itsadug (version 2.5)

compareML: Function for comparing two GAMM models.

Description

Function for comparing two GAMM models.

Usage

compareML(
  model1,
  model2,
  signif.stars = TRUE,
  suggest.report = FALSE,
  print.output = TRUE
)

Value

Optionally returns the Chi-Square test table.

Arguments

model1

First model.

model2

Second model.

signif.stars

Logical (default = TRUE). Whether or not to display stars indicating the level of significance on 95% confidence level.

suggest.report

Logical (default = FALSE). Whether or not to present a suggestion on how one could report the information. If print.output is set to FALSE, suggest.report will set to FALSE too. Please inspect yourself whether the label between square bracket fits your own standards. Note: the X2 should be replaced by a proper Chi-Square symbol \(\chi^2\).

print.output

Logical: whether or not to print the output. By default set to true, even if the the messages are not allowed by a global package option using the function infoMessages.

Notes

For suppressing the output and all warnings, set infoMessages to FALSE (infoMessages('off') ), set the argument print.output to FALSE, and use the function suppressWarnings to suppress warning messages.

Author

Jacolien van Rij. With many thanks to Simon N. Wood for his feedback.

Details

As an Chi-Square test is performed on two times the difference in minimized smoothing parameter selection score (GCV, fREML, REML, ML), and the difference in degrees of freedom specified in the model. The degrees of freedom of the model terms are the sum of 1) the number of estimated smoothing parameters for the model, 2) number of parametric (non-smooth) model terms including the intercept, and 3) the sum of the penalty null space dimensions of each smooth object.

This method is preferred over other functions such as AIC for models that include an AR1 model or random effects (especially nonlinear random smooths using bs='fs'). compareML also reports the AIC difference, but that value should be treated with care.

Note that the Chi-Square test will result in a very low p-value when the difference in degrees of freedom approaches zero. Use common sense to determine if the difference between the two models is meaningful. A warning is presented when the difference in score is smaller than 5.

The order of the two models is not important. Model comparison is only implemented for the methods fREML, REML, and ML, for other models the log-likelihood is being used (function logLik.gam).

See Also

For models without AR1 model or random effects AIC can be used.

Other Testing for significance: plot_diff(), plot_diff2(), report_stats(), wald_gam()

Examples

data(simdat)

if (FALSE) {
infoMessages('on')
# some arbitrary models:
m1 <- bam(Y~Group + s(Time, by=Group), method='ML', data=simdat)
m2 <- bam(Y~Group + s(Time), method='ML', data=simdat)

compareML(m1, m2)

# same models, but now with fREML (default):
m1 <- bam(Y~Group + s(Time, by=Group), data=simdat)
m2 <- bam(Y~Group + s(Time), data=simdat)

compareML(m1, m2)

# exclude significance stars:
compareML(m1, m2, signif.stars=FALSE)

# same model as m1, increased k:
m3 <- bam(Y~Group + s(Time, by=Group, k=25), method='fREML', 
    data=simdat)
compareML(m1, m3)

# do not print output, but save table for later use:
cml <- compareML(m1, m2, print.output=FALSE)$table
cml

# Use suppressWarnings to also suppress warnings:
suppressWarnings(cml <- compareML(m1, m2, print.output=FALSE)$table)

}