Reflection on modern methods: generalized linear models for prognosis and intervention-theory, practice and...

Area of Science:

• Health Data Analysis
• Biostatistics
• Epidemiology

Background:

• Prediction (prognosis) and causal explanation (prevention/treatment) are often conflated in health research.
• This confusion can lead to inappropriate application and interpretation of data analysis models.
• Generalized linear models (GLMs) provide a framework to illustrate these distinctions.

Purpose of the Study:

• To differentiate between predictive and causal inference within the GLM framework.
• To identify specific areas where predictive and causal GLMs diverge.
• To provide recommendations for appropriate use of prediction and causal modeling in health research.

Main Methods:

• Utilizing the framework of generalized linear models (GLMs).
• Identifying five key differences between GLMs for prediction and GLMs for causal inference.
• Analyzing implications for machine learning (ML) methods.

Main Results:

• Five primary distinctions between predictive and causal GLMs were identified: covariate consideration, covariate selection methods, covariate selection and parameterization, model evaluation, and model interpretation.
• Failure to distinguish between these approaches can have significant consequences.
• Implications for machine learning (ML) methods were also considered.

Conclusions:

• Prediction and causal inference are fundamentally different tasks requiring distinct methodologies.
• Appropriate application and interpretation of GLMs are crucial for both predictive and causal health research.
• Recommendations are provided to ensure effective and appropriate use of prediction and causal modeling in health research.