Missing genotype data in a candidate gene association study can make it difficult to model the effects of multiple genetic variants simultaneously. In particular, when regression models are used to model phenotype as a function of SNP genotypes in several different genes, the most common approach is a complete case analysis, in which only individuals with no missing genotypes are included. But this can lead to substantial reduction in sample size and thus potential bias and loss in efficiency. A number of other methods for handling missing data are applicable, but have rarely been used in this context. The purpose of this paper is to describe how several standard methods for handling missing data can be applied or adapted to this problem, and to compare their performance using a simulation study. We demonstrate these techniques using an Alzheimer's disease association study. We show that the expectation-maximization algorithm and multiple imputation with a bootstrapped expectation-maximization sampling algorithm have the best properties of all the estimators studied.