A time series design is a research method that tracks the same outcome repeatedly over time to detect whether a pattern, trend, or intervention effect has occurred. It belongs to the family of quasi-experimental designs, meaning it evaluates cause and effect without randomly assigning participants to groups. Instead, it relies on the natural sequence of measurements collected before and after some event or intervention to determine whether something meaningfully changed.
This approach is especially common in public health, policy evaluation, and social science, where running a traditional randomized experiment would be impractical, too expensive, or ethically impossible.
How Time Series Designs Work
At its core, a time series is simply a sequence of data points arranged in chronological order. What makes it a “design” rather than just a dataset is the deliberate use of that temporal structure to answer a research question. Researchers collect repeated measurements of an outcome (hospital admissions, crime rates, prescription volumes) at regular intervals, then look for shifts in the level or trend of those measurements that coincide with a known event.
The most widely used version is the interrupted time series (ITS). In an ITS study, researchers gather data for a sustained period before an intervention, then continue collecting the same data afterward. The “interruption” is the intervention itself: a new law, a policy change, a public health campaign. By comparing the post-intervention pattern to what would have been expected based on the pre-intervention trend, researchers can estimate whether the intervention had an effect and how large that effect was.
This differs sharply from a simple before-and-after comparison, which only looks at two snapshots. A time series design uses dozens or even hundreds of data points, making it far more sensitive to real changes and far less likely to mistake normal fluctuation for a meaningful shift.
Single-Group vs. Multiple-Group Designs
Time series designs come in two broad categories. In a single-group design, every unit being studied is exposed to the intervention. You track one city, one hospital system, or one population over time. The pre-intervention data serves as the comparison, essentially asking: did the outcome deviate from what its own historical pattern predicted?
In a multiple-group design, sometimes called a controlled interrupted time series, researchers add a comparison group that was not exposed to the intervention. For example, if a state passed a new opioid prescribing law, researchers might track prescription rates in that state alongside rates in a neighboring state that didn’t pass the law. This strengthens the design considerably because it helps rule out the possibility that some other event, not the intervention, caused the observed change. The key assumption in these designs is that both groups would have followed parallel trends if the intervention had never happened.
Where Time Series Designs Are Used
Public health and health policy research rely heavily on this method. Real-world examples include evaluating the impact of pay-for-performance programs on diabetes outcomes, measuring whether hospital safety regulations changed opioid prescribing after emergency department visits, and assessing whether drug restriction policies reduced inappropriate antibiotic use in elderly populations. In Cambodia, researchers used this approach to determine whether a financial incentive scheme for midwives actually increased facility-based deliveries.
The design is particularly well suited to evaluating community-level or policy-level interventions where you’re trying to affect an ongoing, measurable process. You can’t randomize which countries adopt a new drug regulation or which cities experience a public health campaign, but you can track the outcomes before and after those changes take effect.
How Many Data Points You Need
There is no precise formula for calculating the minimum sample size in a time series analysis. Most experts recommend at least 50 observations as a baseline, though more complex analytical methods may require 100 or more. These observations are time points, not people. If you’re measuring monthly hospital admissions, 50 observations means roughly four years of data. This requirement is one reason the design works best when routine administrative data or surveillance records already exist.
Ideally, those data points are split fairly evenly before and after the intervention, giving the statistical model enough information to establish a reliable pre-intervention trend and enough post-intervention data to detect a change.
Statistical Analysis
Two main approaches dominate time series analysis. The simpler one, segmented regression, fits a straight line to the pre-intervention data and another to the post-intervention data, then measures whether there’s a jump in level or a change in slope at the point of intervention. This works well when the trend is reasonably linear and the data points are independent of one another.
In practice, time series data often violates that independence assumption. Measurements taken close together in time tend to be more similar than measurements taken far apart, a property called autocorrelation. Seasonal patterns can also complicate things: flu-related hospital visits spike every winter regardless of any policy change. When these patterns are present, researchers turn to a more flexible approach called ARIMA modeling, which accounts for autocorrelation and seasonality by predicting each data point based on previous data points rather than on time itself. Modern statistical software can automatically identify the best-fitting model, making this technique increasingly accessible.
Strengths of the Design
Time series designs fill a critical gap between observational studies and randomized trials. They enable researchers to evaluate the effects of community interventions and policies in circumstances where randomized controlled trials are too expensive, premature, or simply impractical. Because the design tracks change over time rather than comparing different groups of people at a single moment, it naturally accounts for stable differences between populations that might confound a cross-sectional study.
The design also captures not just whether something changed, but how it changed. Researchers can distinguish between an immediate, sharp effect (a sudden drop in prescriptions the month a law takes effect) and a gradual shift (a slow decline over the following year). They can measure changes in both the average level of an outcome and its trajectory, providing a richer picture than a simple yes-or-no answer.
Key Limitations
The biggest threat to a time series design is what researchers call history bias. Because there’s no randomization, any event that happens around the same time as the intervention could be the real explanation for the observed change. If a city launches an anti-smoking campaign in the same month a beloved local figure dies of lung cancer and makes a public plea to stop youth smoking, it becomes impossible to separate the campaign’s effect from the emotional impact of that event.
Adding a control group helps, but only for events that affect all groups equally. Events unique to the treated group remain a vulnerability. Researchers typically address this by documenting what else was happening during the study period and arguing, based on context, that no plausible alternative explanation exists. This requires judgment, not just statistics.
Seasonality and autocorrelation can also distort results if not properly handled in the analysis. Misidentifying a seasonal pattern as an intervention effect, or failing to account for the natural clustering of similar values over time, can lead to false conclusions. Careful model selection and diagnostic checking are essential to avoid these pitfalls.