difference-in-differences with Decennial Census Data

[Gene Korku]

Hello @FernandoRios ,

Is it possible to use csdid, jwdid and other did packages with decennial census data (Counties are observed every 10 years)? If not can synth did work?

[Jared Greathouse]

It's certainly possible mathematically, the issue is, is it DESIRABLE. Think of how much stuff happens over a 10 year period that synthetic controls would be missing out on. Even for DID. I mean it's possible, but how could we justify parallel trends in such a setting when the data are in 10 year increments? What if PTA would not hold in the between 10 years? Not impossible at all, but you'd need to be very very very sure parallel trends is plausible.

[Gene Korku]

It's certainly possible mathematically, the issue is, is it DESIRABLE. Think of how much stuff happens over a 10 year period that synthetic controls would be missing out on. Even for DID. I mean it's possible, but how could we justify parallel trends in such a setting when the data are in 10 year increments? What if PTA would not hold in the between 10 years? Not impossible at all, but you'd need to be very very very sure parallel trends is plausible.

In that case, assuming I proceed with such data would it suffice to say that making causal claims may probably not be ideal, is it also possible to estimate some long differences using panel data approaches and just argue that these findings are just correlations?

[Jared Greathouse]

Well, it depends on context. Suppose we're dealing with election data that only come every 4 years anyways. Then this is okay because the data only come in 4 year increments (in my country anyways, for the time being). My point is, you'd need to have quite a lot of pre-intervention data, and since you're working with 10 year gaps, it's just hard to argue PTA (to me, I could be wrong) in this instance. I'm not familiar with long differencing from the top of my head, that would be a question Fernando is more equipped to answer.

[Gene Korku]

Well, it depends on context. Suppose we're dealing with election data that only come every 4 years anyways. Then this is okay because the data only come in 4 year increments (in my country anyways, for the time being). My point is, you'd need to have quite a lot of pre-intervention data, and since you're working with 10 year gaps, it's just hard to argue PTA (to me, I could be wrong) in this instance. I'm not familiar with long differencing from the top of my head, that would be a question Fernando is more equipped to answer.

Thank you.