Test the change in the trend instead of comparing two averages

Define the interruption and its plausible effect over time

Record when the intervention was announced, became available and reached meaningful implementation. A gradual rollout is not the same as an instantaneous switch. Ask what outcome could plausibly change, after what delay and for how long. Distinguish a one-time level change from a change in the ongoing trend, and do not choose the expected shape only after seeing the chart.

UK government guidance describes interrupted time series as using repeated observations before and after an intervention, with analysis of the underlying patterns rather than only pre/post averages. It highlights seasonality and autocorrelation among relevant issues. This supports a structured time-series question, not a claim that the method eliminates every concurrent influence.

Check the series before fitting a model

Confirm the outcome definition, observation interval and denominator across the full period. Look for reporting-system changes, missing dates, delayed entries and changing population coverage. An apparent launch effect can arise when the team starts counting a previously unrecorded channel on the same day.

Ask whether the available history covers the patterns important to the business. A seasonal retailer cannot establish its usual annual cycle from a few ordinary weeks. More frequent observations do not automatically create independent information. Have the analyst assess the series length and structure needed for the proposed model rather than borrowing a universal minimum from an unrelated example.

Build the competing-event calendar

List promotions, pricing changes, service outages, holidays, distribution shifts and external events that could affect the same outcome. Identify changes occurring at or near the intervention date. A time-series model cannot automatically distinguish two simultaneous business changes simply because one has a campaign label.

Consider whether a credible unaffected comparison series, outcome or staggered introduction could strengthen the design. Review why the comparison would share relevant pressures while avoiding the intervention. Do not call a conveniently flat metric a control without a substantive rationale. Record spillovers and known exceptions, then ask the analyst how the proposed comparison changes the assumptions and interpretation.

Review uncertainty and the durability of the conclusion

Ask the analyst to show fitted and observed values, residual diagnostics and sensitivity to defensible timing or model choices. Check treatment of serial dependence and seasonal patterns. A narrow uncertainty interval from an inappropriate independence assumption is not stronger evidence. Distinguish an immediate novelty response from a persistent change if the available follow-up permits that question.

Report what the design can separate and what remains entangled. If a major co-intervention prevents attribution, describe the observed change without crediting it solely to marketing. Keep the intervention calendar and data revisions with the analysis so later updates are interpretable. The useful output is an evaluated trend interruption with explicit assumptions, not a percentage lift created by selecting convenient before and after dates.

Make the reasoning concrete.

A fictional booking service introduces a reminder campaign on the same week that it expands appointment capacity and repairs an analytics feed. Bookings rise, and a slide compares the previous month with the following month.

The team reconciles the measurement change and documents the capacity expansion before selecting an analysis. If the effects cannot be separated, the report names the combined observed change rather than inventing isolated campaign impact.

Take the questions into your next brief.

Use these prompts in your own notes. Nothing is submitted from this page.

1. Timing

   When did exposure actually change?

2. Shape

   What response timing is plausible before seeing results?

3. Series

   Did outcome and denominator remain comparable?

4. Co-event

   What else could change the outcome at the same time?

5. Limit

   What attribution remains uncertain after analysis?

Watch for the wrong conclusion.

• Replacing a time-series analysis with two convenient averages.
• Assuming daily observations are independent.
• Ignoring a reporting change that coincides with launch.

Sources & scope

Workflows, worksheets and fictional examples are original editorial analysis. Named research guidance supports identified method principles. This is design-review preparation for a qualified analyst, not an estimation recipe or a claim of observed business results.

• UK Office for Health Improvement and Disparities: Interrupted time series study →Primary research-method guidance · Supports: Repeated pre/post observations, trend-based analysis and the need to address seasonal patterns and autocorrelation. Business-readiness and claim-review steps are original analysis.