Logistic regression with incompletely observed categorical covariates--investigating the sensitivity against violation of the missing at random assumption.

Missing values in the covariates are a widespread complication in the statistical inference of regression models. The maximum likelihood principle requires specification of the distribution of the covariates, at least in part. For categorical covariates, log-linear models can be used. Additionally, the missing at random assumption is necessary, which excludes a dependence of the occurrence of missing values on the unobserved covariate values. This assumption is often highly questionable. We present a framework to specify alternative missing value mechanisms such that maximum likelihood estimation of the regression parameters under a specified alternative is possible. This allows investigation of the sensitivity of a single estimate against violations of the missing at random assumption. The possible results of a sensitivity analysis are illustrated by artificial examples. The practical application is demonstrated by the analysis of two case-control studies.