vcd (version 1.4-3)

binreg_plot: Binary Regression Plot

Description

Creates a display of observed and fitted values for a binary regression model with one numeric predictor, conditioned by zero or many co-factors.

Usage

binreg_plot(model, main = NULL, xlab = NULL, ylab = NULL, xlim = NULL, ylim = NULL, pred_var = NULL, pred_range = c("data", "xlim"), group_vars = NULL, base_level = NULL, subset, type = c("response", "link"), conf_level = 0.95, delta = FALSE, pch = NULL, cex = 0.6, jitter_factor = 0.1, lwd = 5, lty = 1, point_size = 0, col_lines = NULL, col_bands = NULL,  legend = TRUE, legend_pos = NULL, legend_inset = c(0, 0.1), legend_vgap = unit(0.5, "lines"), labels = FALSE, labels_pos = c("right", "left"), labels_just = c("left","center"), labels_offset = c(0.01, 0), gp_main = gpar(fontface = "bold", fontsize = 14), gp_legend_frame = gpar(lwd = 1, col = "black"), gp_legend_title = gpar(fontface = "bold"), newpage = TRUE, pop = FALSE, return_grob = FALSE)
grid_abline(a, b, ...)

Arguments

model

a binary regression model fitted with glm.

main

user-specified main title.

xlab

x-axis label. Defaults to the name of the (first) numeric predictor.

ylab

y-axis label. Defaults to the name of the response - within either 'P(...)' or 'logit(...)', depending on the response type.

xlim

Range of the x-axis. Defaults to the range of the numeric predictor.

ylim

Range of the y-axis. Defaults to the unit interval on probability scale or the fitted values range on the link scale, depending on type.

pred_var

character string of length 1 giving the name of the numeric predictor. Defaults to the first one found in the data set.

pred_range

"data", "xlim", or a numeric vector. If "data", the numeric predictor corresponds to the observed values. If "xlim", 100 values are taken from the "xlim" range. A numeric vector will be interpreted as the values to be predicted.

group_vars

optional character string of conditioning variables. Defaults to all factors found in the data set, response excluded. If FALSE, no variables are used for conditioning.

base_level

vector of length one. If the response is a vector, this specifies the base ('no effect') value of the response variable (e.g., "Placebo", 0, FALSE, etc.) and defaults to the first level for factor responses, or 0 for numeric/binary variables. This controls which observations will be plotted on the top or the bottom of the display. If the response is a matrix with success and failure column, this specifies the one to be interpreted as failure (default: 2), either as an integer, or as a string ("success" or "failure"). The proportions of successes will be plotted as observed values.

subset

an optional vector specifying a subset of the data rows. The value is evaluated in the data environment, so expressions can be used to select the data (see examples).

type

either "response" or "link" to select the scale of the fitted values. The y-axis will be adapted accordingly.

conf_level

confidence level used for calculating confidence bands.

delta

logical; indicates whether the delta method should be employed for calculating the limits of the confidence band or not (see details).

pch

character or numeric vector of symbols used for plotting the (possibly conditioned) observed values, recycled as needed.

cex

size of the plot symbols (in lines).

jitter_factor

argument passed to jitter used for the points representing the observed values.

lwd

Line width for the fitted values.

lty

Line type for the fitted values.

point_size

size of points for the fitted values in char units (default: 0, so no points are plotted).

col_lines, col_bands

character vector specifying the colors of the fitted lines and confidence bands, by default chosen with rainbow_hcl. The confidence bands are using alpha blending with alpha = 0.2.

legend

logical; if TRUE (default), a legend is drawn.

legend_pos

numeric vector of length 2, specifying x and y coordinates of the legend, or a character string (e.g., "topleft", "center" etc.). Defaults to "topleft" if the fitted curve's slope is positive, and "topright" else.

legend_inset

numeric vector or length 2 specifying the inset from the legend's x and y coordinates in npc units.

legend_vgap

vertical space between the legend's line entries.

labels

logical; if TRUE, labels corresponding to the factor levels are plotted next to the fitted lines.

labels_pos

either "right" or "left", determining on which side of the fitted lines (start or end) the labels should be placed.

labels_just

character vector of length 2, specifying the relative justification of the labels to their coordinates. See the documentation of the just parameter of grid.text for more details.

labels_offset

numeric vector of length 2, specifying the offset of the labels' coordinates in npc units.

gp_main

object of class "gpar" used for the main title.

gp_legend_frame

object of class "gpar" used for the legend frame.

gp_legend_title

object of class "gpar" used for the legend title.

newpage

logical; if TRUE, the plot is drawn on a new page.

pop

logical; if TRUE, all newly generated viewports are popped after plotting.

return_grob

logical. Should a snapshot of the display be returned as a grid grob?

a

intercept; alternatively, a regression model from which coefficients can be extracted via coef.

b

slope.

...

Further arguments passed to grid.abline.

Value

if return_grob is TRUE, a grob object corresponding to the plot. NULL (invisibly) else.

Details

The primary purpose of binreg_plot() is to visualize observed and fitted values for binary regression models (like the logistic or probit regression model) with one numeric predictor. If one or more categorical predictors are used in the model, the fitted values are conditioned on them, i.e. separate curves are drawn corresponding to the factor level combinations. Thus, it shows a full-model plot, not a conditional plot where several models would be fit to data subsets. The implementation relies on objects returned by glm, as it uses its "terms" and "model" components.

The function tries to determine suitable values for the legend and/or labels, but depending on the data, this might require some tweaking.

By default, the limits of the confidence band are determined for the linear predictor (i.e., on the link scale) and transformed to response scale (if this is the chosen plot type) using the inverse link function. If delta is TRUE, the limits are determined on the response scale. Note that the resulting band using the delta method is symmetric around the fitted mean, but may exceed the unit interval (on the response scale) and will be cut off.

grid_abline() is a simple convenience wrapper for grid.abline with similar behavior than abline in that it extracts coefficients from a regression model, if given instead of the intercept a.

References

Michael Friendly (2000), Visualizing Categorical Data. SAS Institute, Cary, NC.

Examples

## Simple model with no conditioning variables
art.mod0 <- glm(Improved > "None" ~ Age, data = Arthritis, family = binomial)

binreg_plot(art.mod0, "Arthritis Data")
binreg_plot(art.mod0, type = "link") ## logit scale

## one conditioning factor
art.mod1 <- update(art.mod0, . ~ . + Sex)
binreg_plot(art.mod1)
binreg_plot(art.mod1, legend = FALSE, labels = TRUE, xlim = c(20, 80))

## two conditioning factors
art.mod2 <- update(art.mod1, . ~ . + Treatment)
binreg_plot(art.mod2)
binreg_plot(art.mod2, subset = Sex == "Male") ## subsetting

## some tweaking
binreg_plot(art.mod2, gp_legend_frame = gpar(col = NA, fill = "white"), col_bands = NA)
binreg_plot(art.mod2, legend = FALSE, labels = TRUE,
            labels_pos = "left", labels_just = c("left", "top"))

## model with grouped response data
shuttle.mod <- glm(cbind(nFailures, 6 - nFailures) ~ Temperature,
                   data = SpaceShuttle, na.action = na.exclude, family = binomial)
binreg_plot(shuttle.mod, xlim = c(30, 81), pred_range = "xlim",
            ylab = "O-Ring Failure Probability", xlab = "Temperature (F)")