causal-inference-in-R
causal-inference-in-R copied to clipboard
malcolmbarrett
Expanding on continuous exposures
Balance assessment: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6484705/ Binning: https://journals.lww.com/epidem/Fulltext/2014/03000/Constructing_Inverse_Probability_Weights_for.21.aspx
also looking for information on estimating different estimands for continuous exposures (cc @LucyMcGowan if you find anything)
Getting the weighted correlations for covariate balance assessment is a bit of a pain and might be benefited by a function in a package. Using the packages and models in the book:
wait_times_wts |>
# prep data for `cov.wt()`
drop_na(avg_spostmin, close, extra_magic_morning, weather_wdwhigh, wdw_ticket_season) |>
select(avg_spostmin, close, extra_magic_morning, weather_wdwhigh, wdw_ticket_season) |>
mutate(close = as.numeric(close), wdw_ticket_season = as.numeric(factor(wdw_ticket_season))) |>
# get weighted correlations
cov.wt(wt = wait_times_wts$swts, cor = TRUE) |>
pluck("cor") |>
# tidy results
as.data.frame(rownames = TRUE) |>
rownames_to_column("variable") |>
filter(variable != "avg_spostmin") |>
select(variable, correlation = avg_spostmin)
# variable correlation
# 1 close 0.57380383
# 2 extra_magic_morning 0.03553053
# 3 weather_wdwhigh 0.30051169
# 4 wdw_ticket_season -0.13616836
I thought we had this linked somewhere but couldn't find it in the issues. @malcolmbarrett this is a nice paper for generalized propensity scores with different targets: https://arxiv.org/pdf/1808.05339.pdf
Code to do weighted correlation by hand if helpful:
causal_collider_wts |>
summarise(
exposure_w = sum(swts * exposure) / sum(swts),
covariate_w = sum(swts * covariate) / sum(swts),
numerator = sum(swts * (exposure - exposure_w) * (covariate - covariate_w)),
denom = sqrt(sum(swts * (exposure - exposure_w)^2) * sum(swts * (covariate - covariate_w)^2)),
weighted_cor = numerator/denom
) |>
select(weighted_cor)