Can something like the " gelman.plot" from the "coda" package be added? Simple function below

[ltrainstg]

I found this package recently and I like how it uses ggplot instead of base R for graphs. One of the only graphs that seemed to be missing that was in "coda" that I used frequently was 'gelman.plot'. The current 'ggs_Rhat' function only shows the end result. It is useful to see how this changes over time to see if the parameter has truly converged and not simply lucked into an early convergence.

I attempted to create a function that mimics the gelman.plot from coda as much as possible below. Could something similar be added to this package?

library(coda)
library(ggmcmc)
library(dplyr)
data(s.binary)
gelman.plot(s.binary)
gg_gelmanPlot <- function(x){

  x <- as.mcmc.list(x)
  y <- coda:::gelman.preplot(x, bin.width = 10, max.bins = 50, confidence = 0.95, 
                             transform = FALSE, autoburnin = TRUE)

  df <- NULL
  for(i in 1:2){
    df.piece <- cbind(
      y$last.iter,
      y$shrink[, , 1]) %>%
      as.data.frame()
    baseParams <- colnames(y$shrink[,,i])
    names( df.piece) <- c("iterations", baseParams)
    df.piece <- 
      gather(df.piece, condition, measurement,
             baseParams[1]:rev(baseParams)[1],
             factor_key=TRUE)
    df.piece$i <- i
    df <- rbind(df, df.piece)
  }

  df <- df %>% mutate(Variable = ifelse(i==1, "Median", "97.5%"))
  ggplot(dat = df,
         aes(x=iterations, y=measurement, col= Variable)) +
    geom_line(aes(linetype = Variable)) + 
    geom_hline(yintercept = 1.05, linetype = "dashed") +
    facet_grid(vars(condition), scales = "free") +
    xlab("last iteration in chain") + 
    ylab("shrink factor")

}
gg_gelmanPlot(s.binary)

[xfim]

Dear @ltrainstg ,

Thank you very much for your suggestion and the code provided.

I adapted it and produced a new function ggs_grb() (the name includes the whole group of contributors to the plot idea).

Let me know what you think, see commit f6dddd455116d07e909dfe1676c3f526073889d0.

It is already packed in version 1.5.0.

I would love to hear suggestions specifically for:

• Confidence band. Although the original plot in "coda" comes with the confidence interval, I have opted to take it out to emphasize the main idea of rolling Rhat. But if there is really the need to have it, I can adapt it easily.

• All figures in a single layout. I have decided to use a single layout for all the parameters by using a facets_wrap() approach. In case of many, many parameters, this can be problematic. But there is always the option of limiting them using 'family'.