Interrupted Time-Series (ITS) methods are essential quasi-experimental models for evaluating interventions at the population level when there is no control group and traditional difference-in-differences methods are unfeasible. ITS methods use historical data to construct a counterfactual scenario (what would have happened without the intervention), model the observed scenario (what happened), and compute the average treatment effect by comparing the counterfactual with the actual data for the whole population. Despite various implementations of ITS methods, a framework providing best-practice guidance for practitioners is lacking. Additionally, traditional ITS methods do not incorporate recent advancements in causal machine learning for flexible estimation of counterfactual post-reform outcomes.

We develop a new methodological approach for estimating interrupted time-series for impact evaluation at the population level in the absence of a control group, leveraging ensemble learners and insights from the causal machine learning literature. We extend the use of ITS to investigate heterogeneous treatment effects.

As an illustration, we analyze the sequential implementation of two policies targeting low-value care in Switzerland. In June 2020, the Swiss Society of General Internal Medicine recommended against prescribing vitamin D screening tests, followed by the removal of these tests from the health insurance coverage in July 2021. We model the causal effects of these two population-level interventions using claims data from a major Swiss healthcare company. This example is used to showcase our methodological contribution.

We present a list of recommended practices for researchers using ITS to uncover causal effects, including placebo and specification tests. Our approach is complemented by visualization recommendations, demonstrated through our illustration.