Inferential Assessments About Model Performance

Methods for making inferences about differences between models

# S3 method for resamples
diff(x, models = x$models, metric = x$metrics,
  test = t.test, confLevel = 0.95, adjustment = "bonferroni", ...)

# S3 method for diff.resamples
summary(object, digits = max(3,
  getOption("digits") - 3), ...)

compare_models(a, b, metric = a$metric[1])

Arguments

x	an object generated by resamples
models	a character string for which models to compare
metric	a character string for which metrics to compare
test	a function to compute differences. The output of this function should have scalar outputs called estimate and p.value
confLevel	confidence level to use for dotplot.diff.resamples. See Details below.
adjustment	any p-value adjustment method to pass to p.adjust.
...	further arguments to pass to test
object	a object generated by diff.resamples
digits	the number of significant differences to display when printing
a, b	two objects of class train, sbf or rfe with a common set of resampling indices in the control object.

Value

An object of class "diff.resamples" with elements:

call

the call

difs

a list for each metric being compared. Each list contains a matrix with differences in columns and resamples in rows

statistics

a list of results generated by test

adjustment

the p-value adjustment used

models

a character string for which models were compared.

metrics

a character string of performance metrics that were used

or... An object of class "summary.diff.resamples" with elements:

call

the call

table

a list of tables that show the differences and p-values

...or (for compare_models) an object of class htest resulting from t.test.

Details

The ideas and methods here are based on Hothorn et al. (2005) and Eugster et al. (2008).

For each metric, all pair-wise differences are computed and tested to assess if the difference is equal to zero.

When a Bonferroni correction is used, the confidence level is changed from confLevel to 1-((1-confLevel)/p) here p is the number of pair-wise comparisons are being made. For other correction methods, no such change is used.

compare_models is a shorthand function to compare two models using a single metric. It returns the results of t.test on the differences.

References

Hothorn et al. The design and analysis of benchmark experiments. Journal of Computational and Graphical Statistics (2005) vol. 14 (3) pp. 675-699

Eugster et al. Exploratory and inferential analysis of benchmark experiments. Ludwigs-Maximilians-Universitat Munchen, Department of Statistics, Tech. Rep (2008) vol. 30

See also

resamples, dotplot.diff.resamples, densityplot.diff.resamples, bwplot.diff.resamples, levelplot.diff.resamples

Examples

if (FALSE) {
#load(url("http://topepo.github.io/caret/exampleModels.RData"))

resamps <- resamples(list(CART = rpartFit,
                          CondInfTree = ctreeFit,
                          MARS = earthFit))

difs <- diff(resamps)

difs

summary(difs)

compare_models(rpartFit, ctreeFit)
}