Composite variable bias: causal analysis of weight outcomes.

Area of Science:

• Epidemiology
• Biostatistics
• Causal Inference

Background:

• Composite variables, such as Body Mass Index (BMI) and change scores, are frequently used in research.
• Combining multiple variables into a single composite outcome can obscure the individual causal roles of component variables.
• Composite variable bias, previously noted for exposures, is also a concern for composite outcomes, potentially leading to misleading causal inferences.

Purpose of the Study:

• To illustrate composite variable bias in composite weight outcomes.
• To examine how different composite weight measures (BMI, weight change, BMI change, relative change) affect causal inferences.
• To provide recommendations for analyzing weight-related outcomes to avoid biased estimates.

Main Methods:

• Utilized data from the National Child Development Study (NCDS) cohort (n=9223).
• Estimated causal effects of baseline characteristics (ethnicity, sex, economic status, malaise score, height/weight) on weight-related outcomes at follow-up.
• Employed a directed acyclic graph (DAG) to visualize and address composite variable bias.

Main Results:

• Causal effect estimates varied significantly depending on the specific weight outcome analyzed.
• Analyses using follow-up BMI, weight change, BMI change, or relative body size change produced potentially divergent conclusions regarding interventions.
• The choice of composite weight outcome influences the interpretation of causal relationships.

Conclusions:

• This study is the first to demonstrate that causal estimates derived from composite weight outcomes can vary and be misleading.
• Recommends analyzing only follow-up weight, conditioned on baseline weight, for robust causal estimates.
• The method for conditioning on baseline weight depends on its temporal relationship with the exposure, with DAGs aiding strategy selection.