Error when both n.chain and start.val are given if n.chain>1

[hdarjus]

Hi @ArdiaD,

thanks for the package. For improved MCMC diagnostics for my case study, I wanted to run multiple chains and also provide starting values. I got into an error that I was able to overcome only by reading the source code of the package.

MWE:

model_control <- list(n.chain = 2,
                      start.val = matrix(c(0.01, 0.1, 0.7, 500,
                                           0.3, 3, 0.01, 500),
                                         2, 4, byrow = TRUE))
fit <- bayesGARCH(rep(0, 10), control = model_control)

The following error is thrown before sampling starts:

Error in dimnames(x) <- dn : 
  length of 'dimnames' [2] not equal to array extent

The problem is in this line:

bayesGARCH/R/sampler.R

Line 82 in becc562

else con$start.val <- matrix(rep(con$start.val, con$n.chain),

It is true that providing n.chain is redundant if start.val is given but I personally found the behavior counterintuitive (I think in JAGS and Stan you have to provide both) and the solution should be simple.
I hope you find this ticket useful.

Cheers,
Darjus

[ArdiaD]

Hi @hdarjus cannot you simple make a loop around the fit and collect the results?

[hdarjus]

Yes, that's one potential workaround. However, that doesn't solve the issue that I provided consistent control input (since n.chain is equal to NROW(start.val)) to bayesGARCH and received an error.

A package-wide solution could be to add something like

else if (NROW(con$start.val) > 1) {if (NROW(con$start.val) != con$n.chain) stop("Arguments 'n.chain' and 'start.val' inconsistent; please double-check your input.")}

to the above linked line before the else.

[ArdiaD]

Simply put n.chain = 1. I won't change the package for this design feature. But thx for your proposal

[ArdiaD]

Oh, sorry @hdarjus, I missunderstood your idea. Yes, will make the fix for the next release. Thx for pointing this to me !

[hdarjus]

I am glad it works out!