Analyzing Intervention Effects in a Natural Experiment with Uneven Measurement Points

Question

I am currently working on an observational study aimed at understanding the impact of a certain intervention within a natural setting. Our dataset includes two distinct groups: a treatment group that received an intervention and a control group that did not. Below is an outline of my study's structure and the challenges I'm facing:

Treatment Group: This group includes around 50 individuals who were exposed to an intervention at different times, leading to a varied timeline of intervention across subjects. The dataset comprises multiple measurements of several features for each individual, taken at numerous time points both before and after the intervention. Notably, the number of measurements is not consistent across subjects. A key challenge here is the inconsistency in the number of measurements across subjects.

Control Group: This group did not receive the intervention but is comparable to the treatment group in all other respects. The data structure for the control group mirrors that of the treatment group, albeit without the intervention events.

Data Characteristics: I am treating the data as a multivariate time series, acknowledging that each subject's timeline varies and the sampling across these timelines is inconsistent. The covariates are also correlated with one another.

My main question is: Given the variability in intervention timelines and the inconsistent sampling rates across subjects, what statistical/casual models or techniques are best suited for analyzing the intervention's effects within this framework?

Any help will be appreciated,

Answer 1

Multilevel models do not require either that everyone get the same number of measurements or that they are equally spaced. Which MLM you should choose depends on other things such as the nature of the outcome (is it continuous? Dichotomous? or what?), the structure of the covariances (e.g. AR1), and whether you want to fit random slopes, random intercepts, or both.

All these issues have been discussed here many times. If you can't find answers for your particular case, you can ask a new question, but I hope this will get you started.