Abstract:
|
Our previous work demonstrated that nonlinear hierarchical models (NLHMs) offers superior power, relative to traditional approaches, for detecting a true difference in response to infection in experimental infection (EI) studies. However, in small-sample simulations, a classical maximum likelihood (ML) approach to NLHM fitting and inference yielded poor control of the type I error rate. This is a potentially serious obstacle to adoption of the approach in practice, as small sample sizes are common in EI studies. Here, we investigate two approaches to improving small-sample type I error control when applying the NLHM to EI studies: a bootstrapped omnibus test in the ML framework and a Bayesian NLHM. We demonstrate via simulation that both the Bayesian and bootstrap refinements improve the type I error rate of the NLHM, with the Bayesian approach performing slightly better for sample sizes of 10 or less. The two approaches differed in their power to detect differences in intensity, duration and time to peak response. We also demonstrate the use of these methods on real data from actual EI studies.
|