I am trying to estimate the causal effects of the The Working Income Tax Benefit (WITB) on the labour supply of married women in Canada. The WITB is essentially equivalent to the EITC.

I am looking at two years in the Survey of Labour Income and Dynamics (SLID), 2006 and 2011. The WITB was implemented in 2007. The SLID is essentially equivalent to the Panel Study of Income Dynamics (PSID) in that it reports the labour market activity of individuals, which is what I'm concerned with.

I was able to retrieve the datasets online using ODESI. (I used "person files") I am using the .dta files.

I would like to conduct a Difference-in-Differences (D-i-D) regression and 1) estimate the probability that the individual was working, and 2) estimate how many hours the individual worked, given that the individual was working.

For 1) I will be using "alhrp28" which denotes annual hours paid during the year of observation.

For 2) I am interested in looking at "alfst28" which tells us labour force status during the year of observation. alfst28 corresponds to four values, "em" (employed), "Un" (unemployed), "No" (Not in Labour force), and ".c"

So the first thing I did was that I loaded both datsets into STATA using the "use" command. I noticed that the number of observations from 54,000 to 47,000 once I loaded the data for both years.

From here, I am stuck...not sure how to move forward.

I found some notes online and they say I need to generate a time dummy. But, I am a little confused. I am using panel data for the years 2006 and 2011. Do I need to generate a time dummy and run preliminary regressions before doing DiD?

The notes also say to separate treatment from control. I have been asked to do this using high school education. (I realize this doesn't work so I need to discuss why in my write-up).

Do I generate a "treated" variable in STATA?

And, for 2), what commands can I use to condition on whether the individual was employed?

**Note : I have taken graduate level econometrics, and time series. But, my knowledge of causal inference and things like maximum likelihood is quite poor.


  • $\begingroup$ It is impossible to say if those would be good solutions to do without knowing details of the dataset - there might be some alternative variables that will do the job better. In addition your control group should be the group that did not received treatment. If in this case EITC was just rolled out for everyone you could use something like regression discontinuity to exploit that different people had different likelihood of receiving treatment (if that empirically was case - but that often happens), but you cant say that there was a control group if everyone gets treatment $\endgroup$
    – 1muflon1
    Feb 13, 2021 at 15:40
  • $\begingroup$ @1muflon1 Thanks. I'll edit the post and elaborate more. $\endgroup$ Feb 13, 2021 at 15:43
  • $\begingroup$ @1muflon1 right in this case the "control" group isn't a control group at all. $\endgroup$ Feb 13, 2021 at 15:45
  • $\begingroup$ @1muflon1 updated my post. I loaded both datasets, but I am not sure how to move forward. I have never done a DiD before. $\endgroup$ Feb 14, 2021 at 13:07
  • $\begingroup$ for DiD you need some clear cut control group- eg if one Aust. Territory would implement EITC and another did not. Here you could probably use regression discontinuity or some other method but DiD would likely not work. $\endgroup$
    – 1muflon1
    Feb 14, 2021 at 13:37


Your Answer

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