Causal analysis of individual change using the difference score.

Area of Science:

• Epidemiology
• Longitudinal Studies
• Biostatistics

Background:

• Analyzing changes in health or disease over time is crucial in epidemiology.
• Difference scores from two-wave longitudinal data are commonly used as outcomes in regression models.
• Causal analysis of change requires careful consideration of confounding and individual trajectories.

Purpose of the Study:

• To demonstrate how causal analysis of change can be validly performed using simple linear regression models.
• To highlight the critical adjustments needed for confounding bias and individual growth curves.
• To provide a framework for clearly stating and scrutinizing assumptions in causal inference for change.

Main Methods:

• Utilizing simple linear regression models with continuous difference scores as the outcome.
• Implementing adjustments for confounding bias.
• Accounting for the shape of individual growth curves to model change over time.
• Illustrating the approach with data from the Whitehall II study.

Main Results:

• The type of individual growth curve significantly influences whether age or start score should be included in the regression model.
• Adjustments for confounding and growth curves rely on untestable assumptions based on prior theory.
• Solely relying on observed associations with the difference score can lead to misleading causal inferences.

Conclusions:

• Valid causal analysis of change necessitates explicit theoretical assumptions for adjustments, which cannot be solely data-driven.
• The proposed framework allows for clear articulation and rigorous scientific evaluation of these assumptions.
• This method enhances the reliability of epidemiological studies analyzing change over time.