PARENT SESSION
Oral Session #84: Statistical Ecology.
Presiding: L. Subhash
Thursday, August 8. 9:00 AM to 11:30 AM. Mesquite Room, Radisson.

A comparison of statistical methods for analyzing paired and block desings with a binary response.

Minton, Mark^*,1, Evans, Marc¹, ¹ Washington State University, Pullman, Washington

ABSTRACT- Binary response data is commonly collected in ecological experiments. For binary response data arising from a completely randomized design, the common test of heterogeneous proportions is the most powerful inferential method. However, the choice of a method for analyzing binary response data from a randomized complete block design (paired data) is not clear. Alternative methods include the test for heterogeneous proportions, Cochran's Q, and generalized linear mixed models (GLMM). The latter is based on a logistic regression model incorporating a random effect for the blocking factor, similar to the standard ANOVA model for block designs. Several competing methods of estimation for GLMMs include: generalized estimating equations (GEE); direct maximization (DM); Markov chain Monte Carlo (MCMC); and expectation maximization (EM). Through Monte Carlo simulation we compared the power and nominal type I error rate of these methods for different sample sizes, block variation and treatment effects. For a type I error of = 0.05 the simulated type I error rate of these method were close to the nominal 5% value. SAS's Proc NLMIXED and GENMOD, which implement the DM and GEE methods, respectively, performed poorly as did the test for heterogeneous proportions. Consequently these procedures could yield unwarranted conclusions. However, our simulations indicated Cochran's Q and the GLMM based on the MCMC or EM methods provided the highest statistical power, and could be beneficial in the analysis of ecological data.

KEY WORDS: Binary Response, Block Designs, Monte Carlo Simulation , Statisitics