`runLaplace` with uncompiled Laplace object fails (`gr_logLik_pTransformed` not found)

[paciorek]

It's fine for compiled Laplace object, so this is simply a minor cleanup issue.

model_code <- nimbleCode({
# priors
intercept ~ dnorm(0, sd = 100)
beta ~ dnorm(0, sd = 100)
sigma ~ dhalfflat()
# random effects and data
for(i in 1:10) {
# random effects
ran_eff[i] ~ dnorm(0, sd = sigma)
for(j in 1:5) {
# data
y[i,j] ~ dpois(exp(intercept + beta*X[i,j] + ran_eff[i]))
}
}
})

set.seed(123)
X <- matrix(rnorm(50), nrow = 10)
model <- nimbleModel(model_code, constants = list(X = X), calculate = FALSE,
buildDerivs = TRUE) # Here is the argument needed for AD.
model$intercept <- 0
model$beta <- 0.2
model$sigma <- 0.5
model$calculate()
model$simulate(model$getDependencies('ran_eff'))
model$calculate()
model$setData('y')
glmm_laplace <- buildLaplace(model, c('intercept','beta','sigma'))
runLaplace(glmm_laplace, c(0,0,1))
#  [Warning] Running an uncompiled Laplace or AGHQ algorithm. Use compileNimble() for faster execution.
#Error in optimize(pStart = pStart, method = method, hessian = hessian,  : 
 # object 'gr_logLik_pTransformed' not found

[paciorek]

This is very odd. Some (but not all!) of the methods in the ref class produced by buildAGHQ don't seem to be in scope when one tries to access them from the optimize method. I will continue to investigate.

[paciorek]

@perrydv said this is a known bug/aspect of referenceClasses, relating to something about how elements are only kept track of if they are used at least once.

Suggestion is to invoke gr_logLik_pTransformed and AGHQuad_nfl[[i]]$gr_inner_logLik and similar ones before using runLaplace.

I've done this in runAQHG in the hessian-for-inner-laplace branch. Not clear we can deal with this in findMLE, so that is still a gap here.

[paciorek]

Also noting that I made an equivalent fix for runNestedApprox in nimbleQuad.