Missing confounding information in counterfactual prediction models: a simulation study on model-based treatment...

Area of Science:

• Medical Physics
• Radiation Oncology
• Biostatistics

Background:

• Model-Based Clinical Evaluation (MBCE) estimates causal effects of radiotherapy techniques like proton vs. photon therapy on toxicity.
• Normal Tissue Complication Probability (NTCP) models guide patient selection for proton therapy in the Netherlands.
• Current NTCP models may omit confounders, potentially impacting their suitability for MBCE.

Purpose of the Study:

• To investigate how omitted confounders in NTCP models and patient selection influence the validity of MBCE.
• To assess the suitability of NTCP models used for patient selection in MBCE.

Main Methods:

• Simulated head and neck photon therapy patients with varying confounder effects on dose and toxicity.
• Applied Dutch National Indication Protocol for model-based proton therapy selection.
• Estimated Average Treatment Effect in the Treated (ATT) using current (confounder-excluded) and extended (confounder-included) NTCP models.
• Compared estimated ATT to true effect and explored modified patient selection scenarios.

Main Results:

• The current NTCP model yielded unbiased ATT estimates when the omitted confounder was conditionally independent of patient selection.
• Bias emerged when the omitted predictor was associated with patient selection, which the extended model successfully avoided.
• Simulations with modified patient selection confirmed the impact of confounder-selection association.

Conclusions:

• Omitting confounders in NTCP models does not inherently bias MBCE estimates.
• MBCE validity hinges on whether omitted predictors correlate with the patient selection mechanism.
• Including outcome predictors linked to treatment allocation in NTCP models is recommended for MBCE.