PSMatch Optimal Match Performance

[mftuchman]

I have yet to get an optimal match to run without a failure.   If I have control with M (#cols=VC) and Treatment with N(# cols=VT) (observations, how can I estimate how much resources the optimal match will take?   Every optimal match that I've tried  failed due to non-existence of a good match.  Most variables are binary variables, and distance is defined by LPS (log propensity score).

Obviously there are a LOT of factors that would affect the ability to create an optimal match.  Without knowing much more about my data, are there any general guidelines we should be aware of?  Are there any bounds above which an optimal match would be extremely unlikely?

Here's my error:

ERROR: A feasible optimal fixed ratio matching that has the specified parameters does not exist.

NOTE: The data set WORK.OUTGS has 0 observations and 87 variables.

NOTE: PROCEDURE PSMATCH used (Total process time):

[mhtoto]

I am getting the same error? Any suggestions for troubleshooting tips? Many thanks.

proc psmatch data=Preop
region=allobs;
class first20 diseaseseverity site surgeon left_1 female;
psmodel first20= age ;

match method=optimal(k=1) stat=lps caliper=0.9;

assess lps var=(age ) / weight=none plots=(boxplot barchart);
output out(obs=all)=Outgs lps=_Lps matchid=_MatchID;

[MichaelL_SAS]

You might try specifying "caliper=." in the match statement to see if the caliper requirement is making the optimal matching infeasible. By default a value of caliper=0.25 is used, so specifying the missing value is necessary to remove the requirement. You can also look at the plots produced by the ASSESS statement to compare the range and distribution of LPS values between the treatment and control condition to see if the caliper might be the cause. 

[KK_Mangla]

Hi,

I am facing the similar issue with both OPTIMAL and VARRATIO matching. Is there any method to change the caliper value for optimal with fixed/variable ratio matching. The issue with default value of 0.25 is - it does not give balanced cohorts. Can you suggest me something on this?