recompute makes R crash

[SanVerhavert]

After running:

DC_BF <- anovaBF( DC_mean ~ Clusternummers, data = SchoolCultureAggr, whichRandom = "Clusternummers", rscaleRandom = "medium" )
DC_chains <- posterior( DC_BF, iterations = 10000 )
DC_chains <- recompute( DC_chains, iterations = 10000, thin = 3 )

R crashes with the following log entry (pulled from RStudio): ERROR system error 109 (pipe ended); OCCURRED AT: virtual void rstudio::session::NamedPipeHttpConnection::close() C:\Users\Administrator\rstudio\src\cpp\session\http\SessionNamedPipeHttpConnectionListener.hpp:198; LOGGED FROM: virtual void rstudio::session::NamedPipeHttpConnection::close() C:\Users\Administrator\rstudio\src\cpp\session\http\SessionNamedPipeHttpConnectionListener.hpp:198

R version 3.3.0 (2016-05-03)
Platform: x86_64-w64-mingw32/x64 (64-bit)
Running under: Windows 7 x64 (build 7601) Service Pack 1

EDIT1:
other attached packages:
[1] BayesFactor_0.9.12-2 Matrix_1.2-6 coda_0.18-1
[4] plyr_1.8.3

loaded via a namespace (and not attached):
[1] magrittr_1.5 tools_3.3.0 pbapply_1.2-1 MatrixModels_0.4-1
[5] Rcpp_0.12.5 mvtnorm_1.0-5 stringi_1.0-1 grid_3.3.0
[9] stringr_1.0.0 gtools_3.5.0 lattice_0.20-33

EDIT2: This issue appears to occur only (especially?) when posterior and/or recompute are ran several times. If the R session is restarted and every command is executed only once no problems occur.
As an example: several BF's are computed and from that posteriors are sampled. Then calling recompute several times (i.e. to have different levels of thinning) the R session hangs/freezes until it aborts or is aborted by RStudio.

[richarddmorey]

Thanks for this. Do you have a reproducible example with a built-in or public data set?

[SanVerhavert]

Unfortunately it appears hard to reproduce. It appears to happen only when a large dataset is in the "BFBayesFactor" object. Only in that case the function posterior creates a "BFmcmc" object in stead of again a "BFBayesFactor" object. Saving the predictor and predicted variable that I will use in the anovaBF function in a separate data frame appears to solve the problem.
I am thus unable to provide a reproducible example.

[richarddmorey]

hmmm, maybe a memory issue. I'll try to see if I can get it to happen. How many rows/columns is your dataset?

[SanVerhavert]

My dataset has 65 observations (rows) and 18 variables (columns).

It must be remarked that in the anovaBF function my formula only includes one predictor variable (factor) which is specified as a random effect (no nuisance variable).
[As a side note: In this case, might it be more (memory) efficient to just include the data used in the formula in the "BFBayesFactor" object? Or is there a reason why the whole dataset is included in that object class?]