Missing data are common in many clinical studies. When missingness is nonignorable, a full likelihood analysis of the data requires incorporating a missing data model into the observed data likelihood function. In this article, we study the bias of the ML estimator when the corresponding maximum likelihood is obtained under a misspecified missing data model. We further explore a likelihood ratio statistic for testing the missing data mechanism in binary longitudinal data. The empirical level and power of the test are investigated in small simulations. We also present an example using some real data obtained from a longitudinal study.