R中的分层svyglm

Question

I'm looking for advice on how to conduct a weighted logistic regression analysis, stratified by gender, in R.

For my main, unstratified analysis, I generated inverse probability weights (invp) and ran a weighted logistic regression as follows:

complete_cases_weighted <- svydesign(id=~1, weights=~invp, data=complete_cases)
svyglm(outcome ~ exposure, design=complete_cases_weighted, family="binomial")

I now want to essentially re-run this analysis twice, once for females only, and once for males only.

What's the best way of doing this with correct weighting? Do I re-do the entire process of generating weights and running a weighted logistic regression having subset my data into males and females? Or is there some fancy footwork I can do with svydesign / svyglm to do this?

Answer 1

Since strata are independent data layers, your approach seems to be correct. You build svydesign separately for men and separately for women. Then perform svyglm separately for them. If I understood correctly, then see the example. The weight coefficients are calculated separately for each stratum.

library(tidyverse)
library(survey)
library(broom)

df <- mtcars %>%
  mutate(weight = ifelse(vs == 1, 3, 4))

df %>%
  group_nest(vs) %>%
  mutate(design = map(data, ~ svydesign(
    ids = ~ 1,
    weights = .x$weight,
    data = .x
  )),
  model = map(design, ~ svyglm(disp ~ hp, design = .x) %>% glance)) %>%
  unnest(model)
#> # A tibble: 2 x 9
#>      vs      data design null.deviance df.null   AIC    BIC deviance df.residual
#>   <dbl> <list<ti> <list>         <dbl>   <int> <dbl>  <dbl>    <dbl>       <dbl>
#> 1     0 [18 x 11] <srvy~       193780.      17  3.01 1.31e5  131261.          16
#> 2     1 [14 x 11] <srvy~        42079.      13  1.51 2.68e4   26791.          12

^{Created on 2021-07-29 by the reprex package (v2.0.0)}