Interrupted Time-Series Analysis in Clinical Research: Segmented Regression, Intervention Effects, and Credible Policy Evaluation

Clinical research often evaluates interventions that cannot be randomized at the point of implementation. A hospital introduces a prescribing policy, a health system changes a referral pathway, a regulator modifies a reimbursement rule, or a public-health program begins on a defined date. In each situation, investigators may have repeated outcome measurements over time but no randomized allocation. A simple comparison of the average outcome before and after the intervention can be misleading because it discards the trajectory that preceded the change.

Interrupted time-series analysis addresses this problem by treating the intervention as an interruption in a longitudinal outcome series. The central question is not simply whether the post-intervention average differs from the pre-intervention average. It is whether the observed post-intervention pattern departs from a credible counterfactual continuation of the pre-intervention trend. The design is therefore useful for clinical quality improvement, medication-use research, health-services evaluation, regulatory changes, and population-level interventions.

ITS does not convert an uncontrolled observational series into a randomized experiment. Its value comes from using time as part of the design and from making the counterfactual trend explicit. The credibility of the conclusion depends on the number and spacing of observations, the stability of measurement, the timing of the intervention, the plausibility of other changes at the same time, and the statistical treatment of time-series dependence.

When is an interrupted time series appropriate?

A strong ITS begins with a clearly defined intervention point and an outcome that is measured repeatedly using a consistent or well-documented process. The intervention may be a clinical protocol, service reorganization, reimbursement rule, law, safety warning, or implementation program. The time unit should reflect the mechanism and expected response: daily measurements may be appropriate for an emergency-department process, whereas monthly or quarterly measurements may be more realistic for prescribing, readmission, or population outcomes.

The design is more informative when there are enough observations on both sides of the interruption to characterize the baseline trajectory and assess whether the post-intervention pattern is sustained. The exact minimum depends on the outcome, sampling frequency, seasonality, expected effect, and analytic model. Researchers should not use a fixed observation-count rule as a substitute for design judgment. A short series with a single pre-intervention observation and a single post-intervention observation is a before-after comparison, not a robust ITS.

Before modeling, specify the population, intervention, comparator, outcome, time horizon, and expected lag. The comparator is usually the counterfactual continuation of the pre-intervention trend. If an intervention is expected to take several weeks to operate, an immediate level change may be implausible; a lagged effect or ramp-up specification may be more appropriate. The Bernal, Cummins, and Gasparrini tutorial emphasizes proposing the impact model a priori and examining the time-series features that can distort inference.

The segmented regression model

Segmented regression is a commonly used implementation of ITS. In a basic continuous-outcome formulation, the model includes a time variable, an indicator for whether the observation occurs after the intervention, and a post-intervention time variable. In conceptual terms, the model estimates three components:

• Baseline level: the expected outcome at the chosen time origin.
• Baseline slope: the underlying trend before the intervention.
• Intervention effects: the immediate level change and the change in slope after implementation.

The immediate level change asks whether the outcome jumps or drops at the intervention boundary. The slope change asks whether the rate of change after implementation differs from the rate that would have continued without the intervention. These effects answer different clinical questions. A new policy may produce a rapid change in prescribing level but no sustained trend difference, or it may have little immediate effect but gradually alter the trajectory.

The Wagner and colleagues paper on segmented regression provides a foundational framework for estimating immediate intervention changes and post-intervention trend changes in medication-use time series. In practice, the analyst should define the time origin, intervention index, and post-intervention slope coding clearly enough that another researcher can reconstruct the counterfactual and calculate the reported effect.

For count outcomes, rates, proportions, or over-dispersed measurements, the error distribution and link function should match the outcome. A linear model may be adequate for a continuous outcome under appropriate conditions, while a count or rate outcome may require a generalized linear model with an offset. The choice should be driven by the measurement process and estimand, not by a preference for a familiar software command.

Level change and slope change are not interchangeable

One of the most common interpretation errors is treating the intervention indicator as the complete intervention effect. That interpretation depends on the parameterization, time origin, and model structure. A level coefficient may represent an immediate effect at the intervention point, but the meaning can change when the intervention is coded with a different time origin or when the model includes a lag, ramp, or temporary effect.

A 2025 methodological study in BMC Medical Research Methodology showed that two common segmented-regression parameterizations can represent the same model while giving different interpretations to a key coefficient. The authors specifically caution that the immediate intervention effect should be calculated according to the chosen parameterization rather than read mechanically from a binary indicator. Researchers should therefore report the equation or a precise description of each term, including the time reference used for the intervention effect. Read the 2025 coefficient-interpretation study.

Autocorrelation, seasonality, and over-dispersion

Time-series observations are often correlated with nearby observations. A high outcome this month may make a high outcome next month more likely even in the absence of a new intervention. Ignoring this autocorrelation can produce standard errors that are too small and confidence intervals that are too narrow. The analyst should inspect residual patterns and use an approach appropriate to the series, such as a model with an autoregressive structure, robust variance estimation, or another prespecified correction.

Seasonality is a separate concern. Emergency visits, respiratory infections, medication use, and many service outcomes follow recurring patterns. If the intervention occurs just before a seasonal peak, an unadjusted model may attribute a predictable seasonal movement to the intervention. Seasonal indicators, harmonic terms, or a design that includes comparable seasonal cycles may be considered when justified by the data and clinical context. Seasonality should be assessed using the time scale and observation density available, not added automatically to every model.

Over-dispersion occurs when the variability of a count or rate outcome exceeds what a simple distribution assumes. The response may involve a different variance structure, a negative binomial model, quasi-likelihood, or robust inference. The important reporting point is to explain how the outcome distribution and variance were selected, how model fit was assessed, and whether conclusions changed under a reasonable alternative.

The original ITS tutorial identifies autocorrelation, over-dispersion, and seasonal trends as methodological issues that should be considered in analysis. These are not merely technical details. They determine whether an apparent intervention effect reflects a genuine departure from the pre-intervention trajectory or an unmodeled feature of the outcome series.

Concurrent changes and time-varying confounding

ITS attribution is threatened when another event occurs near the intervention date. A staffing change, new diagnostic technology, outbreak, formulary change, reimbursement adjustment, or media event may influence the outcome at the same time. The closer the competing event is to the interruption, the harder it is to distinguish the effects using a single series.

Researchers should create an implementation timeline before fitting the final model. Record policy announcements, rollout dates, changes in eligibility, data-system migrations, coding revisions, staffing changes, and other plausible influences on the outcome. If the intervention was phased across sites or introduced with a delay, the analytic intervention date should reflect the mechanism actually experienced by the population rather than only the date of the announcement.

Time-varying confounding cannot be solved by adding every available time-varying variable to a regression model. Some variables may be affected by earlier intervention exposure and become mediators or colliders. The causal role and timing of each variable should be considered before adjustment. When the scientific question involves dynamic treatment or repeated exposure, an ordinary single-interruption model may not be sufficient, and a different causal design may be needed.

Controlled interrupted time series

A concurrent control series can strengthen an ITS when it is exposed to the same broad background trends but not to the intervention. The control may be a comparable clinical service, geographic area, facility, outcome, or population. The purpose is to help distinguish an intervention-associated departure from a broader change that affected both series.

Control selection requires more than finding a series that is available. The control should have a plausible relationship with the outcome, a stable measurement process, and a defensible reason for not receiving the intervention at the same time. Researchers should compare pre-intervention trajectories and explain whether the control is expected to respond similarly to seasonality, secular changes, and external shocks. A control series does not automatically satisfy parallel-trend assumptions or remove unmeasured confounding.

The Cochrane EPOC methods resource distinguishes ITS and controlled ITS as designs used for evaluating health-system and policy interventions. When a credible control exists, the design can improve attribution, but the manuscript should still show the pre-intervention series, define the comparison, state the interaction terms, and discuss residual threats to validity.

Evidence summary table

Actionable steps: Build a defensible ITS analysis

Common reporting failures

The first failure is presenting ITS as a simple before-after comparison. Without the pre-intervention trajectory, readers cannot assess whether the post-intervention pattern departs from the expected trend. The second is reporting only a p-value for the intervention indicator while omitting the estimated level change, slope change, confidence interval, time scale, and parameterization.

The third failure is ignoring the intervention mechanism. A policy announcement may not be the same as implementation, and a gradual rollout may not produce an immediate effect. The fourth is treating a control series as a guarantee of causal identification without explaining comparability, exposure to concurrent events, and pre-intervention trajectory.

The fifth failure is assuming that a visually persuasive chart proves an intervention effect. Time-series plots are valuable for understanding the data, but visual separation can be created by seasonality, autocorrelation, measurement changes, or a coincident event. The final failure is hiding sensitivity analyses that materially alter the result. A credible report makes uncertainty and design limitations visible rather than presenting the most favorable specification as the only analysis.

Conclusion

Interrupted time-series analysis is a useful design for evaluating clinical, service, regulatory, and population interventions introduced at a defined time. Segmented regression can distinguish an immediate level change from a change in post-intervention trend, but those coefficients are meaningful only when the time origin, intervention timing, outcome process, and parameterization are clear. Autocorrelation, seasonality, over-dispersion, concurrent events, measurement changes, and time-varying confounding must be treated as design and inference issues rather than afterthoughts.

A defensible ITS manuscript shows the raw time series, explains the counterfactual, states the model before interpreting it, reports the level and slope effects with uncertainty, and tests plausible alternatives. When a comparable control series exists, controlled ITS may strengthen attribution, but it does not remove the need for clinical and statistical judgment. The method is most valuable when it makes the intervention question, assumptions, and limits of inference transparent.