Interrupted Time Series Analysis

Interrupted Time Series (ITS) analysis is a powerful quasi-experimental method used to evaluate the impact of an intervention when a randomized controlled trial is impractical or unethical. By analyzing data points collected over time before and after an event—like a new law, a marketing campaign, or a clinical guideline change—you can estimate whether the intervention caused a meaningful shift in an outcome trend. This technique is indispensable in policy, healthcare, and business, where decisions must be based on observational, real-world data rather than laboratory conditions.

Core Concepts: Segmented Regression and Intervention Effects

At its heart, ITS uses segmented regression to model changes in both the level and the trend of a time series. The "interruption" is the intervention point. You model the pre-intervention segment to establish a baseline trend, and then the post-intervention segment to see how that trend changed.

Formally, a basic ITS model can be expressed as: $$Y_t={\beta }_0+{\beta }_1{T}_t+{\beta }_2{X}_t+{\beta }_3{X}_t{T}_t+{ϵ}_t$$

Where:

• $Y_t$ is the outcome at time $t$.
• $T_t$ is a continuous variable indicating time (e.g., week 1, 2, 3...).
• $X_t$ is a dummy variable representing the intervention (0 = pre, 1 = post).
• ${ϵ}_t$ is the error term.

The coefficients tell the story:

• ${\beta }_0$: The starting level of the outcome at $T_t=0$.
• ${\beta }_1$: The pre-intervention slope or trend.
• ${\beta }_2$: The immediate level change following the intervention. This is the difference between the observed post-interruption level and the level predicted by the pre-intervention trend line.
• ${\beta }_3$: The slope change. This shows if the trend became steeper or flatter after the intervention.

For example, imagine a hospital implements a new patient safety checklist. Your outcome ($Y_t$) is the monthly rate of post-surgical infections. The ${\beta }_2$ coefficient would estimate the immediate drop in the infection rate the month after implementation, while ${\beta }_3$ would show if the rate continued to decline more rapidly over subsequent months.

Handling Autocorrelation and Model Refinement

A critical assumption in standard regression is that errors are independent. In time series data, this is often violated due to autocorrelation, where the error at one time point correlates with errors at previous points. Ignoring autocorrelation leads to underestimated standard errors, making you too likely to find a "significant" effect where none exists.

To diagnose this, you examine the autocorrelation function (ACF) and partial autocorrelation function (PACF) plots of the model's residuals. If significant autocorrelation is present, you must adjust your model. A common and robust approach is to model the error structure explicitly using an ARIMA (AutoRegressive Integrated Moving Average) framework. Essentially, you fit your segmented regression model but allow the errors (${ϵ}_t$) to follow an ARIMA process (e.g., AR(1)), which corrects the standard errors and provides valid statistical inference. Modern software can estimate these combined regression-ARIMA models efficiently.

Advanced Modeling: Multiple Interventions and Control Series

Real-world evaluations are often more complex than a single intervention. You might need to model multiple interventions. For instance, a public health campaign may have an initial launch ($Interventio{n}_1$) followed by a boosted social media push six months later ($Interventio{n}_2$). Your model would include separate dummy and interaction terms for each event, allowing you to isolate the level and slope change attributable to each phase.

To strengthen causal claims, incorporating a control series is highly recommended. This involves analyzing a parallel time series that did not receive the intervention but is subject to similar external trends. For example, when evaluating a new tax on sugary drinks in one state, you could use data from a demographically similar state without the tax as a control. You would fit a model that includes terms for the control series' trend and an interaction between the intervention dummy and a "group" dummy (treatment vs. control). A significant interaction term suggests the change was unique to the intervention group, strengthening the argument that the intervention itself, and not a broader trend, caused the effect.

Applications in Practice

ITS is versatile, with key applications across domains:

• Healthcare Policy Evaluation: Assessing the effect of a vaccine mandate on disease incidence, or a payment reform policy on hospital readmission rates. ITS can separate the policy's impact from seasonal flu trends or other ongoing healthcare changes.
• Marketing Campaign Evaluation: Determining the true impact of a major advertising campaign on weekly sales. The control series might be sales in a region where the ad did not air, helping to account for overall market shifts or holidays.
• Operational Change Evaluation: Measuring how a new manufacturing process affects daily product defect rates, or how a software update influences system downtime. The pre-intervention trend establishes the baseline performance against which the change is judged.

Common Pitfalls

1. Ignoring Autocorrelation: As discussed, this is the most common and serious error. Always check residuals for autocorrelation and use appropriate modeling techniques (like ARIMA errors) to adjust for it. Failing to do so will produce misleadingly optimistic p-values.
2. Insufficient Pre- and Post-Intervention Data Points: Having too few data points, especially pre-intervention, makes it difficult to reliably establish the underlying trend and seasonality. As a rule of thumb, aim for at least 8-12 observations per segment to estimate trends and check model assumptions.
3. Confounding by a Co-occurring Event: An ITS can only attribute change to the intervention if no other major event happened at the same time. For example, if you evaluate a safety law that took effect in January 2020, the COVID-19 pandemic is a massive confounding event. Using a control series is the best defense against this pitfall.
4. Misinterpreting Level vs. Slope Change: A non-significant immediate level change (${\beta }_2$) does not mean the intervention failed. The effect might be a gradual slope change (${\beta }_3$). Conversely, a significant level change that quickly reverts to the old trend may indicate only a temporary shock. You must interpret both coefficients together within the context of your research question.

Summary

• Interrupted Time Series (ITS) analysis is a crucial method for evaluating the real-world impact of interventions using longitudinal data, estimating both immediate level changes and long-term slope changes.
• The core technique is segmented regression, but you must always check for and model autocorrelation (often with ARIMA errors) to ensure valid statistical tests.
• Using a control series significantly strengthens causal inference by accounting for external trends, while modeling multiple interventions allows for the analysis of complex, multi-phase programs.
• ITS has direct, practical applications in evaluating policies, campaigns, and operational changes across healthcare, marketing, and business, providing evidence for decision-making when experiments are not feasible.