"convergence not achieved" issue in PSM

[Jason Li]

Thank you for viewing my question!

Now, I'm using Stata 14.0. The problem here is that I want to run a propensity score matching method to deal with my data.

The code I've used is "psmatch2 indicator login test friend discuss,outcome(contest) n(1) ate ties logit common noreplacement".

The running time is more than an hour, and the note is"
Note: 0 failures and 4 successes completely determined.
convergence not achieved
r(430);"

So, how can I solve this problem?

Thank you very much!!!

[Henrik Dalriksson]

I have the same problem! I am running a regression on EU data then I run the same regression on US data and I have this problem.

Did you find a solution?

[David Radwin]

For the warning that "0 failures and 4 successes completely determined," see this FAQ: http://www.stata.com/support/faqs/st...ic-regression/ .

For the error that "convergence not achieved," that means Stata cannot estimate the logit equation used to create the propensity scores. So you have two options:

1. Change the estimation method used to get propensity scores. For example, change the selection of covariates, increase the maximum number of iterations (set maxiter), transform one or more covariates in some fashion, or use probit instead of logit. This approach probably won't lead to a satisfactory solution, but it's worth considering.
2. Try a different matching method. You can use one of the other matching options in psmatch2 (Leuven and Sianesi, available from SSC), the built-in teffects nnmatch, or another user-written program like cem (coarsened exact matching; Iacus, King, and Porro; findit cem).

[David Radwin]

One more idea is to try the newly released kmatch by Ben Jann, described at http://www.statalist.org/forums/foru...score-matching .

[Charles Ehmat]

David Radwin I have a related question so I hope it is okay to tag on to this post.

My dataset has two different treatments, so I run 4 pairwise comparisons: control vs any treatment, control vs treatment 1, control vs treatment 2, treatment 1 vs treatment 2.

Due to some unbalancedness, I use PSM using the following command with a binary treatment variable for each of the four pairwise combinations

psmatch2 any_treatment indep1 indep2 indep3 ... indepN, outcome(depvar) kernel common

Afterwards, I run pstest to compare before and after matching differences in the dependent variables

pstest indep1 indep2 indep3 ... indepN, both

For the binary treatment variables comparing control vs any treatment and comparing control vs treatment 2, the code runs smoothly. When looking at control vs treatment 1, the psmatch2 command runs, however, the pstest command returns the convergence not achieved r(430) error. Comparing treatment 1 vs treatment 2 already returns this error after the psmatch2 command.

Why is that?

I am using Stata18 on Windows.

Thanks!

[David Radwin]

Sorry Charles Ehmat, I do not know the answer.