Interrupted Time Series: Strong Quasi-Experiments Need More Than a Before-and-After Plot

What Interrupted Time Series Actually Estimates

ITS asks whether the observed outcome trajectory changed when an intervention, policy, shock, or implementation event occurred. The change can appear in two main ways:

Good ITS work usually estimates both, because interventions can cause an immediate shift, a gradual shift, or both. If a paper only says “post period was lower than pre period,” it is leaving the most important structure on the floor.

Why Plain Before-and-After Comparisons Are Weak

Suppose a hospital launches a sepsis bundle in January and 30-day mortality declines afterward. Great. But would mortality have been declining anyway? Was there a simultaneous staffing change, a new EHR alert, case-mix drift, seasonal pathogen variation, or a broader secular improvement across hospitals? A simple pre/post comparison cannot tell you.

ITS is stronger because it uses the pre-intervention trend as part of the counterfactual. That does not make it bulletproof, but it is vastly better than throwing away time.

The Basic Segmented Regression Logic

The standard ITS model includes time, an indicator for the post-intervention period, and an interaction between time and the post-intervention period. In plain English, you estimate:

• the baseline trend before the intervention,
• the immediate step change at the interruption, and
• the difference in trend after the interruption.

That is the conceptual model. The real work is in making sure the error structure, timing assumptions, and competing explanations are not garbage.

A Clinical Example: Antibiotic Stewardship Policy

Imagine monthly broad-spectrum antibiotic prescribing rates across 48 months in a health system. At month 25, a stewardship policy requires infectious disease approval for selected agents. A credible ITS analysis would ask:

• Did prescribing drop immediately after policy activation?
• Did the monthly trend change afterward?
• Was there evidence that prescribing was already dropping before the policy?
• Did anything else happen around month 25 that could explain the break?

If you have enough time points, stable outcome definition, and no obvious coincident intervention, ITS can be excellent here. If the policy rolled out gradually across sites, was announced months early, and changed documentation rules at the same time, your “interruption” is already fuzzy.

What Makes ITS Credible

The Assumptions People Forget

ITS does not require parallel trends like DID, but it absolutely requires assumptions. The most important are usually ignored because they are less famous.

Single-Series ITS vs Controlled ITS

A single interrupted series can be useful, but adding a comparison series usually makes the design much stronger. A controlled ITS asks whether the break in the intervention series differs from what happened in a comparable untreated series over the same period.

My bias: if a sensible control series exists, use it. Not because single-series ITS is invalid, but because reviewers should not have to trust your historical counterfactual more than necessary.

Where Researchers Usually Screw This Up

• They pick the intervention date after seeing the data.
• They ignore anticipation effects even though the policy was announced months earlier.
• They treat a gradual rollout like a sharp interruption.
• They omit seasonality in strongly seasonal outcomes.
• They never assess whether residuals are autocorrelated.
• They change coding definitions at the intervention point and call the resulting jump causal.
• They report only p-values instead of absolute changes, trend changes, and visual diagnostics.

What Good Reporting Looks Like

At minimum, a serious ITS paper should report:

• the exact intervention timing and whether rollout was sharp or gradual,
• the number and spacing of pre and post observations,
• the outcome and denominator definitions, including any changes over time,
• the modeled pre-trend, level change, and slope change,
• how seasonality, autocorrelation, and nonlinearity were handled,
• what other events occurred near the interruption date,
• visual plots with observed and fitted values, not just coefficients,
• sensitivity analyses for alternative interruption dates or lag structures.

If a paper claims ITS and skips half of that, the design is being used as branding rather than evidence.

When ITS Is a Strong Choice

When It Is the Wrong Tool

ITS is weak when the intervention date is fuzzy, the outcome definition changed at the same time, the series is too short, or the environment was hit by multiple overlapping changes. It is also the wrong tool when your real question is about patient-level treatment effects rather than system-level trajectory shifts.

And no, slapping Newey-West standard errors on a flimsy pre/post comparison does not suddenly make it an interrupted time series study.

The Bottom Line

Interrupted time series is one of the best quasi-experimental designs in the real world, especially for policies, implementation changes, and health-system interventions. But it earns that status only when time is modeled carefully and alternative explanations are taken seriously.

The honest question is not “did the line move after the intervention?” It is “would a comparable trajectory likely have continued differently if the intervention had not happened?” That is a much harder question, and that is exactly why ITS can be powerful when done well.

My blunt version: if your ITS design does not make me believe the interruption matters more than the calendar, the method did not fail. The study design did.