How to choose covariates for Coarsened Exact Matching?

[Fred Lee]

How to choose covariates for Coarsened Exact Matching?
Should we put all control variables into covariates or only choose some? Then how to explain why didn't include all control variables?
I know this is not a tech issue, could someone give me some guidance or reference, or discussion?
Thanks so a ton!

[George Ford]

start with -covbal- (you may need to ssc install covbal).

It will tell you how different the means for the variables listed.

You can start by balancing the vars with large standardized differences (>0.25). But note that when you CEM on some variables, it may affect the covariate balance for others.

There's a lot of flexibility with cem (scott, binning, etc), so you can work it until you get a well-balanced sample for any variable you intend to include in the model.

[Fred Lee]

Thanks so much, George Ford
I have about 20 variables, however, using -covbal-, I found only two vars with large standardized differences (>0.25). Should I only include these two as covariate? The two think is not enough.

[George Ford]

Start there. Then do covbal again using the cem_weights to see if it worked and didn't cause imbalance elsewhere.

The goal is a balanced sample. If one variable gets you there, then fine. If its all balanced without matching, then even better.

[Fred Lee]

That's great, thanks a ton!

[Moomal Khan]

Hi George Ford ,
could you please tell me where does the algorithm of CEM and psmatch2 differ?
I have been working on a project using CEM and psmatch2, using the same covariates, the result for psmatch2 appear to be significant whereas the results for cem are insignificant for the same variables? could you please let me know why is it so? or what is causing it?
thanks in advance

[George Ford]

Treatment insignificant or covariates?

[Moomal Khan]

covariates insignificant

[George Ford]

That's quite common. Often covariates are left out after CEM. See the King paper for discussion.

[George Ford]

psmatch2 is propensity score matching. CEM just makes sure the covariates are balanced. I'd do both if this is a school project.

[Moomal Khan]

so that would be okay if i use some covariates in psmatch2 and some others in CEM? or it has to be same?