Bayesplot :: CHEAT SHEET chains_trace

library("bayesplot")
ppd %>%
library("rstanarm") (Intercept) σ
options(mc.cores = parallel::detectCores()) Chain ppc_stat_grouped(y = mtcars$mpg, group = mtcars$am,
library("ggplot2") 5 stat = "median", binwidth=0.5)
50 1
library("dplyr") 4
2 0 1
0 3
Model Parameters 3
To showcase bayesplot, we’ll fit linear regression using rstanarm::stan glm and use this −50 2
4
model throughout. 0 200 400 600 800 1000 0 200 400 600 800 1000 T = median
T (y rep)
model <- stan_glm(mpg ˜ ., data=mtcars, chains=4) The pairs plot is helpful in determining if you have any highly correlated parameters.
posterior <- as.matrix(model)
posterior_chains %>% T (y )
Chances are good you’re most interested in the posterior distributions for select parameters. mcmc_pairs(pars = c("(Intercept)", "wt", "sigma"))

(Intercept) 13 15 17 19 2120.0 22.5 25.0 27.5 30.0

plot_title <- ggtitle("Posterior distributions", 50 50
"medians and 80% intervals")
0 0
mcmc_areas(posterior, Diagnostics
pars = c("drat", "am", "wt"), −50 −50 Bayesplot makes it easy to check diagnostics specific to the NUTS sampling method that
prob = 0.8) + plot_title −50 0 50 −10 −5 0 2 3 4 5 rstanarm uses by default.
4 wt
Posterior distributions 0 0
mcmc_scatter(posterior, pars = c("wt", "sigma"),
np = nuts_params(model$stanfit))
medians and 80% intervals −4
−5
−8
−12 −10 5
−50 0 50 −12 −8 −4 0 4 2 3 4 5
drat 5 5 sigma 4

sigma
4 4
3 3
3
am 2 2
−50 0 50 −12 −8 −4 0 4 2 3 4 5
2

wt Posterior Predictive Checks −12 −8 −4 0 4

Check how well the model covers your data with draws from the posterior predictive density.
−10 −5 0 5 10 wt
Diagnosing convergence with traceplots is simple.
ppd <- posterior_predict(model, draws=500) np <- nuts_params(model$stanfit)
mcmc_trace(posterior, pars=c("sigma", "wt")) ppd %>% mcmc_nuts_energy(np, binwidth=1) +
ppc_intervals(y = mtcars$mpg, yrep = ., x = mtcars$wt, prob = 0.5) + ggtitle("NUTS Energy Diagnostic")
labs(x = "Weight (1000 lbs)", y = "MPG",
sigma wt title = "50% posterior predictive intervals of MPG by weight")
5 4 NUTS Energy Diagnostic
4
0 50% posterior predictive intervals of MPG by weight 1 2
−4
3 ●
●
−8 ●●
2 30 ●●
−12 ● ●● ●
●●
0 1000 2000 3000 4000 0 1000 2000 3000 4000 ● ● ● πE
●●
●
3 4
●●
MPG

● ●●
●
●
●
Using as.array, you can extract each of the four chain’s posterior draws, different from 20 ●●
● ●
●●
●
●
● ● ●
● y
π∆E
above. This allows you to see each chain’s traceplot for selected parameters. ●
●
●
● ●●
●
●
●
●● ●
●●
●
●
● ● ●
● y rep
posterior_chains <- as.array(model) ●●
●●
●●●
fargs <- list(ncol = 2, labeller = label_parsed) 10
pars <- c("(Intercept)", "sigma") −20 −10 0 10 20 −20 −10 0 10 20
chains_trace <- mcmc_trace(posterior_chains, pars = pars, E−E
n_warmup = 300, facet_args = fargs) 2 3 4 5
Weight (1000 lbs)
CC BY-SA 4.0 ◦ Edward A. Roualdes ◦ roualdes.us ◦ version 0.0.2 ◦ updated: 2020-05-19