Estimating the effects of multiple interventions based on Interrupted Time Series Analysis

Abstract

We introduce a novel approach for analyzing the impact of repeated low-frequency interventions on observational data, focusing on situations like routine factory maintenance. This method utilizes Interrupted Time Series Analysis (ITSA) to estimate intervention effects by comparing the observed data with counterfactual predictions based on trends observed prior to the intervention. Traditional ITSA faces challenges with global non-linearity and difficulties in interpreting results from disparate interventions. Our solution models a concise time frame surrounding each intervention, aiming to linearly approximate global non-linear effects and provide consistent, stable estimates across different interventions. We conducted experiments using synthetic data to validate our method. Findings suggest that treating multiple interventions as a unified group under the assumption of homogeneity leads to effective outcomes, highlighting the significance of selecting appropriate intervals for analysis.