Survey data: Probability weights and adjusting for clustering in a logistic regression (non-random sample)

[Shivani Gaiha]

Hi everyone,

I hope to get some timely help on a dataset I am working on. I am struggling a bit with the code and everyone around me seems to be unfamiliar with Stata codes. I would really appreciate it!

Description of the dataset:

• 4351 Observations
• We oversampled for e-cigarette users and balanced other variables based on age, sex, LGBTQ, race/ethcnicity. Then we created one weight variable to be able to match the population target proportions (outwgt).
• Not a random sample, but an online national convenience sample.
• We also collected data on state (53 states) and region of participants (5 regions). The primary sampling unit is id, or individual participants. There were no multi-stage clusters or data collection involved by state or region.
• Outcome variables are binary and would prefer Odds Ratio as the output.

What I want to do? Assess associations between the outcome variable and my dependent variables.

Approach 1.

Steps I took:

1. I am setting the survey weight and entering cluster information through the code below:

svyset id [pweight=outwgt], strata(state) vce(linearized) singleunit(missing) || id, strata(region)

I think this code is inaccurate as it may be telling Stata that I first sampled by region, state and then id? I simply want to adjust for clustering effects, but I think the code I have entered tells Stata that multi-stage sampling is happening.

2. Next I conducted a logistic regression

svy linearized : logistic covidsymptoms i.totalvape i.agecat i.gender i.lgbtq i.newre i.complysip

Issues I need help with:

I get a an error message that:

Note: Stage 1 is sampled with replacement; further stages will be ignored for variance estimation.

And so I am not sure if Stata is ignoring region data altogether.


Need feedback on Approach 1 and what it is doing.


Approach 2.

I tried to change my approach to the following steps:

1. I removed the state and region information from the svy command:

 svyset id [pweight=outwgt]

      pweight: outwgt
          VCE: linearized
  Single unit: missing
     Strata 1: <one>
         SU 1: id
        FPC 1: <zero>

2. And, then I ran a logistic regression:

. logistic symptoms totalvape i.agecat i.gender i.lgbtq i.newre i.complysip, vce(cluster state)

I think this code looks after both the weight and the state-level clustering; BUT I am not sure how to add region to this regression?

Please help.
Shivani