When Is an Interrupted Time-Series Design Appropriate?

Can You Evaluate an Intervention When You Cannot Randomize Who Receives It?

A university introduces a new academic advising policy for all undergraduate students. A school district adopts an electronic attendance system across its schools. A government implements a nationwide education reform. In each situation, researchers may want to determine whether the intervention changed an important outcome, but random assignment may be impossible or inappropriate.

Comparing outcomes before and after implementation seems straightforward. Yet a difference between two measurements could reflect an existing trend, seasonal variation, or another event rather than the intervention itself.

An interrupted time-series design addresses part of this problem by examining repeated observations before and after a defined intervention. Instead of asking only whether the outcome changed, it asks whether the observed pattern departed from what would reasonably have been expected without the intervention.

The central question is whether the available time series, intervention timing, and causal assumptions are sufficient to support that comparison.

Use Interrupted Time Series When an Intervention Has a Definable Starting Point and Repeated Outcome Data Exist

What Makes Interrupted Time Series a Credible Research Design?

What Is an Interrupted Time-Series Design?

An interrupted time-series design is a longitudinal quasi-experimental approach in which an outcome is measured repeatedly before and after an intervention or event. The intervention divides the observed series into pre-intervention and post-intervention periods.

The design uses the pre-intervention pattern to help estimate what might have happened if the intervention had not occurred. The observed post-intervention outcomes are then compared with this estimated counterfactual trajectory.

Unlike a simple before-and-after comparison, ITS explicitly considers the trajectory of the outcome over time. The pre-intervention series may reveal that the outcome was already increasing, decreasing, fluctuating seasonally, or remaining approximately stable.

The term interrupted does not require the outcome to stop or reverse. It refers to the intervention's potential interruption of the underlying time-series process.

When Is Interrupted Time Series Particularly Useful?

ITS is well suited to interventions implemented at an identifiable time and affecting a population, institution, or service system for which routinely collected longitudinal data exist.

Examples include introducing a university-wide academic support policy, changing hospital prescribing procedures, implementing a school attendance system, or introducing a public health regulation.

The approach is especially attractive when withholding the intervention from a comparison group would be infeasible or when the intervention has already been implemented and historical observations are available.

However, the existence of an intervention date and repeated measurements does not automatically make the design credible. Researchers must examine whether the pre-intervention process provides a reasonable basis for estimating the outcome's counterfactual trajectory.

What Is the Difference Between Interrupted Time Series and a Before-and-After Study?

ITS can provide stronger evidence than a simple before-and-after comparison because it uses more information about the outcome's history. It does not, however, automatically establish causation.

Why Must the Intervention Have an Identifiable Time Point?

The researcher needs to know when the intervention could plausibly begin affecting the outcome. Without an identifiable intervention period, separating the pre-intervention trajectory from the post-intervention response becomes difficult.

Some interventions begin abruptly, such as a regulation taking effect on a specified date. Others involve training, phased implementation, gradual adoption, or delayed behavioral responses.

For gradual interventions, researchers may need to model an implementation period, delayed effect, or changing intervention intensity rather than impose an artificial instantaneous interruption.

The intervention's mechanism should inform the expected timing of its effect. A digital attendance system might change recording practices immediately, whereas a curriculum reform may take several academic terms to influence learning outcomes.

Where implementation occurs in stages, the researcher should consider how rollout logistics influence the research design.

What Is the Difference Between a Level Change and a Trend Change?

Two common effects examined in ITS are changes in level and changes in slope.

Suppose a university's monthly student withdrawal rate was increasing before a retention policy was introduced. An immediate reduction at implementation would represent a level change. If the rate subsequently declined more rapidly than its earlier trajectory predicted, that would represent a slope change.

An intervention may produce either effect, both effects, or neither. The appropriate impact model should reflect the intervention's expected mechanism rather than being chosen solely because one specification produces a statistically significant result.

How Does Segmented Regression Estimate an Intervention Effect?

Segmented regression is a commonly used analytical approach for ITS. It models the outcome over time while allowing specified changes at the intervention point.

A basic linear segmented regression model can be written as follows:

The interpretation of coefficients depends on the time coding and model specification. For nonlinear outcomes, counts, proportions, or complex temporal patterns, generalized linear models or other time-series approaches may be more suitable.

Segmented regression also assumes that the specified model adequately represents the relevant counterfactual pattern. A straight-line extrapolation may be inappropriate when the pre-intervention outcome follows a nonlinear or unstable trajectory.

What Is the Counterfactual in an Interrupted Time Series?

The counterfactual is the outcome trajectory that would have occurred in the absence of the intervention.

Because that trajectory cannot be observed after implementation, the researcher estimates it using the pre-intervention pattern and the assumptions of the chosen model.

For example, if an outcome was declining by approximately two units each month before an intervention, the counterfactual projection might continue that decline. A post-intervention outcome below the projection could indicate an additional reduction associated with the intervention.

The difference between observed and projected outcomes can be calculated at particular post-intervention periods or accumulated over a specified interval.

However, this interpretation depends on the assumption that the pre-intervention trajectory would have continued sufficiently similarly without the intervention. If other changes would have altered the trajectory anyway, the projection may be misleading.

How Many Observations Are Needed Before and After the Intervention?

ITS requires enough observations to characterize the pre-intervention pattern and estimate the post-intervention response with useful precision.

There is no universal number of observations that guarantees a valid analysis. The required series length depends on variability, autocorrelation, seasonality, effect size, the complexity of the model, and the timing of the expected effect.

Cochrane's Effective Practice and Organisation of Care guidance has used at least three observations before and three after intervention as a minimum eligibility criterion for certain evidence reviews. This is a screening threshold, not a general recommendation for adequately powered or methodologically robust ITS research.

In practice, substantially more observations are often desirable, particularly when estimating seasonal effects or distinguishing gradual intervention effects from ordinary fluctuations.

The question of how many pre- and post-intervention observations are sufficient should therefore be addressed through design-specific considerations and, where feasible, simulation-based power assessment.

Why Do Autocorrelation and Seasonality Matter?

Successive observations in a time series are often correlated. A school's attendance rate this month may resemble last month's rate because underlying conditions persist.

This dependence is called autocorrelation or serial correlation. If ignored, it can lead to incorrect standard errors, confidence intervals, and hypothesis tests.

Seasonality introduces another challenge. Student attendance may vary by academic term, examination period, or holiday schedule. A policy implemented near a predictable seasonal shift could appear effective even when the observed change follows an established calendar pattern.

Researchers should inspect the series visually and consider appropriate autocorrelation structures, seasonal terms, and other time-dependent components. The choice should reflect the data-generating process and available observations.

For educational data, the academic calendar deserves particular attention. Monthly observations spanning school closures or vacation periods may not be directly comparable without suitable outcome definitions and adjustments.

How Do Concurrent Events Threaten Causal Interpretation?

A central threat to uncontrolled ITS is that another event occurs around the same time as the intervention and independently affects the outcome.

Suppose a university launches a new advising policy while simultaneously expanding financial assistance. A subsequent decline in withdrawal rates might reflect advising, financial support, or both.

Repeated measurements can reveal when the outcome changed, but they cannot by themselves determine which concurrent event caused that change.

This is one form of time-related confounding. Researchers should document relevant policy changes, institutional disruptions, measurement revisions, and other plausible influences.

When a suitable comparison series exists, a controlled interrupted time-series design may strengthen inference by helping distinguish intervention-specific changes from broader temporal influences.

What If the Intervention Changes How the Outcome Is Measured?

An ITS design requires comparable measurements across time. If the intervention changes the measurement process itself, the observed discontinuity may reflect recording practices rather than a substantive change in the underlying outcome.

For example, introducing an electronic attendance platform might increase recorded absences because previously unrecorded absences become visible. The measured attendance rate could decline even if actual student behavior remains unchanged.

Researchers should investigate changes in definitions, instruments, data completeness, reporting requirements, and institutional coverage. When measurement discontinuities cannot be separated from substantive effects, causal claims should be correspondingly limited.

Can Interrupted Time Series Be Used With Individual-Level Data?

Yes. ITS may use repeated aggregate outcomes, repeated cross-sectional samples, or longitudinal individual-level observations, provided the temporal structure supports the design.

However, the unit of observation and dependence structure must be handled correctly. Repeated measurements from the same individuals or institutions may introduce within-unit correlation in addition to temporal autocorrelation.

Collecting thousands of individual records at each time point does not replace the need for an adequate number of distinct time points. The time-series dimension is central to estimating temporal patterns and intervention-associated changes.

Evaluating a University-Wide Student Retention Policy

If the model assumptions are reasonable and no important competing explanation is identified, the findings may support an interpretation that the advising policy contributed to improved retention.

However, suppose the university simultaneously introduced a major scholarship program. Without additional evidence, the researcher could not confidently attribute the estimated change specifically to advising. The appropriate conclusion would acknowledge that the observed interruption coincided with multiple potentially influential changes.

The example illustrates why ITS is more informative than simply comparing average withdrawal rates before and after implementation, while also showing why the design's causal interpretation remains conditional.

Common Mistakes When Designing and Interpreting Interrupted Time Series

How to Decide Whether Interrupted Time Series Fits Your Research Question

Start by identifying the intervention mechanism and the outcome it is expected to influence. Determine whether implementation occurred at a sufficiently identifiable time and whether the available observations extend far enough before and after that point to support the proposed analysis.

Next, evaluate the measurement process. Routinely collected administrative data can be valuable, but their usefulness depends on consistent definitions, adequate completeness, and stable reporting practices.

Researchers evaluating educational technology should pay particular attention to adoption timing, implementation fidelity, software updates, changes in platform-generated indicators, and the academic calendar. A system may be technically deployed on one date but used meaningfully only several weeks later.

It is also useful to specify the expected intervention response before examining the post-intervention results. This reduces the temptation to search retrospectively for whichever interruption point or model produces the most favorable estimate.

When the assumptions required for ITS are not credible, a carefully qualified descriptive time-series analysis may be more defensible than an unsupported causal claim. The broader decision should reflect the actual strength of the evidence rather than the apparent sophistication of the design label.

Before Choosing an Interrupted Time-Series Design, Verify These Conditions

Frequently Asked Questions About Interrupted Time-Series Research

Interrupted Time Series Is Appropriate When the Historical Trend Provides a Credible Comparison

How to Cite This Guide

This guide is intended to be read, shared, and used in research, teaching, and academic work. If you draw on its ideas, explanations, or other content, please acknowledge the source by citing the guide. Doing so gives appropriate credit and helps your readers locate the original resource.