Marginal regression models with a time to event outcome and discrete multiple source predictors.

Area of Science:

• Biostatistics
• Epidemiology
• Psychiatry

Background:

• Multiple informants are crucial for assessing psychopathology.
• Previous research shows varied relationships between self-report, physician report, and mortality prediction.

Purpose of the Study:

• To develop and compare a maximum likelihood (ML) approach with generalized estimating equations (GEE) for marginal regression models using multiple informants and time-to-event outcomes.
• To evaluate the performance of ML and GEE in the presence of missing data.

Main Methods:

• Fitting marginal regression models to predict mortality using self-report and physician report of psychiatric disorders from the Stirling County Study.
• Developing an ML approach and comparing it to GEE for handling multiple informant covariates.
• Investigating the impact of missing data (monotone and non-monotone) on both ML and inverse probability weighted (IPW) GEE methods through simulation.

Main Results:

• ML estimates can match GEE estimates in simple saturated models.
• Inverse probability weighted (IPW) GEE shows minimal efficiency loss compared to ML with monotone missingness.
• ML offers a modest variance decrease over IPW GEE for non-monotone missingness, particularly for covariate estimation.
• ML likelihood parameters may not directly interpret as GEE parameters in more general settings.

Conclusions:

• GEE is recommended for fitting marginal models due to its flexibility, ease of interpretation, and comparable efficiency to ML with missing data.
• The choice between ML and GEE may depend on the specific data structure and analytical goals, particularly concerning parameter interpretation and missing data patterns.