Repository navigation
Use ggdist::curve_interval for plotting curves in Bayesian models #455
Description
This gives an overall smoother line / interval (less "lumpy").
Read more, here >>
Linear regression plot
m <- rstanarm::stan_glm(mpg ~ hp, data = mtcars)
plot(modelbased::estimate_relation(m, length = 100))
Looks lumpy...
library(ggplot2)
emt <- emmeans::emmeans(m, ~ hp, at = list(hp = seq(min(mtcars$hp), max(mtcars$hp), length.out = 100)))
tidybayes::gather_emmeans_draws(emt) |>
ggdist::curve_interval(.value, .along = hp, .width = 0.95) |>
ggplot(aes(hp, .value)) +
geom_ribbon(aes(ymin = .lower, ymax = .upper), alpha = 0.4) +
geom_line()
Mmmmm... smoothhhhh
JN-plot smooth() plot
formula <- brms::brmsformula(
mpg ~ s(wt),
sigma ~ s(wt),
beta ~ s(wt),
family = brms::exgaussian()
)
m <- brms::brm(formula, data = mtcars, chains = 1, iter = 200)
plot(modelbased::estimate_slopes(m, trend = "wt", at = "wt", length = 100))
library(ggplot2)
emt <- emmeans::emtrends(m, ~ wt, "wt", at = list(wt = seq(min(mtcars$wt), max(mtcars$wt), length.out = 100)))
tidybayes::gather_emmeans_draws(emt) |>
ggdist::curve_interval(.value, .along = wt, .width = 0.95) |>
transform(.sig = .lower > 0 | .upper < 0) |>
transform(.part.group = {function(x) rep(seq_along(x$lengths), x$length)}(rle(.sig))) |>
ggplot(aes(wt, .value)) +
geom_hline(yintercept = 0) +
geom_ribbon(aes(ymin = .lower, ymax = .upper, group = .part.group, fill = .sig), alpha = 0.4) +
geom_line()
Activity
DominiqueMakowski commentedon Aug 15, 2021
Mmh interesting, I'll try to include that... I suppose it will go somewhere here:
https://github.com/easystats/modelbased/blob/eed1a1572fd71197aaa979dbcb857c6d0be5f459/R/visualisation_recipe.estimate_predicted.R#L292-L293
this one looks funny tho
DominiqueMakowski commentedon Aug 15, 2021
oh actually no, this should be something that must be done at the level of bayestestR, when the computation of the CI is done. Maybe we could add an option to describe_posterior like
smooth_cior something?DominiqueMakowski commentedon Aug 15, 2021
or in
ci_methodadd "curve" forcurve_intervalmethodmattansb commentedon Aug 15, 2021
Yeah, that sounds right (:
mattansb commentedon Aug 15, 2021
A very hurtful thing to say to the poor graph....
mattansb commentedon Aug 15, 2021
Actually, do we need this implemented in bayestestR? It would essentially be a wrapper around
ggdist::curve_interval, no?DominiqueMakowski commentedon Aug 15, 2021
Well it seems to be like it belongs there, because if I understand it's a different method of computing CIs, different from hdi, quantile etc. Perhaps mostly used for plotting, true, but still
Moreover, modelbased now simply runs insight:
https://github.com/easystats/modelbased/blob/3ae98f03d5311b883276c67c18dd2cf5b49836ee/R/estimate_predicted.R#L219-L226
which in turn calls bayestestR ^^
https://github.com/easystats/insight/blob/f4e5c6d312f96fd671bc1cdc91d274bc567ad41e/R/get_predicted_ci.R#L435-L437
so yeah for generalizability (for people that use insight::get_predicted directly) it'd be convenient to have that feature in bayestestR
bwiernik commentedon Aug 15, 2021
I think it's okay if the curve estimation and the tabling have slightly different estimation. It would be okay to restrict the ggdist calls to see
mattansb commentedon Aug 15, 2021
Another example with moderation:
Created on 2021-08-15 by the reprex package (v2.0.0)
mjskay commentedon Aug 18, 2021
Chiming in since @DominiqueMakowski pinged me on mjskay/ggdist#95. I'm not sure in what context you're planning to use curve_interval, but it's worth pointing out a few things you may want to ponder:
.interval = "bd-mbd"but this requires the {fda} package, which pulls in a bunch of dependencies. Worth pondering if you are thinking about defaults.mjskay commentedon Aug 19, 2021
One last thought from me: I was trying to figure out why your pointwise intervals were so noisy looking in the first place and realized you are using HDIs. HDIs can be very noisy, as you have discovered. If you use quantile intervals instead you'll find the pointwise bands are much smoother, besides having other benefits (like being invariant under transformation). FWIW if I were picking a default pointwise interval type I would lean in that direction.