| Summary: | This report introduces a trend test for binary data that accommodates both treatment-affected survivability and clustering within treatment groups. The test is motivated by chronic rodent carcinogenicity assays that begin exposure in utero and continue exposing postweaning siblings at the same dose level as their dams. The new test modifies the Poly-3 trend test introduced by Bailer and Portier1 to include clustering by adjusting the variance estimate of the lifetime incidence rate of findings. The weighted least squares linear regression approach to the Cochran-Armitage test with weights equal to the inverse of the variance is used to determine the initial statistic. Since sparse findings are common in low-dose groups and may be present in higher dose groups, the variance estimate is pooled across dose groups following Bieler and Williams2 to increase robustness. The new method was first evaluated with simulated data using distributional models for tumor onset and mortality1 with sibling correlation added through copulas. The simulations show that in the absence of positive sibling correlation, the false positive rate and power are similar for the Poly-3 test and the Poly-3 test modified for sibling correlation. However, with positive sibling correlation, the false positive rate is lower using the modified Poly-3 test than with the Poly-3 test. The two methods are also compared using real data from a National Toxicology Program perinatal chronic study, and the results reinforce the conclusion that failing to account for sibling correlation sometimes leads to inflated statistical significance. Keywords: Littermates, siblings, rat, chronic toxicology testing, developmental carcinogenicity, statistical analysis of tumor counts, Poly-3 test, cluster analysis, Rao-Scott
|
|---|