KW-ML for R

R package for "Boosted Kernel Weighting - Using Statistical Learning to Improve Inference from Nonprobability Samples"

Implements functions to compute pseudo-weights for nonprobability samples, including inverse propensity score weighting (ipsw.lg()), kernel weights based on logistic regression (kw.lg()), and kernel weights based on machine learning methods (kw.mob(), kw.crf(), kw.gbm()).

Installation

if (!require("devtools")) install.packages("devtools")
devtools::install_github("chkern/KWML")

Example

Calculate IPSW (ipsw.lg()) and KW-LG (kw.lg()) pseudo-weights with example data. simu_dat is a stacked data frame with a simulated probability and non-probability sample. ipsw.lg() and kw.lg() need a data frame, the name of the weight variable (wt, weights of 1 for non-prob, survey weights for prob sample), the name of the sample membership indicator (trt, 1 for non-prob, 0 for prob sample) and a formula for the propensity model as input.

library(KWML)

ipsw_w <- ipsw.lg(simu_dat, "wt", "trt", 
                  "trt_f ~ x1+x2+x3+x4+x5+x6+x7")

kwlg_w <- kw.lg(simu_dat, "wt", "trt", 
                "trt_f ~ x1+x2+x3+x4+x5+x6+x7")$pswt

For the KW-ML functions (kw.mob(), kw.crf(), kw.gbm()), tuning parameter grids and covariate names for covariate balance calculation need to be specified additionally.

kwmob <- kw.mob(simu_dat, "wt", "trt", 
                "trt_f ~ x1+x2+x3+x4+x5+x6+x7 | x1+x2+x3+x4+x5+x6+x7",
                tune_maxdepth = c(2, 3), 
                covars = c("x1","x2","x3","x4","x5","x6","x7"))

kwcrf <- kw.crf(simu_dat, "wt", "trt", 
                "trt_f ~ x1+x2+x3+x4+x5+x6+x7",
                tune_mincriterion = c(0.95, 0.9), 
                covars = c("x1","x2","x3","x4","x5","x6","x7"))

kwgbm <- kw.gbm(simu_dat, "wt", "trt", 
                "trt ~ x1+x2+x3+x4+x5+x6+x7",
                tune_idepth = 1:3,
                tune_ntree = c(250, 500),
                covars = c("x1","x2","x3","x4","x5","x6","x7"))       

Select KW-ML pseudo-weights with best covariate balance.

kwmob_w <- kwmob$pswt[, kwmob$best]
kwcrf_w <- kwcrf$pswt[, kwcrf$best]
kwgbm_w <- kwgbm$pswt[, kwgbm$best]

Compare weighted mean of y in prob sample and pseudo-weighted means in non-prob sample.

sum((simu_dat$y[simu_dat$trt == 0]*simu_dat$wt)/sum(simu_dat$wt))
sum((simu_dat$y[simu_dat$trt == 1]*ipsw_w)/sum(ipsw_w))
sum((simu_dat$y[simu_dat$trt == 1]*kwlg_w)/sum(kwlg_w))

sum((simu_dat$y[simu_dat$trt == 1]*kwmob_w)/sum(kwmob_w))
sum((simu_dat$y[simu_dat$trt == 1]*kwcrf_w)/sum(kwcrf_w))
sum((simu_dat$y[simu_dat$trt == 1]*kwgbm_w)/sum(kwgbm_w))

Citation

@Article{KernLiWang2020,
  title    = {Boosted Kernel Weighting - Using Statistical Learning to Improve Inference from Nonprobability Samples},
  author   = {Kern, C. AND Li, Y. AND Wang, L.},
  journal  = {Journal of Survey Statistics and Methodology},
  year     = {2020},
  doi      = {10.1093/jssam/smaa028},
  url      = {https://doi.org/10.1093/jssam/smaa028}
}