README.md

In njtierney/mmcc: tidy mcmc.list using data.table

mmcc

Tidying up MCMC output can be a real pain. There are plenty of packages that help with summarising MCMC and providing their own summaries, but sometimes you just want a tidy data structure so you can do your own thing. And quickly.

mmcc provides tidying functions that return tidy data structure from mcmc.list objects. It uses data.table as the backend for speediness, it also provides broom tidiers to assist in some quick summaries.

Installation

Install from github using:

# install.packages("remotes")
remotes::install_github("njtierney/mmcc")

Using mmcc

mcmc_to_dt takes an mcmc.list object and turns it into a data.table of the format:

library(coda)
data(line)

head(data.frame(line$line1))
#>     alpha      beta     sigma
#> 1 7.17313 -1.566200 11.233100
#> 2 2.95253  1.503370  4.886490
#> 3 3.66989  0.628157  1.397340
#> 4 3.31522  1.182720  0.662879
#> 5 3.70544  0.490437  1.362130
#> 6 3.57910  0.206970  1.043500
library(mmcc)

mcmc_dt <- mcmc_to_dt(line)

mcmc_dt
#>       iteration chain parameter    value
#>    1:         1     1     alpha 7.173130
#>    2:         2     1     alpha 2.952530
#>    3:         3     1     alpha 3.669890
#>    4:         4     1     alpha 3.315220
#>    5:         5     1     alpha 3.705440
#>   ---                                   
#> 1196:       196     2     sigma 1.306930
#> 1197:       197     2     sigma 0.846828
#> 1198:       198     2     sigma 0.465129
#> 1199:       199     2     sigma 0.672417
#> 1200:       200     2     sigma 0.639787

tidy.mcmc.list takes an mcmc.list, turns it into a data.table and summarises it in terms of each parameter’s mean, median, standard deviation and credible interval with level given by conf.level:

tidy(line)
#>    parameter      mean        sd      2.5%    median    97.5%
#> 1:     alpha 2.9875644 0.4983950 1.9650403 3.0188300 3.876589
#> 2:      beta 0.7991864 0.3366834 0.1430713 0.7962500 1.469723
#> 3:     sigma 0.9680519 0.7413014 0.4249618 0.7911975 2.559520

We can also optionally ask for a subset of the parameters with a vector of colnames and summarise for each chain:

tidy(line, 
     chain = TRUE, 
     colnames=c("alpha"))
#>    parameter chain     mean        sd     2.5%   median    97.5%
#> 1:     alpha     1 2.982615 0.5313900 2.085719 2.973115 3.838839
#> 2:     alpha     2 2.992514 0.4643476 1.965040 3.063630 3.890256

This may be useful if we want to make a plot that shows how a given parameter varies from chain to chain.

library(ggplot2)

line_tidy <- tidy(line, chain = TRUE)

ggplot(data = line_tidy, 
       aes(x = factor(chain), 
           y = mean)) +
    geom_pointrange(aes(ymin = `2.5%`,
                        ymax = `97.5%`)) +
    facet_wrap(~parameter, 
               nrow = 1, 
               scales = "free_y") +
    theme_bw() +
    xlab("Chain") +
    ylab("Value")

Helpers

mmcc provides helpers to extract the number of chains, iterations, simulations, and variables in an MCMC object:

n_chain(line)
#> [1] 2
n_iter(line)
#> [1] 200
n_sims(line)
#> [1] 400
n_var(line)
#> [1] 3

glance helper

General model summary providing information on:

• Number of chains
• Number of iterations
• Number of variables/parameters
• Lower and upper effective sample size (ESS)
• lower and upper estimates of rhat measurement of convergence assessment

glance(line)
#>   n_chains n_iter n_var ess_lower ess_upper rhat_lower rhat_upper
#> 1        2    200     3  167.5889  455.3178   1.000695   1.037599

Diagnostics

diag_autocorr(line)
diag_ess(line)
#>    chain parameter       ess
#> 1:     1     alpha 200.00000
#> 2:     1      beta 249.43130
#> 3:     1     sigma  90.16969
#> 4:     2     alpha 255.31778
#> 5:     2      beta 200.00000
#> 6:     2     sigma  77.41925
diag_mc_stderr(line)
#> Warning in bmeans(v$value, b_size, warn): likely too few samples (less than
#> 1000) for estimating Monte Carlo standard errors

#> Warning in bmeans(v$value, b_size, warn): likely too few samples (less than
#> 1000) for estimating Monte Carlo standard errors

#> Warning in bmeans(v$value, b_size, warn): likely too few samples (less than
#> 1000) for estimating Monte Carlo standard errors

#> Warning in bmeans(v$value, b_size, warn): likely too few samples (less than
#> 1000) for estimating Monte Carlo standard errors

#> Warning in bmeans(v$value, b_size, warn): likely too few samples (less than
#> 1000) for estimating Monte Carlo standard errors

#> Warning in bmeans(v$value, b_size, warn): likely too few samples (less than
#> 1000) for estimating Monte Carlo standard errors
#>    parameter chain      mean      value
#> 1:     alpha     1 2.9826146 0.03667355
#> 2:     alpha     2 2.9925142 0.03001358
#> 3:      beta     1 0.7866946 0.02285799
#> 4:      beta     2 0.8116781 0.02778074
#> 5:     sigma     1 0.9544249 0.10259971
#> 6:     sigma     2 0.9816789 0.04737521

Thinning

thin_dt(mcmc_to_dt(line), thin = 10)
#>      iteration chain parameter    value
#>   1:         1     1     alpha 7.173130
#>   2:        11     1     alpha 3.065240
#>   3:        21     1     alpha 2.876460
#>   4:        31     1     alpha 2.943310
#>   5:        41     1     alpha 3.087570
#>  ---                                   
#> 116:       151     2     sigma 0.764629
#> 117:       161     2     sigma 1.315890
#> 118:       171     2     sigma 1.034110
#> 119:       181     2     sigma 2.362830
#> 120:       191     2     sigma 1.009290

Why the name, “mmcc”?

Full credit does to Sam Clifford for the name.

To quote Sam:

…it’s all about reshaping and manipulating mcmc chains…therefore, mmcc

Future work

• Summaries for each parameter
• More diagnostic summaries for convergence

Code of Conduct

Please note that this project is released with a Contributor Code of Conduct. By participating in this project you agree to abide by its terms.

njtierney/mmcc documentation built on Oct. 5, 2021, 12:14 a.m.