A causal inference perspective on the analysis of compositional data.

Area of Science:

• Epidemiology
• Biostatistics
• Causal Inference

Background:

• Compositional data, representing parts of a whole, are common in epidemiology.
• Previous analyses of compositional data challenges are extended within a formal causal framework.
• Directed acyclic graphs (DAGs) are utilized to formally represent and analyze compositional data.

Purpose of the Study:

• To depict compositional data using DAGs and identify distinct causal effect estimands.
• To explore the total and relative effects in three specific compositional data scenarios.
• To clarify the interpretation and identifiability of different causal effects in compositional data analysis.

Main Methods:

• Compositional data were represented using directed acyclic graphs (DAGs).
• Two distinct effect estimands, total and relative effects, were identified.
• Analysis involved three example scenarios: economic activity/GDP, fat consumption/body weight, and sedentary time/body weight.

Main Results:

• Both total and relative effects can be identifiable and causally meaningful for certain compositional data scenarios (e.g., GDP, fat consumption).
• In specific scenarios (e.g., sedentary time), only the relative effect is identifiable.
• The relative effect in compositional data analysis represents a joint effect requiring careful interpretation.

Conclusions:

• Directed acyclic graphs (DAGs) provide a valuable tool for analyzing causal effects in compositional data.
• Researchers must explicitly define and interpret the specific causal effect of interest in all compositional data analyses.
• Clear declaration of the sought causal effect and its interpretation is crucial for rigorous research involving compositional data.