Propensity Score Analysis: view matching results (suitability) for each variable & display matching as scatter plot (using teffects psmatch)

[Yvonne Schmiedel]

Dear all,

Using Stata V15, I am intending to do a propensity score analysis for observational data.

I would like to see 1) how well each variable matches 2) create a scatter plot on the paired matches

One of my collegues using R showed me an example, but despite all seaching I could not find the function in Stata how to do it.

I cannot download packages (i.e. psmatch2), I have to use the teffects psmatch command.

The data analysis is a 2 step analysis: 1) main exposure variable BC_candida_all (128 vs 84 patients) 2) main exposure variable antifungal_yn (54 vs 74 patients).


This is how far I got:

teffects psmatch (death_yn) (BC_candida_all antifungal_yn Livercirrhosis Haemodialysis cancer Lungdisease ICUyn, logit)
teffects psmatch (death_yn) (BC_candida_all antifungal_yn Livercirrhosis Haemodialysis cancer Lungdisease ICUyn, logit), atet
teffects overlap


I would be grateful for any suggestions how to get a statistical output on matching result as well as a scatter plot (not only overlap).

Many thanks,

Yvonne

[Mike Lacy]

You wrote:
"I would like to see 1) how well each variable matches 2) create a scatter plot on the paired matches"

Perhaps other people will understand well what you mean, but I'm not clear about what you mean.

1) How well each variable matches *what*? Perhaps you mean you want to check the so-called "balancing property," that is, whether there are substantial mean differences between treatment and matched control subjects?
2) A scatter plot of what vs. what?

I would say, though, that I have found the documentation of -teffects- to be obscure, at least by comparison to that for other Stata commands, so I can appreciate your difficulties.

[Yvonne Schmiedel]

Dear Mike,

thanks for your reply.

I try and be clearer:

I would like to get an output, which gives me a p-value on each variable regarding its matching score / quality.

In R (which I dont know how to use) there is a function for a "scatter" / "jetter" plot.

I cannot copy it here, but it shows the distribution of the propensity score:

Unmated treatment units:

Space

Matched Treatment Units

....... ...................... ............


Matched Control Units

.... .............. ..... ...............

Unmatched Control Units

....... ......



Each dot represents a patient, the more match treatment and control units are in line the better the matching.

Of course, R is a different programme and there may not be the exact same function in Stata, but at least something representative.

I have been searching for some time, but neither found a statitical output on matching quality of the variables nor an appropriate graph.

The overlap graph leaves a lot of room for "free interpretation", so I do not find it very useful.


I hope this has made somewhat clearer what I am looking for.

Many thanks again.