The Impact of Missing Risk Factor Data on Semiparametric Group-Based Trajectory Models

Abstract

To investigate how missing data (Missing Completely at Random [MCAR] vs. Missing Not at Random [MNAR]) on risk factors can impact trajectory solutions (i.e., latent class probabilities) and coefficient estimates capturing the relationship between covariates and trajectory group solutions using a semiparametric group-based trajectory modeling (GBTM) approach. To address this issue, we conducted a systematic investigation using Monte Carlo simulation. Data were generated from a population with known growth parameters and risk factors. Observations for risk factors were then systematically deleted in a way that reflects key missing data assumptions (MCAR and MNAR). Models were then estimated to test the sensitivity of the estimates to each missing data scenario. Two key findings emerged: (1) trajectory solutions were largely unaffected by missing data on risk factors; and, (2) there was some degree of bias in estimating relationships between risk factors and trajectory group membership when data were missing on those risk factors. GBTM may be useful for testing etiological explanations of long-term patterns of offending. Missing data on risk factors poses a threat to this approach, however.