I am interested in using interrupted timeseries analysis on real world electronic healthcare records. My understanding of interrupted timeseries analysis is that data is time-ordered and gathered at uniformed and regular intervals:

  1. at least three data points are required on both sides of the interrupt.
  2. the data points must be separated at equal distances in time.
  3. all samples report per time-point.

Of the 120-odd journal publications I’ve read, the observational data is always well structured. On the other hand, clinical and therapy data from primary care Electronic healthcare records, are very irregular and prone to social and personal effects.

An example of irregular time points would be a patient’s drug prescription record. One might expect regular prescriptions but one finds some patients might start using over the counter drugs (so not I record), or the didn’t take the drugs, etc. such behaviour are a confound, as treatment may continue but in a different form until they decide to visit their doctor.

Can anyone describe, or direct, how one can include irregular time points in ITS? Or perhaps some kind before-after intervention distributions method?

If ITS is inappropriate, then I would like to know what techniques are appropriate at measuring an outcome (Y) over time, before and after an intervention, from a cohort of heterogeneous subjects of which each subject have inconsistent event-time data points.

Update: what is being reported could be eg related to drug prescription, a patient characteristic such as weight, something related to a diagnostic or even how long they waited for a phone consultation. The type of data, discrete, continuous, categorical etc., for now doesn’t matter. At the moment I am just interested in the irregular nature of the data points in time. All ITS studies I’ve seen have a regular and uniformed placement with respect to time.

  • $\begingroup$ How is the drug prescription record measured? is it a binary variable for each time point or what? $\endgroup$
    – utobi
    Sep 6, 2022 at 22:43
  • $\begingroup$ Please see updates $\endgroup$ Sep 7, 2022 at 6:07
  • $\begingroup$ this post seems vague to me. The 'interruption' depends on many factors. At time $t$ you might have no observation because of a measurement problem or because of the data generation mechanism. Depending on what's the aim you might try a Poisson process approach. $\endgroup$
    – utobi
    Sep 7, 2022 at 20:18

1 Answer 1


Electronic healthcare records provide longitudinal data that don't have the regular observation times of panel or standard time-series data. Even if formal methods of interrupted (regularly spaced) time series analysis aren't appropriate, you can still model such longitudinal data as functions of time. Chapter 7 of Frank Harrell's course notes shows ways to handle data irregularly spaced in time from a set of individuals. Use an adequately flexible function to describe the time course of the data so that the irregularity of observation times doesn't matter.

For example, you can define the reference time = 0 as the intervention time for each patient, transform calendar dates into time relative to that reference, and evaluate levels or rates of change before and after the intervention. That's the basis of the difference-in-difference approach to trying to infer causality from observational data, which doesn't necessarily require identical observation times.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.