Treatment effect estimation by comparing observed and predicted outcomes: conditions for valid inference and...

Area of Science:

• Causal Inference
• Health Services Research
• Biostatistics

Background:

• Non-randomized studies often estimate new treatment effects by comparing observed outcomes to predictions from a standard treatment model.
• This approach, related to standardization and g-methods, lacks formal theoretical conditions for unbiased estimation.
• Applications include model-based clinical evaluation, such as in radiotherapy.

Purpose of the Study:

• To formalize the theoretical framework for estimating treatment effects using prediction models.
• To clarify the necessary conditions for achieving unbiased treatment effect estimates.
• To illustrate these conditions with a radiotherapy case study.

Main Methods:

• Formalization within the potential outcomes framework for causal inference.
• Explanation of methodology, necessary conditions, and validity assessment.
• Case study: comparing proton and photon therapy for head and neck cancer dysphagia.

Main Results:

• Five sufficient conditions for unbiased estimation were identified: transportability, ignorability, consistency, positivity, and correct model specification.
• These conditions are largely untestable but can be supported by empirical evidence.
• Model prediction accuracy in relevant subpopulations is crucial.

Conclusions:

• The approach can provide unbiased treatment effect estimates if the prediction model performs well.
• Potential sources of bias necessitate systematic consideration of all conditions.
• Domain expertise and empirical evidence are recommended to support condition plausibility.