The LAGDA generates three forms of data to be statistically analysed: (1) Quantitative continuous data, (2) Time-to-event data for developmental rates (Time to NF stage 62) and (3) Ordinal data in the form of severity scores or developmental stages from histopathology evaluations. The recommended statistical analysis decision tree for the LAGDA is shown in Figure 1. Also, some annotations which might be needed to conduct statistical analysis for the measurements from the LAGDA are indicated below. For the analysis decision tree, the results of measurements for mortality, growth (weight and length) and liver-somatic-index (LSI) should be analysed according to the “Other endpoints” branch.
Data for continuous endpoints should first be checked for monotonicity by rank transforming the data, fitting to an ANOVA model and comparing linear and quadratic contrasts. If the data are monotonic, a step-down Jonckheere-Terpstra trend test should be performed on replicate medians and no subsequent analyses should be applied. An alternative for data that are normally distributed with homogeneous variances is the step-down Williams’ test. If the data are non-monotonic (quadratic contrast is significant and linear is not significant), they should be analysed using a mixed effects ANOVA model. The data should then be assessed for normality (preferably using the Shapiro-Wilk or Anderson-Darling test) and variance homogeneity (preferably using Levene’s test). Both tests are performed on the residuals from the mixed effects ANOVA model. Expert judgment can be used in lieu of these formal tests for normality and variance homogeneity, though formal tests are preferred. If the data are normally distributed with homogeneous variance, then the assumptions of a mixed effect ANOVA are met and a significant treatment effect is determined from Dunnett’s test. Where non-normality or variance heterogeneity is found, then the assumptions of Dunnett’s test are violated and a normalising, variance stabilising transform is sought. If no such transform is found, then a significant treatment effect is determined with a Dunn’s test. Whenever possible, a one-tailed test should be performed as opposed to a two-tailed test, but it requires expert judgment to determine which is appropriate for a given endpoint.
Mortality data should be analysed for the time period encompassing the full test and should be expressed as proportion that died in any particular tank. Tadpoles that do not complete metamorphosis in the given time frame, those tadpoles that are in the larval sub-sample cohort, those juvenile frogs that are culled, and any animal that dies due to experimenter error should be treated as censored data and not included in the denominator of the percent calculation. Prior to any statistical analyses, mortality proportions should be arcsin-square root transformed. An alternative is to use the step-down Cochran-Armitage test, possibly with a Rao-Scott adjustment in the presence of overdispersion.
Males and females are not sexually-dimorphic during metamorphosis so larval sub-sampling growth data should be analysed independent of gender. However, juvenile growth data should be analysed separately based on genetic sex. A log-transformation may be needed for these endpoints since log-normality of size data is not uncommon.
Liver weights should be normalised as proportions of whole body weights (i.e., LSI) and analysed separately based on genetic sex.
Time to metamorphosis data should be treated as time-to-event data, with any mortalities or individuals not reaching NF stage 62 in 70 days treated as right-censored data (i.e. the true value is greater than 70 days but the study ends before the animals had reached NF stage 62 in 70 days). Median time to NF stage 62 completion of metamorphosis in dilution water controls should be used to determine the test termination date. Median time to completion of metamorphosis could be determined by Kaplan-Meier product-limit estimators. This endpoint should be analysed using a mixed-effects Cox proportional hazard model that takes account of the replicate structure of the study.
Histopathology data are in the form of severity scores or developmental stages. A test termed RSCABS (Rao-Scott Cochran-Armitage by Slices) uses a step-down Rao-Scott adjusted Cochran-Armitage trend test on each level of severity in a histopathology response (Green et al., 2014). The Rao-Scott adjustment incorporates the replicate vessel experimental design into the test. The “by Slices” procedure incorporates the biological expectation that severity of effect tends to increase with increasing doses or concentrations, while retaining the individual subject scores and revealing the severity of any effect found. The RSCABS procedure not only determines which treatments are statistically different from controls (i.e., have more severe pathology than controls), but it also determines at which severity score the difference occurs thereby providing much needed context to the analysis. In the case of developmental staging of gonads and reproductive ducts, an additional manipulation should be applied to the data since an assumption of RSCABS is that severity of effect increases with dose. The effect observed could be a delay or acceleration of development. Therefore, developmental staging data should be analysed as reported to detect acceleration in development and then manually inverted prior to a second analysis to detect a delay in development.
Green JW, Springer TA, Saulnier AN, Swintek J. 2014. Statistical analysis of histopathology endpoints. Environmental Toxicology and Chemistry 33, 1108-1116.