Heterogeneous Treatment Effect Interrupted Time Series (HTE-ITS)
Heterogeneous Treatment Effect Interrupted Time Series Analysis · Also known as: HTE-ITS, Subgroup ITS, Effect-modifier ITS, Segmented ITS with interaction
Heterogeneous Treatment Effect Interrupted Time Series extends the standard ITS design to detect whether an intervention's effect on a time series differs systematically across subgroups or in response to unit-level moderators. Where ordinary ITS yields a single level-change and slope-change estimate, HTE-ITS adds interaction terms for a moderating variable, revealing who benefits more or less from the intervention and by how much.
Read the full method
Sign in with a free account to read this section.
Method map
The neighbourhood of related methods — select a node to explore.
When to use it
Use HTE-ITS when you have a clearly defined intervention time point in a longitudinal or aggregate time series, and a theoretical reason to expect the effect to vary across subgroups or along a continuous moderator. Minimum data requirements are typically 10-12 observations both before and after the intervention per subgroup; with fewer, power to detect heterogeneity is very low. The method is appropriate for health-care, education, and policy data aggregated over regular time intervals. Do not use it when the intervention date is ambiguous, when the pre-intervention series shows non-stationarity that cannot be modelled, when subgroup sample sizes are too small for stable estimates, or when the moderator itself is endogenous to the intervention.
Strengths & limitations
- Reveals differential intervention effects across subgroups or continuous moderators that an average ITS estimate would conceal.
- Exploits the full time-series structure, which provides greater statistical power than cross-sectional subgroup analyses.
- Does not require randomisation: the pre-intervention period serves as its own control, making it feasible with routinely collected administrative data.
- Can accommodate count outcomes, rates, and continuous measures through generalised linear model variants (Poisson, negative binomial).
- Directly visualisable: predicted trajectories for each subgroup can be plotted against observed data to communicate heterogeneity clearly.
- Requires sufficient observations per subgroup in both the pre- and post-intervention periods; sparse data produce unstable interaction estimates.
- The causal interpretation of heterogeneity rests on the assumption that the moderator Z is not itself changed by the intervention — an assumption that is hard to verify.
- Serial autocorrelation in the residuals, common in time series, must be explicitly modelled or corrected; ignoring it inflates the precision of heterogeneity estimates.
- Multiple moderator testing increases the probability of spurious findings; pre-registration of hypothesised moderators is strongly advisable.
- The standard ITS assumption of a single sharp intervention point is maintained; gradual rollout or staggered implementation requires more complex specifications.
Frequently asked
How is HTE-ITS different from running separate ITS models per subgroup?
Running separate models per subgroup provides subgroup-specific estimates but does not formally test whether the effects differ. HTE-ITS includes an interaction term that directly tests the null hypothesis of equal effects, uses the pooled residual variance for greater efficiency, and produces a single p-value for heterogeneity that is properly calibrated.
How many time points do I need per subgroup?
A commonly cited minimum is 10-12 observations in both the pre- and post-intervention periods per subgroup. Fewer points make it difficult to estimate the baseline trend reliably, and the interaction estimate becomes highly sensitive to individual observations.
What if autocorrelation is present in the residuals?
Autocorrelation violates the independence assumption and inflates the apparent precision of all estimates, including the heterogeneity interaction. It should be addressed either by fitting an ARIMA error structure, by using Newey-West standard errors, or by including lagged outcome terms in the model.
Can I include more than one moderator?
Yes, but each additional moderator multiplies the number of interaction terms and the risk of spurious findings. Pre-register the hypothesised moderators, correct for multiple comparisons, and treat exploratory moderators as hypothesis-generating rather than confirmatory.
Is the moderating variable Z allowed to change over time?
The simplest HTE-ITS formulation treats Z as fixed (a baseline characteristic measured before the intervention). Time-varying moderators are technically possible but require careful specification to avoid confounding the moderator with the intervention itself.
Sources
- Lopez Bernal, J., Cummins, S., & Gasparrini, A. (2017). Interrupted time series regression for the evaluation of public health interventions: a tutorial. International Journal of Epidemiology, 46(1), 348-355. DOI: 10.1093/ije/dyw098 ↗
- Kontopantelis, E., Doran, T., Springate, D. A., Buchan, I., & Reeves, D. (2015). Regression based quasi-experimental approach when randomisation is not an option: interrupted time series analysis. BMJ, 350, h2750. DOI: 10.1136/bmj.h2750 ↗
How to cite this page
ScholarGate. (2026, June 3). Heterogeneous Treatment Effect Interrupted Time Series Analysis. ScholarGate. https://scholargate.app/en/causal-inference/heterogeneous-treatment-effect-interrupted-time-series
Which method?
Set this method beside its closest kin and read them side by side — the library lays the books on the table; the choice is yours.
- Difference-in-DifferencesEconometrics↔ compare
- Heterogeneous Treatment Effect Difference-in-DifferencesCausal inference↔ compare
- Interrupted Time SeriesCausal inference↔ compare
- Panel Data Interrupted Time SeriesCausal inference↔ compare