Last chance! 50% off unlimited learning
Sale ends in
This function calculates a model's calibration performance (reliability) with the Hosmer & Lemeshow goodness-of-fit statistic, which compares predicted probability to observed occurrence frequency at each portion of the probability range.
HLfit(model = NULL, obs = NULL, pred = NULL, bin.method,
n.bins = 10, fixed.bin.size = FALSE, min.bin.size = 15,
min.prob.interval = 0.1, quantile.type = 7, simplif = FALSE,
verbosity = 2, alpha = 0.05, plot = TRUE, plot.values = TRUE,
plot.bin.size = TRUE, xlab = "Predicted probability",
ylab = "Observed prevalence", ...)a model object of class "glm".
a vector of observed presences (1) and absences (0) or another binary response variable. This argument is ignored if model is provided.
a vector with the corresponding predicted probabilities as given e.g. by logistic regression. A warning is emitted if it includes values outside the [0, 1] interval. This argument is ignored if model is provided.
argument to pass to getBins specifying the method for grouping the records into bins within which to compare predicted probability to observed prevalence; type modEvAmethods("getBins") for available options, and see Details for more information.
argument to pass to getBins specifying the number of bins to use if bin.method = n.bins or bin.method = quantiles. The default is 10.
argument to pass to getBins, a logical value indicating whether to force bins to have (approximately) the same size. The default is FALSE.
argument to pass to getBins specifying the minimum number of records in each bin. The default is 15, the minimum required for accurate comparisons within bins (Jovani & Tella 2006, Jimenez-Valverde et al. 2013).
argument to pass to getBins specifying the minimum interval (range) of probability values within each bin. The default is 0.1.
logical, wheter to perform a faster simplified version returning only the basic statistics. The default is FALSE.
integer specifying the amount of messages or warnings to display. Defaults to the maximum implemented; lower numbers (down to 0) decrease the number of messages.
alpha value for confidence intervals if plot = TRUE.
logical, whether to produce a plot of the results. The default is TRUE.
logical, whether to report measure values in the plot. The default is TRUE.
logical, whether to report bin sizes in the plot. The default is TRUE.
label for the x axis.
label for the y axis.
further arguments to pass to the plot function.
HLfit returns a list with the following components:
a data frame of the obtained bins and the values resulting from the hosmer-Lemeshow goodness-of-fit analysis.
the value of the Chi-squared test.
the number of degrees of freedom.
the p-value of the Hosmer-Lemeshow test. Note that this is one of those tests for which higher p-values are better.
the root mean squared error.
Most of the commonly used measures for evaluating model performance focus on the discrimination or the classification capacity, i.e., how well the model is capable of distinguishing or classifying presences and absences (often after the model's continuous predictions of presence probability or alike are converted to binary predictions of presence or absence). However, there is another important facet of model evaluation: calibration or reliability, i.e., the relationship between predicted probability and observed occurrence frequency (Pearce & Ferrier 2000; Jimenez-Valverde et al. 2013). The HLfit function measures model reliability with the Hosmer & Lemeshow goodness-of-fit statistic (Hosmer & Lemeshow 1980).
Note that this statistic has strong limitations and caveats (see e.g. http://www.statisticalhorizons.com/hosmer-lemeshow, Allison 2014), mainly due to the need to group the values into bins within which to compare probability and prevalence, and the strong influence of the binning method on the results. The 'HLfit' function can use several binning methods, which are implemented and roughly explained in the getBins function and can be accessed by typing 'modEvAmethods("getBins")'. You should try 'HLfit' with different binning methods to see how if the results are robust.
Allison P.D. (2014) Measures of Fit for Logistic Regression. SAS Global Forum, Paper 1485
Hosmer D.W. & Lemeshow S. (1980) A goodness-of-fit test for the multiple logistic regression model. Communications in Statistics, A10: 1043-1069
Jimenez-Valverde A., Acevedo P., Barbosa A.M., Lobo J.M. & Real R. (2013) Discrimination capacity in species distribution models depends on the representativeness of the environmental domain. Global Ecology and Biogeography 22: 508-516
Jovani R. & Tella J.L. (2006) Parasite prevalence and sample size: misconceptions and solutions. Trends in Parasitology 22: 214-218
Pearce J. & Ferrier S. (2000) Evaluating the Predictive Performance of Habitat Models Developed using Logistic Regression. Ecological Modeling, 133: 225-245
# NOT RUN {
# load sample models:
data(rotif.mods)
# choose a particular model to play with:
mod <- rotif.mods$models[[1]]
# try HLfit using different binning methods:
HLfit(model = mod, bin.method = "round.prob",
main = "HL GOF with round.prob (n=10)")
HLfit(model = mod, bin.method = "prob.bins",
main = "HL GOF with prob.bins (n=10)")
HLfit(model = mod, bin.method = "size.bins",
main = "HL GOF with size.bins (min size=15)")
HLfit(model = mod, bin.method = "size.bins", min.bin.size = 30,
main = "HL GOF with size.bins min size 30")
HLfit(model = mod, bin.method = "n.bins",
main = "HL GOF with 10 bins")
HLfit(model = mod, bin.method = "n.bins", fixed.bin.size = TRUE,
main = "HL GOF with 10 bins of fixed size")
HLfit(model = mod, bin.method = "n.bins", n.bins = 20,
main = "HL GOF with 20 bins")
HLfit(model = mod, bin.method = "quantiles",
main = "HL GOF with quantile bins (n=10)")
HLfit(model = mod, bin.method = "quantiles", n.bins = 20,
main = "HL GOF with quantile bins (n=20)")
# }
Run the code above in your browser using DataLab