Introduction to bqmm

bqmm fits Bayesian mixed-effects quantile regression using the asymmetric Laplace working likelihood and Stan.

A first model

library(bqmm)

fit <- bqmm(distance ~ age + (1 | Subject),
            data = nlme::Orthodont,
            tau  = 0.5)

summary(fit)

Several quantiles at once

Passing a vector of quantiles fits each one and returns a bqmm_multi:

fit3 <- bqmm(distance ~ age + (1 | Subject),
             data = nlme::Orthodont,
             tau  = c(0.1, 0.5, 0.9))

coef(fit3)      # tau-by-coefficient matrix
plot(fit3)      # coefficient-versus-tau paths

Valid inference

By default bqmm applies the Yang, Wang and He (2016) correction so that fixed-effect intervals are asymptotically valid despite the misspecified asymmetric Laplace likelihood:

vcov(fit, adjusted = TRUE)
vcov(fit, adjusted = FALSE)   # naive posterior covariance

See vignette("bqmm-inference") for details.