Professional Documents
Culture Documents
J. A. Hoekstra'
ABSTRACT
In this paper methods to estimate the median lethal concentration
(LC50) are reviewed. Estimation from data sets with limited occurrence
of partial response (>O%, <loo%) receives particular attention.
1. INTRODUCTION
2. METHODSBASEDON TOLERANCE
DISTRIBUTIONS
in which f(u) is the density function of the log tolerance in the population.
The probit and logit model assume a normal and logistic density, respec-
tively. Both functions are symmetric. Symmetry can be enhanced by using
an appropriate transformation. Mostly, the log transformation is chosen, as
is described here, but sometimes there may be reasons t o prefer another or
no transformation (Finney 1971).
Some asymmetric and more heavy tailed models are discussed by Mor-
gan (1988). Generally, these models make high demands on the data for
adequate estimation, and are not often applied in routine toxicity testing.
More wideiy used is the extension of the probit or the logit model with one
~~ ~ ~ ~~~~~ ~~ ~~~~
number tested 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10
number killed 0 0 1 0 0 0 0 0 0 0 0 0 9 7 1 0 1 0 1 0 1 0
Table 1. Toxicity of 2.4.5-tricholorophenol applied to the earthworm Eisenia andrei (C.A.M. Van Gertel, unpublished data).
142 J. A. HOEKSTRA
2.2 ESTIMATION
Maximum likelihood estimation (MLE) of the log(LC50) = p = -a/@
is the prevalent method. The estimates are obtained either numerically
(Finney 1971) or graphically (Litchfield and Wilcoxon 1949). The latter
method seems rather outdated at least if computer facilities are available.
An alternative estimation criterion is minimum x2 (Ashton 1972).
Confidence intervals can be constructed by applying Fiellers theorem
(see Finney 1978) or the delta-method (see, e.g.! Hamilton 1980). The first
method assumes a normal distribution for & and p. The second uses a normal
approximation for &/p itself, resulting in symmetric intervals. Goedhart
(1986) found for 2 investigated data sets that the assumption for & and p
was the better.
Maximum likelihood breaks down if there is just one concentration at
which partial kill occurs: the MLE of ,d is then infinite. (Computer programs
do not always warn against this situation. A convergence criterion can be
met before ,O causes overflow, in which case faulty estimates will be given.)
Williams (1986)points out that with one partial kill a confidence inter-
val for the LC50 can still be determined by likelihood methods: the interval
ESTIMATION
OF THE LC50, A REVIEW 143
is given by the set of values not rejected by a likelihood ratio test. The like-
lihood ratio test statistic is obtained by subtraction of the deviance under
the full model (both parameters estimated by MLE) from the deviance of
the model under the null hypothesis (fixed value of 1,/3 estimated by MLE).
In case of one partial kill the full model has zero deviance.
Other parametric methods to obtain confidence limits are the paramet-
ric bootstrap (Efron 1985) and Bayesian procedures (Racine et al. 1986).
The properties of parametric bootstrapping with quantal data are not well
investigated. Confidence intervals are likely to be liberal. Racine et al.
(1986) contrast their Bayesian methods with (the failure of) Fiellers method
in small data sets. However, Fiellers method should not be used in this sit-
uation. This is not to denounce Bayesian analyses: when prior information
is available, the advantages are as expected.
Sanathanan, Gade and Shipkowitz (1987) developed a trimmed logit
method in analogy with the trimmed Spearman-Karber method (see Section
3.1), to diminish the sensitivity of logit analysis to anomalous results in the
tails of the tolerance distribution.
3. TOLERANCE METHODS
DISTRIBUTION-FREE
METHOD
3.1 SPEARMAN-K~RBER
The method of Spearman-Karber (Spearman 1908; Karber 1931; Finney
1971) obtains an estimate of 1as a weighted average of the mid-points be-
tween successive log concentrations. The weights are the estimated tolerance
densities at the midpoints. Let X I < x2 < ..-< Zk be the log concentra-
tions and p 1 , p 2 , . . .,pk the observed proportions mortality. The number
of individuals tested at x, is denoted by n,. If p1 = 0 and pk = 1, the
Spearman-Karber estimator (SK) can be written as:
1=1
f(4.du
Equating the above estimator of the mean to the median lethal dose is correct
if the underlying tolerance distribution is symmetric.
144 J. A. HOEKSTRA
1 .o
W
ul 0.9
Z 0.8
0
a 0.7
cn
w
w 0.6
z 0.5
0
-
t- 0.4
L41
0 0.3
a 0.2
0
C
Y
a 0.1
0.0
LOG CONCENTRATION
tioned above, and are likely to be more efficient than the moving-average
method.
Stephan (1977) suggested that in absence of partial kills, the highest
concentration with complete survival and the adjacent lowest concentration
with complete mortality can be used as limits of an interval with confidence
level (1-2(1/2))x 100%. This procedure, however, has no statistically sound
basis.
4. COMPARISON
4.1 LITERATURE
The merits of MLE and minimum x2 have been discussed extensively
(Berkson 1980). Which method is best seems to depend very critically on the
design and the chosen criterion for performance (Berkson 1955; Engeman,
Otis and Dusenberry 1986; Hamilton 1979; James el al. 1984; Smith et al.
1984). Probit analysis has a computational disadvantage to logit analysis: in
many computer implementations the numerical approximation of the normal
distribution is poor in cases of very high or low percentage points.
Finney (1950, 1953) compared the performance of MLE, SK and MA
with respect to asymptotic variance and bias under the probit (1950) and
logit (1953) models. The experimental design was chosen to be an unlimited
series of equally spaced log concentrations each with n observations. The
asymptotics refer to n -+ 00. Important conclusions are:
0 SK is identical to logit-MLE (apart from discontinuity in the distribu-
tion of estimates).
0 MA can best be applied with the largest possible span, once a data set
has been obtained. It then becomes similar to SK.
0 If the design is asymmetrical about p, SK and MA have a bias indepen-
dent of n. The wider the spacing of the design points, the larger the
bias becomes.
0 The variance of all three estimators is strongly dependent on the dis-
tance from k t o the nearest design point.
Hamilton (1979) and James et al. (1984) carried out large simulation
studies t o compare a number of estimators under a variety of symmetric
OF THE LC50, A REVIEW
ESTIMATION 147
tolerance distribution models. The outcomes of the two studies differ re-
markably, especially for the trimmed SK-estimators. The difference is at-
tributable to the different position of the dose levels relative to p in the
two studies (Hoekstra 1989). The conclusions of the two simulation studies
are as follows. Hamilton (1979) recommends the 10% or 20% trimmed SK
for n = 10,20 because of their favourable MSE under a variety of tolerance
models. However, due to the position of p in the design the reported MSE’s
of heavily trimmed estimators are misleadingly small. James el al. (1984)
find that SK5% performs quite well. Heavier trimmed SK’s were never more
efficient than the untrimmed SK.
Miller and Halpern (1980) study asymptotic properties (n -+ 00, dif-
ference between design points -+ 0 and infinite number of design points) of
several robust estimators and find that SKlO% performs better than SK5%
for two very heavy tailed distributions but less for moderately contaminated
distributions.
Hamilton (1980), using the same design and tolerance distributions as
he did before, compared 95% confidence intervals obtained from logit-MLE
(delta method) and trimmed SK’s in a Monte Carlo study (n = 5,10,20).
The empirical coverages never fell below 91% for any of the methods. With
strongly contaminated steep tolerance distributions MLE, SK, SK5% and
SKlO% all resulted in conservative confidence intervals, MLE giving the
widest intervals in cases where all methods could be applied (>1 partial
kill).
Stephan (1977) advocates the moving average method because of its
calculability in cases when p1 # 0 or P k # 1 (unlike the untrimmed SK
which would need an undesirable fabrication of data) and when there is
just one concentration with partial kill (unlike MLE). However, he does not
consider the efficiency of MA. Engeman et al. (1986) compared MA with
probit and logit methods using 3 to 6 dose levels and n from 2 to 20 by
simulation. MA was clearly less efficient under the logit model, but did not
fail to produce an estimate as often as the other methods.
Williams (1986) also concludes that for small sample sizes MLEintervals
obtained by the Fieller method are conservative even under the logit model
itself. He uses designs with a total of 20 to 30 observations spread over 4 to
6 dose levels and different values of the design centre relative to p. The LR-
intervals were reported to be liberal, but rarely fell below 93% confidence
probability. The simulation study included models leading to substantial
percentages of data sets with one or no concentrations with partial kill.
Using Hamilton’s (1979) design and models, Kappenman (1987) com-
pares a kernel method with SK10%; the earlier performs better, especially
with large n (n = 30). Miller and Schmitt (1988) compare 2 kernel meth-
ods under several models, including non-symmetric ones, with the probit
method. The design involves 48 equidistant log concentrations with n = 1.
The kernel methods do considerably better under the non-probit models.
148 J. A. HOEKSTRA
4.2 ADDITIONAL
SIMULATION
STUDY
An additional study was carried out t o compare some methods in sit-
uations with few partial responses. The methods considered are MLE, LR,
SK and MA with span 8; the design was x = 1,2,. . .,10 with n = 10 at each
level. To avoid undue i n h e n c e of the position of p on the outcomes, p was
drawn from a uniform distribution between 5 and 6 for each of the 1000 sim-
ulated data sets independently. The underlying tolerance distribution was
the logistic distribution with p = 1 , 3 and 6. The average percentages re-
sponse at the 10 design points are given in Table 2. The results are presented
in Table 3.
For each method, all statistics are given with reference t o those data
sets on which the method could be applied. MLE failed in all cases with 0
or 1 partial kill, the other methods in cases with 0 partial kills. The results
are scaled with respect to the variance of the tolerance curve.
No difference of practical relevance exists between SK and MA8. This
agrees with Finney’s (1953) conclusion about the similarity of the 2 methods
in this type of situation. Bias played a negligible role in all methods. The
empirical confidence probabilities are all acceptably close t o the nominal
value. With a steep logistic curve or equivalently a wide spacing of the
concentrations, the methods SK and MA8 give a slightly liberal confidence
interval. Hamilton (1980) found the same with a smaller p ( p = 1.838) and
n = 5 . MLE is not generally applicable in this situation, and is conservative
even for the subset of data to which it can be applied. Fieller’s method
seems t o yield no improvement over the simpler delta-method.
OF THE LC50,A REVIEW
ESTIMATION 149
5 . CONCLUSIONS
REFERENCES
Ashton, W.D. (1972). The Logit Transformotion. New York: Hafner Publishing Co.
Barlow, R.E., D.J.Bartholomew, J.M. Bremner and H.D. Brunk (1972). Stotisticol Znfer-
ence Under Order Restnctions. New Yolk Wiley.
Bennett, B.M. (1952), “Estimation of LD50 by moving averages”. Journal of Hygiene 50,
157-164.
Bennett, B.M. (1963), “Optimum moving averages for the estimation of medium effective
dose in bioassay”. Journol of Hygiene 61, 401-406.
Berkson, J. (1955), “Maximum likelihood and minimum X’ estimates of the logistic func-
tion”. Journal of the American Stotisticul Association 50, 130-162.
Berkson, J. (1980), “Minimum chi-square, not maximum likelihood!” Annals of Stotistics
0 , 447-487.
Efron, B. (1985), “Bootstrap confidence intervals for a class of parametric problems”.
Biometriko 7 2 , 45-58.
Engenman, R.M.,D.L. Otis and W.E. Dusenberry (1986), “Small sample comparison of
Thompson’s estimator t o some common bioassay estimators”. Journol of Stotiaticol
Computotion ond Sirnulotaon 25, 237-250.
Finney, D.J. (1950), “The estimation of the mean of a normal tolerance distribution”.
Sonkhyo 10, 341-360.
Finney, D.J. (1953), “The estimation of the ED50 for a logistic response curve”. Sonkhyo
12, 121-136.
Finney, D.J. (1971). Prubit Anolysis, 3rd edition. Cambridge: The University Press.
Finney, D.J.(1978). Stotistiurl Methods in Biologicol Assoy, 3rd edition. London: Charles
Griffin.
Finney, D.J. (1985), “The median lethal dose and its estimation”. Archives of Toticology
56, 215-218.
Goedhart, P.W. (1986), “Statistical analysis of quantal assay data”. IT1 - T N O rapport,
nr. 86, IT1 A 31.
Govindarajulu, Z. (1988). Stotisticol Technaques an Biwssoy. New York: Karger.
Hamilton, M.A. (1979), ‘Robust estimates of the ED50”. Journol of the Americon Stotis-
ticol Associotion 74, 344-354.
Hamilton, M.A. (1980), “Inference about the ED50 using the trimmed Spearman-Kkber
procedure - a Monte Carlo investigation”. Communicotions an Stotistics B9, 235-
254.
Hamilton, M.A., P.C. Russo and R.V. Thurston (1977), “Trimmed Spearman-Kirber
method for estimating median lethal concentrations in toxicity bioassays”. Enuiron-
mentol Science ond Technology 11, 714-719. Correction (1978), 12, 417.
Hoekstra, J.A. (1987), “Acute bioassays with control mortality”. Woter, Air and Soil
Pollution 35, 311-317.
Hoekstra, J.A. (1989), “Estimation of the LC50”. Letter to the editor. Biornetrics 45,
337-338.
Hoel, D.G. (1980), “Incorporation of background in dose-response models”. Fedemtion
Proceedings 39, 73-75.
James, B.R., K.L. James and H. Westenberger (1984), “An efficient R-estimator for the
ED50”.Journol of the American Statistical Association 79, 164-173.
Kalbfleisch, J.D., D.R. Krewski and J. Van Ryzin (1983), “Dose-response models for time-
to-response toxicity data (with discussion)”. Conodion Journal of Stotistics 11,25-49.
Kappenman, R.F. (1987), “Nonparametric estimation of dodleresponse curves with appli-
cation to ED50 estimation”. Journal of Stotisticol Computotion ond Simulotion 28,
1-13.
Kirber, G. (1931). “Beitrag zur kollektieven Behandlung pharmakologischer Reihenver-
suche”. Archiu f i r Ezpenmentelle Pothologce und Phormokologte 162, 480-487.
Kooijman, S.A.L.M. (1981), “Parametric analyses of mortality rates in bioassays”. Woter
Reseorch 15, 107-119.
Kooijman, S.A.L.M. (1983), “Statistical aspects of the determination of mortality rates in
bioassays”. Water Reseorch 17, 749-759.
Litchfield, J.T. and F. Wilcoxon (1949), “A simplified method of evaluating doseeffect
experiments”. Journol of Phormocology ond Ezperimentol Therapeutics 96, 99-113.
152 J . A. HOEKSTRA
Miller, R.G. and J.W. Halpern (1980), ‘Robust estimator for quantal bioaasay”. Biometriko
67, 103-110.
Moore, D.F. (1987), “Modelling the extraneous variance in the presence of extra-binomial
variation”. Applied Stotistics 36, 8-14.
Morgan, B. J.T. (1988). ‘Extended models for quantal response data”. Statistic0 Neer-
londico 42, 253-273.
Miller, H.-G. and Th. Schmitt (1988), ‘Kernel and probit estimates in quantal bioassay”.
Joumol of the Americun Stotisticol Asaociotion 83, 750-759.
Racine, A., A.P. Grieve, H. Flibler and A.F.M. Smith (1986), ‘Bayesian methods in prac-
tice: Experiences in the pharmaceutical industry with Discussion”. Applied Stotistics
36, 93-150.
Sanathanan, L.P., E.T. Gade and N.L. Shipkowitz (1987), “Trimmed logit method for
estimating the ED50 in quantal bioassay”. Biometrics 43, 825-832.
Smith, K.C., N.E. Savin and J.L. Robertson (1984), ‘A Monte Carlo comparison of max-
imum likelihood and minimum chi-square sampling distributions in logit analysis”.
Biometrics 40, 471-482.
Spearman, C. (1908), ‘The method of ‘right and wrong cases’ (‘constant stimuli’) without
Gauss’s formulae”. British Journol of Psychology 2, 227-242.
Stephan, C.E. (1977), “Methods for calculating an LC50”. In Aguotic tozicology ond hoz-
ord euoluotion, F.L. Mayer and J.L. Hamelink (eds), 65-84. Philadelphia: American
Society for Testing and Materials.
Thompson, W.R. (1947), ‘Use of moving averages and interpolation to estimate median-
effective dose - I - Fundamental formulas, estimation of error and relation to other
methods”. Bocteriologwol R e v ~ e w s11, 115-145.
Williams, D.A. (1982), “Extra-binomial variation in logistic linear models”. Applied Stotis-
tics 31, 144-148.
Williams, D.A. (1986), ‘Interval estimation of the median lethal dose”. Biometrics 42,
641-646.