Abstract: Treatment Effects in Randomized Longitudinal Experiments with Different Types of Non-Ignorable Dropout

Treatment Effects in Randomized Longitudinal Experiments with Different Types of Non-Ignorable Dropout Manshu Yang and Scott E. Maxwell University of Notre Dame In medical and psychological research, randomized longitudinal designs are commonly used, in which participants are randomly assigned to a treatment or control group and measured repeatedly over time. In these longitudinal experiments, missing data problems are inevitable and if data are missing not at random (MNAR), treatment effects estimated from traditional linear mixed-effects models might be severely biased. In such cases, an alternative approach is to use pattern-mixture models. The current study compared the traditional mixed-effects model approach and the pattern-mixture model approach under (1) two types of non-ignorable missingness, namely, random-slope-dependent dropout (missingness depends on latent random slope) and outcome-dependent dropout (missingness depends on observed outcome variables); (2) same missing pattern or different missing patterns across groups; and (3) same missing proportion or different missing proportions across groups. The situation of different missing patterns across groups describes a scenario where participants from the two groups are missing for different reasons (e.g., participants from the control group drop out because their random slopes are higher than average whereas participants from the treatment group drop out because of their lower than average random slopes). Results indicate that under random-slope-dependent dropout, treatment effect estimates were generally unbiased in pattern-mixture models, whereas estimates from mixed-effects models were severely biased if missing patterns differ across groups. Under outcome-dependent dropout, treatment effect estimates from both pattern-mixture models and mixedeffects models were biased when missing patterns differ across groups. Patternmixture models, however, outperformed mixed-effects models with a smaller magnitude of bias. Furthermore, the bias in both types of models can be attenuated by increasing the reliability of random slope measurement. Correspondence concerning this abstract should be addressed to Manshu Yang, 118 Haggar Hall, Department of Psychology, University of Notre Dame, Notre Dame, IN 46556. E-mail: myang@