topic Weighted random stratified sampling in Statistical Procedures

Weighted random stratified sampling

jgtaylor — Tue, 15 Sep 2015 13:37:27 GMT

Hi,

Apologies if this is a silly question, I am a relatively new SAS user currently running 9.3.

I have a dataset (CITIES) of around 20,000 cities, each with population size classification, climate type, and national GDP classification.Some cities in the dataset also have data on transport, health, etc, which I have summarised into a single column called DataCoverage which counts the columns with known data for each city.

I'd like to do further analysis on a sub-sample of cities, and I would like to randomly select them in a manner which reflects the existing proportions of the data. I have done:

proc surveyselect data = CITIES out = samp1 method = srs sampsize=200 seed = 9876;
strata CLIMATE POPULATION_CLASS GDP_CLASS / alloc=proportional;

run;

What I would really like to do is select a subsample, which represents the proportions of the original dataset, but gives more weight to those with a larger DataCoverage (i.e. more known data, so I don't have to go find the data somewhere myself). Is such a thing possible?

Thanks,

Jon

Re: Weighted random stratified sampling

ballardw — Tue, 15 Sep 2015 15:01:59 GMT

If you can provide a numeric variable that represents data coverage, with larger meaning more coverage, you might be able to get this with a PPS selection using that variable for the SIZE.

Depending on how you are defining "reflects the existing proportions" you may need to look at setting sample sizes per strata.

Re: Weighted random stratified sampling

jgtaylor — Wed, 16 Sep 2015 12:31:59 GMT

Thanks, I think this gets me close.

I'm defining "reflects the existing proportions" as, for example:

If CITIES with a Population_Class of <50K with GDP_Class of "LowerMiddle GDP" in CLIMATE "Temperate Humid" comprise 5% of all cities in the world, then I want them to be 5% of my sampled dataset. The percent of each strata in the sample should reflect that in the original dataset.

I have changed to:

proc surveyselect data = CITIES out = samp1 method = pps sampsize=200 seed = 9876;
strata CLIMATE POPULATION_CLASS GDP_CLASS / alloc=proportional;
size DataCoverage;
run;

I've got a few problems, namely:

1) it's not giving me any cities with DataCoverage=0. It's ok to have some in order to maintain proportions, I just want to minimise them if possible

2) Since DataCoverage isn't great, I am not getting a sample size of 200 (97, actually).

Thanks again for your help.

Jon

Re: Weighted random stratified sampling

ballardw — Wed, 16 Sep 2015 15:30:16 GMT

The SIZE has to be non-zero in the basic use of surveyselect. Since we are using something that really isn't a population size counter then I would suggest add 1 to your datasource rate for all variables to get a 1 or greater and then at the end subtract the one out to get back to the original rank.

Also every record should have a datasource value or they would be excluded.

Re: Weighted random stratified sampling

jgtaylor — Thu, 17 Sep 2015 07:36:46 GMT

Very clever! That seems to have done what I wanted it to. Thanks for your help!

Jon