Consequences of Model Misspecification for Maximum Likelihood Estimation with Missing Data