its2s: a Python package for two-stage interrupted time series analysis using machine learning

Area of Science:

• Epidemiology
• Biostatistics
• Data Science

Background:

• Randomized controlled trials (RCTs) are not always feasible for causal inference.
• Natural experiments and quasi-experimental designs, like interrupted time-series (ITS) analysis, offer alternatives.
• Two-stage ITS designs utilize machine learning for flexible counterfactual prediction, but fitting high-dimensional models poses challenges for reproducibility.

Purpose of the Study:

• To introduce its2s, an open-source Python package designed to implement two-stage ITS analysis.
• To provide a flexible and reproducible framework for researchers conducting quasi-experimental studies using machine learning.
• To lower the barriers to entry for rigorous counterfactual modeling in the absence of RCTs.

Main Methods:

• The its2s package implements a two-stage ITS design leveraging machine learning models.
• Users can specify intervention dates, training/testing periods, and select from various built-in model architectures (e.g., Prophet-XGBoost, NeuralProphet).
• Confidence intervals are generated using a moving block bootstrap method to maintain temporal autocorrelation in residuals.

Main Results:

• Validation through a simulation study successfully recovered a known policy effect (11.77% vs. true 12%).
• Application to the 2021 Pacific Northwest heat dome identified a 53% excess injury mortality rate in the subsequent three weeks.
• The its2s package demonstrated flexibility, supporting both default and highly customized analytical workflows.

Conclusions:

• The its2s package offers a robust and user-friendly framework for conducting advanced ITS-based quasi-experimental research.
• It facilitates the application of machine learning for counterfactual modeling, enhancing causal inference capabilities.
• its2s promotes reproducibility and rigor in analyzing real-world events and policy impacts.