Missing Data in Intensive Longitudinal Suicide Research: A Monte Carlo simulation study

Background: Ecological momentary assessment (EMA) studies are increasingly used to study the dynamic psychological processes that underlie affect and suicidal ideation (SI). Although these provide a promising directly for modeling those relationships, they often include substantial missing data and use estimators that rely on data being missing completely at random (MCAR) or missing at random (MAR) as underlying assumptions. However, missing data may be driven by systematic relationships between these psychological processes and disengagement.Methods: We conducted a Monte Carlo simulation study evaluating six estimators of the marginal mean of SI. We generated realistic EMA data under a 4 x 3 x 2 factorial design varying missingness mechanism (MCAR, MAR, missing not at random [MNAR] driven by momentary symptoms, MNAR driven by latent disengagement states), target compliance (40%, 60%, 80%), and study complexity (standard vs. volatility-dependent missingness). Each condition comprised 500 replications of a 21-day EMA protocol with N = 300 participants completing 4 prompts per day. Estimator performance was evaluated via bias, root mean standard error, and empirical confidence interval coverage relative to the actual ground truth estimand.Results: Under MCAR, all estimators exhibited negligible bias, low RMSE, and near-complete coverage. Under MAR, bias remained modest, but coverage degraded at lower compliance, particularly under the complex study type. Under both MNAR mechanisms, all estimators, including doubly robust approaches, exhibited substantial negative bias (~20-65% underestimation of SI), inflated RMSE, and near-zero empirical coverage, even under high (80%) compliance. Conclusions: Conventional missing data estimators may fail to recover valid inferences regarding sample prevalence when compliance is dependent on latent and temporally emergent psychological processes. Even under high compliance with doubly-robust estimators, inferences may be biased under realistic symptom-dependent missingness.