Are there any examples of the difference in differences model being used in medical research? What are the thoughts of statisticians on this quasi-experimental approach?
DiD historically has been used more in econometrics, where panel data (and instrumental variable methods and policy changes) is more common, than in medicine/epidemiology where data is more commonly individualized.
There is a good primer for DiD in healthcare settings by Ellen Caniglia and Eleanor Murray: Difference-in-Difference in the Time of Cholera: a Gentle Introduction for Epidemiologists - PMC
And one recent example is investigating the effect of masking policies in schools on COVID-19 incidence: https://www.nejm.org/doi/full/10.1056/NEJMoa2211029
(note that here, too, the underlying data is aggregated time-series data)
Thank you Ehud. My interest is the before and after effects of a change in regulations regarding drug administration. The change impacted one age group only so there is a possibility of a treatment and control comparison.
Can the difference-in-difference be used if the outcome is BINARY- i.e., yes, I got the treatment (==1) and NO (==0)?
You asked about a binary outcome, yet from your example it sounds like a binary treatment. I hope I got your question right, I’ll try to answer for a binary response.
But tbh, I don’t have a high-confidence answer for that.
On the one hand, simple DiD estimators can be implemented with a simple interaction term between treatment and pre-post-time indicator. Therefore, I don’t see much objections to incorporate that into any generalized linear model (which I think is also the framework Wooldridge advocates for in Simple approaches to nonlinear difference-in-differences with panel data | The Econometrics Journal | Oxford Academic).
On the other hand, I can’t think of any examples for nonlinear DiD, and I suspect it has to do with economists’ unexplained fond to “linear probability models” (basically running OLS when the outcome is binary).