janickweberpals / svycoxphmi

Comparison of coxph and svycoxph to estimate marginal hazard ratios after propensity score matching in multiply imputed datasets

README

svycoxphMI

svycoxph in ps-matched datasets after multiple imputation

This is a reproducible example on how to use coxph and svycoxph in combination with multiple imputation and propensity score matching using a mimids object from the MatchThem package.

Load packages:

suppressPackageStartupMessages(library("dplyr"))
suppressPackageStartupMessages(library("survival"))
suppressPackageStartupMessages(library("mice"))
suppressPackageStartupMessages(library("MatchThem"))
suppressPackageStartupMessages(library("survey"))

Create example dataset:

# load example dataset with missing observations
set.seed(42)
data_miss <- survival::lung |> 
  # randomly assign an exposure
  mutate(exposure = rbinom(prob = 0.3, size = 1, n = nrow(survival::lung))) |> 
  # create an id variable
  mutate(id = paste0("ID", seq(1, nrow(survival::lung), 1)))

Multiple imputation using mice:

# impute data
data_imp <- mice(
  data = data_miss |> select(-id),
  m = 10,
  seed = 42,
  print = FALSE
  )

Apply propensity score matching with replacement within in each imputed dataset:

# apply propensity score matching on mids object
data_matched <- matchthem(
  exposure ~ age + sex,
  datasets = data_imp,
  approach = 'within',
  method = 'nearest',
  caliper = 0.2,
  ratio = 1,
  replace = F
  )

Matching Observations  | dataset: #1 #2 #3 #4 #5 #6 #7 #8 #9 #10

We now want to compare treatment effect estimates for exposure when computed (a) using coxph (survival package) and (b) svycoxph (survey package). More information on estimating treatment effects after matching/weighting is provided in https://kosukeimai.github.io/MatchIt/articles/estimating-effects.html#survival-outcomes

(a) coxph

# coxph result
coxph_results <- with(
  data = data_matched,
  expr = coxph(formula = Surv(time, status) ~ exposure, 
               weights = weights, 
               cluster = subclass,
               robust = TRUE
               )
  ) |> 
  pool() |> 
  tidy(exponentiate = TRUE) |> 
  select(term, estimate, std.error)

coxph_results

      term estimate std.error
1 exposure 1.069502 0.1878997

(b) svycoxph

# svycoxph result
svycoxph_results <- with(
  data = data_matched,
  expr = svycoxph(formula = Surv(time, status) ~ exposure)
  ) |> 
  pool() |> 
  tidy(exponentiate = TRUE, conf.int = TRUE) |> 
  select(term, estimate, std.error)

1 - level Cluster Sampling design (with replacement)
With (76) clusters.
survey::svydesign(ids = ~subclass, weights = ~weights, data = m.data.i)
1 - level Cluster Sampling design (with replacement)
With (76) clusters.
survey::svydesign(ids = ~subclass, weights = ~weights, data = m.data.i)
1 - level Cluster Sampling design (with replacement)
With (76) clusters.
survey::svydesign(ids = ~subclass, weights = ~weights, data = m.data.i)
1 - level Cluster Sampling design (with replacement)
With (76) clusters.
survey::svydesign(ids = ~subclass, weights = ~weights, data = m.data.i)
1 - level Cluster Sampling design (with replacement)
With (76) clusters.
survey::svydesign(ids = ~subclass, weights = ~weights, data = m.data.i)
1 - level Cluster Sampling design (with replacement)
With (76) clusters.
survey::svydesign(ids = ~subclass, weights = ~weights, data = m.data.i)
1 - level Cluster Sampling design (with replacement)
With (76) clusters.
survey::svydesign(ids = ~subclass, weights = ~weights, data = m.data.i)
1 - level Cluster Sampling design (with replacement)
With (76) clusters.
survey::svydesign(ids = ~subclass, weights = ~weights, data = m.data.i)
1 - level Cluster Sampling design (with replacement)
With (76) clusters.
survey::svydesign(ids = ~subclass, weights = ~weights, data = m.data.i)
1 - level Cluster Sampling design (with replacement)
With (76) clusters.
survey::svydesign(ids = ~subclass, weights = ~weights, data = m.data.i)
1 - level Cluster Sampling design (with replacement)
With (76) clusters.
survey::svydesign(ids = ~subclass, weights = ~weights, data = m.data.i)
1 - level Cluster Sampling design (with replacement)
With (76) clusters.
survey::svydesign(ids = ~subclass, weights = ~weights, data = m.data.i)

svycoxph_results

      term estimate std.error
1 exposure 1.069502 0.1891482