Reflection on modern methods: causal inference considerations for heterogeneous disease etiology

Area of Science:

• Epidemiology
• Molecular Pathology
• Biostatistics

Background:

• Molecular pathological epidemiology investigates disease mechanisms and risk factor effects.
• Evaluating heterogeneity of risk factor effects across disease subtypes is a common research goal.

Purpose of the Study:

• To identify potential biases in causal effect estimation when using multinomial regression for disease subtype analysis.
• To explain the mechanism of selection bias in this context.

Main Methods:

• Analysis of multinomial regression as a series of logistic regressions.
• Explanation of bias using directed acyclic graphs (DAGs).
• Demonstration of bias magnitude via hypothetical data and simulation studies.

Main Results:

• Multinomial regression, commonly used for disease subtype analysis, can introduce selection bias.
• This bias can prevent the recovery of true causal effects, even when all confounders are measured.
• The bias arises from the inherent structure of comparing subtypes to controls.

Conclusions:

• The standard multinomial regression approach in molecular pathological epidemiology may not yield valid causal inferences for disease subtypes.
• Researchers should be aware of and account for potential selection bias in such analyses.
• Alternative methods may be needed to accurately assess risk factor effects across disease subtypes.