A methodological framework for model selection in interrupted time series studies.

Area of Science:

• Public Health Research
• Health Services Evaluation
• Statistical Modeling

Background:

• Interrupted time series (ITS) is a robust design for assessing public health and health service interventions.
• ITS analysis estimates intervention effects by comparing observed outcomes to a counterfactual trend.
• Accurate modeling of both the counterfactual and intervention impact is essential.

Purpose of the Study:

• To outline the critical components of interrupted time series (ITS) modeling.
• To emphasize the importance of appropriate model specification for valid intervention evaluation.
• To guide researchers in customizing ITS models based on substantive knowledge.

Main Methods:

• Defining the counterfactual by extrapolating pre-intervention trends.
• Specifying the impact model to capture intervention effects (level change, slope change, lag, transition, ceiling/floor).
• Considering factors like pre-intervention period, confounders, subgroup variations, and trend linearity/nonlinearity.

Main Results:

• Inappropriate model specification can introduce bias in ITS analyses.
• Testing multiple models or using poorly tailored models increases the risk of false positives.
• Customizing ITS models a priori using substantive knowledge is vital for accurate results.

Conclusions:

• Researchers must carefully define the counterfactual and impact models in ITS studies.
• Substantive knowledge should guide a priori model customization to avoid bias.
• Sensitivity analyses or exploratory studies can address model specification uncertainty.