Simulate data from a multilevel VAR model

Generates data from the generative model implied by each Stan model family. Useful for testing parameter recovery and model validation.

Usage

sim_var(
  N,
  T_obs,
  p,
  K = 1L,
  family = c("bernoulli", "ordinal", "gaussian"),
  alpha = NULL,
  gamma = NULL,
  Phi = NULL,
  sigma = NULL,
  kappa = NULL,
  q = 0L,
  x_gen = NULL,
  sd_alpha = 0.5,
  sd_phi = 0.2,
  sd_gamma = NULL,
  re_temporal = FALSE,
  C = 5L,
  burnin = 500L,
  seed = NULL
)

Arguments

N: Integer. Number of subjects (groups).
T_obs: Integer. Number of time points per subject.
p: Integer. Number of outcome nodes.
K: Integer. AR order (default 1).
family: Character. One of "bernoulli", "ordinal", "gaussian".
alpha: Numeric vector of length p. Population intercepts (on logit scale for bernoulli, identity for gaussian). For ordinal, this is absorbed into kappa and should be left NULL. Generated if NULL.
gamma: Matrix q x p. Population covariate effects. Generated if NULL and q > 0.
Phi: Matrix (p*K) x p. Population lag coefficients. Generated if NULL.
sigma: Numeric vector of length p. Residual SD per node (gaussian only). Generated if NULL.
kappa: List of p ordered vectors, each of length C-1. Cutpoints per node (ordinal only). Generated if NULL.
q: Integer. Number of covariates (default 0).
x_gen: Function f(N, T_obs) returning an N x T_obs x q array of covariates. If NULL, default generation is used.
sd_alpha: Numeric. SD of random intercepts (scalar or p-vector). Default 0.5. Set to 0 to simulate a fixed-effects-only model with no between-person variation in intercepts.
sd_phi: Numeric. SD of random lag coefficients (scalar or matrix). Default 0.2.
sd_gamma: Numeric or NULL. SD of random covariate slopes. NULL means no random slopes on covariates.
re_temporal: Logical. Include random slopes on lag predictors? Default FALSE.
C: Integer. Number of ordinal categories (ordinal only, default 5).
burnin: Integer. Number of time points to discard as warmup before recording data (default 500). The VAR process is simulated for burnin + T_obs time points per subject, and the first burnin are discarded. This allows the process to reach its stationary distribution before data collection begins.
seed: Integer or NULL. RNG seed.

Value

A list with two components:

data: A long-format data frame with columns id, t, y_1, ..., y_p, and optionally x_1, ..., x_q.
truth: A list of true generating parameters.

Details

To simulate a VAR without any random effects (i.e. all subjects share identical parameters), set sd_alpha = 0, re_temporal = FALSE (the default), and sd_gamma = NULL (the default).