Piecewise Linear Mixed Models With a Random Change Point

Interfaces brms
2

I haven’t read the paper in detail, but you may try the following model. I built it rather ad-hoc so not sure if that’s the best solution we have in brms. Please also note that the priors are just to get the model running with this non-sensible simulated data.

library(brms)

bform <- bf(
  y ~ b0 + b1 * (age - omega) * step(omega - age) + 
    b2 * (age - omega) * step(age - omega),
  b0 + b1 + b2 + alpha ~ 1 + (1|person),
  # to keep omega within the age range of 0 to 10
  nlf(omega ~ inv_logit(alpha) * 10),
  nl = TRUE
)

df <- data.frame(
  y = rnorm(330),
  age = rep(0:10, 30),
  person = rep(1:30, each = 11)
)

bprior <- prior(normal(0, 3), nlpar = "b0") +
  prior(normal(0, 3), nlpar = "b1") +
  prior(normal(0, 3), nlpar = "b2") +
  prior(normal(0, 3), nlpar = "alpha")

make_stancode(bform, data = df, prior = bprior)

fit <- brm(bform, data = df, prior = bprior, chains = 1)
summary(fit)

# you need the github version of brms for this to run
marginal_effects(fit)
3 Likes