WHO IS GOING TO MISS US? UNDERSTANDING ATTRITION IN MIXED-MODE LONGITUDINAL SURVEYS

This study examines the determinants and consequences of attrition in longitudinal surveys, with a focus on its implications for representativeness and causal inference. Selective attrition may introduce bias and compromise the validity of statistical inference. The study addresses this challenge by investigating whether nonresponse is ignorable under a Missing at Random (MAR) framework and evaluating the effectiveness of methodological and operational strategies to mitigate selection bias. Using data from the DZC Project, the analysis combines three complementary approaches. First, it assesses the relationship between attrition and observed individual characteristics, testing the plausibility of different missing data mechanisms. Second, it evaluates whether standard weighting adjustments are sufficient to restore representativeness across survey waves. Third, it examines the role of survey design, particularly the use of mixed-mode data collection strategies, in reducing selection bias. Additionally, the study distinguishes between different types of nonresponse—refusal, non-contact, and procrastination—and investigates the predictive power of baseline paradata for future attrition. The results indicate that attrition is not consistent with a Missing Completely at Random (MCAR) process but can be plausibly treated as MAR when conditioning on observed characteristics. Weighting strategies are generally effective in maintaining representativeness, although their performance varies across subgroups. Mixed-mode designs improve response rates and help recover harder-to-reach respondents, thereby reducing selection bias. Paradata emerge as valuable predictors of attrition, offering practical insights for cost-effective fieldwork planning. Overall, findings contribute to the methodological literature on missing data and provide evidence-based guidance for the design and implementation of longitudinal surveys in middle-income country contexts.