ASA Connect

Hi all, 
I'm new in sampling methods and I'd like your advice. I have about 210 local offices in 20 regions, with available data for each one of these offices about whether they are rural/urban and their size (large/ small). I'd like to select a sample of these offices according to these strata (urban/rural and large/small) but I don't know how to define the sample size and then how to carry out the stratified sampling Any suggestions or literature guidance?

Best regards, 

Lina

------------------------------
Lina Massou
Research Associate
University of Cambridge
------------------------------

Hi Lina,

If only calculating sample sizes was so simple! To estimate the sample size needed, you need to take into account what level of precision you need for your estimates. For example, are you looking to measure large differences (say 15 percentage points) or small/rare differences (say 2 percentage points)?

When it comes to stratified sampling, you need at least 2 responding PSUs per stratum to be able to estimate variance. If you are looking at 20 regions x 2 urbanicity x 2 size components, you're at 80 strata, so you need a minimum of 160 PSUs (2 per stratum) sampled. This is also assuming that each stratum has at least 2 PSUs (local offices) in it. You might want to collapse strata depending on the types of analyses you need to do and your budget for the number of local offices you can go to. 

As for how to perform stratified sampling, you can look at Chapter 3 of Kish's Survey Sampling. The simple version is that basically once you determine the strata, you can treat each stratum as it's own separate population and you do your sampling for each stratum independently of the other strata. You could do SRS, PPS, systematic random sampling, etc. for each stratum. All of these methods would make the probability of selection known for each PSU within each stratum to be used in weighting. There are also different ways of allocating how many PSUs you should take per stratum when you have determined the total number of PSUs you plan to sample, such as proportional allocation based on some measure of size/population, or optimal allocation based on assumptions of variance measures for each stratum.

I hope this helps!

------------------------------
Martha McRoy
International Research Methodologist
Pew Research Center
------------------------------

Hi
I agree with the approach suggested using systematic random sampling. A thought occurred that if you could not obtain all data on urban/rural or large/ small using an a priori population estimate you could estimate these data from "smoothing" the data or the nearest neighbor approach. Hope it goes well for you.
Thanks
Julie

Sent from my iPhone

------Original Message------

Hi Lina,

If only calculating sample sizes was so simple! To estimate the sample size needed, you need to take into account what level of precision you need for your estimates. For example, are you looking to measure large differences (say 15 percentage points) or small/rare differences (say 2 percentage points)?

When it comes to stratified sampling, you need at least 2 responding PSUs per stratum to be able to estimate variance. If you are looking at 20 regions x 2 urbanicity x 2 size components, you're at 80 strata, so you need a minimum of 160 PSUs (2 per stratum) sampled. This is also assuming that each stratum has at least 2 PSUs (local offices) in it. You might want to collapse strata depending on the types of analyses you need to do and your budget for the number of local offices you can go to. 

As for how to perform stratified sampling, you can look at Chapter 3 of Kish's Survey Sampling. The simple version is that basically once you determine the strata, you can treat each stratum as it's own separate population and you do your sampling for each stratum independently of the other strata. You could do SRS, PPS, systematic random sampling, etc. for each stratum. All of these methods would make the probability of selection known for each PSU within each stratum to be used in weighting. There are also different ways of allocating how many PSUs you should take per stratum when you have determined the total number of PSUs you plan to sample, such as proportional allocation based on some measure of size/population, or optimal allocation based on assumptions of variance measures for each stratum.

I hope this helps!

------------------------------
Martha McRoy
International Research Methodologist
Pew Research Center
------------------------------

Thank you both of you! Still working on it!

------------------------------
Lina Massou
Research Associate
University of Cambridge
------------------------------

Original Message:
Sent: 02-22-2018 09:44
From: Julie Tackett
Subject: Stratified sampling of regions

Hi
I agree with the approach suggested using systematic random sampling. A thought occurred that if you could not obtain all data on urban/rural or large/ small using an a priori population estimate you could estimate these data from "smoothing" the data or the nearest neighbor approach. Hope it goes well for you.
Thanks
Julie

Sent from my iPhone


Original Message------

Hi Lina,

If only calculating sample sizes was so simple! To estimate the sample size needed, you need to take into account what level of precision you need for your estimates. For example, are you looking to measure large differences (say 15 percentage points) or small/rare differences (say 2 percentage points)?

When it comes to stratified sampling, you need at least 2 responding PSUs per stratum to be able to estimate variance. If you are looking at 20 regions x 2 urbanicity x 2 size components, you're at 80 strata, so you need a minimum of 160 PSUs (2 per stratum) sampled. This is also assuming that each stratum has at least 2 PSUs (local offices) in it. You might want to collapse strata depending on the types of analyses you need to do and your budget for the number of local offices you can go to.

As for how to perform stratified sampling, you can look at Chapter 3 of Kish's Survey Sampling. The simple version is that basically once you determine the strata, you can treat each stratum as it's own separate population and you do your sampling for each stratum independently of the other strata. You could do SRS, PPS, systematic random sampling, etc. for each stratum. All of these methods would make the probability of selection known for each PSU within each stratum to be used in weighting. There are also different ways of allocating how many PSUs you should take per stratum when you have determined the total number of PSUs you plan to sample, such as proportional allocation based on some measure of size/population, or optimal allocation based on assumptions of variance measures for each stratum.

I hope this helps!

------------------------------
Martha McRoy
International Research Methodologist
Pew Research Center
------------------------------