Arch Toxicol (1985) 56 : 215-218

Toxicology
9 Springer-Verlag 1985

Archives of

Original investigations
The median lethal dose and its estimation
D. J. Finney
Department of Statistics & AFRUS, University of Edinburgh, Scotland, UK

Abstract. A n important paper by Zbinden and Fluri-Roversi (1981) has shown the many weaknesses in any policy or regulatory system that regards an estimated LDs0 in an animal species as an adequate guide to toxicity in man. The present paper draws attention to some statistical aspects of LDs0 estimation that are too often neglected or misunderstood when this quantity is wanted. It is solely concerned with practice when a LDs0 must be estimated, and deliberately does not approach the broader issues of whether the LDs0 should be estimated. A first need is clear distinction between the true but unknown form of dependence of mortality on dose and the estimate of it (or of a particular property such as the LDs0) that is obtainable from an experiment. Some assumptions are necessary before any estimation is possible. The graphical and semi-graphical methods that once were popular because of their simplicity and speed are today only reasonable as a last resort, when data are wholly inadequate and all that can be found is a very rough preliminary indication. Many "simple" arithmetical methods have been shown to be inherently bad, in that equally simple alternatives are usually more precise and less subject to bias. The Spearman-K/irber method remains as a useful possibility, demanding little knowledge of the form of the response curves but often needing other unverifiable assumptions. For most purposes, maximum likelihood estimation of a parametric formulation of the response curve is the best choice, not only because of theoretical merits but also because it can now be performed on a microcomputer in a very few seconds.
Key words: LDs0 - Dose-response - Efficient estimation Bias - Maximum likelihood

science nor good sense, yet it appears to be a policy widely adopted by national regulatory authorities. This paper is concerned neither with the general question of the practical usefulness of LDs0 estimates nor with the increasing agitation against large-scale exploitation of animals. The author has neither the competence nor the wish to criticize the comprehensive case that Zbinden and Flury-Roversi have so well presented. His purpose is solely to examine some statistical aspects of how to estimate LDs0 from toxicity tests where the need for an estimate is agreed (or an estimate is demanded, wisely or not, by a regulatory authority). No one will question that, in circumstances to which Zbinden and Flury-Roversi's objections do not apply, if the LDs0 is wanted it should be obtained as efficiently as possible in respect of number of test animals used and in relation to optimal extraction of information from the data. Moreover, the arguments against naive interpretation of a single LDs0 often apply less strongly to a "relative potency", a value expressed relative to that of a similar substance adopted as a standard. Zbinden and Flury-Roversi, however, make some statements about estimation that are questionable on statistical grounds, and clarification of these issues seems desirable. Chanter and Haywood (1982) have recently supported the case for abandoning uncritical reliance on LDs0, and have mentioned some of the statistical issues.

2. The basic experiment

The standard form of experiment for estimating LDs0 involves first choosing a series of doses (amounts, concentrations, etc.) of the substance, at each of which tests will be made with all other conditions held constant. Usually, and for good reason, the doses are in a geometric sequence (such as 0.08, 0.12, 0.18, 0.27, . . .) so that the logarithms of dose are equally spaced, but this is not essential. At a dose, n subjects are exposed (in a manner that permits each to react independently of the others), and r, the number dying, is recorded: p (= r/n), the fraction dying, is calculated. Usually, and for good reason, the value of n is the same at every dose, but this is not essential. A n experiment with a geometric series of doses and a constant n will be said to have a symmetric design.

1. Introduction
Zbinden and Flury-Roversi (1981) have discussed many of the logical objections to reliance on "the LDs0" as a single measure of the toxicity of therapeutic and other substances. In particular, they have emphasized that LDs0 depends nor merely upon species of test animal but also upon sex, age, strain, and other factors within species, as well as upon diet, manner of administration, factors of the test environment, etc. To regard even a precise estimation of LDs0 under specific conditions as a guide to the amount of the substance that can safely be administered to or ingested by man is neither good

3. Estimation and assumptions

If all n were very large (say n = 10,000), values of p should approximate closely to the corresponding true mortality rates, P. A n ideal true response curve would be that of P plotted

Offprint requests to: D. J. Finney

216 against dose; if that curve were known sufficiently exactly (by evaluation of P at very many different doses), the LDs0 could be read from it as the dose for which P = 0.50. Of course in practice n is much smaller (perhaps 50 or 10 or even 5), and the true curve can be estimated only imprecisely from values o f p calculated from these small n at only a few different doses. The logical distinction betweenp and P is vital to the discussion that follows. No inference about the LDs0 is possible without some assumption on the dependence of P on dose. The weakest usable assumption is that P increases continuously as dose increases, with P = 0 at zero dose and P = 1 at some sufficiently high dose. (Modifications can take account of a threshold dose such that at all lower doses P = 0, mortality from other causes that makes P > 0 at zero dose, or a proportion of totally resistant subjects that cause P to approach some limit less than 1 at high doses; these will not be discussed beyond the comment that any such additional complexities make the argument of this paper apply more forcefully.) In practice, a stronger asumption is usually made, namely that the curve relating P to logarithm of dose is symmetric above and below the LDs0. More precisely, if 0 is the true LD50 and a any number greater than 1.0, and if the values of P for doses O/a, aO are P1, P2, then P2 = 1 - P I Most methods of estimating 0 subsequently discussed depend explicitly or implicitly on this assumption, which has some empirical support at least as an approximation. The common choice of a geometric sequence of doses is allied to the belief that the dependence of response on dose will take a simpler form with a logarithmic dose scale. equation (4.1), the reader should understand that this is solely for convenience of presentation: its merits and its weaknesses are shared with (4.2), and indeed with other analogous equations. Equation (4.1) is a strong assumption. The dependence of P on dose is asserted to depend on only two unknown parameters, a and r , but there is the inherent plausibility of a smooth and continually increasing relation. Substantial evidence from experiments indicates that many sets of experimental data are in good agreement with it, and that serious deviations are more often due to non-independence of the reactions of subjects at a dose than to systematic departure from the smooth equation. What alternatives can be suggested? Any formulation of the dependence of P on dose must have a qualitative similarity to (4.1), (4.2). Of course, far more complicated equations could be proposed, involving more than two unknown parameters (as is required for complexities mentioned in Section 3), but, without larger sets of data than toxicologists are normally prepared to obtain, neither the need for such equations can be demonstrated nor the parameters (and hence the LDs0) estimated.

5. Estimation
With minimal assumptions about the dose-response curve, scope for estimation of LDs0is severly rest ricted. One possibility is to plot the observed proportions, p, against dose (or, better, log dose), and either join the points by straight lines or draw a smooth curve by eye. Linear segments, with marked changes of direction at each plotted p, correspond to an assumption far more unrealistic than (4.1) or (4.2). Freehand drawing used to be attractive as a speedy alternative to arithmetic, but to-day's easy and cheap computation totally reverses this. The subjectivity of estimation from a freehand curve is a further major objection, unless the response curve is so plainly asymmetric that an equation such as (4.1) obviously will not do and the freehand approach is adopted as a last resort. If the standard experiment described in Section 2 relates to a dose-response curve specifid in terms of unknown parameters, such as (4.1) or something more complicated, the accepted principle for estimating the parameters (and thence the LDs0 ) is m a x i m u m likelihood. The essential iterative calculations were described in 1935 (Bliss; see also Finney 1947; when results have been compared with those from the alternative of m i n i m u m )~ estimation (Finney 1978), they are usually scarcely distinguishable. Maximum likelihood is known to have many optimal properties in respect of efficient use of data. Until recently, the laboriousness of its computations caused it to be regarded as a tool for statisticians but unacceptable to most biologists. Today, any scientific computer service should have software for handling very general maximum likelihood problems. Perhaps even more relevant, even an inexpensive desk-top microcomputer can easily be programmed for all estimation relevant to equation (4.1); it will then obtain an estimate of LDs0 with a small fraction of the time and effort that any graphical method would need. Zbinden and Flury-Roversi imply that the method originated by Litchfield and Wilcoxon (1949) is still widely used. Good though this was when maximum likelihood computations were laborious, it is now totally inappropriate since it was devised solely as a quick approximation to maximum likelihood. This comment applies with equal force to other graphical methods proposed around 1940-1950. If (4.1) or

4. The dose-response function

Two widely used more exact assumptions about P have been (4.1) and P = 1/[1 + exp ( - 2 Y)I, where
Y = a + fix,

(4.2)

(4.3)

x is the logarithm of the dose, and a, fl are parameters determining a particular response function as one of the family corresponding to all possible pairs of a, /3. The historical origins of these matter little; suffice it to say that (4.1) corresponds to a normal distribution for the logarithms of doses just large enough to kill whereas (4.2) adopts the logistic relation that is regarded as an appropriate dose-response equation in many other biological contexts. Despite the marked logical distinction, rarely if ever is knowledge of the mechanism of toxicology such as to make one of them clearly preferable; moreover, it is easily demonstrated that very numerous data would be needed before one could appear empirically a much better fit than the other. Although hereafter frequent reference will be made to the use of

217 (4.2) is applicable, the best of the computationally simple methods is the Spearman-Kfirber; although this is perhaps less dependent upon an assumed equation, it commonly requires unsupported assumptions about doses outside the range of those tested and today can well be abandoned in favour of maximum likelihood. Other methods once popular Read-Muench, Dragstedt-Behrens, and the moving average method (Thompson 1947) commended by Zbinden and Flury-Roversi - have been proved to have serious weaknesses in respect of bias, and to be almost always inferior (practically and theoretically) to Spearman-K~irber (Finney 1950, 1953). If assumptions are limited to that of continuous increase in P as dose increases, objective estimation of the LDs0 is scarcely possible. Spearman-Kgrber is still usable but nothing can be said about its properties. When values of n are fairly large (say 50 or greater), simple interpolation between two consecutive doses for which the calculated values o f p lie on either side of 0.5 may suffice. Even this will be unsatisfactory if consecutive doses are so far apart that P increases very markedly over the interval between them. For smaller n, graphical procedures seem the only choice: the truth is that the experiment is too small for its task. Dixon and Mood (1948) proposed a different experimental approach, with subjects tested in a temporal sequence such that the dose for each subject is determined by the response (death or survival) of its predecessor. Obviously, this is limited in applicability by considerations of time lapse between administration and response and by practical convenience, but it does offer some hope of reducing the number of subjects tested. Since 1948, many statisticians have studied and developed modifications of the so-called up-and-down techniques, most recently Kershaw (1983). When the nature of the testing makes it practicable, one of these probably should be adopted, but it is now recognized that the reduction in number of subjects needed for equivalent precision is not tremendous. A n analogous procedure for testing subjects one at a time at a sequence of doses was proposed by Deichmann and Le Blanc (1943), with the claim that a good approximation to the LD50 could be obtained with as few as six subjects. Although Zbinden and Flury-Roversi write favourably of the procedure, all it does is "to provide a logical (not necessarily the optimal) basis for choosing doses; even the best possible choice with only six subjects will provide very limited information on the LDs0 unless subjects are exceedingly sensitive to small changes in the dose of the substance under test. A n experiment conducted in this way is no more than a simple form of that described in Section 2, with six doses and n = 1. One feature of maximum likelihood estimation is that, in relation to an assumed dose-response equation such as (4.1), it uses the data with maximum efficiency. Obviously judgement must depend upon what level of approximation to the LD50 is regarded as good enough; even with efficient estimation, an experiment like that of Section 2 rarely yields an adequate result unless it uses at least 30 subjects, and alternative estimation techniques can only make things worse. Unless the dose interval is chosen very successfully, Deichmann's scheme of six subjects and more informal estimation will at best give a very crude approximation (see below). Estimates of LDs0 in different laboratories, or even in the same laboratory on different occasions, can be widely discrepant, as Zbinden and Flury-Roversi among others have pointed out. Why then worry about the efficiency of a single estimation? A statistician can reply only that scientists continue to perform isolated experiments for this purpose, whether to meet regulatory requirements or to satisfy a personal interest. When they do so, good experimental planning costs no more than bad, and today efficient estimation can be as cheap and easy as any of the alternatives that have been proposed (some merely inefficient, some wrong in that they estimate a different property). Possibly science would be better served by cooperative efforts planned to include several laboratories and a range of controlled conditions; although this approach would demand many more subjects in all, it should eventually lead to greater understanding. Possibly the LDs0 as an indicator of toxicity should be abandoned in favour of some different method of quantitative characterization; this may be desirable, but it is likely to demand more fundamental biological knowledge rather than arbitrary definition. Such issues deserve discussion among toxicologists and statisticians. This paper has the more modest aim of emphasizing that, as long as the LD50 is used, there is no excuse (scientific or economic) for not estimating according to some accepted criterion of optimality. Zbinden and Flury-Roversi rightly insist that toxic properties of a chemical compound are not fully expressed by statement of the LDs0. They draw attention to a suggestion by Molinengo (1979) for estimating the LDs0 from a curve relating survival time to dose. Unfortunately, this involves definition of LDs0 as the dose killing 50% in infinite time and thereby introduces serious confusion with the natural mortality; moreover Molinengo's method misinterprets a regression by uncritical interchange of independent and dependent variates, and in other respects departs from sound statistical principles. The quantity eventually obtained may indeed sometimes be numerically close to the LDs0 as more usually defined, but exactly what property it estimates cannot readily be stated ; it is certainly not a valid alternative estimate of the LDs0 and in some circumstances it could be very different. Another interesting reference from Zbinden and Flury-Roversi is that to Cobb and Grimshaw (1979). Their paper illustrates well the additional information on symptomatology and on characteristics of subjects that is needed for thorough understanding of toxicity. However, to state that Cobb and Grimshaw were able to "characterize comprehensively" (a phrase of uncertain meaning) the acute intoxication syndrome by using only ten monkeys is surely an exaggeration. Many measurements on blood and urine samples were made on up to seven occasions over 2 weeks and very full histological studies were made on all animals. Nevertheless, the results on these few monkeys, distributed among four doses of diquat, are too irregular to amount to a clear characterization. One may suspect that only a more extensive planned experiment, on a greater number of subjects and perhaps with a factorial structure, will provide any full explanation. If estimating the LDs0 was an important objective, the experiment was certainly inadequate. Indeed, the Cobb and Grimshaw data provide an interesting commentary on Deichmann's approach. They recorded death rates of 1/2, 0/4, 1/2, 2/2 at doses 100, 200, 300, 400 mg/kg. Maximum likelihood indicates a value of about 270 mg/kg for the LDs0 , but the data scarcely suffice to put any limits on 0. Even though the doses are spaced slightly more widely than Deichmann recommended, the experiment gives almost negligible information on the location of the LDs0; in this instance, his ratio of 1.5 between successive doses was evidently much too small "to preclude, practically, the possibility of killing an animal with one dose, while failing to kill with the next higher dose".

218

6. A statistical disclaimer
T h e r e is n o f o u n d a t i o n in statistical t h e o r y or principle for believing t h a t e x p r e s s i o n of dose as " m g / k g " (or in similar units) e n a b l e s t h e toxic effect to b e r e g a r d e d as i n d e p e n d e n t of species! This o u g h t to b e obvious, b u t so f r e q u e n t l y e s t i m a t i o n of LDs0 in a l a b o r a t o r y m a m m a l s e e m s to b e t h o u g h t a basis for c o n f i d e n t a s s e r t i o n a b o u t t h e LDs0 in m a n . A n y m e a n s of t r a n s l a t i n g dose effects f r o m o n e species to a n o t h e r must rest n o t o n a r b i t r a r y a s s u m p t i o n s b u t o n biological/pharmacological theory. Z b i n d e n a n d F l u r y - R o v e r s i h a v e amply illustrated t h e wide discrepancies t h a t occur b e t w e e n species.

References
Bliss CI (1935) The calculation of the dosage-mortally curve. Ann Appl Biol 22:134-167 Chanter DO, Haywood R (1982) The LDs0 test: Some considerations of precision. Toxicol Lett 10:303-307 Cobb LM, Grimshaw P (1949) Acute toxicity of oral diquat (1,1'-ethylene - 2,2'-bipyridinium) in Cynomolgus monkeys. Toxicol Appl Pharmacol 51:277-282 Deichmann WB, Le Blanc TJ (1943) Determination of the approximate lethal dose with about six animals. J Industrial Hygiene Toxicol 25 : 415-417

Dixon WJ, Mood AM (1948) A method for obtaining and analyzing sensitivity data. J Am Statist Assoc 43: 109-126 Finney DJ (1947) Probit analysis. London: Cambridge University Press (3rd edition 1973) Finney DJ (1950) The estimation of the mean of a normal tolerance distribution, Sanhky~ 10:341-360 Finney DJ (1953) The estimation of the EDs0 for the logistic response curve. Sankhyh 12:121-136 Finney DJ (1978) Statistical method in biological assay. London: Charles Griffin & Co. Ltd. (3rd ed) Kershaw CD (1983) Sequential estimation procedures for binary response. Ph. D. Thesis, University of Edinburgh Litchfield JT, Wilcoxon F (1949) A simplified method of evaluating dose-effect experiments. J Pharmacol Ther 96: 99-113 Molinengo L (1979) The curve doses vs survival time in the evaluation of acute toxicology. J Pharm Pharmacol 31:343-344 Thompson WR (1947) Use of moving averages and interpolation to estimate median effective dose. Bacteriol Rev 11:115-145 Zbinden G, Flury-Roversi M (1981) Significance of the LDs0-test for the toxicological evaluation of chemical substances. Arch Toxicol 47: 77 - 99

Received December 21, 1983/Accepted October 29, 1984

Note added in proof." Dr Tang Zhung-Ming (Institute of Radiation Medicine, Beijing, People's Republic of China) has written a good short program in BASIC that should be easily adapted to any microcomputer.