ETD

• attrition
• inverse probability weighting
• longitudinal panel
• MAR
• MCAR
• missing mechanism
• MNAR
• multiple imputation
• pattern-mixture models
• weighting

This thesis addresses the problem of weighting in panel studies in the presence of attrition, analysing how different statistical approaches perform under three missingness mechanisms: Missing Completely At Random (MCAR), Missing At Random (MAR) and Missing Not At Random (MNAR). Using a simulated longitudinal dataset with monotone dropout over time, three main methods are tested: Inverse Probability Weighting (IPW), Multiple Imputation (MI) and Pattern-Mixture Models (PMM). The results show that under MCAR all methods provide unbiased estimates; under MAR, IPW reduces bias, while MI may fail when high variability in the outcome variable is present; under MNAR, neither IPW nor MI can correct the distortion introduced by the missingness, whereas PMM prove useful for sensitivity analysis. Overall, this work highlights that the efficiency of the weighting methods depends on the correct specification of the response model, and that in the presence of non-ignorable missing mechanisms additional assumptions are required to produce reliable estimates.