Causal nonresponse models for repeated categorical measurements

Area of Science:

• Statistics
• Biostatistics
• Econometrics

Background:

• Missing data in repeated categorical outcomes poses analytical challenges.
• Nonrandom missingness requires specialized statistical approaches.
• Existing methods like log-linear models have limitations in parameter interpretability.

Purpose of the Study:

• To extend conditional likelihood procedures for repeated categorical data with nonrandomly missing values.
• To develop a statistically rigorous method for handling complex missing data patterns.
• To provide interpretable parameters within a causal modeling framework.

Main Methods:

• Utilized causal models for nonresponse, extending conditional likelihood procedures.
• Developed an approach analogous to log-linear models but with enhanced interpretability.
• Applied methods to examples with binary responses and covariates, including a complex case.

Main Results:

• The proposed causal model extension allows for direct interpretation of parameters related to outcome distributions.
• Computations are feasible using standard statistical software, similar to log-linear approaches.
• A simulation study demonstrated the properties of the estimates derived from the new method.

Conclusions:

• Causal models provide a robust framework for addressing nonrandomly missing data in repeated categorical outcomes.
• The developed method enhances parameter interpretability compared to existing log-linear approaches.
• The approach is practical for application using readily available statistical software.