-->
Adventures & experiences in contemporary technology
Recently, I needed to calculate geometric means after a break of a number of years and ended racking a weary brain until it was brought back to me. In order that I am not in the same situation again, I am recording it here and sharing it always is good too.
The first step is to take the natural log (to base e or the approximated irrational value of the mathematical constant, 2.718281828) of the actual values in the data. Since you cannot have a log of zero, the solution is either to exclude those values or substitute a small value that will not affect the overall result as is done in the data step below. In SAS, the log function uses the number e as its base and you need to use the log10 equivalent when base 10 logs are needed.
data temp;
set temp;
if result=0 then _result=0.000001;
else _result=result;
ln_result=log(_result);
run;
Next, the mean of the log values is determined and you can use any method of doing that so long as it gives the correct values. PROC MEANS is used here but PROC SUMMARY (identical to MEANS except it defaults to data set creation while that generates output by default; because of that, we need to use the NOPRINT option here), PROC UNIVARIATE or even the MEAN function in PROC SQL.
proc means data=temp noprint;
output out=mean mean=mean;
var ln_result;
run;
With the mean of the log values obtained, we need to take the exponential of the obtained value(s) using the EXP function. This returns values of the same magnitude as in the original data using the formula emean.
data gmean;
set mean;
gmean=exp(mean);
run;
While you could generate data sets containing means and confidence intervals using PROC SUMMARY or PROC MEANS, curiosity and the need to verify a program using a different technique were what drove me to consider using PROC UNIVARIATE for the task. For the record, the PROC SUMMARY code is below and the only difference between it and MEANS is that it doesn’t produce output by default, something that’s not needed in this case anyway. Quite why there are two SAS procedures doing exactly the same thing is beyond me though I do wonder if the NOPRINT options was a later addition than these two procedures. The LCLM and UCLM keywords are what triggers the calculation of confidence limits and the ALPHA option controls the confidence interval used; 0.05 specifies a 95% interval, 0.1 a 90% one and so on.
proc summary data=sashelp.class mean lclm uclm alpha=0.05;
var age;
output out=sasuser.lims mean=mean lclm=lclm uclm=uclm;
run;
Given that I have had PROC UNIVARIATE producing statistics that MEANS/SUMMARY didn’t in previous versions of SAS (I believe that is was standard deviation that was absent from MEANS/SUMMARY), I might have expected the calculation and export of confidence limits to a data set to be straightforward. Sadly, it’s not a case of simply adding LCLM and UCLM keywords in the OUTPUT statement for the procedure and ODS OUTPUT is needed to create the data set instead. An ODS SELECT statement is needed to pick out the BasicIntervals output object (UNIVARIATE creates quite a few, it seems) that is created through specification of the CIBASIC and ALPHA (performs the same role as it does for PROC MEANS/SUMMARY) options on the PROC UNIVARIATE statement. The reason for the ODS LISTING and ODS RTF statements below is to stop output being sent to the output window in a standard SAS session. For some reason, it appears that you need the sending of output to one of the LISTING, HTML or RTF destinations or there will be no data in the data set; I met up with the same behaviour when using ODS PS, an ODS PRINTER destination. The data set will contain statistics for mean, standard deviation and variance so that’s why there is a WHERE clause on the ODS OUTPUT statement.
ods listing close;
ods rtf body="c:\temp\uni_eg.doc";
ods select BasicIntervals;
ods output BasicIntervals=sasuser.stats(where=(lowcase(parameter)="mean") );
proc univariate cibasic alpha=0.05 data=sashelp.class;
var age;
run;
ods output close;
ods rtf close;
ods listing;