Specifying gvar in csdid for repeated cross-section

[Teresa Janz]

Hello Stata community!

I have a question about CSDID, specifically how to specify gvar with repeated cross-sectional data. I am using Stata 18.

I would like to evaluate the impact of a cash transfer programme on poverty. I have yearly, repeated cross-sectional data from 2004-2019. Treatment (the cash transfer) occurs at the household level, and has started in 2009. After that, the programme has been rolled out continuously. Several districts in the country are geographically prioritised, but not all households within the same district are treated. Also, not all districts occur in every year of the repeated cross-sectional household survey.

I have created a treatment variable that takes the value 0 for years 2004-2009, as well as households who were not beneficiaries after, and 1 for households who are beneficiaries (see below, example for 2004-2014):

           |                                  treatment
      Year |         0       2009       2010       2011       2012       2013       2014 |     Total
-----------+-----------------------------------------------------------------------------+----------
      2004 |     3,698          0          0          0          0          0          0 |     3,698
      2005 |     1,896          0          0          0          0          0          0 |     1,896
      2006 |     2,435          0          0          0          0          0          0 |     2,435
      2007 |     2,041          0          0          0          0          0          0 |     2,041
      2008 |     1,986          0          0          0          0          0          0 |     1,986
      2009 |     1,842        119          0          0          0          0          0 |     1,961
      2010 |     1,781          0        174          0          0          0          0 |     1,955
      2011 |     1,662          0          0         95          0          0          0 |     1,757
      2012 |     1,633          0          0          0        221          0          0 |     1,854
      2013 |     1,741          0          0          0          0        176          0 |     1,917
      2014 |     1,609          0          0          0          0          0        219 |     1,828
-----------+-----------------------------------------------------------------------------+----------
     Total |    22,324        119        174         95        221        176        219 |    23,328

Eventually, with this specification,

csdid poor urban head_years_schooling head_literate head_female head_regionbirth head_age language, gvar(treatment) notyet time(year)

gives me 0 observations and all results are omitted.

I have two questions:
1) I understand that this is not correct, and the treatment variable should not be 0 for, e.g., years 2010-2014 when treatment is 2009, right? However, based on my repeated cross-sectional data, how can I identify which households remain in the treatment group from year to year?
2) Does it constitute a problem that the programme was rolled out in specific districts, but not all households in the same districts were treated each year? Eligibility was based on a set of household demographic and living standard measures.

I would appreciate any help regarding the coding of my gvar and how to implement csdid. Thank you very much in advance!
Teresa

[FernandoRios]

Hi Teresa
You are right regarding Gvar.
What you need to be able to answer is:
Is there a way that you could ID a household in 2008 (and any other year earlier) that could have benefited of the treatment f they were observed in 2009?
If you can answer this question, you can create your Gvar.
F

[Teresa Janz]

Thanks for your quick response, Fernando!

Unfortunately, since I only have repeated cross-sectional data, I cannot precisely identify households in 2008 who would have benefited of the treatment in 2009. The targeting formula for this programme is a rather complex mix of geographical prioritisation of districts, proxy-means test as well as some demographic conditions (e.g., having a child). I can approximate that, but not define with certainty that a household would be eligible.
Would that mean that csdid is not appropriate in my case? If so, would you think there would be any other way to estimate the programme's impact?

Best,
Teresa

[FernandoRios]

Not only csdid but did in general
remember the idea is to observe the same units across time (panel) or similar units (rc)

[Teresa Janz]

Thank you!