Psmatch2 - Number of observations under common support different with "ate" option

[Amanda Lee Johnson]

Hello,

I'm using Stata 14 and have two problems when I use the "ate" option and omit the option from psmatch2.


When I run the psmatch2 command using the following options ", logit noreplacement common neighbor(1) ate" I get an equal number of treated and untreated participants "on support" (as expected):

psmatch2 ever_ecigw1 $covars_0 if wave1==1 & wave2==1 & sample_0==1, out(nvr2evr_cigsw2) logit noreplacement common neighbor(1) ate

However, Stata does not display the SE or the t-statistic for the ATT or ATE:

When I remove the ate option: ", logit noreplacement common neighbor(1)", the number of untreated participants is no longer equal to the number of treated participants, even when specifying neighbor(1):

psmatch2 ever_ecigw1 $covars_0 if wave1==1 & wave2==1 & sample_0==1, out(nvr2evr_cigsw2) logit noreplacement common neighbor(1)

However, the SE and t-statistic for the ATT are displayed:

Why does omitting the "ate" option create an unequal number of untreated and treated participants, even when specifying neighbor(1)? Am I able to limit the number of untreated participants to the number of treated participants and still get the standard error and t-statistic for the ATT?

Thank you,

Amanda Johnson

[David Radwin]

Are you certain that the first example yields equal numbers of treated and controls? Did you try excluding cases with missing weights after matching?

Here is a silly example that demonstrates what I mean. It initially looks like 460 treated and 460 controls, but after you exclude cases with missing weights, it's actually 460 treated and 150 controls.

. sysuse nlsw88, clear
(NLSW, 1988 extract)

. set seed 12345

. psmatch2 union age grade, out(south) logit noreplacement common neighbor(1) ate

Logistic regression                             Number of obs     =      1,876
                                                LR chi2(2)        =      16.28
                                                Prob > chi2       =     0.0003
Log likelihood = -1036.7974                     Pseudo R2         =     0.0078

------------------------------------------------------------------------------
       union |      Coef.   Std. Err.      z    P>|z|     [95% Conf. Interval]
-------------+----------------------------------------------------------------
         age |   .0116694   .0177773     0.66   0.512    -.0231734    .0465122
       grade |   .0845844   .0212503     3.98   0.000     .0429345    .1262342
       _cons |  -2.707582   .7643717    -3.54   0.000    -4.205724   -1.209441
------------------------------------------------------------------------------
There are observations with identical propensity score values.
The sort order of the data could affect your results.
Make sure that the sort order is random before calling psmatch2.
----------------------------------------------------------------------------------------
        Variable     Sample |    Treated     Controls   Difference         S.E.   T-stat
----------------------------+-----------------------------------------------------------
           south  Unmatched | .295652174   .467514124   -.17186195   .026248469    -6.55
                        ATT | .295652174   .267391304    .02826087   .024567277     1.15
                        ATU | .506521739   .295652174  -.210869565            .        .
                        ATE |                          -.091304348            .        .
----------------------------+-----------------------------------------------------------
Note: S.E. does not take into account that the propensity score is estimated.

 psmatch2: |   psmatch2: Common
 Treatment |        support
assignment | Off suppo  On suppor |     Total
-----------+----------------------+----------
 Untreated |       956        460 |     1,416
   Treated |         0        460 |       460
-----------+----------------------+----------
     Total |       956        920 |     1,876


. tab _treated _weight

           | psmatch2:
           | weight of
 psmatch2: |  matched
 Treatment |  controls
assignment |         1 |     Total
-----------+-----------+----------
 Untreated |       150 |       150
   Treated |       460 |       460
-----------+-----------+----------
     Total |       610 |       610


. tab _treated _weight, missing

 psmatch2: |  psmatch2: weight of
 Treatment |   matched controls
assignment |         1          . |     Total
-----------+----------------------+----------
 Untreated |       150      1,266 |     1,416
   Treated |       460          0 |       460
         . |         0        370 |       370
-----------+----------------------+----------
     Total |       610      1,636 |     2,246


. which psmatch2
.\psmatch2.ado
*! version 4.0.10 10feb2014 E. Leuven, B. Sianesi

[Amanda Lee Johnson]

Hi David,

Thank you for your feedback. It is good to know that I should utilize "_weight" and "_treated" to assess the number of treated and control participants that were used in the model, instead of relying on the output from psmatch2.

From your example, it looks as thought n=460 treated participants were matched with n=150 control participants, even though you specified no replacement. Does this mean that only n=150 of the treated participants were matched to the n=150 control participants?

Thank you for your help,

Amanda

[David Radwin]

That's a good question. I don't know the answer offhand, nor was I able to figure it out after some effort. I can tell you that others have had concerns about weights in psmatch2, such as http://stephenporter.org/understandi...atas-psmatch2/ .

[Amanda Lee Johnson]

Hi David,

Thank you for your feedback. I have also been searching online and have come up empty-handed. I appreciate your help and found the linked post very helpful in explaining weights in psmatch2.