Monte-Carlo Simulation of Correlated Binary Responses

Abstract

Simulation studies can provide powerful conclusions for correlated or longitudinal response data, particularly for relatively small samples for which asymptotic theory does not apply. For the case of logistic modeling, it is necessary to have appropriate methods for simulating correlated binary data along with associated predictors. This chapter presents a discussion of existing methods for simulating correlated binary response data, including comparisons of various methods for different data types, such as longitudinal versus clustered binary data generation. The purposes and issues associated with generating binary responses are discussed. Simulation methods are divided into four main approaches: using a marginally specified joint probability distribution, using mixture distributions, dichotomizing non-binary random variables, and using a conditionally specified distribution. Approaches using a completely specified joint probability distribution tend to be more computationally intensive and require determination of distributional properties. Mixture methods can involve mixtures of discrete variables only, mixtures of continuous variables only, and mixtures involving both continuous and discrete variables. Methods that involve discretizing non-binary variables most commonly use normal or uniform variables, but some use count variables such as Poisson random variables. Approaches using a conditional specification of the response distribution are the most general, and allow for the greatest range of autocorrelation to be simulated. The chapter concludes with a discussion of implementations available using R software.