Start values be outside prior range in DREAMzs / DEzs

[evodevosys]

I'm repeatedly getting these two warnings:

4: Start values outside prior range
5: Z matrix values outside prior range

What does this mean? It seems to happen repeatedly during the run, so I don't think it's just when I start a chain. Is it something to worry about? Should I modify my priors or runMCMC call to fix it?

I've pasted my priors and call below:

Generic prior density and prior sampler calls. The currentParInfo global variable represents a list with information that is sent into the actual distribution functions. The funcs field is one of the distribution functions below.

prdensities=function(pars){
  ds=sum(sapply(1:length(pars),function(i){
    currentParInfo$funcs[[i]](a=currentParInfo$low[i],b=currentParInfo$high[i],m=currentParInfo$best[i],s=currentParInfo$s[i],p=pars[i])
  }))
  return(ds)
}

prsamplers=function(n=1){
  ds=sapply(1:length(currentParInfo$low),function(i){
    currentParInfo$funcs[[i]](a=currentParInfo$low[i],b=currentParInfo$high[i],m=currentParInfo$best[i],s=currentParInfo$s[i],n=n)
  })
  return(ds)
}

The distribution functions are below. Variables a and b are lower and upper bounds where the distribution requires truncation. These are the same values that are passed into createPrior(). m is the mean and is the standard deviation where the distribution requires it. Again, the same as passed into createPrior() [best=mean]

The beta distribution goes between 0 and 1. Best is passed in as 1. Could this be the problem?

# dtrunc and rtrunc are from the LaplacesDemon package
dist.normtrunc<-function(a,b,m,s,p=NULL,n=NULL){
  if(is.null(n)){
    return(dtrunc(p,'norm',a=a,b=b,mean=m,sd=s,log=TRUE))
  } else {
    return(rtrunc(n,'norm',a=a,b=b,mean=m,sd=s))
  }
}

dist.unif<-function(a,b,...,p=NULL,n=NULL){# the ... lets it take m and s and not throw and error even though it doesn't use them
  if(is.null(n)){
    return(dunif(p,min=a,max=b,log=TRUE))
  } else {
    return(runif(n,min=a,max=b))
  }
}

dist.beta<-function(...,p=NULL,n=NULL){# the ... lets it take a,b,m and s and not throw and error even though it doesn't use them
  # This is fixed at the beta(4,1) distribution
  if(is.null(n)){
    return(dbeta(p,shape1=4,shape2=1,log=TRUE))
  } else {
    return(rbeta(n,shape1=3,shape2=1))
  }
}

The runMCMC call is:

priors=createPrior(density=prdensities,sampler=prsamplers,lower=currentParInfo$low,upper=currentParInfo$high,best=currentParInfo$best)
  bSetup=createBayesianSetup(likelihood=runTheModelAndGetLogLikelihoods,prior=priors)
  settings.1=list(iterations=50000,burnin=10000,nrChains=3,adaptation=9000)#for DREAM

  results=runMCMC(bayesianSetup =bSetup,settings=settings.1,sampler='DREAMzs')

[florianhartig]

The most likely reason is that your prior sampler produces values that are not within your prior or within the lower / upper conditions that you specify in the BayesianSetup.

In particular the lower upper has been an issue in the past - if this is inconsistent with your prior, the z-Matrix could be filled with values that have no prior support.

[evodevosys]

Thanks for the quick reply!! While you were replying I edited my question that gives a little more detail about just that. The upper and lower bounds for the sampler are truncated at the same values of low and high that I pass into createPrior(). [note that I don't use dist.norm() I'll take that out. The uniform, beta, and truncated normal are all bounded]

[florianhartig]

You're sure? I'd bet your sampler creates samples outside the prior range. Try x = bSetup$prior$sampler(1000), and then bSetup$prior$density(x) and check if there are any -Inf values.

[evodevosys]

I tried it with 100000 values. No -Inf

> x=bSetup$prior$sampler(100000)
> any(is.infinite(bSetup$prior$density(x)))
[1] FALSE

[florianhartig]

That is weird. The message is quite clear, and it all sounds as if your prior sampler produces values that are outside the prior range.

If you look into the Z-matrix of your sampler (directly after starting it up), e.g. like that

settings = list(iterations = 4)
out <- runMCMC(bayesianSetup = bayesianSetup, sampler = "DREAMzs", settings = settings)
out$Z

what kind of values do you see? Are they really inside the prior range? Do a visual check of prior and likelihood of these values - anything unusual?

If nothing of this helps, I guess the only thing is to look into the code and see what's going wrong. If you can send me a reproducible example of what you're doing, I could have a look, otherwise switch the debug mode on check why this error is occurring.

To be clear - it's warning, in principle the algorithm should still work, but it's a weird thing that this happens, so I would investigate what's going on.

[evodevosys]

The out$Z values are all inside the prior range. The likelihoods and priors don't look particularly odd to me - but I haven't looked at a ton of these:

I'll do some more investigating, see if I can narrow down exactly when it is happening and if there is anything consistent about the parameter values it is trying at those times, and put together a reproducible example

Thanks

[florianhartig]

Hi, yes, this looks completely normal.

I have no idea why you get this error, and I can't reproduce it with an example model. To investigate further, I would need a reproducible example.