Using Optmatch on data in SAS, Stata, etc

Ben B. Hansen, Mark Fredrickson, Josh Errickson, Josh Buckner

2024-09-20

General comments

If a users preferred data analysis software is other than R, optmatch can still easily be used to perform the matching while all other data analysis can be performed in the preferred software.

In general, the procedure will be

1. In your preferred software, export a data set containing a treatment indicator and all variables to match, exactMatch or caliper on.
   • (Optionally, if you wish to match using a propensity score, fit such a model and include the predicted propensity scores in the data set.)
2. Import the data into R.
3. Perform the matching in R using optmatch.
4. Export the data from R, including information on the matched sets.
5. Import the data back into your preferred software.

The most general way to import the data back-and-forth are using comma separated value files (.csv files), which any statistical software should be able to read & write.

For .csv files, sample R code may be

> externaldata <- read.csv("externaldata.csv", header = TRUE)
> externaldata$match <- fullmatch(..., data = externaldata)
> write.csv(externaldata, file = "externaldata.matched.csv")

An example of doing this sort of operation with SAS is below. Following that, we demonstrate a similar procedure in Stata using the R package haven which reads and writes Stata’s .dta files directly.

Using Optmatch with SAS

For this example, lets say we have some simple demographics. We will treat gender as the treatment indicator, and wish to match on a combination of a propensity score for gender (using both age and height) and age.

data people;
  infile datalines dsd dlm=' ' missover;
  input gender age height;

datalines;
0 25 62
0 41 68
0 38 63
0 22 62
1 33 70
1 35 71
1 47 68
1 23 64
;
run;

Now we can fit a logistic model to predict gender using age and height.

proc logistic data = people;
  model gender (event='1') = age height;
  output out = preddata p=ppty;
run;

Finally, since we want to match only on the new ppty propensity score and age, we can drop height.

data newpeople;
  set preddata;
  keep gender age ppty;
run;

Export the file from SAS into a .csv file.

proc export data=newpeople;
  outfile="/path/to/save/sasout.csv";
run;

Inside R, we can load this data.

> sasdata <- read.csv("/path/to/save/sasout.csv", header = TRUE)

If you have string variables (e.g. race as “White”, “Hispanic”, etc), you may need to include the argument stringsAsFactors = FALSE. (This is the default in current R, but older versions of R had TRUE as the default.)

Now, perform matching as desired, saving the final match to sasdata. For example,

> library(optmatch)
> f <- fullmatch(gender ~ age + ppty, data = sasdata)
> sasdata$match <- f

Save this data back to .csv as follows.

> write.csv(sasdata, "/path/to/saverout.sas.csv", row.names = FALSE)

The use of row.names = FALSE stops R from including the row names (likely 1, 2, 3, etc) as the first column in the data. If you re-arranged the data at any point, you may need to set that to TRUE, but keep in mind to handle it properly in SAS, as the default will be to treat it as a variable.

Now, returning to SAS, we can read the new rout.sas.csv file in. The only catch is that we want to ensure that the match is read as a string by using $, since it may have values like 1.1 and 1.10, representing two different matches, but which are identical if treated as numeric.

data matchedpeople;
  infile "/path/to/save/rout.sas.csv" dsd firstobs=2;
  input gender age ppty match $;
run;

The argument firstobs=2 skips the variable names; alternatively you could pass col.names=FALSE to R’s write.csv, but then the rout.sas.csv file lacks any variable information, which may be useful to have.

A Note about Ordering

If you carry out any additional operations in between steps above which re-order the original data, or the data exported over to R, the two data sets could have mis-matched rows by the end. If this is a concern, please retain or create a unique identifier per row. For example, something like

data people_with_id
    set people;
    rownum = _N_;
run;

When subsetting the data to drop variables irrelevant to the matching, be sure to keep rownum.

After you’ve brought the data with the match information back into SAS, you can sort both data sets and merge with something like

proc sort data=people_with_id out=people_with_id2;
  by rownum;
run;

proc sort data=matchedpeople out=matchedpeople2;
  by rownum;
run;

data matchedmerged ;
  merge people_with_id2 matchedpeople2;
  by rownum;
run;

Using Optmatch with Stata

We’ll use the same example, with some simple demographics. We will treat gender as the treatment indicator, and wish to match on a combination of a propensity score for gender (using both age and height) and age.

input gender age height
0 25 62
0 41 68
0 38 63
0 22 62
1 33 70
1 35 71
1 47 68
1 23 64
end

First, lets fit the logistic regression model.

logit gender age height
predict ppty, xb

At the end we’ll be merging two files together to avoid any ordering issues, and as noted above, to do so we’ll create a unique identifier.

gen rownum = _n

We’ll save only the relevant variables (treatment indicator, anything to be matched on, and the ID variable to merge on) to avoid saving and loading a very large file.

preserve
keep gender age ppty rownum
save "/path/to/save/toR.dta"
restore

Turning to R, this can be read in using the haven package

> library(haven)
> statadata <- read_dta("/path/to/save/toR.dta")

Now, perform matching as desired, saving the final match to statadata. For example,

> library(optmatch)
> f <- fullmatch(gender ~ age + ppty, data = sttadata)
> statadata$match <- f

We’ll use haven again to write the data back to Stata. We do not recommend using .csv files to transfer the data back to Stata, though the write.csv file would be similar to that for SAS.

> write_dta(statadata, "/path/to/save/rout.stata.dta")

Back in Stata, you can merge this into the existing data set by the following commands:

sort rownum
merge 1:1 rownum using "/path/to/save/rout.stata.dta"

The force option may be necessary to overcome type differences. Additional tweaks may be necessary here if you have special variable types.