Biased and unbiased estimation in longitudinal studies with informative visit processes

Area of Science:

• Biostatistics
• Longitudinal Data Analysis
• Clinical Research Methodology

Background:

• Longitudinal studies are crucial for research, but data availability often depends on study characteristics.
• Clinical databases are increasingly used for longitudinal research, posing challenges due to non-random visit schedules.
• Visit times influenced by patient characteristics can lead to biased estimations compared to prospective studies.

Purpose of the Study:

• To investigate the impact of informative visit processes on parameter estimation in longitudinal studies.
• To quantify the bias introduced by visit times related to random effects in generalized linear mixed models.
• To identify which parameters are susceptible to bias and which remain consistently estimable.

Main Methods:

• Utilized generalized linear mixed models (GLMMs) to analyze longitudinal data.
• Incorporated a log link function to connect the visit process with random effects.
• Developed a framework to elucidate parameter bias under informative visit processes.

Main Results:

• Demonstrated that informative visit processes can significantly bias estimators for parameters linked to random effects.
• Showed that parameters not directly associated with random effects can still be consistently estimated.
• Quantified the degree of bias based on the relationship between visit timing and random effects.

Conclusions:

• Informative visit processes in longitudinal data present a significant challenge to accurate statistical inference.
• Careful consideration of visit time mechanisms is essential when analyzing longitudinal data from clinical databases.
• The proposed methodology allows for the identification of biased parameters, aiding in more reliable data interpretation.