Abstract

Ecological Momentary Assessment (EMA) has become an essential tool in psychological research due to its capacity to capture real‑time experiences within adolescents’ natural environments. By minimizing recall bias and providing high‑resolution temporal data, EMA offers a more accurate understanding of fluctuations in emotions, behaviors, and contextual factors as they unfold in daily life. However, EMA studies face important methodological challenges, particularly missingness in assessments. Adolescents may fail to complete prompts due to burden, disengagement, or situational constraints, generating nonrandom missing data patterns that can bias inferences and compromise study validity. Understanding the mechanisms underlying both study dropout and within‑study missing assessments is therefore crucial for improving EMA design, implementation, and the robustness of subsequent analyses, including the choice of appropriate imputation techniques.
Using data from an initial cohort of 115 Swiss adolescents, each invited to complete three EMA waves—each wave consisting of 21 consecutive days with four daily assessments—we investigate missing data processes at two levels. First, leveraging the fact that 108 participants completed the first EMA wave, we examine between‑wave attrition by linking baseline demographic, psychological, and behavioral characteristics to the probability of discontinuing in subsequent waves. This analysis identifies predictors of dropout across the longitudinal EMA design.
Second, among adolescents who remain active within a given wave, we analyze the moment‑to‑moment missingness mechanism. We model the probability of missing an EMA prompt as a function of temporal structure, including time of day, day within the 21‑day sequence (capturing temporal trends), and an autoregressive component reflecting the influence of the previous prompt’s completion status. To characterize these processes, we employ Generalized Linear Mixed Models (GLMM) as well as Hidden Markov Models (HMM) for missingness states. This two‑level framework enables exploration of potential Missing Not at Random (MNAR) mechanisms both across waves (study attrition) and within waves (prompt‑level missingness).