Plan an interrupted time-series analysis

Define the interruption before plotting effects

An interrupted time series compares an outcome trajectory before and after a clearly identified intervention. The central question is how the post-intervention series differs from a credible continuation of the earlier pattern. It is not simply whether the average after the intervention is lower than the average before it.

The UK government interrupted-time-series guidance highlights seasonality and autocorrelation as important issues. Begin with an intervention chronology: announcement, implementation, expected onset and any phased adoption. Record other events that could affect the same outcome at the same time.

Create a stable series

Use consistent observation intervals, definitions and denominators. If the outcome is a count, record changes in population size or exposure time. Identify missing periods, coding changes and data-system transitions. A jump caused by a new reporting system should not be interpreted as an intervention effect.

Specify whether the hypothesised effect is an immediate level change, a gradual slope change or both. Define how time is centred and how the intervention and post-intervention time variables are coded. Check seasonal patterns and serial dependence in residuals, not only in the raw outcome. Choose a model appropriate to the outcome and its uncertainty.

Hypothetical example

A fictional monthly model predicts 100 events at implementation, with an immediate decrease of 10 and a further decrease of two per month relative to the prior trend. Six months after implementation, the modelled difference from that counterfactual is −10 − 2 × 6 = −22 events.

That calculation assumes the post-intervention time variable equals six at the selected point. If the first intervention month is coded as one rather than zero, the interpretation changes. Documenting the coding prevents an off-by-one error in the reported effect.

Challenge the counterfactual

Ask whether another policy, seasonal event, population change or measurement change occurred near the interruption. A suitable comparison series can strengthen interpretation, but only if its relationship to the treated series is defensible. Additional observations do not by themselves remove confounding by concurrent changes.

Report the estimated change at a meaningful time with uncertainty, along with the model specification and diagnostic findings. Avoid extrapolating far beyond observed follow-up. Use the data snapshot worksheet to preserve the exact time series and derivation rules. The conclusion should state what the design supports and which alternative explanations remain.

Source verification: 6 November 2026. Examples are hypothetical.