add_pairwise_contrasts for quantile regressions: library(quantreg) & library(lqmm)

[yuryzablotski]

Support for quantile regressions is only possible for tau = 0.5 (median regressio), however, the whole point of quantile regression is to be able to model quantiles:

works

d <- mtcars %>% mutate(cyl = factor(cyl)) m <- lm(mpg ~ cyl, d) library(gtsummary) tbl_regression(m, add_pairwise_contrasts = T)

works

library(quantreg) qm <- rq(mpg ~ cyl, d, tau = 0.5) tbl_regression(qm, add_pairwise_contrasts = T)

doesnt work

qm10 <- rq(mpg ~ cyl, d, tau = 0.1) tbl_regression(qm10, add_pairwise_contrasts = T)

by the way, emmeans with different quantiles work:

emmeans(qm10, pairwise ~ cyl, tau = 0.10, type = "response")

how can this be integrated into tbl_summary with add_pariwise_contrasts?

Thanks forword!

[yuryzablotski]

After the emmeans author wrote me the possibility to specify tau argument in emmeans, I figured out the solution

tbl_regression(qm10, add_pairwise_contrasts = T, emmeans_args = list(tau = .1))

He also said, that bei the next update of emmeans, the neccessaty to specify tau will be releaved.

sorry for circumstances! and thanks for your work! I am a bit fan and everyday user.

[ddsjoberg]

glad you found a resolution! happy programming!

[yuryzablotski]

Thanks Daniel,

one of the best gtsummary features for me is "add_pairwise_contrasts = T" via the integration with emmeans package, and amazingly it even by default uses the "weights = "prop". Special thanks for that! Hier is a code for non-0.5 quantile in a quantile regression which works perfectly:

library(ISLR) qm90 <- rq(wage ~ jobclass * health_ins, Wage, tau = 0.9)

emmeans(qm90, pairwise ~ jobclass|health_ins, type = "response", tau = 0.9)

emmip(qm90, jobclass ~ health_ins, tau = 0.9, CIs = T)

library(ISLR) qm90 <- rq(wage ~ jobclass + education + health_ins, Wage, tau = 0.9)

wrong, because only the first level of categorical

confounders are taken

emmip(qm90, ~ education, tau = 0.9, CIs = T)+ theme_bw()

correct: because averages over categories of

categorical confounders

emmip(qm90, ~ education, tau = 0.9, CIs = T, weights = "prop")+ theme_bw()

emmeans(qm90, pairwise ~ education, weights = "prop")

tbl_regression( qm90, add_pairwise_contrasts = T, emmeans_args = list(tau = .9, weights = "prop"))

The "weights = "prop" is actually not necessary, which is cool!

Happy Christmal Holidays! Yury

[ddsjoberg]

Hi @yuryzablotski ! Thanks for the details! These calculations are expertly handled in the broom.helpers package. Can you copy the details in an issue on that repo? https://github.com/larmarange/broom.helpers

[yuryzablotski]

done!