Skip to contents

Generate Random Starting Values for OpenMx Models

Source: R/random_starts.R
random_starts.Rd

Generates plausible random starting values for OpenMx models. Starting values are scaled, where possible, to the observed data. Multiple sets of starting values can be evaluated to identify the solution with the lowest minus twice log-likelihood.

Usage

random_starts(
  x,
  nstarts = 20,
  tolerance_minus2ll = 1e-06,
  verbose = TRUE,
  run = TRUE,
  ...
)

Arguments

x

An OpenMx MxModel object.

nstarts

Integer. Number of sets of starting values to generate.

tolerance_minus2ll

Numeric. Maximum difference from the smallest minus twice log-likelihood for solutions to be considered equivalent.

verbose

Logical. Whether to print information about starting values and model fit.

run

Logical. If TRUE, estimate the models and return the best-fitting solution. If FALSE, return a list of models with randomized starting values.

...

Additional arguments passed to the model-specific implementation.

Value

If run = TRUE, the best-fitting OpenMx model with an attribute tab_ll containing information about the starting values and their fit statistics. If run = FALSE, a list of OpenMx models.

Details

Random starting values are generated using R's random-number generator. Use set.seed() for reproducibility.

RAM mixture models

Starting values are derived from the observed data and the model's RAM matrices. The function supports free parameters in matrices named A, S, M, mat_dev, and Thresholds. The F matrix may be present but is not randomized.

For continuous variables, observed ranges and standard deviations are used to determine plausible starting values. Fallback scales are used when the observed range or standard deviation is zero or undefined.

Free paths in A are sampled uniformly after scaling by the ratio of the standard deviations of the dependent and predictor variables: $$A_{ij} \sim U(a,b)\frac{sd_i}{sd_j}$$ where c(a, b) = scale_A.

Starting values for S are generated using a random lower-triangular Cholesky factor \(L\), giving $$S = LL^\top.$$ Diagonal elements of \(L\) are sampled between 0.5 and 1 times the corresponding observed standard deviation. Off-diagonal elements are sampled between -0.25 and 0.25 times that standard deviation.

Free means in M are sampled uniformly between the observed minimum and maximum of the corresponding variable.

Free elements of mat_dev are sampled uniformly from scale_mat_dev. For ordinal variables, empirical latent-normal thresholds are obtained from cumulative category proportions using qnorm(), perturbed by uniform jitter, and sorted to preserve their ordering.

For models containing submodels, initialization is applied recursively. Observed data stored in the model take precedence over externally supplied data.

After random values have been assigned, OpenMx::omxAssignFirstParameters() ensures that parameters sharing labels have consistent starting values.

Additional arguments for RAM mixture models

When attr(x, "tidySEM") == "mixture", the following arguments can be passed through ...:

  • data: Optional observed data.

  • scale_A: Lower and upper multipliers for free elements of the A matrix. Default: c(-0.5, 0.5).

  • scale_mat_dev: Sampling range for threshold deviations. Default: c(0.001, 1.5).

  • jitter_thresholds: Magnitude of random jitter applied to thresholds. Default: 0.25.

Examples

if (FALSE) { # \dontrun{
library(OpenMx)

set.seed(123)
model <- remix(yi, vi, classes = 2, run = FALSE)
fit <- random_starts(model, nstarts = 20)

attr(fit, "tab_ll")
} # }