tef_checkPars: Check for bound hitting and other undesireable outcomes...
In akcochrane/TEfits: Time-evolving model fits

tef_checkPars

R Documentation

Check for bound hitting and other undesireable outcomes within an optim() call

Description

TEfits internal.

Usage

tef_checkPars(
  err,
  guesses,
  curDat,
  pNames,
  evalFun,
  errFun,
  respVar,
  linkFunX = NA,
  y_lim,
  rate_lim,
  shape_lim,
  penalizeRate,
  paramTerms,
  guessGroups = NULL
)

Arguments

`err`	Error
`guesses`	Parameter values
`curDat`	Data being fit
`pNames`	Parameter names
`evalFun`	Function being fit
`errFun`	Function to calculate error
`respVar`	Name of the response variable
`linkFunX`	If relevant, the "x" value for a link function (e.g., Weibull, logistic)
`y_lim`	Limits to fit values
`rate_lim`	Limits to rate parameter
`shape_lim`	If using a Weibull change function, limits to Weibull shape parameter
`penalizeRate`	Logical. Should error be penalized if rate is extremely close to the bounds?
`paramTerms`	parameter-level regressions, to be evaluated for checking y_lim and rate_lim
`guessGroups`	deprecated

Details

Sane boundaries for parameters are the only way that many nonlinear regression optimizations can be identifiable. Fortunately, theory-driven constraints on parameter ranges provide useful a priori restrictions on the possible ranges for parameters and model predictions. This function checks the following:

start and asymptote parameters – all models are parameterized in terms of starting and ending values. This ensures that the starting and ending values comply with the y_lim boundaries; y_lim may be user-defined, defined by another model feature (e.g., bernoulli error function is limited to predicted values of 0 or 1; Weibull link thresholds must be above 0).
rate parameter – If not user-input, then defined by TEfits::tef_getLinkedFun. Defaults, with exponential change, to a minimum that would provide 50 amount of time, and to a maximum value that would provide 80 functions have limits that, with their respective parameterizations, are intended to imitate the limits of the 3-parameter exponential (i.e., imitate the overall shape of the curve's extremes) These are heuristics. The default values are intended to be flexible while maintaining a sufficiently constrained curve such that both starting asymptote parameters are interpretable (i.e., if the rate parameter, which is a time constant, were to be extremely small, the start parameter could become infinitely large or small). If a time-evolving process occurs on a timescale that cannot be fit by the default boundaries, it is likely that the data is unsufficient to characterize that process.
pPrevTime parameter – 4-parameter Power change utilizes a so-called "previous learning time" parameter that assumes that the learning function extents backward through time. This parameter must be greater than 0 and smaller than 1*10^5