Hello all,
I hope you are well.
I've been playing around with the pstest post-estimation command for the psmatch2 command as I have been running a propensity score.
To me, the whole idea of PSM was that we would compare the ATT subject to all characteristics being the same between the treated and control subject to one group having treatment (versus one not).
The command I used is " pstest, both graph "
So although I do have overlap in my test (suggesting very similar covariates among overlapping units of obvs), however by pstest is giving rather opposite results. Firstly, the Pseudo r2 is 1 upon being matched and there is a large degree of bias - in fact, it seems the bias increases following matching.
I'm guessing the pstest balance check applies for all covariates, but as my data has many missing values, I have only generated overlap for a relatively small number of observations.
Should I still pay attention to the lack of 'macro' overlap + is there a way to do the pstest only for the matched observations? I thought that's all we should care about.
Please do correct wherever I am going wrong - much appreciate any insight.
Kind regards,
Rishi
I hope you are well.
I've been playing around with the pstest post-estimation command for the psmatch2 command as I have been running a propensity score.
To me, the whole idea of PSM was that we would compare the ATT subject to all characteristics being the same between the treated and control subject to one group having treatment (versus one not).
The command I used is " pstest, both graph "
So although I do have overlap in my test (suggesting very similar covariates among overlapping units of obvs), however by pstest is giving rather opposite results. Firstly, the Pseudo r2 is 1 upon being matched and there is a large degree of bias - in fact, it seems the bias increases following matching.
I'm guessing the pstest balance check applies for all covariates, but as my data has many missing values, I have only generated overlap for a relatively small number of observations.
Should I still pay attention to the lack of 'macro' overlap + is there a way to do the pstest only for the matched observations? I thought that's all we should care about.
Please do correct wherever I am going wrong - much appreciate any insight.
Kind regards,
Rishi
Comment