51 Appendix 6, Regulation (EU) 2017/735

In forcesince 14 Feb 2017

The observations used to fit a model are replicate means (length and weight) or replicate proportions (egg hatch and larval mortality) (OECD 2006).

Weighted regression using replicate sample size as weight is generally advised. Other weighting schemes are possible, such as weighting by predicted mean response or a combination of this and replicate sample size. Weighting by reciprocal of within-concentration sample variance is not recommended (Bunke et al. 1999, Seber and Wild, 2003, Motulsky and Christopoulos 2004, Huet et al. 2003).

Any transformation of responses prior to analysis should preserve the independence of the observations and ECx and its confidence bounds should be expressed in the original units of measurement, rather than in transformed units. For example, a 20 % change in the logarithm of length is not equivalent to a 20 % change in length (Lyles et.al 2008, Draper and Smith 1999).

The flowchart in figure 4 gives an overview for ECx estimations. The details are described in the text below.

For egg hatch and larval mortality, it is generally best to fit a decreasing model unless one is fitting a probit model as described below. That is, one should model the proportion of eggs that do not hatch or larvae that die. The reason for this is that ECx refers to a concentration at which there is a change equal to x % of the control mean response. If there are 5 % control eggs that fail to hatch and one models failure to hatch, then EC20 refers to a concentration at which there is a change equal to 20 % of the 5 % control failure to hatch, and that is a change of 0,2 × 0,05 = 0,01 or 1 percentage point to 6 % failure to hatch. Such a small change cannot be estimated in any meaningful way from the data available and is not biologically important. Whereas if one models the proportion of eggs that hatch, the control proportion would be 95 % in this example and a 20 % reduction from the control mean would be a change of 0,95 × 0,2 = 0,18, so from 95 % hatch success to 77 % (= 95 – 18) hatching success and that effects concentration can be estimated and is presumably of greater interest. This is not an issue with size measurements, though adverse effects on size generally mean a decrease in size.

Except for the Brain-Cousens hormetic model, all of these models are described and recommended in OECD (2006). What are called OECD 2-5, are also discussed for ecotoxicity experiments in Slob (2002). There are, of course, many other models that might be useful. Bunke, et al. (1999) lists numerous models not included here and references to other models are plentiful. Those listed below are suggested as particularly appropriate in ecotoxicity experiments and widely used.

Bruce-Versteeg

Simple Exponential (OECD 2)

Exponential with shape parameter (OECD 3)

Simple Exponential with Lower Bound (OECD 4)

Exponential with shape parameter and lower bound (OECD 5)

Michaelis-Menten

Hill

Where there is visual evidence of hormesis (unlikely with egg hatch success or larval survival, but sometimes observed in size observations)

Brain-Cousens Hormetic; Brain and Cousens (1989)

Increasing models for these responses can be fit by probit (or logistic) models if there is no evidence of extra-binomial variance and control incidence is estimated in the model fit. This is not the preferred method, as it treats the individual, not the replicate, as the unit of analysis (Morgan 1992, O'Hara Hines and Lawless 1993, Collett 2002, 2003).

Visually compare observed and predicted percent decrease at each test concentration (Motulsky and Christopoulos 2004, Draper and Smith 1999).

Compare regression error mean square against the pure error mean square using an F-test (Draper and Smith 1999).

Check that every term in the model is significantly different from zero (i.e., determine whether all model terms are important), (Motulsky and Christopoulos 2004).

Plots of residuals from regression vs. test concentration, possibly on a log(conc) scale. There should be no pattern to this plot; the points should be randomly scattered about a horizontal line at zero height.

The data should be evaluated for normality and variance homogeneity in the same way as indicated in Appendix 5.

In addition, normality of the residuals about the regression model should be assessed using the same methods indicated in Appendix 5 for the residuals from ANOVA

Use Akiake's AICc criteria. Smaller AICc values denote better fits and if AICc(B)-AICc(A)≥10, the model A is almost certainly better than model B (Motulsky and Christopoulos (2004).

Compare the two models visually by how well they meet the single model criteria above.

The parsimony principal is advised, whereby the simplest model that fits the data reasonably well is used (Ratkowsky 1993, Lyles et.al 2008).

The confidence interval (CI) for ECx should not be too wide. Statistical judgment is needed in deciding how wide the confidence interval can be and ECx still be useful. Simulations for regression models fit to egg hatching and size data show that about 75 % of confidence intervals for ECx (x = 10, 20 or 30) span no more than two test concentrations. This provides a general guide for what is acceptable and a practical guide for what is achievable. Numerous authors assert the need to report confidence intervals for all model parameters and that wide confidence intervals for model parameters indicate unacceptable models (Ott and Longnecker 2008, Alvord and Rossio 1993, Motulsky and Christopoulos 2004, Lyles et al. 2008, Seber and Wild 2003, Bunke et al. 1999, Environment Canada 2005).

The CI for ECx (or any other model parameter) should not contain zero (Motulsky and Christopoulos 2004). This is the regression equivalent the minimum significant difference that is often cited in hypothesis testing approaches (e.g. Wang et al 2000). It also corresponds to the confidence interval for the mean responses at the LOEC not contain the control mean. One should wonder whether the parameter estimates are scientifically plausible. E.g. if the confidence interval for y0 is ± 20 %, no EC10 estimate is plausible. If the model predicts a 20 % effect at a concentration C and the maximum observed effect at C and lower concentrations is 10 %, then the EC20 is not plausible (Motulsky and Christopoulos 2004, Wang et al. 2000, Environment Canada 2005).

ECx should not require extrapolation outside the range of positive concentrations (Draper and Smith 1999, OECD 2006). For example, a general guide might be for ECx to be no more than about 25 % below the lowest tested concentration or above the highest tested concentration.

Alvord, W.G., Rossio, J.L. (1993); Determining confidence limits for drug potency in immunoassay, Journal of Immunological Methods 157, 155-163.

Brain P. and Cousens R. (1989); An equation to describe dose responses where there is stimulation of growth at low doses. Weed res. 29: 93-96.

Bunke, O., Droge, B. and Polzehl, J. (1999). Model selection, transformations and variance estimation in nonlinear regression. Statistics 33, 197-240.

Collett, D. (2002); Modelling Binary Data, second edition, Chapman and Hall, London.

Collett, D. (2003); Modelling Survival Data in Medical Research, second edition, Chapman and Hall, London.

Draper, N.R. and Smith, H. (1999); Applied Regression Analysis, third edition. New York: John Wiley & Sons.

Environment Canada (2005); Guidance Document on Statistical Methods for Environmental Toxicity Tests, Report EPS 1/RM/46

Huet, S., A. Bouvier, M.-A. Poursat, E. Jolivet (2003); Statistical Tools for Nonlinear Regression: A Practical Guide with S-PLUS and R Examples, Springer Series in Statistics, New York.

Lyles, R. H., C. Poindexter, A. Evans, M. Brown, and C.R. Cooper (2008); Nonlinear Model-Based Estimates of IC50 for Studies Involving Continuous Therapeutic Dose-Response Data, Contemp Clin Trials. 2008 November; 29(6): 878–886.

Morgan, B.J.T. (1992); Analysis of Quantal Response Data, Chapman and Hall, London.

Motulsky, H., A. Christopoulos (2004); Fitting Models to Biological Data Using Linear and Nonlinear Regression: A Practical Guide to Curve Fitting, Oxford University Press, USA.

O'Hara Hines, R. J. and J. F. Lawless (1993); Modelling Overdispersion in Toxicological Mortality Data Grouped over Time, Biometrics Vol. 49, pp. 107-121

OECD (2006); Current approaches in the statistical analysis of ecotoxicity data: A guidance to application. Series Testing and Assessment, No. 54, Organisation for Economic Co-operation and Development, OECD, Paris.

Ott, R.L., M.T. Longnecker, An Introduction to Statistical Methods and Data Analysis, sixth edition, 2008, Brooks-Cole, Belmont, CA

Ratkowsky, D.A. (1993); Principles of nonlinear regression, Journal of Industrial Microbiology 12, 195-199.

Seber, G.A.F., C.J. Wild, Nonlinear Regression, Wiley, 2003

Slob W. (2002); Dose-response modelling of continuous endpoints. Toxicol. Sci., 66, 298-312

Wang, Q., D.L. Denton, and R. Shukla (2000); Applications and Statistical Properties Of Minimum Significant Difference-Based Criterion Testing In a Toxicity Testing Program, Environmental Toxicology and Chemistry, Vol. 19, pp. 113–117, 2000.