.0
8.0
7.2
6.0
5.6
5.0
4.1
1,5


SELEN 1 UM
(FRESHWATER)
Rank
13
12
11
10
9
8
7
6
5
4
3
2
1

SMAV
42400.
28500.
26100.
24100.
13600.
12600.
10200.
9000.
6500.
3870.
1460.
710.
340.

CYANIDE
(FRESHWATER!
Rank
15
14
13
12
1 1
10
9
8
7
6
5
4
3
2
1



SMAV
2326.
2240.
639.
431.
318.
167.
147.
137.
125.
125.
103.
102.
102.
83.
57.



SELENIUM
(SALTWATER)
Rank
13
12
11
10
9
3
7
6
5
4
3
2
1

SMAV
17348.
14651 .
9725.
7400.
4600.
4400.
3497,
1 740.
1200.
1040.
300.
600.
599.


-------
 TABLE  1.   Continued
ENOOSULFAN
(SALTWATER)
Rank
12
11
10
9
8
7
6
5
4
3
2
!
SMAV
730.
157.
7.6
1.31
0.83
0.76
0.38
0.30
0.14
o.to
0.09
0.04
ENOOSULFAN
(FRESHWATER)
Rank
10
9
8
7
6
5
4
3
2
1
SMAV
261.
88.
6.0
5.3
3.8
3.7
3.2
2.3
0.83
0.34
ARSENIC f€RCURYb SILVER SILVER
(FRESHWATER) (FRESHWATER) (FRESHWATER) (SALTWATER)
Rank
12
11
10
9
8
7
6
5
4
3
2
!
SMAV Rank SMAV Rank SMAV Rank SHAV
41760. 11 2000. 10 5.77 10 1400.
23130. 10 2000. 9 5,52 9 550.
26042. 9 2000. 8 4.11 8 500.
22040. 8 1000. 7 0.112 7 250.
18096. 7 784. 6 0.0230 6 210,
15660. 6 249. 5 0.015 5 36.
14964. 5 240. 4 0.014 4 33.
13340. 4 50. 3 0.0123 3 21.
5278. 3 20. 2 0.0121 2 20.
1348. 2 10. 1 0.00192 1 4.7
879. 1 5.
812. .
CHLORDANE
(SALTWATER)
Rank
&
7
6
5
4
3
2
1


SMAV
120.
17.5
16.9
11.8
6.4
6.2
4.8
0.4


8 TsHsn from Table 3  in the "Aquatic Life Toxicology"  sections of  the  water  quality  criteria  documents  (1).
  For the purposes of this project, the Species Maan Acut»  Intercepts  for  several  of  tne metais  in  frash  »o
  ware considered to  be Species Mean Acute Values.  All SMAYs are  In
13 Tha acute value for F axone 1 1 a c I y peat a should  have been  published originally  as  20  yg/L,  not 0.02  ^g/L  (3).

-------
TABLE 2.  EXAMPLE  SETS  CF  FAMILY MEAN ACUTE VALUES.8
CADMIUM
(SALTWATER)
Rank
25
24
23
22
21
20
19
18
17
16
15
14
13
12
11
10
9
8
7
6
5
4
3
2
1
FMAV
37600.
21200.
19200.
1 1100.
6600.
5290.
3940.
3800.
3500.
3440.
3260.
2930.
2410.
1800.
1 710.
1670.
1480.
1220.
1080.
760.
645.
320.
156.
78.
75.
CADMIUM
(FRESHWATER)
Rank
18
17
16
15
14
13
12
11
10
9
8
7
6
5
4
3
2
1
FMAV
138.
133.
36.7
85.9
55.9
54.8
30.3
28.5
28.0
19.7
12.2
8.86
7.01
2.87
1.58
1.15
0.50
0.048
COPPER
(FRESHWATER)
Rank
23
22
21
20
19
18
17
16
15
14
13
12
11
10
9
a
7
6
5
4
3
2
1


FMAV
260.
150.
145.
117.
46.5
45.3
38.7
35.2
22.9
14.4
10.0
3.86
3.58
3.56
2.28
2.13
2.12
1.73
1 .42
1.34
0.99
0.76
0.30


ENDRIN
(FRESHWATER)
Rank
17
16
15
14
13
12
11
10
9
a
7
6
5
4
3
2
1

FMAV
109.
64.
60.
32.
4.7
4.3
1..80
1.50
1.30
1.0
0.95
0.85
0.66
0.65
0.49
0.48
0.44

MERCURY
(SALTWATER)
Rank
23
22
21
20
19
18
17
16
15
14
13
12
11
10
9
8
7
6
5
4
3
2
1


FMAV
1680.
1260.
400.
315.
230.
223.
158.
116.
98.
89.
84.
83.
79.
60.
50.
17.
14.
14.
12.
6.6
6.5
4.8
3.5


COPPER
(SALTWATER)
Rank
17
16
15
14
13
12
11
10
9
8
7
6
5
4
3
2
1

FMAV
600.
526.
487.
412.
330.
268.
212.
160.
138.
136.
129.
120.
69.
66.
40.
39.
28.

DOT
(FRESHWATER)
Rank
20
19
18
17
16
15
14
13
12
11
10
9
a
7
6
5
4
3
2
1





FMAV
1230.
92.
67.
54.
36.
33.
32.
25.
19.
17.5
10.
7.0
4.1
4.0
3.2
2.4
2.3
1.7
1.6
1.3





CHROMIUM(VI)
(SALTWATER)
Rank
17
16
15
14
13
12
11
10
9
8
7
6
5
4
3
2
1

FMAV
105000.
93000.
91000.
57000.
32000.
30500.
22000.
17200.
1 5000 .
10000.
7500.
6600.
6300.
4300.
3650.
2970.
2490.

Zl
NC
(SALTWATER)
Rank
20
19
18
17
16
15
14
13
12
It
10
9
8
7
6
5
4
3
2
1





FMAV
70600.
50000.
39000.
9460.
6330.
6330.
4090.
3640.
3380.
2440.
2160.
1 780.
1450.
1000.
543.
525.
400.
321.
310.
166.





0 1 ELDR 1 N
(SALTWATER)
. Rank
16
15
14
13
12
11
10
9
8
7
6
5
4
3
2
1


FMAV
34.0
31.2
23.0
19.7
13.0
16.7
14.2
7.6
7.0
6.0
5.0
4.5
2.3
1.5
0.9
0.7



-------
TABLE 2.   Continued
ENDRIN
(SALTWATER)
Ran*
16
15
14
13
12
1 1
10
9
a
7
6
5
4
3
2
1
FMAV
14.2
12.
3.1
1.7
1.1
1.1
0.63
0.6
0.47
0.3
0.29
0.1
0.094
0.05
0.048
0.037
ALDRIN
(FRESHWATER)
Rank
14
13
12
11
to
9
a
7
6
5
4
3
2
1
FMAV
9650.
180.
143.
50.
27.5
27.
21.
20.
16.
Jo-
lt.
9.
3.
7.4
HEPTACHLOR
(SALTWATER)
Rank
16
15
14
13
12
1 1
10
9
8
7
6
5
4
3
2
t
DOT
FMAV
194.
188.
112.
55.
21.5
10.
a.
3.92
3.77
3.4
3.
3.
1.5
0.86
0.8
0.057
LINDANE
(SALTWATER)
Rank
16
15
14
13
12
1 1
10
9
8
7
6
5
4
3
2
1
FMAV
3680.
450.
66.0
56.0
55.9
47.
35.0
30.6
14.0
9.0
7.3
6.66
6.28
5.0
5.0
0.17
NICKEL
(SALTWATER)
Rank
14
13
12
11
10
9
a
7
6
J
4
3
2
1
FMAV
39.
7.9
7.0
6.0
4.0
2.0
1.6
1.4
0.87
0.&8
0.6
0.53
0.4
0.14
(SALTWATER)
Rank
14
13
12
11
10
9
a
7
6
;
4
3
2
1
FMAV
350000.
320000.
150000.
47000.
35000.
1 7000.
9670.
7960.
6360.
2080,
1180.
600.
366.
310.
NICKEL
(FRESHWATER)
Rank
16
15
14
13
12
11
10
9
a
7
6
5
4
3
2
1
FMAV
2230.
2030.
1540.
1080.
730.
720.
665.
627.
609.
457.
446.
440.
401 .
345.
234.
65.1
CHROMIUM< 111)
(FRESHWATER)
Rank
13
12
11
10
9
8
7
6
5
4
3
2
1

FMAV
885.
633.
224.
224.
211.
207.
153.
138.
136.
132.
123.
47.
33.4

ZINC
(FRESHWATER)
RanK
15
14
15
12
11
10
9
8
7
6
5
4
3
2
I

FH*¥
2260.
1019.
716.
708.
531 .
463.
315.
251 .
213.
161 .
136.
92.8
48.8
42.0
13.7

ALORIN
(SALTWATER)
Rank
13
12
11
10
9
8
7
6
5
4
3
2
1

FMAV
100.0
36.0
33.0
33.0
25.0
13.0
12.0
9.8
8.0
7.2
5.0
5.0
5.7

                                                    10

-------
'ABLE  2.   Continued
TOXAPHENE
(SALTWATER)
Rank
13
12
11
10
9
8
7
6
5
4
3
2
1
FMAV
1 120.
824.
43.8
16.
9.6
3.2
4.5
4.4
1.4
1.1
1.1
0.5
0.11
HEPTACMLCfi
(FRESHWATER)
Rank
10
9
3
7
6
5
4
3
2
1
FMAV
180.
148.
58.6
37.0
29.5
24.8
7.8
2.8
1,8
1.0
SILVER
/ ^ » 1 Tt J
\ J/-IU ( *
Rank
10
9
8
7
6
5
4
3
2
1
A -ren %
FMAV
1400.
550.
500.
250.
210.
36.
33.
21.
20.
4.7
DIELDRIN
(FRESHWATER)
RanK
12
11
10
9
8
7
6
5
4
3
2
1

FMAV
740.
593,
191.
39.
30.
24.
20.
11.
8.
5.5
5.0
4.5

CHROMIUM(VI >
(FRESHWATER)
Rank
10
9
3
7
6
5
4
3
2
1
S
FMAV
1 62000.
71900.
63800.
59900.
30400.
30000.
25000.
6400.
4600.
67.
ILVER
K r Rc.ormn i en f
Rank
9
8
7
6
5
4
3
2
1

FMAV
5.77
5.52
4.11
0.112
0.0230
0.015
0.013
0.0123
0.00192

TOXAPHENE
(FRESHWATER)
Rank
12
11
10
9
3
7
6
5
4
3
2
1

FMAV
180.
28.
21.
20.
13.
12.0
8.0
5.8
4.7
3.5
2.6
1.3

SELENIUM
(FRESHWATER)
RanK
10
9
8
7
6
5
4
3
2
1
FMAV
42400.
28500.
24100.
13600.
12600.
9580.
6500.
6170.
1660.
340.
MERCURY
( FRESHWATER )
Rank
9
a
7
6
5
4
3
2
1

FMAV
2000.
2000.
2000.
1000.
784.
244.
32.
10.
5,

SELEN 1 UM
(SALTWATER)
RanK
12
11
10
9
8
7
6
5
4
3
2
1

FMAV
17348.
14651.
9725.
7400.
4600.
4400.
3497.
1200.
1180.
1040.
600.
599.

LINOANE
(FRESHWATER)
RanK
10
9
8
7
6
5
4
3
2
1
FMAV
532.
207.
138.
94.8
68.
53.1
52.9
22.4
22.
10.
ENOOSULFAN
(FRESHWATER)
Rank
9
8
7
6
5
4
3
2
1

FMAV
261.
38.
5.9
3.8
3.7
3.2
2.3
0.83
0.34

ENOOSULFAN
(SALTWATER)
RanK FHAV
11 730.
10 157.
9 3.16
3 0.33
7 0.76
6 0.38
5 0.30
4 0.14
3 0.10
2 0.09
1 0.04


CYANIDE
(FRESHWATER)
Rank FMAV
10 2326.
9 2240.
8 431 .
7 306.
6 199.
5 167.
4 125.
3 118.
2 83.
1 77.
ARSENIC
(FRESHWATER)
Rank FMAV
8 41760
7 29130.
6 22040.
5 20190.
4 13096,
3 1 4 1 30 .
2 1794.
1 379.


                                                       11

-------
TABLE 2.  Continued
CHLOROANE
(FRESHWATER)
Rank
8
7
6
5
4
3
2
1
FMAV
190.
39.
58.
44.
32.
21.
15.
6.3
CHLOROANE
(SALTWATER)
Rank
3
7
6
5
4
3
2
!
FMAV
120.
17.5
16.9
11.8
6.4
6.2
4.8
0.4
a Calculated from the Sp«c!«s M§an Acuta Values In Table 1.  All FMAVs ar«  In \ig/L.
                                                        12

-------
                                  CONCLUSIONS




 1.   Calculation of a FAV from a typically small set of MAVs  requires  chat  che




     set be considered a sample from a statistical population and  chat  che  FAV




     be considered an estimate of the fifth percentile of that  population.




2,   The set of MAVs must be assumed to have been obtained from the statistical




     population by a specific sampling procedure; of reasonably simple sampling




     procedures, an assumption of random sampling appears most  consistent with




     actual data selection,




3.   Available sets of MAVs suggest that the statistical populations are highly




     positively skewed and that estimation would be benefitted  by  logarithmic




     transformation of MAVs.




4.   Available sets of ln(HAV)s suggest that the statistical populations are




     significantly and variably skewed and that FAV calculation should be based




     on a subset of ln(MAV)s nearest the fifth percentile.




5.   Available sets of ln(MAV)s suggest that FAV estimation is better served by




     the assumption of a triangular distribution of the populations of ln(MAV)s




     than by the assumption of a normal, rectangular, or biexponential




     distribution.




6.   Simulations using a triangular distribution indicate that 'parametric'




     methods for percentile estimation and "graphical1 methods in which ranked




     data are assigned cumulative probabilities PR*P(E(XR)) produce




     undesired biases in the true cumulative probabilities corresponding to




     fifth percentile estimates.




7.   Simulations also indicate that a graphical method  with  (a) ranked data,




     XR=»ln(MAV), assigned cumulative probabilities PR"R/(N+l), (b) ?g




     transformed to its corresponding standard variate, ZR"/PR, and



                                       13

-------
      (c)  a line fitted to % versus XR by the geometric mean functional
      relationship produces the least bias among alternatives examined.
8.    These simulations also suggest that it is appropriate to restrict  the
      calculation procedure to the  four X^s with P^s nearest 0,05,  because
      (a)  in the absence of skewness the precision of fifth percentile  estimates
      is little worsened by this and (b) in the presence of skewness this avoids
     the  introduction of substantial bias.
9.   The  old procedure used in the 11/28/80 Guidelines has some aspects which
     are contraindicated either theoretically or empirically and the new
     procedure described here should replace  it; however, FAVs calculated  from
     example data sets by the two  procedures  usually do not differ by more than
     a factor of 2.
10.  Modifications of the recommended new procedure with  respect to assumed
     distribution, general percentile estimation method,  and subset size  (up to
     half the set size) rarely cause FAVs calculated  from example  data  sets to
     vary by more than a factor of 2.  Therefore, even if it is debatable
     whether optimal decisions were made in developing the recommended  new
     procedure,  it is unlikely that any alternative procedure, within reason,
     would produce substantially different results.
II.  Modification of the recommended new procedure to use the entire data  set
     often produced  substantially  different FAVs from example data sets,  but
     many of these FAVs were sufficiently at  variance with the  lowest MAVs  in
     the data sets to reject this  modification.
12.  Modification of the recommended new procedure to reflect an extreme
     deviation from  random sampling consistently produced higher FAVs  from
     example data sets, but the average increase was only about fifty percent;
                                       14

-------
     therefore, questions about the propriety of applying methods based on




     random sampling to a system in which sampling is not strictly random




     are probably not of great importance.




13.   Recommendations about FAV calculation are the same whether HAVs  are for




     species or families.
                                    15

-------
                               STATEMENT OF PROBLEM




      A FAV is  defined  as  an  estimate  of the concentration corresponding to the




 fifth percentile  of  a  suitable set of MAVs for a material; i.e., the FAV exceeds




 five percent o£ the  MAVs  and  is  exceeded by ninety-five percent.  Because the




 number of  species tested  with  any particular material is usually rather small,




 most sets  of MAVs will  not have  a datum which can reasonably be designated as




 the  fifth  percentile;  rather,  the set of MAVs must be assumed to be a  sample of




 a  population (in  the statistical sense) that is  large enough that a fifth




 percentile is  defined.  For  example,  if resources permitted, the MAVs  of  a




 material  for many hundreds of  aquatic taxa could be determined.  Such  a set  of




 MAVs could reasonably be  considered to tvave a fifth percent lie  that could be




 obtained by inspection  and is  the type of statistical population of hfaich the




 available  sets  of MAVs  are assumed to be samples.  Of course, the population of




 MAVs would need to be  determined using a mix of  taxa that is acceptable to




 toxicologists  for calculating  a  FAV.  The above  assumption is inherent to the




 definition of  the FAV and any  objection to it, or modification  of it,  was not a




 subject of this project.




      An additional assumption  is necessary because any estimation method  using  a




 sample  from  a  population  requires that the manner in which the  sample  was




 obtained be  adequately  specified.  The toxicity  of a material is measured by




 many  independent  investigators,  who select test  species based on a  poorly




 defined combination of  tradition, convenience, happenstance, and  intent to




 diversify  the mix  of species.  All available data meeting certain quality




 standards  (I)  are  incorporated  into the aets of  MAVs.  This  incorporation step




does  not affect the nature of  the sampling process, except that a FAV  will not




 be calculated  unless the  set of  MAVs  is of a minimum size (eight)  and  contains



                                       16

-------
 representatives  of  certain categories of species (1).  It is not possible to




 represent  this  process  in  a  form  suitable  for applying appropriate exact




 estimation methods.   The issue then becomes what feasible description of




 sampling  (e.g.,  random, systematic) most closely approximates this process,



 Random  sampling  was  selected  for  the following reasons:



 (1)  Although  they  meet certain minimal diversity standards, the available sets



     of MAVs vary markedly,  and apparently haphazardly, in the species and




     higher  taxonomic levels  they contain.  Such variation is not compatible



     vith  entirely  systematic sampling schemes and suggests that an appropriate




     sampling  assumption should contain a  strong random element.



 (2)  Even  where  some  elements of  systematic sampling are  evident in the




     available sets  of MAVs,  a high correlation  of  toxicity  with these elements




     is usually  not  apparent.  Without such a correlation,  an assumption of




     systematic  sampling is  not particularly needed  because  for practical pur-




     poses it can be  approximated by an assumption  of random sampling,




     In addition to  being as  much, or more, in accord with  actual sampling




procedures than  other tractable sampling assumptions, the assumption of random




sapling nay b*  justified, in part, by noting chat,  in general,  deviations from




this assumption  may occur without seriously compromising  results.   Methods based




on random  sapling do not lose all ch*ir utility if  it is not possible  to




rigorously define *  population and to formally conduct random  sampling  from  it.




The population may even be somewhat hypothetical, being defined, in part by  the




data selection process.  Sampling may be nonrandom,  but as  long  as  the  sampling




process has a low enough correlation with  response,  results  under an  assumption




of random  sampling will not deviate by more than a certain  amount  from results
                                       17

-------
under more appropriate assumptions.  Consideration will be given below to what




errors would be  introduced if random sampling were assumed for percentile




estimation when  sampling is actually nonrandora.




     Finally,  if a procedure adopted for calculating FAVs results in criteria




chat are somehow independently validated, the procedure can be considered to be




entirely empirical and the assumptions become part of the definition of  che FAV




needed to produce the desired criteria.  This, however, is speculative and the




question remains as to whether the procedure developed here employs the  most




appropriate assumptions and, if not, whether this has any substantial impact  on




FAVs.
                                        18

-------
                   DESCRIPTION OF PERCENTILE ESTIMATION METHODS
      Method*  for  eseimacing, from  random samples, a specified percentile of a
 population  can generally be placed into one of two categories.  These categories
 are  presented here primarily to  facilitate discussion and are not meant to imply
 that  methods  in different categories do not have some important coramon features
 or are  not  sometimes nearly equivalent.  One notable feature of any method for
 estimating  percentiles is the need to make at least some distributional
 assumptions about  the population from which the sample was drawn.
      For methods  in the first category, the parameters in the general
 mathematical  equation for the assumed distribution are estimated  from  the  sample
 by mathematical procedures formulated to-produce estimates with desired
 properties, such  as being unbiased, having minimum variance, or having maximum
 likelihood.  The common formulas for estimation of mean  and variance  from  a
 sample  from a normal population  is an example of such a  method.   Once  the
 parameters are estimated, it is a  simple matter to substitute then  into the
general equation  for the distribution and to estimate a  desired  percentile.
 This category will be referred to  as 'parametric methods'.
     The second category of methods involves ranking the data in  a  sample  and
 then  plotting the  ranked data (X^) versus a cumulative probability  (PR)
 assigned to each rank (R).  For calculation simplicity,  plotting  is  usually  on  a
coordinate system  for .which the cumulative form of the assumed  distribution  is  a
 straight line.  A  line is fitted to the plotted data by  eye or  by some
appropriate mathematical curve-fitting technique and the estimate of the  desired
 percentile is read off the plotted line or computed from the equation  for  the
 line.  This category will be referred to as 'graphical methods',  although
explicit graphing  is never strictly necessary.  In most  cases,  graphical methods
                                       19

-------
 are  not  mathematically  rigorous and do not produce the unbiased, maximum
 likelihood,  or minimum  variance estimates that parametric methods are designed
 co  produce.   This  does  not mean, however, that graphical methods will not
 perform  adequately in practice; in  face, in some cases their performance is  very
 similar  to  that  of parametric methods.  Furthermore, for some cases suitable
 parametric methods do not exist or  are unreasonably cumbersome; graphical
methods  thus  might be a very useful alternative.
      For both parametric and graphical methods, discussion here will be
 restricted  to a  class of distributions which have only two parameters, these
being  a  location parameter and a scale parameter.  By this it is meant that,  for
 each  distribution  type,  there exists  a standard distribution with standard
variate denoted  'z', such that for  any distribution of this type wich variate
denoted  'x'  there  exists a location parameter  'L1 and a  scale parameter  'S1 such
that  x"L+S"z.  For example, for the normal distribution, the mean and standard
deviation are usually used as the location and scale parameters, respectively,
and the  standard normal  variate (also called the "standard normal deviate'  (4))
 is then as usually tabulated.
Paratastric Method*
     The only parametric method considered here will be  a general one,  termed
 'best  linear unbiased estimation' (5,6), which can be applied to any distribu-
tion characterized by location and  scale parameters.  This method  is  'unbiased1
in that the parameter estimates will, on the average, equal the  true  population
parameter values.   It is 'linear' in  that the  parameter  estimates are linear
functions of the data.   It is 'best1  in that the parameter estimates  have the
lowest variances of all  linear unbiased techniques.  There may  be nonlinear or
biased methods that have smaller variances, but, in general, the performance,
                                       20

-------
with  respect to bias and variance, of this technique cannot be much improved,

Parameter  estimates (L, S)  are obtained by minimizing the value of the matrix
expression:
                         ,    ^   /N  , T  - 1  ,    /^   A  x
                         ( c - L - S-)T'V K(c - L - S-z
where:  N  is  the  sample  size;
        jc  is  a  (Nxl) matrix  consisting of the  ranked sample;
        j^  is  a  (Nxl) matrix  of the expected values of ranked standard variates
           of  random samples  of size N from  the  assumed distribution;
        V_  is  the  (NxN) varianee/eovariance matrix for ranked standard variates
           of  random samples  of size N from  the  assumed distribution; and
        T  denotes matrix  transposition.
                                              /\ A A
This method has the additional advantage that  Xp"L*S-zp is  also  the best

linear unbiased estimate  for x_, the pc" percentile of the  population.

This can be demonstrated  in  a variety of ways,  but is mo»t  obvious when it is

realized that any particular percentile could,  quite legitimately, be designated

the location  parameter.

     This  method  also has the advantage that  it can be applied  to  an arbitrary

sub sample  of  the data and still produce the best linear unbiased estimate that

can be obtained from that subsaaple.  Applying  the method  to a  subsample  simply

requires eliminating, from matrices JK, _z_, and  V^" , the elements  referring to

data not in the desired sub sample.  (Note:  The  calculation  and  inversion  of V_ is

not affected  by these deletions; rather, deletions are made after  inversion.)  A

notable property of such  'censoring1 of data  is that, if the remaining data are

those nearest the percentiie of interest, the  variance of  the  percentile

estimate is little worsened  as the number of data used is  reduced.   This

suggests that using all the  data, other than  to determine  ranks  and  define V_,

has relatively little utility in this kind  of  estimation.

     The ability of this method to use only a  subsample of  the  data  has

particular significance when the distribution  of the population  from which the

                                       21

-------
data  are drawn ii not perfectly characterized.  For example, when concerned  with




che  fifth  percentile, deviations  from the assumed distribution that are




restricted  to the upper part of the distribution will impact the calculations




little  if  only the  lowest  few data in the sample are used.  Even if the




distributional assumption  is violated near the fifth percentile, the impact  o£




this  violation will be reduced as the number of data formally used in the above




equations  ii reduced, as long as  the data used are those nearest the percentile




of interest.  Of course, the method still makes distributional assumptions about




both  the data used and those not  used and errors will arise if these assumptions




are  incorrect, but as long as the distributional assumptions are not grossly




violated in the range of the selected subsample, theae errors will generally be




minimal.   The question then arises as to the optimal sub sample size (n) ,  a  small




size  having the advantage of reducing the effects of deviation from the assumed




distribution and a large size having the advantage of reducing the variance  of




estimates when the distributional assumptions are correct.  The  answer  to this




question is specific to the problem of concern and will be  considered  below.




      One troubling aspect of this methodology is, ironically, its  lack of bias.




This  is a  problem because  the lack of bias is in the variate rather than  in  the




cumulative  probability; i.e., in  repeated sampling, x_ will average x_, but




the true cumulative probability (P(xp)) corresponding to  x_ will  not average




p, unless cumulative probability  is linearly related to variate,  which  ia only




true  for rectangular distributions (simulated sampling  from a variety  of  popula-




tions is presented below to demonstrate this point).  Because che  definition of




the PAV is based on protecting a  certain percentage of  a  specified taxon, this




method is inappropriate; rather,  a method that is unbiased  with  respect to  the
                                       22

-------
desired cumulative probability is desired.   We  art  aware  of  no  published ueehods

of this sort.  It ii for this reason that graphic*! methods  are now considered.



Graphical Methods

     Graphical methods for examining cumulative distributions  inherently have  four

isaucs that must be resolved:

(I) Cuaulative Probabilities Assigned to Ranked Data

    Formulae reported (4,7,8) for calculating the cumulative probability  P^ to
    assign to a datura XR with rank R in a sample of size  N include R/*f,
    
-------
    (d) ?a"P(S(%)) has, by definition, obvious theoretical  foundations  because
        it denotes the cumulative probability corresponding  to  cha  expected value
        of XR.  (S(XR) ii 4iso called 'rankit'  (4)).   this formula, is  a
        counterpart co PR"R/(H+l),  differing by being based  on  the  expected value
        of ranked data rather than  the true cumulative probabilities corresponding
        co ranked data.  Unlike PR*R/(H*l), its values are distribution-dependent,
        Its use will also b« further explored below,  but  because  it  is based on the
        expected value of the variate, it is anticipated  that  it  will  show  the same
        prob lea of bias as the parametric method above.

(2) Transformation of Axes
    Thisis dictatedby the assumed distribution and  by the reecriction  adopted
    her* chat che assumed distribution should produce a linear  plot on the
    selected axea.  In general, it  is the axis against which cumulative
    probabilities are plotted that  is transformed and the transformation is based
    on the standard distribution of the assumed distribution;  in fact, this can  be
    treated as a transform of PR to a corresponding standard variate ZR.  In
    such a case, the plot becomes one of a ZR assigned to etch rank versus  the
    observed datum X^.   The slope dX/dZ is the scale parameter and the intercept
    on the X axis is the location parameter.

(3) Subeaaple Site
    This issue is identical to that discussed for the parametric method.  A
    later  section vill  consider how the sub sample size (n) can beec be determined,
    baaed  on simulations under various assumptions.

(4) Fitting a Line to PlottedData

    Because the restriction of a linear plot has already been made, this issue
    reduces to how to compute the slope of the line nose appropriate to the data.
    Because,  for any N, che 2R or PR assigned to a ranked datum is  fixed, and
    thus nay be an analogy to an independent variable, and because the  line to be
    fitted can be expressed as Xgl+S'Zg, it may be  thought that the  standard
    I ease-squares regression formula with XR as the dependent variable  and ZR
    as the independent  variable would b« chs preferred choice.  As will be seen
    below,  this turns out to be the case when ?n"P(E(XR)) is used and when
    ipi  rather JChan P(Yp), is desired to be unbiased.  Aa before, achieving an
    unbiased  P(Yp) is not amenable to exact techniques and  an empirical approach
    muse be used.   To this end, three different, but simple, slope  formulas were
    considered (again,  this approach is strictly empirical, employing these
    formulas  ae representing a range within which a reasonable slope night lie;
    nothing is implied  here about a theoretical justification  for one or che other
    formula and it is not implied chat this application meets the assumptions on
    which  any of the formulas are based):

    (a)  L3-X  - standard bivariaee least-squares with Xg as  che dependent variable
        (residuals minimized in X-direccion):
                                        24

-------
 (b) LS-Z - standard bivariate least-squares  with  ZR as  Che dependent variable
    (residual* minimized in Z-direction):
                          a  -
(c) GMFR - geofoetric mein functional  relationship (residuals are minimized in
    th« direction of ch« arithmetic  reciprocal of the §lop«; this method
    produces the geometric meen of the slopes by the two previous methods and
    has seen some application in regression  where both variables are in error
    (9,10)):
                             •\
Whatever slope formula is used,  the line always  passe*  throufh the
mean X& and the mean ZR.   The location parameter estimate  U, which  is the
intercept on the X axis,  is therefore
                            A
                            L
                                     25

-------
                           SELECTION OF DISTRIBUTION
      All methods  for  estimating the fifth percentile of a population from a
 sample  require at  least some assumptions about the distributional
 characteristics of the  population.  Few data sets from which an FAV will be
 calculated will be large enough that such characteristics can be inferred from
 the  individual data set.  However, the large number of sets available (Tables I
 and  2)  provides an opportunity for evaluating these characteristics and for
 determining which  characteristics can be reasonably applied to  all data sets and
 which parameters must be estimated individually from each set.
      It is desirable  to keep the number of unknown distributional parameters as
 low  as  possible, not  only because analysis becomes markedly more complicated as
 the  number of parameters increases, but also because data sets  of the minimum
 size  (H»8) may be  overly fitted if the number of parameters is  not small.  The
example data sets  (Tables I and 2) vary widely in their means  and coefficients
of variation.  Therefore, at least two parameters, a location  parameter (e.g.,
mean) and a scale  parameter (e.g., standard deviation), are required.
     Because these two parameters relate to the first and second moments of the
samples, an obvious third parameter to consider is skewness, which is related  co
the third moment of the samples.  Skewness is also strongly indicated by
inspection of the  example data sets.  A skewness measure (4),  the normalized
third central moment,  Was estimated for each example data set.   All sets showed
positive skewness.  The skewness was substantial enough to  reject,  at  Che 0.10
 level of significance, the hypothesis that the set was a random sample  from a
normally distributed  population for 35 of 3? SMAV sets and  for 34 of  37 FMAV
sets; at the 0.01  level of significance, this hypothesis was  rejected  for  30
SMAV sets and 25 FMAV sets.
                                       26

-------
      Because of this strong positive skewness, a logarithmic (base e)
 cran8formation was  applied  to each MAV, so that discussion will now relate to
 che distribution of In(MAV).  The skewness measure for each data set was
 recomputed and the  average  skewness decreased  from 2,39 for SMAVs and 2.08 for
 FHAVs  to  0.06 for SMAVs  and 0.07 for FMAVs.
      The  small average skewness does not, however, mean that individual sets can
 be considered to be samples from nonskewed populations.  When the skewness
 measures  of individual data sets were  tested  under the same null hypothesis as
 above, the hypothesis was rejected at  the 0.10 significance level for 8 SMAV
 sets  and  7 FMAV seta and  at the 0.01 signficance level for 3 of  the  SMAV  sets
 and 2 of  the FMAV sets.   Although this is substantially fewer than before
 logarithmic transformation, it still indicates that skewness in  some sets might
be too large to ignore.   Furthermore,  among the sets with significant skewness,
 the skewness was sometimes  positive and sometimes negative, indicating  that the
 populations these sets represent vary  substantially in skewness.
      Therefore, despite  logarithmic transformation, skewness in  the  data  must
 still be dealt with by the methodology adopted for the estimation of the  fifth
 percentile.  Two general  approaches were considered for this.  Firse.
distributions with a third  parameter that affects skewness and which can  be
 estimated from a sample  could be used.  This  approach  greatly  increases the
difficulty of parameter  estimation and it  is  questionable whether the  smaller
data  sets reliably have  enough information to  make this effort  appropriate  or
 worthwhile.  The second  approach is to limit  the estimation method  to  a subset
of the data near the percentile of interest.   By doing this, the effects  of
having non-zero skewneas  are markedly  reduced  and the  location  and  scale
 parameter estimates apply only locally, incorporating  the effects of skewness  at
                                       27

-------
chat  locality.  This approach allows Che use of relacively simple estimation

methods  and,  as will be  further discussed below, has very little impact on the

precision of  fifth  percentile estimates even if a population is not skewed.   The

second approach will therefore be employed here,

      Higher moments of the data sets were not directly examined because (a)  the

decision to limit analysis to a subset of the data makes such an examination

complicated and (b) the effects of higher moments should be adequately accounted

for  either by this  limitation or by the examination of specific distributions

that  follows,

      Inference of distributional characteristics from the example  data sets was

therefore limited to symmetric distributions with just location and scale

parameters to be estimated; furthermore, the most relevant  information in the

data  sets is that nearest the fifth percentile.  The approach  followed here was

to examine the fit  of specific distributions to the example data sets.  Four

distributions were considered:

(D R « c t an gu 1 a r Pi s t r i b u t ion

         This was included as an extreme case because it assumes that  the
    relative frequency of ln(MAV)g remains constant between some lower and  upper
    limits, whereas theoretical considerations  and  inspection  of the data sets
    suggest that the frequency declines as the  lower and upper  lisits  are
    approached (i.e., very sensitive and very resistant taxa are rarer than
    those with moderate sensitivity).  The standard probability density function
    for this distribution is:
                       .'  f(z) -  1//TT   ;  -/T<  z <
                          f ( z) -  0       ;  z < -JT,  z >

(2 ) Triangular Distribution

         This was included because  it is the simplest distribution  that  incor-
    porates two basic properties  that the  frequency  of  ln(MAV)s  should have:  (a)
    sensitivity should have lower and upper limits (no  species  succumbs  to
    infinitesimal concentrations of a material and no species  tolerates  infinite
    concentrations) and (b) the  frequency  of ln(MAV)s should  decline  to  zero  as

                                       28

-------
     Che  Haiti  are approached (fewer species are near the limits than are near
     the  raidrange).  The scandard probabiLicy density function for this  distribu-
     cion is:
                  f(z) -  (1- z|)//6   ;   -/6 < z < /6

                  f(z) -  0            ;   z < -/6 , z > /6

 ( 3 )  Normal Distribution

          This  was included due  Co  its broad applicability and Co provide a
     curved alcernacivt to the linear frequency trend of the triangular
     d iscribucion ; chis curvicure causes relatively rare sensitive or resistant
     taxa Co have somewhat more  extreme ln(HAV)i (relative co the range  of the
     majority of the caxa  with moderate sensitivity) than does the triangular
     distribution.  No lover or  upper limits exist, but the frequency becomes so
     small at reasonably moderate deviations from  the mean chat  Chis  deficiency
     is probably of limited consequence.  The scandard probability density
     function for chis distribution  is:


                  f(z) -  — e~z /2   ;  -- < z < +-
(4) Biexponenti al Distribution

         This was included  as an extreme case  in which  the most  sensitive  and
    resistanc taxa have greatly different  ln(MAV)s than  the majority of  the  taxa
    with moderate sensicivicy.  The  scandard probability density function  for
    this distribution is:

                  f(2) .-L-.-^M   .  .„<,<+.
                          /2

Shapes of these distributions when  chey have mean * 0  and  standard  deviacion «  I

are displayed on Figure  1.

     The best linear unbiased estimation method* discussed  earlier  was  used  to

estimate location and seal*  parameters  from each example data  sec  for each

combination of the four distributions  above  and  four  subset  sizes  (n*4,  N/4,

N/2, and N, where N-daca  set size;  n  also  was  required  co  be  ac  least 4,  which
  The matrix V_ for  this method  was  calculated  by exact integrals for all

  distributions except normal,  for  which  approximate  formulas  were  used (11)
                                        29

-------
                1.  Standard probability cUn«lty plot* for rectangular C
                    triangular (•-•-•), noraal (	),  and bIexponential  (
                    distribution*.
O
       0.60-
     m
     «0.40-
                      .>-V
                                                 „%,

-O
D
_Q
o
L.0.20
Q-





n n




—


h- S
/'
'*'*'
**< ''
mmA*****\' 1
/•': \x>
« .* \ *.
^ * ' \
f* * *. .
/f •' "' v*
x"/ / *• v\
.• t : : \x.
• / •' '-. ^ v.
/ / '*. v^
•*"
.

i 1 i 1 . 1 .







*
\
>• \
"• v,*\
* **^»x
| :X*T4rf'--«Li_.,..r
              -3.0
-2.0       -1.0        0.0       1.0        2.0
 Number of Standard Deviations from Mean
3.0

-------
 was  considered  a ainimui number to use co test distribution fits),  Frcra each
 such  estimation,  Che  expected  value  (E(X^)) of each ranked datum was estimated




 as  L+S '£(£3) i  'rfhere L is  che  location  parameter estimate, S is the scale
 parameter esciaata,  and  E(Z^)  is  ch* expected value of the datum of rank R in



 snples  of size  N  fcoa ch*  asinawi  standard discribucion.  th« raeio:
(£.•., th«  fraction  of  ch* varianct of  ch*  subiae not  explained by  fitting the



data  co  ch* distribution) was  adopt wi as a o*aiur* of  good n*aa-of- fit of ch* data




CO th* aasua*d distribution ov*r  ch*  sis* (R) of th« subs*C ua«d.   Avtraf*



good n*ii-of- fits for all SMAV  s*cs ar*  r*port*d in Tabl* 3 and for  FMAV sats ia




Table 4.



     Th*  criangular  distribution  was  j«L*et*
-------
TABLE 3.  AVERAGE GOODNESS-OP-FITS3 FOR  In(SMAV) DATA SETS
ASSUMED DISTRIBUTION
RECTANGULAR
TRIANGULAR
NORMAL
BIEXPONENTIAL
a "Goodness-of-f It" Is the
not explained by fitting
N
0.169
0.082
0.081
O.L06
fraction
SUBSET SIZE (n)
•M/2 N/4b >4
0.150 0.144 0.135
0.099 0.114 0.118
0.104 0.134 0.137
0.230 0.235 0.206
of variance of 'n1 data points
data to assumed distribution; lower values
indicate better fits; values should be compared only within columns.
b n =» 4 when N <_ 16.

TABLE 4. AVERAGE GOODN!SS-QF-FITSa
ASSUMED DISTRIBUTION
RECTANGULAR
TRIANGULAR
NORMAL'
BISXPONEOTIAL
N
0.145
0.095
0.087
0.109

FOR In(FMAV) DATA SETS.
SUBSET SIZE (n)
N/2 N/4b 4 •
0.127 0.121 0.125
0.103 0.108 0.116
0.120 0.129 0.143
0.277 0.218 0.238
a "Goodness-of-f1t" Is the fraction of variance of  'n'  data points
  not explained by fitting data to assumed distribution;  lower  values
  indicate better fits;  values should be compared only  within columns.

b n - 4 when N < 16.
                                  32

-------
           SELECTION OF PERCENTILE ESTIMATION METHOD AND SUBSET SIZE

     Ten  thousand computer-generated random samples from a standard triangular

distribution  (location parameter " mean * 0; scale parameter =• standard devia-

tion =•  1) were used to estimate the  fifth percentile of the distribution for

each combination of the following methods, sample sizes, and subsample sizes.

(I) Percentile Estimation Methods

    Seven methods were examined.  These included one parametric method (best
    linear unbiased estimation) and  six graphical methods (all possible
    combinations of the two  formulas for  assigning cumulative  probabilities
    (P(S(XR)), E(P(XR)) and  the three  formulas  for computing slope  (LS-X,
    LS-Z, GMFR)).

(2) Sample Sizes

    Sample sizes (N) of 8,  15, and 30 were selected as  being representative of
    the minimum size, a moderate size, and a large size that are  found in  the
    available sets of SMAVs  and FMAv"s .

(3) Subsample Sizes

    Subsample sizes (n) of  the 4, N/4, N/2, and N points closest  to the  fifth
    percentiie were considered; for  N/4,  an additional  restriction  of  n>4  was
    imposed; n"4 was considered to be  the minimum reasonable size,  a lesser
    number making analysis  too sensitive  to a spurious  datum.

From location and scale parameter estimates, the estimate of the  fifth

percentile was calculated aa xg*L-l.675*S, -1,675 being the  fifth percentiie

for the standard triangular  distribution.  The  average  xg over the  10,000

simulations was designated  as xg and should equal -1.675 for methods unbiased

with respect to the variate.  Because  the parameters of the  population  from

which the samples were drawn are known, the true cumulative  probability  P(xg)

corresponding to each xg was calculated.  The average  P(x5)  over  the 10,000

simulations was designated  Pg and should  equal  0.050 for methods  unbiased  with

respect to cumulative probability.  "x"g is tabulated  in  Table 5 and  Pg  is
                                        33

-------
 tabulated in Table 6.   Table 5  also  includes  the  standard  deviations  for '£5 in
 order to indicate  the  relative  precision  of the various methods.
      As expected,  the  best  linear  unbiased  estimation  method  did  produce  an
 essentially unbiased loj,  as  did the  graphical  method using P(E(X^)) co
 assign  cumulative  probability and  LS-X to calculate slope  (Table  5),  However,
 it  is bias  in P<$ that  is  of  paramount  concern  here.  The best  linear  unbiased
 estimation  method  and  all graphical  methods using  P(E(X^)) were substantially
 more  biased  than the graphical  methods using  E(P(XR))  (Table  6) and were
 therefore dropped  from consideration.   In addition, the standard  deviations of
^5  by the best  linear  unbiased  method  were  usually no  better  than 10% less
 than  those  of the  graphical  methods  using E(P(Xa)) (Table  5),  indicating  that
 the better  precision of the  best  linear unbiased method is of little
 consequence.
      Although they did have  lower  biases  than  the  other methods,  none of  the
 graphical methods  using E(P(X^))  to  assign  cumulative  probability had an
 unbiased  Pj  and the bias  varied with n and  N  (Table 6).   Furthermore, none  of
 the formulas  for calculating slope had the  lowest  bias for  ail combinations of  n
 and N.   The  zeotnetric  mean  functional  relationship was selected as having the
 lowest  average bias over  all combinations.
      Selection of  Che  most appropriate subsample  size  required consideration  of
 the precision of xj (Table 5)  for  the  selected percentile  estimation  method
 (graphical method  using E(P(X^))  to  assign  cumulative  probability and GMFR  to
 calculate slope).   For N»8,  the standard  deviation of  x
-------
TABLE 5-  ME^3 AND STANDARD DEVIATIONS3 OF ESTIMATES OF FIFTH PERCENTILE  (x5)
          BY VARIOUS METHODS,  FOR 10,000 SAMPLES  FROM A STANDARD TRIANGULAR
          DISTRIBUTION.
N n
PARAMETRIC
METHOD


-- r,
LS-X
84-1
(0
8 -1
(0
15 4 -1
(0
8 -1
(0
15 -1
(0
30 4 -1
(0
8 -1
(0
15 -1
(0
30 -1
(0
,68
.57)
.68
.52)
.67
.39)
.67
.38)
.67
.36)
.67
.25)
.67
.25)
.67
.25)
.67
.24)
-1
(0
-1
(0
-1
(0
-1
(0
-1
(0
-1
(0
-I
(0
-1
(0
-1
(0
.68
.57)
.68
.53)
.67
.39)
.67
.39)
.67
.38)
.67
.25)
.67
.26)
.67
.26)
.67
.26)
n-p(E(xR):
LS-Z
-1
(0
-1
(0
-1
(0
-1
(0
-1
(0
-1
(0
-1
(0
-1
(0
-1
(0
.79
.61)
.81
.55)
.73
.40)
.75
.40)
.76
.39)
.69
.25)
.71
.26)
.72
. 26)
.72
.26)
\— — «.
-GRAPHICAL-
/
GMFR
-1
(0
•™ 1
(0
-1
(0
-1
(0
-" 1
(0
— 1
(0
-1
(0
-1
/ rs
w
-1
(0
.73
•59)
.74
.54)
.70
.39)
.71
.39)
.71
.38)
.68
.25)
.69
.26)
.70
"i r \
.io;
.70
.26)


rF
LS-X
-1
(0
-1
(0
-1
(0
-1
(0
-1
(0
-1
(0
-1
(0
-1
f n
\v
-1
(0
.83
.62)
.83
.56)
.77
.41)
.77
.41)
.77
.39)
.73
.26)
.73
.26)
.73
.27)
.73
.27)

l=E(P(XR))
LS-Z
-1.95
(0.67)
-1.98
(0.58)
-1.84
(0.44)
-1.85
(0.43)
-1.86
(0.40)
-1.75
(0.26)
-1.77
(0.27)
-1.77
(0.27)
-1.77
(0.27)


GMFR
-1.89
(0.64)
-1.90
(0.57)
-1.80
(0.42)
-1.81
(0.42)
-1.81
(0.39)
-1.74
(0.26)
-1.75
(0.27)
-1.75
t A 1 1 \
\\j-t-i /
-1.75
(0.27)
 Standard deviations  of  estimates  in parentheses.
                                      35

-------
TABLE 6.  MEAN TRUE CUMULATIVE PROBABILITIES  (P ) OF ESTIMATES OF  FIFTH
          PERCENTILE (P(x5)) BY VARIOUS METHODS, FOR 10,000 SAMPLES
          FROM A STANDARD TRIANGULAR DISTRIBUTION.
  M    n                              METHOD
          PARAMETRIC  	GRAPHICAL	
                      	PR-P(E(XR))	  	PR=E(P(XR))	
                         LS-X     LS-Z     GMFR       LS-X     LS-Z     GMFR
8 4
8
15 4
8
15
30 4
8
15
30
0.076
0.072
0.063
0.062
0.061
0.055
0.055
0.055
0.055
0.076
0.073
0.063
0.063
0.062
0.056
0.056
0.056
0.056
0.066
0.058
0.057
0.054
0.052
0.054
0.051
0.051
0.051
0.071
0.065
0.060
- 0.058
0.057
0.055
0.053
0.053
0.053
0.063
0.057
0.053
0.053
0.052
0.048
0.049
0.050
0.050
0.054
0.044
0.047
0.044
0.042
0.046
0.044
0.044
0.044
0.058
0.050
0.050
0.049
0.048
0.047
0.047
0.047
0.047
                                       36

-------
 lowest  at  n*4  and  did  not  vary  among  Che  n  by more  Chan  3%,   There  is thus no




 substantial,  advantage,  with  respect to  precision, to  using  n>4.




      Another  factor  in  the selection  of the subsample size  is  the possibility of




 skewness.  Therefore,  simulations  were  also conducted  using a  skewed triangular




 distribution  for  which  the mode was 20%,  rather  than  50%, of  the distance  from




 the  lower  to the  upper  limit.   For  all  sample sizes,  using  n*N resulted  in




 ?5<0.02  and,  for  N-15  and  N-30,  using n-N/2 resulted  in  Pj  being about 0.03,




 Using n"4, however,  resulted  in ?5  being  between 0.04 and 0.05 for  all N.




 Using n*4  also  resulted  in substantially  lower  standard  deviations  for X5  than




 using n*N/2 and n*N,




      Because  for  a nonskewed  population the smallest  subsample size considered




 performed  little,  if any,  worse than  larger subsamples,  and because the




 possibility and consequences  of skewness  give good  reason  to  restrict  the




 subsample  size, a  subsample  size n™4  is recommended .   Limited consideration  was




 given to a smaller subsaraple  (n™2), but for N*8  this  resulted in  substantial
bias  in Pg  (Pg^T.OZ)  and  in  a  202  increase  in  the  standard  deviation of




*5 ; also, the  use  of  so  few  data markedly  increases  the  sensitivity of results




to an occasional unusually low datum.




      Consideration was also  given  to  the possible  effects  of nonrandonj sampling




by determining  the bias  introduced  into Pg  if  the  method  recommended above is




used when samples  are not obtained  randomly.   Two  nonrandom sampling schemes




ware  investigated.  First, samples  were  taken  in  an  entirely systematic fashion




highly correlated  with variate, data  being  uniformly distributed over percen-




tiles with  the  ic^ datum  (Xj_)  being set equal  to  x_. , where p^*(i-0. 5)/N.




With  such a  scheme, PCxg) equals 0.015, 0.024,  and 0.034 for N equal to 8, 15,
                                        37

-------
 and  30,  respectively.   Although PCx;)  is  therefore  substantially biased, this
 is  an  extremely  unrealistic depiction  of  the sampling and the_ biases are extreme
 upper  limits.  Sampling  schemes with a more  realistic systematic component would
 result  in  much lower biases.
     The  second  nonrandotn  sampling  scheme used  was  stratified sampling  in which
 each member of the  sample  was assumed  to  be  randomly sampled  from a restricted
 percentile range,   The  range  for  the ith  datum  of a sample  was  [(IOOS)(i-Q.5)/N]
 * 25%;  i.e., a fifty percentile range  centered  on the value  used  for the  ic^
datum of the systematic  sample discussed  above.  For low  and  high i, the  range
 was compressed so  that  percentiies  were maintained  between  0  and  100 and  so
 that, over the entire sample, each  member of the population  had  an equal  chance
of being drawn.  Specifically, for  low i,  if a  percentile was computed by the
 above  formula to be <0,  its absolute value was  used.  This  results  in  the  ranges
 for low  i  being  narrower than the nominal  50Z  (as small  as  27Z  for  i*i and N»3Q)
 and sampling within the  ranges being somewhat skewed to  low percentiies;
analogous  compression and  skewing occurred  for  high i.   Because the  recommended
 procedure  would  heavily  employ data with  low i, the systematic  component  of  this
 procedure  is therefore  even greater than  implied by the  restriction  of sampling
to fifty percentile ranges.  Using  this sampling scheme,  ten-thousand  samples of
 size 8, 15, and  30  from  a  standard  triangular distribution  were computer
generated  and FAVs were calculated  from each sample by the  procedure recommended
above based on random sampling.   ?5 was 0.040,  0.042, and 0.044 for  N  equal  to
8, 15, and 30, respectively.  This  small  bias suggests that,  even with a  strong
systematic element  in the  actual  sampling,  an  assumption of random  sampling
 performs well as long as there is also a  substantial random element  in the
sampling,  or at  least an element  that  is  not correlated  with variate.
                                       38

-------
        APPLICATION  OF  RECOMMENDED  PROCEDURE AND ALTERNATIVES TO DATA SETS

      In che  previous sections  all  issues necessary for Che recommendacion of  a

 procedure  for  FAV calculation  have been considered.  The recommended new

 procedure  assumes the  sec of ln(MAV)s is a random sample from a triangular

 distribution with unknown location and  scale parameters.  The mechanics of che

 procedure  can  be summarized as  follows:

      The  ln(MAV)s are  ranked and each assigned a cumulative probability
      PR*R/(N+I), where R is the rankAand N the number of data  in the  aet.  A
      line  of the form  ln(MAV)«¥s/Pji+L is fit to th«  four points with  PR
      nearest 0.05.  (The square root of PR constitutes transformation to  the
      variate of a standard triangular distribution somewhat different than, but
      equally valid  to,  that used above; it is used here because it  is simpler to
      calculate when PR<0.5).   S, L, and FAV are calculated as  follows:
                       i  -
\
                                                     2/4
                           £ -

                            InFAV  - Sno.05 +  £

     Example calculations ar*  provided  in Appendix  I.

This procedure was applied to  che example data sets  in Tables  I and 2.  The  FAVs

thus calculated for each data  set are included in Tables 7  and 8,  along with the

lov*st MAV as a reference.

     Also included in Tablet 7 and 8 are PAVs calculated  for  each  data  set  using

che old procedure presented in che November 28, 1980 version  of the Guidelines

(1),  This procedure can be described as follows:

     The ln(MAV)s are ranked and assigned to  fixed  intervals  with  width -  0.25
     and with the first interval starting at  the  low«st  In(HAV).   Each  interval
     is assigned a cumulative  proportion P*Ro«x/N,  where N  is  che  number  of
     ln
-------
TABLE 7.  FAVs CALCULATED  FROM  SMAV  DATA  SETS Bt OLD PROCEDURE,  RECOMMENDED NEW PROCEDURE, AND VARIOUS  MODIFICATIONS
          OF RECOMMENDED NEW PROCEDURE,
MATERIAL
COPPER
DOT
CADMIUM
CADMIUM
TOXAPHENE
ZINC
ENORIN
MERCURY
ZINC
LINOANE
COPPER
NICKEL
DIELORIN
0 ALORIN
ENDRIN
HEPTACHLOR
DIELORIN
L1NDANE
CHROMIUM! VI)
CHROMIUMU II )
HEPTACHLOR
NICKEL
DOT
ALDRIN
CYANIDE
TOXAPHENE
CHROMIUM
-------
Table  7.  Continued
MATERIAL
ARSENlCd II)
MERCURY
SILVER
SILVER
ENOOSULFAN
CHLQRDANE
WATER
FRESH
FRESH
FRESH
SALT
FRESH
SALT
N
12
11
10
10
10
8
LOWEST
SHAV
810
5.0
0.0019
4.7
0.340
0.400
OLD
PROCEDURE
440
3.7
0.0014
3.3
0.218
0.090
NEW
PROCEDURE
340
2,6
0.0014
3.3
0.183
0.200

r»=N/2
260
1.6
0.0017
3.9
0.214
0.200

n=N
620
2.9
0.0003»
2.8
0.1 25»
0.378
NATIONS OF
UNIFORM
OIST.
430
3.5
0.0019
4.4
0.258
0.352
RECOMMENDED NEW PROC
NORMAL SLOPE
DIST. CHANGE
310
2.3
0.0012
3.0
0.158
0.162
430
2.7
0.0018
3.7
0.186
0.278

•tuuRE — —
PARAM.
METHOD
560
3.6
0.0018
4.0
0.251
0.313

NONRANDOM
SAMPLING
820
5.2
0.0019
4.7
0.340
0.280
GEOMETRIC MEAN OF  RATIOS  OF  FAV BY
MODIFIED PROCEDURE  TO  THAT BY
RECOMMENDED PROCEDURE:

NUMBER OF DATA SETS FOR WHICH  FA¥
BY MODIFIED PROCEDURE  DIFFERS  FROM
THAT BY RECOMMENDED PROCEDURE  BY
MORE THAN A FACTOR  OF  1.4:

NUMBER OF DATA SETS FOR WHICH  FAV
BY MODIFIED PROCEDURE  DIFFERS  FROM
THAT BY RECOMMENDED PROCEDURE  BY
MORE THAN A FACTOR  OF  2.0:
0.89      1.00      1.04      0.93      1.1 I
                                26
                                12
0.96      1.07     1.19      1.48
                                                                                  13

-------
Table 8,  FAVs CALCULATED FROM FMAV  DATA  SETS  BY  OLD FttOCEOURE,  RECOMMENDED NEW PROCEDURE,  AND VARIOUS  M30IFICATIOHS
          Of RECOMMENDED MEW PROCEDURE.
MATERIAL
CADMIUM
COPPER
MERCURY
DOT
ZINC
CADMIUM
ENDRIN
COPPER
CHROMIUMiVi )
DIELDRIN
ENDRIN
HEPTACHLOR
LINOANE
NICKEL
ZINC
ALDRIN
DDT
NICKEL
CHROMtUMUm
ALDRIN
TOXAPHENE
OIELDRIN
TOXAPHENE
SELENIUM
ENDOSULFAN
HEPTACHLOR
CHROMIUM(VI)
SELENIUM
LINOANE
CYANIUE
SILVER
SILVER
WATER
SALT
FRESH
SALT
FRESH
SALT
FRESH
FRESH
SALT
SALT
SALT
SALT
SALT
SALT
FRESH
FRESH
FRESH
SALT
SALT
FRESH
SALT
SALT
FRESH
FRESH
SALT
SALT
FRESH
FRESH
FRESH
FRESH
FRESH
SALT
FRESH
N
25
23
23
20
20
18
17.
17
17
16
16
16
16
16
15
14
14
14
13
13
13
12
12
12
11
10
10
10
10
10
10
9
LOWEST
FMAV
n
0.30
3.5
1.30
166
0.048
0.44
28
2490
0.70
0.037
0.057
0.170
65
13.7
7.4
0.140
310
33
3.7
0.110
4.5
1.30
600
0.040
1.00
67
340
10.0
77
4.7
0.0019
OLD
PROCEDURE
45
0.34
3.7
1.22
166
0.038
0.33
27
1940
0.67
0.036
0.044
0.121
50
10.4
4.0
0.102
160
30
3.5
0.065
1.6
0.99
440
0.033
0.75
a
154
8.2
58
3.3
0.0011
NEW
PROCEDURE
70
0.38
3.8
1.28
182
0.058
0.40
25
2370
0.53
0.031
0.061
0.192
66
12.3
6.7
0.130
210
23
3.3
0.087
3.7
1.07
440
0.033
0.50
23
167
6.4
63
3.3
0.0013

n*N/2
79
0.45
2.5
0.98
146
0.086
0.38
23
2130
0.55
0.021
0.106
0.398
93
13.7
5.8
0.149
130
28
3.2
0.094
2.7
1.12
330
0.028
0.34
38
209
7.0
60
3.9
0.0013
	 MODIFICATIONS OF
n*N UNIFORM
OIST.
119
0.26
2.7*
0.55**
109*
0.234
0.09**
26
1550*
0.76
0.017*
0.117
0.395
122
19.1
1.1**
0.092
110*
33
2.3*
0.047
1.0**
0.87
320*
0.003**
0.52
243
513
7.0
25"
2.8
0.0003*
69
0.39
3.8
1.30
187
0.069
0.41
27
2440
0.58
0.032
0.076
0.248
75
14.1
6.9
0.149
240
27
3.5
0.112
3.9
1.24
490
0.039
0.68
56
267
8.1
68
4.4
0.0018
RECOMMENDED NEW PRC
NORMAL SLOPE
OIST. CHANGE
70
0.38
3.8
1.27
180
0.054
0.40
25
2340
0.51
0.030
0.055
0.170
62
11.5
6.6
0.122
200
21
3.2
0.077
3.6
1.00
420
0.031
0.44
16
136
5.8
61
3.0
0.0011
73
0.39
3.8
1.29
187
0.063
0.42
26
2370
0.55
0.032
0.073
0.272
70
12.7
6.8
0.138
220
24
3.4
0.095
3.8
1.08
470
0.034
0.53
31
181
6.9
64
3.7
0.0016
i/~*fTAl f3C ^ ___
PAR AM.
METHOD
97
0.42
4.0
1.37
192
0.075
0.44
28
2540
0.68
0.036
0.077
0.263
76
14.2
7.2
0.148
270
31
3.5
0.113
4.1
1.20
540
0.036
0.67
39
256
7.6
71
4.0
0.0017

NQt«ANDQM
SAMPLING
77
0.58
4.4
1.46
236
0.143
0.46
32
2680
0.77
0.041
0.143
0.578
103
19.3
7.5
0.182
320
36
3.9
0.147
4.6
1.42
600
0.042
1.00
67
340
10.0
77
4.7
0.0017

-------
Table 8.  Continued
MATERIAL
MERCURY
ENDOSUCFAN
ARSENICH I 1)
CHLORDANE
CHLORDANE
WATER
FRESH
FRESH
FRESH
FRESH
SALT
N
9
9
8
a
a
LOWEST
FHAV
5.00
0.340
880
6.3
0.400
OLD
PROCEDURE
3.42
0.208
570
3.7
0.090
NEW
PROCEDURE
0.94
0.169
220
4.0
0.200

n=N/2
0.94
0.169
220
4.0
0.200

n*H
2.32
0.094»
730
4.6
0.378
;ATIONS TO RECOMMENDED NEW PRC
UNIFORM NORMAL SLOPE
DIST. DIST. CHANGE
1.79
0.248
410
5.3
0.352
0.73
0.144
170
3.6
0.162
1.12
0.172
250
4.0
0.278

PARAM.
METHOD
1.89
0.234
450
4.6
0.313

NOtttANOOM
SAMPLING
4.76
0.319
790
5.6
0.280
GEOMETRIC MEAN OF  RATIOS  OF  FAV  BY
MODIFIED PROCEDURE TO  THAT BY
RECOMMENDED PROCEDURE:

NUMBER OF DATA SETS  FOR WHICH  FA¥
BY MODIFIED PROCEDURE  DIFFERS  FROM
THAT BY RECOMMENDED  PROCEDURE  BY
MORE THAN A FACTOR OF  1.4:

NUMBER OF DATA SETS  FOR WHICH  FAV
BY MODIFIED PROCEDURE  DIFFERS  FROM
THAT BY RECOMMENDED  PROCEDURE  BY
MORE THAN A FACTOR OF  2.0:
0.91
1.00      1.01
0.89     1.23      0.92      1.08      1.25        1.58
                                 29
                                 13
                                                                                   17

-------
    interval with the lowest P if no interval has a P less Chan 0.05)  is  desig-
    nated Interval A.  The next highest nonempty interval is designated  Interval
    B.  The FAV is then computed as ln(FAV)=VA+(VB-VA)/(PB-PA)-(0.05-PA).

Criticisms of this procedure include:

(1) The formula PaR/N is positively biased as discussed earlier and chus
    results in negative bias in the FAV.

(2) The positive bias in the cumulative proportions is increased by using the
    maximum rank in an interval rather than the average rank, when there  is  more
    than one In(MAV) in the interval.

(3) Often only one In(MAV) is  in Interval A or Interval 1 or both, making the
    method quite sensitive to data variation.

(4) A linear relationship of P versus  V is assumed, which is equivalent to
    assuming a rectangular distribution; as discussed earlier this is
    contraind icated by the available data sets.

(5) The use of intervals is meant to causa pooling of ln(MAV)s which are
    indistinguishable; the interval width of 0.25 was selected because it is a
    typical value for the standard deviation of replicate acute toxicity tests;
    this value may not be appropriate  for all species and materials and  is
    strictly appropriate only when ln(MAV)s are based on only one  toxicity test.
    More importantly, this pooling method works effectively  only for Interval A,
    because tht starting point for Interval B is fixed by that  for Interval A
    and therefore does not necessarily, properly oool data in  the vicinity of
    Interval B.  In any event, this pooling serves no useful purpose except to
    prevent the slope for interpolations and extrapolations  from being
    inappropriately calculated based on two identical, or nearly identical,
    ln(MAV)s, a purpose which can be better served by routinely using more
    points to assess data trends.

(6) The use of intervals containing variable numbers of ln(MAV)s makes the
    method sensitive to minor changes  in the data set which may move ln(MAV)s
    into or out of intervals; this sensitivity can be quite  marked and can even
    be anomalous.   For example, in the SMAV data set for heptachlor in fresh
    water  (Table 1), Interval A would  contain the lowest two  SMAVs and Interval
    8  would  consist of the third lowest SMAV,  However, if the  lowest SMAV was
    just 6Z lower (for Example, because a new toxicity  test  for that species
    lowered  th« mean slightly), the two lowest SMAVs would be sufficiently
    separated to be in separate intervals, which wauid  become  the  new, and
    markedly different,  Intervals A and B.  The calculated FAV  would change  from
    0.52 to 0.83,  a change that not only is much larger  than  the change  in the
    SMAV that caused it (60% versus 6%), but also is in the  opposite direction
    to the change in the S>1AV.
                                       44

-------
 These  criticisms  are  sufficient  to  warrant  the  replacement of this old procedure
 with the  new  procedure  recomnended  above.   However, the two procedures generally
 do  not produce  markedly different results (Tables 7 and 8).  On the average,
 FAVs calculated using the old procedure  are only HZ lower than those calculated
 using  the new procedure  for  the  SMAV data sets  and 9Z  lower for the FMAV data
 sets.   Individual FAVs  were  within  a factor of  1.4 for over 75Z of che FMAV data
 sets and  over 85Z of  the SMAV data  sets  and within a factor of 2.0 for over 85X
 of  the  FMAV sets  and  94% of  SMAV data  sets.  Where differences are greater  than
 twofold,  comparison of  the FAVs  with the data sets does not clearly indicate
 that one  or the other of these procedures results in more questionable FAVs.
     The  consequences of modifying  the major features  of the recommended new
 procedure  were  also explored to determine how sensitive FAVs are  to such
changes.   If  the  sensitivity is  low, any objection to  compromises  or  approxima-
 tions  used in arriving  at the recommended new procedure are largely irrelevant,
because more  exact analysis  or different compromises (within reason)  would  have
 little effect in  practice.    If sensitivity  is high, the basis  for  the recommended
new procedure becomes more critical and  further examination is warranted.
     Three features of che recotn"l6nded new  procedure were modified co sngn  the
range over which  they could  reasonably be varied.  The assumed distribution was
changed to rectangular  and to normal.  The  subset size (n) was changed to N/2
and N.   The pcrcentile estimation method was changed to the graphical method
with slope formula LS-X, but still  with  P^^E(P(X^)), and  to the  parametric
method  (best linear unbiased estimate).  The FAVs for  these modifications are
 included  in Tables 7 and 8.   Also included  in these tables  are (a) the geometric
mean of the ratios of the FAV by each  modified  procedure  to that  by the
recommended procedure, (b)  the number  of data sets for which the  FAV  by  each
                                       45

-------
modified  procedure differs by more than a faccor of 1.4 from that by che
recommended  procedure,  and (c) che number of data sets for which the FAV by each
modified  procedure differs by more than a factor of 2.0 from that by the
recommended  procedure.
      Changes in  the  assumed distribution, in the percentile estimation method,
and  in  the subset size  to N/2 had only minor effects on results.  The geometric
means of  the ratios  of  the FAVs by these modifications to that by the
recommended  procedure were close to  1.0 (0.92-1.25),  For individual data sets,
FAVs  by these modifications differed by more than a factor of 1.4 from the FAVs
by the  recommended new  procedure for no more than 20% of the data sets and by
more  than a  factor of 2.0 for no more than 6t of the data sets.
      Modification of subset size to n"N, however, caused major changes.  The
geometric means  of the  ratios of the FAV by this modification to that by the
recommended  procedure were close to  1,0 (0.93 for SMAVs, 0.89 for iFMAVs), but
individual FAVs  changed by more than a factor of 1.4 for over 70% of the SMAV
daca  sets and for nearly SOX of the FMAV data sets and by more than a factor of
2.0 for about one-third of both the SMAV and FMAV data sets.  Such differences
do not demonstrate,  per se, thac this isodified procedure is less appropriate
than the recotnaended new procedure, but it does raise such a suspicion.  In
particular,   when n"N, an unusually large number of sets have a FAV that  is well
below both the lowest MAV and the FAV calculated by the recommended new
procedure.   Using the location and scale parameter estimates from the modified
procedure with n="N,   the fiducial probability that the lowest MAV could  be  so
high was evaluated for each data set; where this probability is  <0.20 a  single
asterisk is  placed next to the FAV for n"N and where the probability is  <0.10  a
double asterisk is used.  The frequency of these marked entries  suggests that
                                       46

-------
using  n"H  is  in  fact inappropriate.  This is directly rtLated to the  existence
of  statistically significant skewness in the data sets, positive skewness
resulting  in  inappropriately low FAVs when n"N and negative skewness  resulting
In  high  FAVs,
     A related observation that also contraindicates the use of n*N is  that,
when compared to the recommended new procedure and the diverse modifications
already  mentioned, th* modification with n*N both frequently produces the  lowest
FAV (for 20 SMAV data sets and 19 FMAV data sets) and frequantly produces  the
highest  FAV (for 14 SMAV data sets and 12 FMAV data sets).  Such frequent
occupation of both extremes is again due to the variable skewness of the sets
and is indicative of the error of assuming all data are equally useful  in
estimating low percentiles when distributional assumptions are not completely
met.   Furthermore, th«s« problem* with using n"S art not restricted to the
modification with n"{f already discussed.  Graphical methods with the other slope
formulas and other distributional assumptions (e.g., normal) were  tested using
n*N with similar results.  Likewise, the best linear unbiased method using n"N
and assuming a normal distribution showed similar problems.  (This latter method
is equivalent to the simple approach of calculating a sample mean  and unbiased
standard deviation (12) and estimating the fifth parcentile  as  lying 1.645
standard deviations below the m«an, -1.645 being th« fifth percentile of a
standard normal  distribution.)
     Another advantage of not using n"N is that certain semiquantitative data
can be used.   Acute tests on some materials with some species produce 'greater
than'   values because concentrations high enough to cause effects were not, or
could   not be used (because of solubility or time constraints).   Regardless of
the reason, because such data are usually for resistant species, they can
                                       47

-------
usually be used  if n-4, but, if n-N, either they must be excluded, thereby




biasing the data  set, or additional acute tests must be conducted, thereby




increasing costs.  Finally, it should be noted that the use of na4 does  not




constitute "not  using all  the data1, because ail data are used in setting  ranks




and cumulative probabilities and thus in selecting which four daca will  be used




explicitly in  final calculations; rather, the use of n-4 is more properly




interpreted as a  simple scheme of giving greater weight to those MAVs which




provide the most  information about the  fifth percentile.




     The consequences of nonrandom sampling can also be partly addressed here.




Assuming that  the available data sets somehow resulted  from the strictly




systematic sampling scheme discussed earlier, FAVs were calculated by assigning




cumulative probabilities P^»(R-0.5)/N to the ranked  ln(MAV)s  and  interpolating




between the two data with  P^ nearest 0.05 (or extrapolating using the lowest




two points if N<10), the interpolation being based on the assumption of a




triangular distribution.   The results of this exercise  are included  in the  last




columns of Tables 7 and 8  and indicate that higher FAVs are produced than by the




recommended new procedure, but the differences average  only about a  factor  of




1,5 for SMAV sets and 1.6  for FMAV sets and are less than a factor of 1.4 for




65Z of the SMAV sets and 55% of the FMAV sets.  Considering that  this alterna-




tive sampling scheme is so extreme, and chat therefore  a scheme with a more




realistic systematic component would produce results much nearer  those obtained




by the recommended new procedure, this  further suggests that  the  issue of the




sampling assumption is not of great importance.




     A final  point that should be emphasized is that, whether SMAVs  or FMAVs  are




used, the same conclusions are reached  regarding  the appropriate  attributes of
                                       48

-------
Che  procedure  for estimating  fifth percent!Lea.  Also, although  it does not bear




on Che recommendations made here, it is interesting to note that FAVs calculated




from SMAVs are similar to those calculated  from FMAVs,  The geometric mean of




the ratios of the FAV computed from FMAVs to that computed from  SMAVs, by the




recommended procedure, was 1.04.  The two FAVs differ by a factor of  1.4 for




only 12 sets, by a factor of  2.0 for only 6 sets, and by more than a  factor o£




2.8 for no sets.
                                       49

-------
                                   DISCUSSION

     The  recommended new procedure uses linear extrapolation or interpolation to

estimate  the  fifth percentile of a statistical population of mean acute values

(MAVs)  from which the  available MAVs are assumed to have been randomly obtained.

The  available MAVg are ranked from low to high and the cumulative probability

for  each  is calculated as PR*R/(N+l), where R * rank and N » number of MAVs in

the  sec,  Extrapolation or interpolation is based on an assumed linear

relationship  between>/P^ and  In(MAV), and uses only the four points with P^

closest to 0.05 because this subset  provides the most useful information

concerning the  fifth percentile.

     The  bases  for the new procedure are mostly mathematical, with some  input

from toxicological and practical considerations.  The FAV, however, is basically

a eoxicological value, and the acceptability of any calculation  procedure  to

toxicologists will be based on the acceptability of che resulting FAVs.  Most

aquatic toxicologists will judge the acceptability of an FAV by  comparing  it

with che  lowest MAVs in the data set, and thus it is quite appropriate that the

four MAVs with  estimated cumulative  probabilities closest to 0,05 be  given  the

most weight in  calculating the FAV";  in fact, Cue reconsn ended procedure is

largely a formalization of the way one would obtain a FAV by 'eyeballing'  the

data.

     An important property of the new procedure is that the  resulting FAV  is  not

very sensitive  to modifications in the procedure or slight changes in the  data

set.  A variety of calculation procedures all produced FAVs  that  were quite

similar for most data sets.   In addition, the recommended new  procedure  rarely

produced  either the highest or lowest of the FAVs obtained  with  the procedures

examined.   Another important property of a procedure is its  performance  with
                                       50

-------
data  sets  which contain apparent discontinuities in the Lower tail,   For




heptachlor,  lindane,  and  chlordane in salt water and chromium(VI) in fresh




water,  the lowest  SMAV  is  at  least a factor of 10 lower than the second lowest




SMAV  (Table  1).   In  addition,  for these  four and cadmium in fresh water, the




lowest  FMAV  is at  least a  factor of 10 lower than the second lowest  FMAV (Table




2).   Of chese, only  chromium(VI) in fresh water has a very large range of FAVs




in Tables  7  and 8.   Even  though the two  lowest MAVs are far apart, most of chese




daca  sets  seem to  provide  adequate information about the FAV because similar




FAVs  were  obtained using  a variety of procedures.  These examples support the




idea  that  the best approach to take toward calculating the FAV is to select a




procedure  that is best on  the  average and then use it with ail data sets, except




possibly in  extraordinary  cases.




      An unfortunate  aspect of  the methodology for calculating the FAV is  the




necessity  of extrapolating to  estimate the 0.05 cumulative probability  for small




data  sets; if extrapolations become too  great, the FAVs will be  suspect.  For




only  5  of  the 74 data sets is  the FAV more than a factor of 2 lower than the




lowest  MAV.  Thus, for the available data sets this procedure rarely




extrapolates much below the lowest value in the data set.




      Overall, the recommended  new procedure is the best of the procedures




examined,  regardless of whether the FAV  is calculated from SMAVs or FHAVs.  It




is a  straightforward procedure for interpolation or extrapolation based  on




fitting a  line to the most useful points.  It produces results similar  to and




usually intermediate to those  obtained by other reasonable procedures.   In




addition, the calculations are relatively easy to perform with the  aid  of a hand




calculator, as described  in Appendix 1.  The major weakness of this procedure  is
                                       51

-------
that ie as«uai«» the same degree of tailing  for  all daca  sets.   Fortunately,  for




most data sets the FAV is not very dependent on  the  assumed  degree  of  calling




and deviation from the assumed intermediate degree of  tailing  is  not too




critical.  Other procedures wauld suffer  as much or  more from  the same  or  other




weaknesses.
                                       52

-------
                                 REFERENCES






  1,    U.S.  EPA.   1980.  Federal Register.  45:79318-79379.  November 28.






  2.    Javitz,  H.  and J. Skurnick.  1980.  Analyses of relaCtonships among




       various  aquatic toxicity test: results.  Final Report, Task 9, EPA




       Contract 68-01-3887.






  3.    Kelt, M.   1981,  Letter to J. H. McCormtck.  U.S. EPA, Duluch, MN.






  4.    Sokal, R.R. and F.J. Rohlf.  1969.  Biometry^  Freeman, San Francisco.






  5.    Lloyd, E.H.   1952.  Least squares estimation of location and scale




       parameters  using order statistics.  Biometrika, 39:88-95.






 6.    Johnson, N.L. and S. Kotz.   1970.  Distributions in statistics:  Continuous




       unvariate distributions - 1.  Wiley, New York.






  7.    Box, G.E.P., W.G. Hunter, and J.S. Hunter.  1979.  Statistics for




       Experimenters.  Wiley, New York.






 8.    Cunnane, C.   1978.  Unbiased plotting positions - a review.  J_. Hydro 1.




       37:205-232.






 9.    Sprsnt, P,  and G.R, Dolby.   1980.  Response to query.  Biometrics,




       36:547-550.






10.   Barker, P., Y.C.  Soh, and R.J. Evans.  1988.  Properties of  the geometric




      mean functional relationship.  Biometrics, 44:279-281.






11.   David, F.N. and N.L. Johnson.  1954.  Statistical treatment  of censored




      data.   Part I:  Fundamental  formulae.  Biometrika, 41:228-240.






12.   Zar, J.H.   1974.   Biostatistical Analyses.  Prentice-Hall, Englewood




      Cliffs, New Jersey.




                                     53

-------
                                 APPEHDIX  1
A.
General Instructions for Recotsnended flew Procedure for FAV Calculat
                                                                       ion .
     1.   Based  on data  ate  size  (N), determine  four  ranks  (R) with cumulative
         probabilities  (?R»R/(N+O)  closest  to 0.05;  for N<60,  this  will be
         R-l through 4;  for 60
-------
 C.   Example Computer  Program  in  BASIC  Language for  Calculating  che  FAV

     10  REM THIS  PROGRAM  CALCULATES  THE  FAV  WHEN  THERE  ARE  LESS  THAN
     20  REM 59 MAVS  IN  THE  DATA  SET.
     30  X-0
     40  X2-0
     50  Y-0
     60  Y2-0
     70  PRINT "HOW MANY MAVS  ARE IN  THE  DATA SET?"
     80  INPUT N
     90  PRINT "WHAT ARE THE FOUR LOWEST  MAVS?"
     100  FOR R-l TO 4
     110  INPUT V
     120  X-X+LOGCV)
     130  X2-X2+(LOG(V))*(LOG(V))
     140  P-R/CN+1)
     150  Y2-Y2+P
     160  Y-Y+SQR(P)
     170  NEXT R
     180  S-SQR((X2-X*X/4)/(Y2-Y*Y/4))
     190  L-(X-S*Y)/4
     200  A-S*SQR(0.05)+L
     210  F-EXP(A)
     220  PRINT "FAV «  "F
     230  END
D.  Example Printout  from Program

    HOW MANY MAVS ARE IN THE DATA SET?
    ? 8
    WHAT ARE THE FOUR LOWEST MAVS?
    ? 6.4
    ? 6.2
    ? 4.8
    ? .4
    FAV » 0.1998
                                     55

-------