Answers are grounded in this paper and cite their sources.
Intensive longitudinal designs are increasingly popular for assessing moment-to-moment changes in mood, affect, and interpersonal or health behavior. Compliance in these studies is never perfect given the high frequency of data collection, so missing data are unavoidable. Nonetheless, there is relatively little existing research on missing data in the specific context of intensive longitudinal data. The few studies that exist tend to focus on methods appropriate for data that are missing at random (MAR). However, missing not at random (MNAR) data are prevalent in these designs, particularly when the interest is sensitive topics like mental health, substance use, or sexual behavior. As a motivating example, we consider a study on people with binge eating disorder that has large amounts of missingness in a self-report item related to overeating. Missingness may be high because participants felt shame reporting engaging in this behavior, which is a clear case of MNAR and for which methods like multiple imputation and full information maximum likelihood are much less effective. To improve handling of such MNAR intensive longitudinal data, we propose embedding a Diggle-Kenward selection model into the dynamic structural equation modeling framework. This approach is straightforward to apply in popular software like Mplus, only requires a few extra lines of code relative to models that assume MAR, and can be fit expediently. We contrast results from our proposed approach to models that assume MAR and conduct a simulation study to systematically study performance of the proposed model with continuous or binary outcomes.
No takes yet. Share an insight, caveat, or question.
Daniel McNeish (2024) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: