Options for the Randomized Parametric Expectation Maximization (RPEM) estimation method (Chen et al. 2024). This is the K=1 minimal core; see `design/rpem/` for the full roadmap.
Usage
rpemControl(
nGauss = 1000L,
nMH = 50000L,
mhBurn = 5000L,
niter = 50L,
collect = 15L,
seed = 42L,
atol = 1e-08,
rtol = 1e-08,
cores = 1L,
impInflate = 0,
likLbfgs = TRUE,
lbfgsLmm = 5L,
lbfgsFactr = 1e+07,
lbfgsPgtol = 0,
lbfgsMaxIter = 20L,
print = 1L,
printNcol = NULL,
useColor = NULL,
...
)Arguments
- nGauss
Number of Monte Carlo samples per subject in the E-step (`m_Gauss`).
- nMH
Number of Metropolis-Hastings trials collected in the M-step.
- mhBurn
Number of M-step MH burn-in trials (discarded).
- niter
Maximum number of E-M iterations.
- collect
Number of terminal iterations averaged for the final estimate.
- seed
RNG seed for the threefry sampler (reproducible for a fixed thread count).
- atol, rtol
ODE solver tolerances.
- cores
Number of cores for the threefry draw (solve threading is set by rxode2).
- impInflate
Opt-in mode-centered importance sampling for the E-step. `0` (default) keeps the paper's prior sampling (draw eta ~ N(0, Omega)). A value `>= 1` draws instead from N(EBE, impInflate*Omega) – centered at the previous iteration's posterior mean with that variance-inflation factor – and importance-weights, improving posterior-tail coverage for high-variance random effects in multi-eta models (whose largest Omega prior sampling under-estimates). Experimental: a partial mitigation, not a full fix (see design/rpem/04).
- likLbfgs
For a general log-likelihood (`ll()`) endpoint, refine the fixed-effect likelihood parameters each iteration by a box-constrained L-BFGS-B optimization of the importance-weighted observation log-likelihood (mirrors the saem/saemix ind.fix10 step), respecting the parameter bounds from the model, rather than the default single damped-Newton re-solve step. `TRUE` (default) for `ll()` models; ignored for standard residual-error models.
- lbfgsLmm, lbfgsFactr, lbfgsPgtol, lbfgsMaxIter
L-BFGS-B tuning for the `likLbfgs` likelihood-parameter refinement (number of corrections, convergence `factr`/`pgtol`, and max iterations).
Iteration-print frequency: display the parameter walk (population estimates + omega, with the back-transformed row) every `print` iterations (saem/focei/vae style), streamed live as the loop runs. `1` (default) prints every iteration; `0` captures the parameter history silently. The walk is *always* saved to the fit object's parameter history (`fit$parHist` / `fit$parHistStacked`) regardless. May also be an `iterPrintControl()` object.
- printNcol, useColor
Iteration-print formatting (columns per row, ANSI color); passed through to `iterPrintControl()`.
- ...
Ignored (reserved for future options).