Plot the distribution of predictions for each class

Description

Draw a plot of the distribution of the predicted values for each class. Can be used for earth models, but also for models built by lm, glm, lda, etc.

Usage

plotd(object, hist = FALSE, type = NULL, nresponse = NULL, dichot = FALSE,
      trace = FALSE, xlim = NULL, ylim = NULL, jitter = FALSE, main=NULL,
      xlab = "Predicted Value", ylab = if(hist) "Count" else "Density",
      lty = 1, col = c("gray70", 1, "lightblue", "brown", "pink", 2, 3, 4),
      fill = if(hist) col[1] else 0,
      breaks = "Sturges", labels = FALSE,
      kernel = "gaussian", adjust = 1, zero.line = FALSE,
      legend = TRUE, legend.names = NULL, legend.pos = NULL,
      cex.legend = .8, legend.bg = "white", legend.extra = FALSE,
      vline.col = 0, vline.thresh = .5, vline.lty = 1, vline.lwd = 1,
      err.thresh = vline.thresh, err.col = 0, err.border = 0, err.lwd = 1,
      xaxt = "s", yaxt = "s", xaxis.cex = 1, sd.thresh = 0.01, ...)

Arguments

To start off, look at the arguments object, hist, type.
For predict methods with multiple column responses, see the nresponse argument.
For factor responses with more than two levels, see the dichot argument.

data(etitanic)
earth.mod <- earth(survived ~ ., data=etitanic)
plotd(earth.mod, vline.col=1, err.col=c(2,3,4))
      

Note

This function calls predict with the data originally used to build the model, and with the type specified above. It then separates the predicted values into classes, where the class for each predicted value is determined by the class of the observed response. Finally, it calls density (or hist if hist=TRUE) for each class-specific set of values, and plots the results.

This function estimates distributions with the density and hist functions, and also calls plot.density and plot.histogram. For an overview see Venables and Ripley MASS section 5.6.

Partitioning the response into classes

Considerable effort is made to partition the predicted response into classes in a sensible way. This is not always possible for multiple column responses and the nresponse argument should be used where necessary. The partitioning details depend on the types and numbers of columns in the observed and predicted responses. These in turn depend on the model object and the type argument.

Use the trace argument to see how plotd partitions the response for your model.

Degenerate densities

A message such as
Warning: standard deviation of "male" density is 0, density is degenerate?
means that the density for that class will not be plotted (the legend will say "not plotted").

Set sd.thresh=0 to get rid of this check, but be aware that histograms (and sometimes x axis labels) for degenerate densities will be misleading.

Using plotd for various models

This function is included in the earth package but can also be used with other models.

Example with glm:

      library(earth); data(etitanic)
      glm.model <- glm(sex ~ ., data=etitanic, family=binomial)
      plotd(glm.model)
  

Example with lm:

      library(earth); data(etitanic)
      lm.model <- lm(as.numeric(sex) ~ ., data=etitanic)
      plotd(lm.model)
  

Using plotd with lda or qda

The plotd function has special handling for lda (and qda) objects. For such objects, the type argument can take one of the following values:

"response" (default) linear discriminant
"ld" same as "response"
"class" predicted classes
"posterior" posterior probabilities

Example:

    library(MASS); library(earth); data(etitanic)
    lda.model <- lda(sex ~ ., data=etitanic)
    plotd(lda.model) # linear discriminant by default
    plotd(lda.model, type="class", hist=TRUE, labels=TRUE)
    

This handling of type is handled internally by plotd and type is not passed to predict.lda (type is used merely to select fields in the list returned by predict.lda). The type names can be abbreviated down to a single character.

For objects created with lda.matrix (as opposed to lda.formula), plotd blindly assumes that the grouping argument was the second argument.

plotd does not yet support objects created with lda.data.frame.

For lda responses with more than two factor levels, use the nresponse argument to select a column in the predicted response. Thus with the default type=NULL, (which gets automatically converted by plotd to type="response"), use nresponse=1 to select just the first linear discriminant. The default nresponse=NULL selects all columns, which is typically not what you want for lda models. Example:

    library(MASS); library(earth);
    set.seed(1)      # optional, for reproducibility
    example(lda)     # creates a model called "z"
    plot(z, dimen=1) # invokes plot.lda from the MASS package
    plotd(z, nresponse=1, hist=1) # equivalent using plotd
                                 # nresponse=1 selects first linear discr.
    

The dichot=TRUE argument is also useful for lda responses with more than two factor levels.

TODO

Handle degenerate densities in a more useful way.
Add freq argument for hist.

See Also

density, plot.density
hist, plot.histogram
earth, plot.earth

Examples

if (require(earth)) {
    old.par <- par(no.readonly=TRUE);
    par(mfrow=c(2,2), mar=c(4, 3, 1.7, 0.5), mgp=c(1.6, 0.6, 0), par(cex = 0.8))
    data(etitanic)
    mod <- earth(survived ~ ., data=etitanic, degree=2, glm=list(family=binomial))

    plotd(mod)

    plotd(mod, hist=TRUE, legend.pos=c(.25,220))

    plotd(mod, hist=TRUE, type="class", labels=TRUE, xlab="", xaxis.cex=.8)

    par(old.par)
}