ref_dynamics

Usage

ref_dynamics(
  reference,
  data,
  id,
  time,
  process = ref_process(),
  crossfit = 5,
  uncertainty = c("total", "conditional"),
  n_draw = NULL
)

Attaches a Gaussian-copula process on marginal normal scores $Z=\Phi^{-1}(F(y\mid x))$. The default kernel is stable rank plus a Matern-3/2 process plus a measurement nugget, which is positive semidefinite for every irregular visit schedule.

Arguments

Argument	Description
reference	A ref_fit.
data	Longitudinal reference data.
id	Subject identifier.
time	Time variable (typically age).
process	A kernel from ref_process().
crossfit	Number of subject-level folds used to obtain out-of-fold Z scores via ref_crossfit() (whole subjects stay in one fold). 0 or NULL uses in-sample scores from reference.
uncertainty	Marginal predictive estimand used to create every Z score and every downstream transition or forecast. Defaults to "total", matching predict.ref_fit().
n_draw	Coefficient draws for total marginal uncertainty. The value is stored and reused downstream.

Details

Because the marginal scores have unit variance, the three variance components are estimated on the simplex $\tau_b^2 + \tau_g^2 + \sigma_e^2 = 1$ and the Matern length-scale is bounded by the observed lags. When all within-subject lags are (nearly) identical the length-scale is not identified and the model is automatically reduced to the stable-rank-plus-nugget kernel (process = "stable"), which has a single parameter: the correlation at the common lag. Optimisation is multi-start L-BFGS-B; when no start converges the process is reported as identified = FALSE and every history-conditioned quantity downstream is NA.

Value

An object of class ref_dynamics with, per outcome, a fitted process ($processes), and a components table of normalised variance fractions with approximate standard errors.