For each sampled MCMC generation, the values estimated for a pair of parameters are logged and plotted against one another. Points are color coded by when in the analysis they were sampled, so that users can visually assess mixing.
plot_joint_marginal(parameter1, parameter2, percent.burnin = 0, thinning = 1,
param.name1 = deparse(substitute(parameter1)),
param.name2 = deparse(substitute(parameter2)))
One of the two parameters for which the joint marginal is being plotted.
The other of the two parameters for which the joint marginal is being plotted.
The percent of the sampled MCMC generations to be discarded as "burn-in." If the
MCMC is run for 1,000,000 generations, and sampled every 1,000 generations, there
will be 1,000 sampled generations. A percent.burnin of 20 will discard
the first 200 sampled parameter values from that sample.
The multiple by which the sampled MCMC generations are thinned. A thinning of
5 will sample every 5th MCMC generation.
The name of one of the two parameters for which the joint marginal is being plotted.
The name of the other of the two parameters for which the joint marginal is being plotted.
Gideon Bradburd
Visualizations of the joint marginal distribution allow users to (1) assess how well the MCMC is mixing, and (2) potentially diagnose instances of non-identifiability in the model. Strong linear trends in the joint marginal, or visible "ridges" in the likelihood surface, may be indicative of parameter non-identifiability, in which multiple combinations of values of these two parameters provide equally reasonable fits to the data.