Measuring the Effects of Population-level Interventions Using Healthcare Claims Data

Talk at the 2026 Symposium of Causal Inference in the Health Sciences, Fribourg

How do you evaluate the causal impact of a health policy when everyone is treated at the same time? This talk at the 2026 Causal Inference Symposium in Fribourg presents interrupted time series applied to Swiss health insurance claims data — and shows how real-world evidence from routine administrative data can guide policy in the absence of a control group.
Author

Aurélien Sallin

Published

March 18, 2026

Keywords

real-world evidence, interrupted time series, causal inference, health policy evaluation, Switzerland, Swiss healthcare, claims data, administrative data, routine data, low-value care, itscausal, SWICA, population-level intervention, Krankenkassendaten, Schweizer Gesundheitssystem

The challenge: Public health authorities, insurers, and researchers need to know whether their policies work. But population-level interventions affect everyone simultaneously! They rule out randomized experiments, difference-in-differences designs, and synthetic control methods.

The question is: How do you evaluate causally what happened?

The setting: Swiss health insurance claims data — rich, longitudinal, and collected for billing purposes rather than research. Administrative data at its most routine, and yet often the only window into what actually changes at the population level.

The method: Interrupted time series (ITS). The illustration: Two successive Swiss policies targeting the same low-value practice — unnecessary Vitamin D testing — with strikingly different results.

📊 Full presentation: Fribourg 2026 slides


The Setting: Causal Analysis Without a Control Group

Evaluating the impact of public health policies is essential for accountability, quality improvement, and better policy design. Yet many real-world settings make this hard:

  • No control group: national clinical guidelines, reimbursement restrictions, or public health campaigns affect all units simultaneously.
  • Routine data: health insurance claims data are collected for billing and administrative purposes, not designed for causal research. They are available, but they were not built for this.

This is the setting I work in daily at SWICA, and the focus of my talk at the 2026 Symposium of Causal Inference in the Health Sciences in Fribourg (March 18, 2026).


The Method: Interrupted Time Series

Interrupted time series (ITS) is a pragmatic response to the no-control-group problem. The core idea: use the pre-intervention trend to project a counterfactual — what would have happened without the intervention — and compare this to what was actually observed.

ITS is widely used in real-world evidence and health policy evaluation. It is also widely misimplemented. The key risks:

  • Strong causal assumptions: trend continuity, no concurrent interventions, and no anticipation effects must all hold.
  • Functional form: the shape of the counterfactual matters enormously. Misspecification leads to biased estimates.

Sound implementation requires careful attention to both. This is the problem that itscausal, an R package I developed with colleagues of the Berner Fachhochschule, is designed to address.


The Illustration: Vitamin D Testing and Low-Value Care

Switzerland offers a natural experiment. Two successive policies targeted the same unnecessary practice — routine Vitamin D testing for low-risk patients:

  1. April 2021: Smarter Medicine (Switzerland’s Choosing Wisely™ equivalent) listed Vitamin D testing in its Top 5 procedures to avoid. A clinical recommendation directed at physicians.
  2. 2022: The Federal Office of Public Health introduced a coverage restriction (limitatio), directly limiting reimbursement for such tests under mandatory health insurance.

Using claims data from SWICA covering more than 3’600 general practitioners, we estimated the causal effect of each intervention using interrupted time series:

  • The clinical recommendation reduced unnecessary testing by approximately 6%.
  • The financial incentive reduced testing by approximately 58% within six months.

The contrast is striking. Soft nudges and hard levers can both work — but they do not work equally.


What We Can Do With Routine Data

The core message is encouraging: no control group is not the end of the world. When rich longitudinal claims data are available, interrupted time series can produce credible causal estimates of policy effects — as long as the analysis is done carefully.

Administrative data from health insurance systems like SWICA’s are a powerful resource for real-world evidence. They cover large populations, track individuals over time, and record actual healthcare utilization rather than self-reported behavior. These are precisely the properties that make ITS viable.

The challenge is to move from widely used to soundly implemented. itscausal is our contribution to that goal.


📊 The full presentation is available here: Fribourg 2026 slides

🔗 itscausal on GitHub: github.com/ASallin/itsCausal