Sensitivity analysis for non-monotone missing binary data in longitudinal studies: Application to the NIDA...

Area of Science:

• Biostatistics
• Longitudinal Data Analysis
• Substance Use Disorder Research

Background:

• Conventional "worst value" imputation in substance use disorder (SUD) trials assumes missing data equals drug use, leading to biased treatment effect estimates.
• This deterministic approach fails to account for the uncertainty and varying likelihood of drug use when data is missing.

Purpose of the Study:

• To develop and present a novel probabilistic imputation method for handling non-monotone missing binary data in longitudinal SUD studies.
• To replace biased deterministic imputation with a more accurate approach that incorporates prior beliefs about missing data.

Main Methods:

• A joint model combining a not missing at random (NMAR) selection model with a generalized linear mixed model for longitudinal binary data.
• The NMAR selection model links missing outcomes to observed outcomes using a sensitivity parameter.
• The mixed model utilizes bridge distributions for random effects, providing subject-specific and marginal interpretations.

Main Results:

• The proposed probabilistic method offers a transparent and statistically sound alternative to conventional imputation techniques.
• This approach allows for sensitivity analysis by explicitly stating assumptions about missing data.

Conclusions:

• Probabilistic imputation is superior to deterministic "worst value" imputation for analyzing missing data in SUD trials.
• The presented joint modeling framework provides unbiased estimates and a basis for robust sensitivity analyses in longitudinal studies.