Colin

Update: if I set a pretty strong prior on the par_int parameter as par_int ~ dnorm(5, 1.0E0), the parameters do converge and prediction is good. The parameter fits to ~7.5, with 5 being outside the 95% credible interval. Inspection of parameter correlations shows this introduced enough noise in the correlations among intercept terms to allow parameters to converge (though, intercept parameters are still strongly correlated). However, I still don't know if this is a defensible solution. My only reason...

Update: if I set a pretty strong prior on the par_int parameter as par_int ~ dnorm(5, 1.0E0), the parameters do converge and prediction is good. The parameter fits to ~7.5, with 5 being outside the 95% credible interval. Inspection of parameter correlations shows this introduced enough noise in the correlations among intercept terms to allow parameters to converge (though, intercept parameters are still strongly correlated). However, I still don't know if this is a defensible solution. My only reason...

Update: if I set a pretty strong prior on the par_int parameter as par_int ~ dnorm(5, 1.0E0), the parameters do converge and prediction is good. The parameter fits to ~7.5, with 5 being outside the 95% credible interval. Inspection of parameter correlations shows this introduced enough noise in the correlations among intercept terms to allow parameters to converge. However, I still don't know if this is a defensible solution. My only reason for setting such a tight prior is "this is what makes it...

Update: if I set a pretty strong prior on the par_int parameter as par_int ~ dnorm(5, 1.0E0), the parameters do converge and prediction is good. The parameter fits to ~7.5, with 5 being outside the 95% credible interval. Inspection of parameter correlations shows this introduced enough noise in the correlations among intercept terms to allow parameters to converge. However, I still don't know if this is a defensible solution. My only reason for setting such a tight prior is "this is what makes it...

I am modeling microbiome count data, using a multinomial dirichlet. The number of times I observe each microbial "species" depends on its fractional abundance within a microbial community, and the number of DNA sequences observed per sample, referred to as sequencing depth. I am finding I get great agreement between predicted vs. observed values, but parameters do not converge. I have tried "fixing" the parameters for one species at a constant value, which solved this problem when I was fitting a...

It is worth noting that the dirichlet-only model does fit. This requires modifying the jd object to be: jd <- list(K = ncol(prob), N = nrow(prob), seq.prb=prob, x = preds, N.preds = ncol(preds)) And removing the line: seq.cnt[i,1:K] ~ dmulti(seq.prb[i,1:K],seq.depth[i]) from the model.

I am modeling microbiome count data, using a multinomial dirichlet. The number of times I observe each microbial "species" depends on its fractional abundance within a microbial community, and the number of DNA sequences observed per sample, referred to as sequencing depth. I am finding I get great agreement between predicted vs. observed values, but parameters do not converge. I have tried "fixing" the parameters for one species at a constant value, which solved this problem when I was fitting a...

I am modeling microbiome count data, using a multinomial dirichlet. The number of times I observe each microbial "species" depends on its fractional abundance within a microbial community, and the number of DNA sequences observed per sample, referred to as sequencing depth. I am finding I get great agreement between predicted vs. observed values, but parameters do not converge. I have tried "fixing" the parameters for one species at a constant value, which solved this problem when I was fitting a...