secr.refit: Repeat or Resume Model Fit

Description

From 5.2.1 it is possible to save a part-fitted model to a file and resume fitting with secr.refit. The function may also be used to re-fit a completed model, or to repeat just the variance estimation (Hessian) step.

Usage

secr.refit(object, ...)

Value

A fitted secr model.

Arguments

object: secr object (a fitted model) or character value naming a saved progress RDS file (e.g. progress.RDS)
...: arguments of secr.fit to override those in object

Details

There are two common uses for the function.

1. To resume fitting of a partly fitted model from an RDS file saved with saveprogress.

2. To recompute variances of a fitted model. The numerical method is fdHess from package nlme. Often this delivers variance estimates when the default algorithm failed.

The ... argument is used to override previous settings (arguments of secr.fit); commonly method = "none" to recompute variances. A new 'details' argument may be incomplete: specified components replace those in the original (see example resetting 'saveprogress' below).

The 'start' argument may not be overridden: it is set by secr.refit from the previous fit. The 'binomN' argument is elevated automatically from details$binomN if required.

A third use of secr.refit is to vary the model structure. This requires that 'object' is a fitted model. A change in model structure is implied when the ... argument replaces any of 'model', 'CL', 'detectfn', 'link', 'fixed', 'hcov', or 'groups'. New starting values are then determined in secr.fit by a call to makeStart, and may not precisely follow the previous fit.

Also, secr.refit will compute a full model from a relative density model (i.e. estimate the intercept) - just set CL = FALSE. If method = "none" then new variances are computed using fdHess. Note this does not work if the model includes individual covariates, which are prohibited in a full-likelihood model.

Examples

Run this code


# \donttest{

# Repeat variance step using nlme::fdHess
# method = "none" skips likelihood maximization
secr.refit(secrdemo.0, method = "none")

# Save data and coefficients after every 10 evaluations to a file
# with the default name 'progress.RDS'.
fit1 <- secr.fit(captdata, CL = TRUE, details = list(saveprogress = 10))

# Review progress. In this example, fit1 ran to completion, 
# but the log stops at a multiple of 10 evaluations.
attr(readRDS('progress.RDS'), 'log')

# Restart at last line in log, suppressing overwrite of progress log
secr.refit('progress.RDS', details = list(saveprogress = FALSE))

# cleanup
unlink("progress.RDS")

# Example using secr.refit() to complete a relative density model
# fitCL <- secr.fit(OVpossumCH[[1]], buffer = 120, detectfn = 'HEX', 
#    CL = TRUE, model = D~x+y)
# fit <- secr.refit(fitCL, CL = FALSE, method = "none")    
# coef(fitCL)    
# coef(fit)
# predict(fit)

# }