To give a background of the problem: I'm performing a difference-in-differences analysis to estimate the effect of a treatment, say A. I have a panel data with 5 rounds and about 3000 observations in each round. The treatment was introduced between rounds 2 and 3, thus treated rounds are 3,4,5 while untreated rounds are 1,2. The treatment intensity varies at the level of locality, the data for which have been obtained from a separate administrative source and matched with the survey data. The introduction of the treatment was staggered at the level of locality so I also have a measure of years of exposure to treatment. I intend to obtain the difference-in-difference estimator from interaction between treatment dummy,years of exposure, and treatment intensity.

Now the problem is data for some years (corresponding to round 5 of the survey) are missing in the administrative records. For all the localities, treatment intensity data corresponding to round 5 are missing. Out of these localities, some localities have data on treatment intensity at a later year.

I'm considering using the later year's treatment intensity as a proxy for round 5 treatment intensity for the localities available (which would introduce measurement errors), and leaving the rest of the localities missing on treatment intensity in that round. Can this be done?


  • $\begingroup$ A bit more info would probably be helpful. Is there a reason we could expect treatment intensity in year 6 (or later) to be a rough match for year 5? Are there systematic differences in which localities report intensity in later years? Regarding importance: how much do outcome variables tend to lag the treatment (i.e. do you observe most of your effect in years 3 and 4, or is year 5 critical for your estimate)? Also, to clarify: are all localities in your sample being treated at some point, or do you have a control group that's never treated? $\endgroup$
    – AndrewC
    Jul 29, 2021 at 16:13
  • $\begingroup$ From the data it can be seen that treatment intensity has fallen significantly in round 4 and the round 6 value for treatment intensity is somewhat higher than round 4 value, but still closer to round 4 value than value in earlier rounds. So one can expect a trend from round 4 to round 6 such that round 5 value can be approximated by round 6 value. As for importance, I am looking at outcomes that manifest over the longer term so round 5 would be critical to my estimate. In my sample, all the localities are being treated so I do not have a control group that is never treated. $\endgroup$
    – Jan3
    Jul 30, 2021 at 6:54


Your Answer

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