Inverse probability weighting combined with survey design weights

DuGoff, Schuler and Stuart recommend multiplying the propensity-score weight by the survey design weight and estimating the treatment effect as the difference of the two weighted (Hajek) group means. Kish's effective sample size is reported because the combined weights are usually far more variable than either component alone.

Usage

Ipwsrv(y, T, weights, propensity)

Arguments

y: Outcome vector.
T: Treatment indicator, 0 or 1.
weights: Survey design weights; NULL means all ones.
propensity: Propensity scores, strictly inside (0, 1).

Value

List with estimate, mu1, mu0, se, var1, var0, sum_w, ess, n1, n, method.

Details

Formula: w = d * (T/pi + (1 - T)/(1 - pi)).

References

DuGoff, Schuler and Stuart (2014), Generalizing observational study results, Health Services Research 49(1):284-303. doi:10.1111/1475-6773.12090

Examples

set.seed(1)
Ipwsrv(y = rnorm(20), T = rbinom(20, 1, 0.5), weights = runif(20, 1, 2),
       propensity = runif(20, 0.3, 0.7))
#> $estimate
#> [1] 0.1157841
#> 
#> $mu1
#> [1] 0.2557062
#> 
#> $mu0
#> [1] 0.139922
#> 
#> $se
#> [1] 0.3780691
#> 
#> $var1
#> [1] 0.05089439
#> 
#> $var0
#> [1] 0.09204183
#> 
#> $sum_w
#> [1] 64.11693
#> 
#> $ess
#> [1] 18.49712
#> 
#> $n1
#> [1] 11
#> 
#> $n
#> [1] 20
#> 
#> $method
#> [1] "w = d (T/pi + (1-T)/(1-pi)); Hajek difference, DuGoff, Schuler & Stuart (2014)"
#> 
#> attr(,"class")
#> [1] "morie_rich_result" "list"