Using matching with DID for repeated cross sectional data

[Bhawna Taneja]

I am new to STATA and working on repeated cross-sectional data for Difference-in-Difference analysis. I have treatment and control groups before and after the intervention. My treatment and control groups are, however, unbalanced (the covariates are significantly different across the two groups both in pre-and post-intervention). I, therefore, want to apply the matching technique to match the three groups (pre-treatment, pre-control, post-control) to my post-treatment group and then use DID. Any guidance about the STATA commands and references to papers that have applied this technique would be helpful.

[Maxence Morlet]

Econometrically speaking, I wouldn't resort to matching on covariates if I were you. You may want to check out Zeldow and Hatfield (2021) and Arkhangelsky et al. (2021) to see why.

You may want to check out community contributed commands sdid or scul.

[Bhawna Taneja]

Thank you for replying. I will look at these references.

[Jared Greathouse]

I am the author of scul, and if you decide to use it, I do not recommend you use covariates in the estimation process. I give the user the option to do so should they want to, but it really isn't needed and is arguably undesirable. I'm also not understanding the matching technique you discuss. You write

I, therefore, want to apply the matching technique to match the three groups (pre-treatment, pre-control, post-control) to my post-treatment group

This isn't how it works! If this were machine learning, we'd say this is training your model on your test data. You want to match on pre-intervention qualities only. In fact, let me give an example of doing just this. This will be an excerpt from my SJ paper, which has recieved very good feedback from the reviewers. Consider the Proposition 99 example for synthetic controls. If you haven't, install lassopack if you wanna follow along with me.

version 17
//set varabbrev off

clear *

set scheme sj

use "https://github.com/scunning1975/mixtape/blob/master/smoking.dta?raw=true", clear

xtset state year, y

loc int_time = 1989

local lbl: value label `r(panelvar)'


loc unit ="California":`lbl'

g treat = cond(`r(panelvar)'==`unit' & `r(timevar)' >= `int_time',1,0)
lab var cigsale "Cigarette Sales"
lab var year "Year"

lab var treat "Proposition 99"


/*#########################################################

    * Figure 1

    Creates a SC using OLS and SCM
*########################################################*/

*** SCUL
keep cigsale state year

label var year "Year"

qui greshape wide cigsale, j(state) i(year)

tsset year

order cigsale`unit', a(year)
cls
cvlasso cigsale`unit' cigsale1-cigsale39 if year < `int_time', lopt h(5) lglmnet roll

predict cf, lopt

keep year cf cigsale`unit'

label var cigsale "Real California"

label var cf "SCUL California"

twoway (tsline cig, lcolor(black)) ///
(tsline cf, lcolor(red) lwidth(medium) lpattern(shortdash)) //

What I am doing here, is I am fitting a LASSO to the treated unit of interest (California) predicting from 1970 to 1988, since the intervention took place in 1989. If you do any matching here, you'll wanna match on pre-treatment characteristics only, not on post outcome values.

[Geremew Worku]

Dear @Bhawna Taneja; I face also the same problem. If someone has a way to find its application using STATA.

I am new to STATA and working on repeated cross-sectional data for Difference-in-Difference analysis. I have treatment and control groups before and after the intervention. My treatment and control groups are, however, unbalanced (the covariates are significantly different across the two groups both in pre-and post-intervention). I, therefore, want to apply the matching technique to match the three groups (pre-treatment, pre-control, post-control) to my post-treatment group and then use DID. Any guidance about the STATA commands and references to papers that have applied this technique would be helpful.