Quasi-experimental designs
- Study design
They evaluate a programme or policy without randomisation, using comparisons over time or with similar groups to come close to an experiment.
What it is
A quasi-experimental design evaluates an intervention that the investigator does not randomise, often because it is already in place (a law, a programme, a protocol). The main ones are the controlled before-after study (difference in differences) and the interrupted time series, which uses many measurements before and after to project the earlier trend and compare it with what was observed. They are the usual method for evaluating public health policies.
How it is built
- The intervention and its exact start date are defined.
- Repeated outcome measurements before and after are gathered (ideally 8 or more on each side).
- A comparison group or series that did not receive the intervention is sought.
- The earlier trend is modelled and compared with what was observed (change in level and slope).
- Other events occurring on the same dates are reviewed.
Worked example
- A hospital launches a catheter-infection programme at month 24. Earlier rate: 10 per 1 000 catheter-days at start, falling 0.1 per month.
- Expected rate from the trend at month 24: 10 − 0.1 × 24 = 7.6; at month 36: 10 − 0.1 × 36 = 6.4.
- Observed rate at month 36: 4.0 per 1 000 catheter-days.
- Effect from the series: 6.4 − 4.0 = 2.4 per 1 000 catheter-days below what was expected.
- A simple before-after (month 24 vs month 36) would give 7.6 − 4.0 = 3.6: it includes 1.2 from the earlier trend (0.1 × 12).
- Reading: the programme is linked to 2.4 fewer infections per 1 000 catheter-days; a third of the simple drop was already happening.
Most used quasi-experimental designs
| Design | What it compares | What it controls | Main threat |
|---|---|---|---|
| Uncontrolled before-after | One group, 2 measurements | The group's fixed traits | Trends and regression to the mean |
| Controlled before-after | Change in treated versus untreated | Shared trends | Non-parallel trends |
| Interrupted time series | Projected versus observed trend | Earlier trend | Another event on the same date |
| Interrupted series with control | The above in two series | Shared trends and events | Poorly comparable groups |
Markers
- ITS
- DiD
- Level change
- Slope change
When to use it
- Evaluating laws (sugary drink taxes, helmet use), campaigns and coverage changes.
- Measuring the effect of a hospital protocol already rolled out across the whole service.
- Relies on routine data: death, surveillance or billing records.
Common pitfalls
- Too few measurements before the intervention: the earlier trend is poorly estimated.
- Ignoring a simultaneous event (another campaign, a change in how cases are recorded).
- Assuming parallel trends without looking at both groups' earlier data.
Test yourself
Think of the answer, then tap the question to see it.
What is the key assumption of difference in differences?
That, without the intervention, the two groups would have followed parallel trends.
What advantage does an interrupted time series have over a simple before-after?
It uses many measurements to project the earlier trend, so it does not credit the intervention with an improvement already under way.
What is the main threat to an interrupted time series without a control group?
That another event on the same date explains the change (history bias).
Connections
Educational content on epidemiological and statistical methods; the examples use made-up numbers to teach the calculation. Not a clinical or legal guide: current Colombian rules are in the official sources.
Sources: Chapter 24: Including non-randomized studies on intervention effects (Cochrane, 2024); Chapter 25: Assessing risk of bias in a non-randomized study (Cochrane, 2024); Difference-in-Difference Estimation (Columbia University Mailman School of Public Health, 2026).
Spotted a mistake?We are always improving for the community. Tell us what is wrong and our team will review it.Report a mistake
Representative model, for educational purposes only.