nimbleMacros (version 0.1.1)

LINPRED: Macro to build code for linear predictor from R formula

Description

Converts an R formula into corresponding code for a linear predictor in NIMBLE model code. Options are available to specify a link function and to also generate code for priors corresponding to the parameters in the linear predictor.

Value

NIMBLE code for the linear predictor specified by the formula, and optionally the associated priors.

Arguments

formula

An R formula, possibly with the parameters followed by brackets containing indices. If there are no indices, the macro attempts to guess the correct indices from the context. The formula must be right-hand side only (e.g. ~x). This must always be the first argument supplied to LINPRED.

link

A link function that will be applied to the left-hand-side (the response) in the final linear predictor. Default is none.

coefPrefix

All model coefficient names will begin with this prefix. default is "beta_" (so 'x' becomes 'beta_x', etc.)

sdPrefix

All dispersion parameters will begin with this prefix. Default is no prefix.

priors

Prior specifications, generated with setPrior()

modelMatrixNames

Logical indicating if parameters should be named so they match the names one would get from R's model.matrix. Default is FALSE.

noncentered

Logical indicating whether to use noncentered parameterization for random effects. Default is FALSE. Under the noncentered parameterization, random effects have a standard normal prior (beta_x_raw ~ dnorm(0, sd = 1)) and are then scaled by the hyperparameters (mean and SD), i.e., beta_x = mu_beta + sd_beta * beta_x_raw. This parameterization can improve mixing in some cases.

centerVar

Grouping variable (covariate) to 'center' the random effects on. By default (when NULL), random effects come from normal distributions with mean 0 as with lme4. For example, for random intercepts by grouping variable x, the linear predictor would be beta_Intercept + beta_x[x[i]] and the prior for the random effects would be beta_x ~ dnorm(0, sd_x). When centerVar = x, the linear predictor would be beta_x[x[i]] and the random effect prior would be beta_x ~ dnorm(beta_Intercept, sd = sd_x). That is, the mean of the random effects is now beta_Intercept. These two formulations should yield the same results. Note that this option is unrelated to the noncentered argument despite the similar names.

modelInfo

Used internally by nimbleMacros; a list of model information such as constants and dimensions

.env

Used internally by nimbleMacros; the environment where the model was created

Author

Ken Kellner

Examples

constants <- list(x = rnorm(3), x2 = factor(letters[1:3]))
code <- nimbleCode({
  mu[1:3] <- LINPRED(~x + x2)
})

mod <- nimbleModel(code, constants = constants)
mod$getCode()