Propensity Score Matching in Stata using teffects

Note: readers interested in this article should also be aware of King and Nielson's 2019 paper Why Propensity Scores Should Not Be Used for Matching.

For many years, the standard tool for propensity score matching in Stata has been the psmatch2 command, written by Edwin Leuven and Barbara Sianesi. However, Stata 13 introduced a new teffects command for estimating treatments effects in a variety of ways, including propensity score matching. The teffects psmatch command has one very important advantage over psmatch2: it takes into account the fact that propensity scores are estimated rather than known when calculating standard errors. This often turns out to make a significant difference, and sometimes in surprising ways. We thus strongly recommend switching from psmatch2 to teffects psmatch, and this article will help you make the transition.

An Example of Propensity Score Matching

Run the following command in Stata to load an example data set:

use http://ssc.wisc.edu/sscc/pubs/files/psm

It consists of four variables: a treatment indicator t, covariates x1 and x2, and an outcome y. This is constructed data, and the effect of the treatment is in fact a one unit increase in y. However, the probability of treatment is positively correlated with x1 and x2, and both x1 and x2 are positively correlated with y. Thus simply comparing the mean value of y for the treated and untreated groups badly overestimates the effect of treatment:

ttest y, by(t)

(Regressing y on t, x1, and x2 will give you a pretty good picture of the situation.)

The psmatch2 command will give you a much better estimate of the treatment effect:

psmatch2 t x1 x2, out(y)

----------------------------------------------------------------------------------------
        Variable     Sample |    Treated     Controls   Difference         S.E.   T-stat
----------------------------+-----------------------------------------------------------
               y  Unmatched |  1.8910736  -.423243358   2.31431696   .109094342    21.21
                        ATT |  1.8910736   .871388246   1.01968536   .173034999     5.89
----------------------------+-----------------------------------------------------------
Note: S.E. does not take into account that the propensity score is estimated.

The teffects Command

You can carry out the same estimation with teffects. The basic syntax of the teffects command when used for propensity score matching is:

teffects psmatch (outcome) (treatment covariates)

In this case the basic command would be:

teffects psmatch (y) (t x1 x2)

However, the default behavior of teffects is not the same as psmatch2 so we'll need to use some options to get the same results. First, psmatch2 by default reports the average treatment effect on the treated (which it refers to as ATT). The teffects command by default reports the average treatment effect (ATE) but will calculate the average treatment effect on the treated (which it refers to as ATET) if given the atet option. Second, psmatch2 by default uses a probit model for the probability of treatment. The teffects command uses a logit model by default, but will use probit if the probit option is applied to the treatment equation. So to run the same model using teffects type:

teffects psmatch (y) (t x1 x2, probit), atet

Treatment-effects estimation                    Number of obs      =      1000
Estimator      : propensity-score matching      Matches: requested =         1
Outcome model  : matching                                      min =         1
Treatment model: probit                                        max =         1
------------------------------------------------------------------------------
             |              AI Robust
           y |      Coef.   Std. Err.      z    P>|z|     [95% Conf. Interval]
-------------+----------------------------------------------------------------
ATET         |
           t |
   (1 vs 0)  |   1.019685   .1227801     8.30   0.000     .7790407     1.26033
------------------------------------------------------------------------------

The average treatment effect on the treated is identical, other than being rounded at a different place. But note that teffects reports a very different standard error (we'll discuss why that is shortly), plus a Z-statistic, p-value, and 95% confidence interval rather than just a T-statistic.

Running teffects with the default options gives the following:

teffects psmatch (y) (t x1 x2)

Treatment-effects estimation                    Number of obs      =      1000
Estimator      : propensity-score matching      Matches: requested =         1
Outcome model  : matching                                      min =         1
Treatment model: logit                                         max =         1
------------------------------------------------------------------------------
             |              AI Robust
           y |      Coef.   Std. Err.      z    P>|z|     [95% Conf. Interval]
-------------+----------------------------------------------------------------
ATE          |
           t |
   (1 vs 0)  |   1.019367   .1164694     8.75   0.000     .7910912    1.247643
------------------------------------------------------------------------------

This is equivalent to:

psmatch2 t x1 x2, out(y) logit ate

----------------------------------------------------------------------------------------
        Variable     Sample |    Treated     Controls   Difference         S.E.   T-stat
----------------------------+-----------------------------------------------------------
               y  Unmatched |  1.8910736  -.423243358   2.31431696   .109094342    21.21
                        ATT |  1.8910736   .930722886   .960350715   .168252917     5.71
                        ATU |-.423243358   .625587554   1.04883091            .        .
                        ATE |                           1.01936701            .        .
----------------------------+-----------------------------------------------------------
Note: S.E. does not take into account that the propensity score is estimated.

The ATE from this model is very similar to the ATT/ATET from the previous model. But note that psmatch2 is reporting a somewhat different ATT in this model. The teffects command reports the same ATET if asked:

teffects psmatch (y) (t x1 x2), atet

Treatment-effects estimation                    Number of obs      =      1000
Estimator      : propensity-score matching      Matches: requested =         1
Outcome model  : matching                                      min =         1
Treatment model: logit                                         max =         1
------------------------------------------------------------------------------
             |              AI Robust
           y |      Coef.   Std. Err.      z    P>|z|     [95% Conf. Interval]
-------------+----------------------------------------------------------------
ATET         |
           t |
   (1 vs 0)  |   .9603507   .1204748     7.97   0.000     .7242245    1.196477
------------------------------------------------------------------------------

Standard Errors

The output of psmatch2 includes the following caveat:

Note: S.E. does not take into account that the propensity score is estimated.

A recent paper by Abadie and Imbens (2012. Matching on the estimated propensity score. Harvard University and National Bureau of Economic Research) established how to take into account that propensity scores are estimated, and teffects psmatch relies on their work. Interestingly, the adjustment for ATE is always negative, leading to smaller standard errors: matching based on estimated propensity scores turns out to be more efficient than matching based on true propensity scores. However, for ATET the adjustment can be positive or negative, so the standard errors reported by psmatch2 may be too large or to small.

Handling Ties

Thus far we've used psmatch2 and teffects psmatch to do simple nearest-neighbor matching with one neighbor (and no caliper). However, this raises the question of what to do when two observations have the same propensity score and are thus tied for "nearest neighbor." Ties are common if the covariates in the treatment model are categorical or even integers.

The psmatch2 command by default matches with one of the tied observations, but with the ties option it matches with all tied observations. The teffects psmatch command always matches with all ties. If your data set has multiple observations with the same propensity score, you won't get exactly the same results from teffects psmatch as you were getting from psmatch2 unless you go back and add the ties option to your psmatch2 commands. (At this time we are not aware of any clear guidance as to whether it is better to match with ties or not.)

Matching With Multiple Neighbors

By default teffects psmatch matches each observation with one other observation. You can change this with the nneighbor() (or just nn()) option. For example, you could match each observation with its three nearest neighbors with:

teffects psmatch (y) (t x1 x2), nn(3)

Postestimation

By default teffects psmatch does not add any new variables to the data set. However, there are a variety of useful variables that can be created with options and post-estimation predict commands. The following table lists the 1st and 467th observations of the example data set after some of these variables have been created. We'll refer to it as we explain the commands that created the new variables. Reviewing these variables is also a good way to make sure you understand exactly how propensity score matching works.

      
      +-------------------------------------------------------------------------------------------------------+
      |        x1          x2   t          y   match1        ps0        ps1          y0         y1         te |
      |-------------------------------------------------------------------------------------------------------|
   1. |  .0152526   -1.793022   0   -1.79457      467   .9081651   .0918349    -1.79457   2.231719   4.026289 |
 467. | -2.057838    .5360286   1   2.231719      781    .907606    .092394   -.6012772   2.231719   2.832996 |
      +-------------------------------------------------------------------------------------------------------+

Start with a clean slate by typing:

use http://ssc.wisc.edu/sscc/pubs/files/psm, replace

The gen() option tells teffects psmatch to create a new variable (or variables). For each observation, this new variable will contain the number of the observation that observation was matched with. If there are ties or you told teffects psmatch to use multiple neighbors, then gen() will need to create multiple variables. Thus you supply the stem of the variable name, and teffects psmatch will add suffixes as needed.

teffects psmatch (y) (t x1 x2), gen(match)

In this case each observation is only matched with one other, so gen(match) only creates match1. Referring to the example output, the match of observation 1 is observation 467 (which is why those two are listed).

Note that these observation numbers are only valid in the current sort order, so make sure you can recreate that order if needed. If necessary, run:

gen ob=_n

and then:

sort ob

to restore the current sort order.

The predict command with the ps option creates two variables containing the propensity scores, or that observation's predicted probability of being in either the control group or the treated group:

predict ps0 ps1, ps

Here ps0 is the predicted probability of being in the control group (t=0) and ps1 is the predicted probability of being in the treated group (t=1). Observations 1 and 467 were matched because their propensity scores are very similar.

The po option creates variables containing the potential outcomes for each observation:

predict y0 y1, po

Because observation 1 is in the control group, y0 contains its observed value of y. y1 is the observed value of y for observation 1's match, observation 467. The propensity score matching estimator assumes that if observation 1 had been in the treated group its value of y would have been that of the observation in the treated group most similar to it (where "similarity" is measured by the difference in their propensity scores).

Observation 467 is in the treated group, so its value for y1 is its observed value of y while its value for y0 is the observed value of y for its match, observation 781.

Running the predict command with no options gives the treatment effect itself:

predict te

The treatment effect is simply the difference between y1 and y0. You could calculate the ATE yourself (but emphatically not its standard error) with:

sum te

and the ATET with:

sum te if t

Regression on the "Matched Sample"

Another way to conceptualize propensity score matching is to think of it as choosing a sample from the control group that "matches" the treatment group. Any differences between the treatment and matched control groups are then assumed to be a result of the treatment. Note that this gives the average treatment effect on the treated—to calculate the ATE you'd create a sample of the treated group that matches the controls. Mathematically this is all equivalent to using matching to estimate what an observation's outcome would have been if it had been in the other group, as described above.

Sometimes researchers then want to run regressions on the "matched sample," defined as the observations in the treated group plus the observations in the control group which were matched to them. The problem with this approach is that the matched sample is based on propensity scores which are estimated, not known. Thus the matching scheme is an estimate as well. Running regressions after matching is essentially a two stage regression model, and the standard errors from the second stage must take the first stage into account, something standard regression commands do not do. This is an area of ongoing research.

We will discuss how to run regressions on a matched sample because it remains a popular technique, but we cannot recommend it.

psmatch2 makes it easy by creating a _weight variable automatically. For observations in the treated group, _weight is 1. For observations in the control group it is the number of observations from the treated group for which the observation is a match. If the observation is not a match, _weight is missing. _weight thus acts as a frequency weight (fweight) and can be used with Stata's standard weighting syntax. For example (starting with a clean slate again):

use http://ssc.wisc.edu/sscc/pubs/files/psm, replace
psmatch2 t x1 x2, out(y) logit
reg y x1 x2 t [fweight=_weight]

Observations with a missing value for _weight are omitted from the regression, so it is automatically limited to the matched sample. Again, keep in mind that the standard errors given by the reg command are incorrect because they do not take into account the matching stage.

teffects psmatch does not create a _weight variable, but it is possible to create one based on the match1 variable. Here is example code, with comments:

gen ob=_n //store the observation numbers for future use
save fulldata,replace // save the complete data set

keep if t // keep just the treated group
keep match1 // keep just the match1 variable (the observation numbers of their matches)
bysort match1: gen weight=_N // count how many times each control observation is a match
by match1: keep if _n==1 // keep just one row per control observation
ren match1 ob //rename for merging purposes

merge 1:m ob using fulldata // merge back into the full data
replace weight=1 if t // set weight to 1 for treated observations

The resulting weight variable will be identical to the _weight variable created by psmatch2, as can be verified with:

assert weight==_weight

It is used in the same way and will give exactly the same results:

reg y x1 x2 t [fweight=weight]

Obviously this is a good bit more work than using psmatch2. If your propensity score matching model can be done using both teffects psmatch and psmatch2, you may want to run teffects psmatch to get the correct standard error and then psmatch2 if you need a _weight variable.

This regression has an N of 666, 333 from the treated group and 333 from the control group. However, it only uses 189 different observations from the control group. About 1/3 of them are the matches for more than one observation from the treated group and are thus duplicated in the regression (run tab weight if !t for details). Researchers sometimes use the norepl (no replacement) option in psmatch2 to ensure each observation is used just once, even though this generally makes the matching worse. To the best of our knowledge there is no equivalent with teffects psmatch.

The results of this regression leave somewhat to be desired:

------------------------------------------------------------------------------
           y |      Coef.   Std. Err.      t    P>|t|     [95% Conf. Interval]
-------------+----------------------------------------------------------------
          x1 |    1.11891   .0440323    25.41   0.000      1.03245    1.205369
          x2 |    1.05594   .0417253    25.31   0.000       .97401     1.13787
           t |   .9563751   .0802273    11.92   0.000     .7988445    1.113906
       _cons |   .0180986   .0632538     0.29   0.775    -.1061036    .1423008
------------------------------------------------------------------------------

By construction all the coefficients should be 1. Regression using all the observations (reg y x1 x2 t rather than reg y x1 x2 t [fweight=weight]) does better in this case:

------------------------------------------------------------------------------
           y |      Coef.   Std. Err.      t    P>|t|     [95% Conf. Interval]
-------------+----------------------------------------------------------------
          x1 |   1.031167   .0346941    29.72   0.000     .9630853    1.099249
          x2 |   .9927759   .0333297    29.79   0.000     .9273715     1.05818
           t |   .9791484   .0769067    12.73   0.000     .8282306    1.130066
       _cons |   .0591595   .0416008     1.42   0.155    -.0224758    .1407948
------------------------------------------------------------------------------

Other Methods of Estimating Treatment Effects

While propensity score matching is the most common method of estimating treatment effects at the SSCC, teffects also implements Regression Adjustment (teffects ra), Inverse Probability Weighting (teffects ipw), Augmented Inverse Probability Weighting (teffects aipw), Inverse Probability Weighted Regression Adjustment (teffects ipwra), and Nearest Neighbor Matching (teffects nnmatch). The syntax is similar, though it varies whether you need to specify variables for the outcome model, the treatment model, or both:

teffects ra (y x1 x2) (t)
teffects ipw (y) (t x1 x2)
teffects aipw (y x1 x2) (t x1 x2)
teffects ipwra (y x1 x2) (t x1 x2)
teffects nnmatch (y x1 x2) (t)

Complete Example Code

The following is the complete code for the examples in this article.

clear all
use http://www.ssc.wisc.edu/sscc/pubs/files/psm

ttest y, by(t)
reg y x1 x2 t

psmatch2 t x1 x2, out(y)
teffects psmatch (y) (t x1 x2, probit), atet

teffects psmatch (y) (t x1 x2)
psmatch2 t x1 x2, out(y) logit ate
teffects psmatch (y) (t x1 x2), atet

use http://www.ssc.wisc.edu/sscc/pubs/files/psm, replace

teffects psmatch (y) (t x1 x2), gen(match)

predict ps0 ps1, ps
predict y0 y1, po
predict te
l if _n==1 | _n==467

use http://www.ssc.wisc.edu/sscc/pubs/files/psm, replace

psmatch2 t x1 x2, out(y) logit
reg y x1 x2 t [fweight=_weight]

gen ob=_n
save fulldata,replace

teffects psmatch (y) (t x1 x2), gen(match)
keep if t
keep match1
bysort match1: gen weight=_N
by match1: keep if _n==1
ren match1 ob

merge 1:m ob using fulldata
replace weight=1 if t

assert weight==_weight

reg y x1 x2 t [fweight=weight]
reg y x1 x2 t

teffects ra (y x1 x2) (t)
teffects ipw (y) (t x1 x2)
teffects aipw (y x1 x2) (t x1 x2)
teffects ipwra (y x1 x2) (t x1 x2)
teffects nnmatch (y x1 x2) (t)

Last Revised: 2/16/2015