Methods for Hypothesis Testing in Animal Carcinogenicity Experiments

Thesis (Ph.D.)--University of Washington, 2016-08

Animal carcinogenicity experiments are conducted by private entities and government agencies to investigate whether a substance causes cancer. Since most tumors are occult and it is necessary to conduct a necropsy to detect the presence of cancer, producing unbiased tests for differences in tumor incidence is challenging. The highest doses given in these experiments often have toxic effects in the animals or might affect tumor lethality, and many common statistical methods for survival data perform poorly when such effects induce differences in mortality across treatment groups. The poly-k test, developed by Bailer & Portier (1988) as a modification of the Cochran-Armitage test, was designed to avoid problems from differential mortality, but is based on strong parametric assumptions, specifically that underlying tumor hazard may be modeled as Weibull with shape parameter k (usually set to 3). Existing literature that examines the performance of the poly-k and competitor tests assumes Weibull tumor hazards, and finds that the poly-k can be biased under treatment toxicity when its shape parameter assumption is not met. Given that this test has become a standard for government agencies such as the National Toxicology Program and the FDA, closer examination is warranted. Our simulations examine the performance of the poly-k under non-Weibull tumor hazards, and find that our own parametric tests can outperform it under these conditions. For tests like the poly-k, our goal is to develop and examine the performance of adaptive testing algorithms that estimate a test's parameter based on the data; one such estimation was suggested under Weibull assumptions by Moon et al. (2003), though their approach based on estimating lifetime cumulative tumor incidence rates does not make full use of the available data, and requires that the experiment include multiple interim sacrifices for adequate performance. Under most of our simulation settings, our poly-k_MLE test, based on estimating the shape parameter k by maximizing the full likelihood, has type I error and power very similar to the poly-k test with correctly-specified k, and it maintains size better, with comparable or higher power, than the test based on the Moon et al. estimate. Our MLE-based test does not require interim sacrifices, although it may perform better under some interim sacrifice experimental designs. We compare these tests under a variety of parametric assumptions and serial sacrifice designs to examine which experimental settings are most optimal for test performance.