topic Re: Computing geometric mean by different cohorts in Statistical Procedures

Computing geometric mean by different cohorts

Sean_OConnor — Mon, 14 Sep 2020 11:18:19 GMT

Folks,

I have been computing mean and medians using the following statement.

proc summary data = Quantity_of_Properties_1 nway completetypes;
	;
	format class $class. year_filed year_filed. county_y $county_y.  child_filed $child_filed.;
	class class year_filed county_y  child_filed/mlf;
	var realprice1;
	output out = Table_0_fil (drop=_TYPE_ /*_freq_*/) 
	median= med_price mean=mean_price;
run;

However, I would like to compute the geometric mean, but I understand I can only do it through proc surveymeans. Is there anyway to get the same kind of output as above with proc surveymeans?

I've tried to the following but it does not work so any help would be great.

PROC SURVEYMEANS data=prices geomean;
	format class $class. year_filed year_filed. county_y $county_y.  child_filed $child_filed.;
	class class year_filed county_y  child_filed;
	var realprice1;
	ods output statistics=i;
run;

Re: Computing geometric mean by different cohorts

PeterClemmensen — Mon, 14 Sep 2020 11:38:15 GMT

Have you read this?

Compute the geometric mean, geometric standard deviation, and geometric CV in SAS

Re: Computing geometric mean by different cohorts

ChrisHemedinger — Mon, 14 Sep 2020 11:41:14 GMT

PROC UNIVARIATE also can also be used -- see this article. And the TTEST procedure.

Re: Computing geometric mean by different cohorts

Sean_OConnor — Mon, 14 Sep 2020 11:54:46 GMT

Thanks for this. It doesn't look like proc univariate can handle multi level formats though or am I incorrect on this?

Re: Computing geometric mean by different cohorts

SteveDenham — Mon, 14 Sep 2020 12:24:49 GMT

Try a two-step process. First, log transform realprice1 in a DATA step (for illustration purposes, let's call it lnrealprice1). Then feed it through PROC SURVEYMEANS just as you have the untransformed value. Finally, using the output dataset, exponentiate the mean, SD, confidence bounds, etc. Make sure that you include the sampling weights in the PROC SURVEYMEANS call, using either the WEIGHTS or REPWEIGHTS statement.

Thanks go to @PeterClemmensen and @ChrisHemedinger for the referral to Rick Wicklin's ( @Rick_SAS ) blog.

SteveDenham

Re: Computing geometric mean by different cohorts

Sean_OConnor — Mon, 14 Sep 2020 12:58:14 GMT

Hi Steve,

It appears that surveymeans can't handle multilevel formats either. For instance, proc summary /mlf option works, but this doesn't appear to be the case with the other functions.

Would I be correct in this assumption?

Re: Computing geometric mean by different cohorts

mkeintz — Mon, 14 Sep 2020 14:26:07 GMT

You could also make a temporary dataset with a new variable equal to the log10(realprice1) or log(realprice1), followed by proc summary with the class statements you want, and adding the new variable to the var statement.

Then take the output dataset TABLE_0_FIL produced by the proc summary, and get the antilog (either 10**x or exp(x)) of the mean of the new variable. You will have the desired subgroup geometric means alongisde the other stats you requested.

Re: Computing geometric mean by different cohorts

ballardw — Mon, 14 Sep 2020 17:51:06 GMT

@Sean_OConnor wrote:

Hi Steve,

It appears that surveymeans can't handle multilevel formats either. For instance, proc summary /mlf option works, but this doesn't appear to be the case with the other functions.

Would I be correct in this assumption?

You really should have defined in your initial question that the groups were defined by a multilabel format, if that is what you mean by "multilevel". Most formats are "multilevel" unless you are mapping all values to a single one. The documentation for Multilabel has the procedures that support the feature.

To use SURVEYMEANS you would make additional DOMAIN variables for each level and use a DOMAIN statement such as:

domain Level1 Level1*Level2; to get the combined mean of Level1 and Level2 plus an overall for each Level1 value.