Forest plots for model parameters about see HOT 19 CLOSED

Or plot multiple CIs one on top of each other?
See how something similar is now done in emmeans here.

from see.

The only edge case that I see is in the case of multiple CIs, but we can easily detect that by checking the length of ci which is stored in the parameters' attributes, and if length > 1 then pass the whole parameter table to the new reshape_ci function, and then simply add the CI as a facet... easy peasy

from see.

Yup, if tastefully implemented it could be nice

from see.

(tastefully = in a readable and aesthetically pleasing fashion ^^)

from see.

if tastefully implemented it could be nice

Like in sjPlot?

from see.

I don't know, does sjPlot handle multiple CIs 😬 ?

Also, by default, I very strongly favour geom_pointrange over error bars (i.e., I find simple lines ------x------, as opposed to brackets ][------x------][, sleeker) ☺️

from see.

@strengejacke I have misread your link, sjPlot also uses pointrange by default and sort-of boxplots to handle 2 different CIs for is that correct?

from see.

But yes, sjPlot is tasteful ^^

Nevertheless, it would be good to deal with multiple CIs in an automatic and systemic way

from see.

For Bayesian models, two HDIs are plotted by default:

I use geom_errorbar(), but you can use the width-argument to remove the "brackets" from the line-ends.

from see.

Oh, we need to have information about the link-function (as atribute?), in order to choose the correct scale (see again https://strengejacke.github.io/sjPlot/articles/plot_model_estimates.html).

from see.

For model parameters? Why the link? It's always identity for model_parameters isn't it

from see.

Ok, I probably was thinking a bit too far ahead... Usually, in social science, you would rather want to present (e.g.) Odds or Incidence Rate Ratios instead of log-estimates.

from see.

We can add this feature in the future for sure

from see.

The plot might also benefit from the clean_names attributes (currently used by parameters_table).

Note to myself: as matter of fact, format_parameters() might be improved to deal with cases of s(x), poly(x) and bs(x).

from see.

This looks also nice, a "completely" different style...

from see.

Quick sketch:

library(ggplot2)
library(parameters)

model <- lm(mpg ~ wt + cyl + gear + hp, data = mtcars)
mp <- model_parameters(model, standardize = "refit")

ggplot(mp, aes(x = Parameter, y = Coefficient)) +
  geom_errorbar(aes(ymin = CI_low, ymax = CI_high), alpha = .3, size = 4, width = 0, color = "#2196F3") +
  geom_segment(aes(y = CI_low, yend = CI_low + .4, x = Parameter, xend = Parameter), size = 4, color = "#2196F3", alpha = .5) +
  geom_segment(aes(y = CI_high, yend = CI_high - .4, x = Parameter, xend = Parameter), size = 4, color = "#2196F3", alpha = .5) +
  see::geom_point2(color = "#2196F3", size = 2.5) +
  coord_flip()+
  see::theme_lucid()

^{Created on 2019-07-26 by the reprex package (v0.3.0)}

from see.

Minimalist is the new maximalist:

library(ggplot2)
library(parameters)

model <- lm(mpg ~ wt + cyl + gear + hp, data = mtcars)
mp <- model_parameters(model, standardize = "refit")

# mp <- mp[mp$Parameter != "(Intercept)", ]
max_value <- max(abs(mp[["Coefficient"]]))

ggplot(mp, aes(x = Parameter, y = Coefficient, color = Coefficient)) +
  geom_hline(aes(yintercept = 0), linetype = "dashed") +
  geom_pointrange(aes(ymin = CI_low, ymax= CI_high), size = 1) +
  scale_color_gradientn(colours = c("red", "orange", "green"), limits=c(-max_value, max_value)) +
  coord_flip() +
  see::theme_modern(legend.position = "none")

^{Created on 2019-07-26 by the reprex package (v0.3.0)}

from see.

For Multiple CIs is kinda tough to display it on one plot

library(ggplot2)
library(parameters)
library(rstanarm)

model <- stan_glm(mpg ~ wt + cyl + gear + hp, data = mtcars, refresh = 0)
mp <- model_parameters(model, ci = c(0.5, 0.7, 0.9))
#> Possible multicollinearity between hp and cyl (r = 0.78). This might lead to inappropriate results. See 'Details' in '?rope'.

max_value <- max(abs(mp[["Median"]]))

ggplot(mp, aes(x = Parameter, y = Median, color = Median)) +
  geom_hline(aes(yintercept = 0), linetype = "dashed") +
  geom_pointrange(aes(ymin = CI_low_90, ymax= CI_high_90), size = 0.5) +
  geom_pointrange(aes(ymin = CI_low_70, ymax= CI_high_70), size = 1) +
  geom_pointrange(aes(ymin = CI_low_50, ymax= CI_high_50), size = 2) +
  scale_color_gradientn(colours = c("#f44336", "#FF9800", "#4CAF50"), limits=c(-max_value, max_value)) +
  coord_flip() +
  see::theme_modern(legend.position = "none")

^{Created on 2019-07-26 by the reprex package (v0.3.0)}

from see.

Current implementation:

library(parameters)
library(see)

model <- glm(mpg ~ wt + cyl + gear + hp, data = mtcars)
mp <- model_parameters(model)
plot(mp)

mp <- model_parameters(model, ci = c(0.5, 0.7, 0.9))
plot(mp)

^{Created on 2019-07-30 by the reprex package (v0.3.0)}

from see.