psmatch2 diff in diff

[Faiza Zafar]

Hi Community,

I am working on the psmtach2 command and trying to match firms with on a set of firm characteristics. I am quite new to this and want to confirm if my understanding of the matching is right.

My question is : Once i run the following command

 psmatch2 treated ROA Beta HRscore BS Size LEV in_id , out(WACC) n(1) caliper( 0.01) logit

I obtain the following results , as an eg please see the picture attached.

1 ) The unmatched difference in means should or should not be significant? does it matter?
2 )For an ideal match we want the ATT difference to be insignificant?
-> The interpretation is that once matched the control and treated groups should have similar approx similar means, and if there is a difference in the means it should be statistically insignificant?

Any help would be great

[George Ford]

Matching produces comparable means for the explanatory variables, not the outcome variable.

Your results suggest that the difference seen in the unmatched data is perhaps due to comparing unlike groups.

[Faiza Zafar]

Please correct me if i am wrong but ATT is showing the groups vs control variables and the difference between them for the main dependent variable? The pstest is used for checking the balancing of the independent variable

[Faiza Zafar]

Matching produces comparable means for the explanatory variables, not the outcome variable.

Your results suggest that the difference seen in the unmatched data is perhaps due to comparing unlike groups.

Does that mean that the unmatched sample t -p-value should be insignificant and once matched the ATT (p-value) should be significant?

[George Ford]

The unmatched difference is just the simple difference between the control and treated groups. The ATT is the means difference between the treated and control group once the sample is balanced/weighted (by the propensity score).

There is no "should be". It is what it is. In your case, there's a means difference in the raw data, but no difference in the balanced/weighted data. Thus, the raw means differences is presumably due to confounding.