PSMatching: "The Balancing Properity is not satisfied"

[kristian kristiansen]

Hi!
I having a problem with the PSMatching (Nearest Neighbor)

I have a sample of public and private firms, which I want to match on size and industry.
Variable: Bors_aks(Dummy variable: public/private), gross_inv_scaled = measure the firms investment level, sumeiend = size of the firm, bransje = industry

I got the message: "The balancing property is not satisfied Try a different specification of the propensity score"

I have tried to read about it, but not sure how I should change the specification. Should I use ln(sumeiend)? should I specify that "bransje(industry" is a dummy variable?

Stata code:

* Define treatment, outcome and independent variables
global treatment bors_aks
global ylist gross_inv_scaled
global xlist sumeiend bransje
global breps 5

* PSM with common support

pscore $treatment $xlist, pscore(myscore) blockid(myblock) comsup

* matching methods

* nearest neighbor matching

attnd $ylist $treatment $xlist, pscore(myscore) comsup boot reps($breps) dots

[Melissa Garrido]

Hi Kristian,
Transformations of covariates, including log transformations, are a possibility for improving covariate balance. However, if you only have 2 covariates, you might consider exact matching over propensity score matching. If you haven't already, you should also check whether one of your covariates perfectly predicts treatment assignment.
Hope this helps,
Melissa

[Steve Samuels]

Kristian, the FAQ ask that you not post photos for good reasons:; photos 1) are usually too small to read; 2) they often require that would-be viewers open a new browser window that blocks the original post, a real nuisance; and, even in the new window 3) photos may be still be too small; or 4) they may be blurry. . Your photo falls under reasons 1, 2, and 4. Please next time follow the FAQ #12 request and post code and results between CODE delimiters. The opening delimiter is "[C O D E]" and the closing delimiter is "[/C O D E]", but with spaces removed.

[kristian kristiansen]

What is the consequences of the balancing property problem?
- Is it possible to use the results?
- Or do you suggest to use the coarsened exact matching method instead?

[Enrique Pinzon (StataCorp)]

Dear Kristian,

I do not know the answer to your last question but we have a good discussion in the Stata documentation about balancing. Also, there is some additional discussion that may be helpful in http://www.stata.com/stata-news/news...tment-effects/

[Melissa Garrido]

Propensity scores are used to create treatment and comparison groups that are as similar as possible on a set of observed covariates hypothesized to be associated with both treatment and outcome. If the balancing property is not satisfied, the treatment and comparison groups are unlikely to be sufficiently similar to reduce selection bias in your treatment effect estimate. Now, if you have a single unbalanced covariate in one block of a propensity score constructed from several covariates, you may still be able to proceed with your analysis. In your case, where imbalance exists but you only have 2 covariates, it seems that you may not have an adequate comparison group to use for your analyses (whether via propensity scores or matching).