ROC {epiDisplay}

ROC curve

Description

Receiver Operating Characteristic curve of a logistic regression model and a diagnostic table

Usage

lroc(logistic.model, graph = TRUE, add = FALSE, title = FALSE, 
    line.col = "red", auc.coords = NULL, grid = TRUE, grid.col = "blue", ...)
roc.from.table(table, graph = TRUE, add = FALSE, title = FALSE, 
	line.col = "red", auc.coords = NULL, grid = TRUE, grid.col = "blue", ...) 

Arguments

logistic.model

A model from logistic regression

table

A cross tabulation of the levels of a test (rows) vs a gold standard positive and negative (columns)

graph

Draw ROC curve

add

Whether the line is drawn on the existing ROC curve

title

If true, the model will be displayed as main title

line.col

Color of the line

auc.coords

Coordinates for label of 'auc' (area under curve)

grid

Whether the grid should be drawn

grid.col

Grid colour, if drawn

...

Additional graphic parameters

Details

'lroc' graphs the ROC curve of a logistic regression model. If ‘table=TRUE’, the diagnostic table based on the regression will be printed out.

'roc.from.table' computes the change of sensitivity and specificity of each cut point and uses these for drawing the ROC curve.

In both cases, the area under the curve is computed.

Author(s)

Virasakdi Chongsuvivatwong [e-mail]

See Also

'glm'

Examples

# Single ROC curve from logistic regression
# Note that 'induced' and 'spontaneous' are both originally continuous variables
model1 <- glm(case ~ induced + spontaneous, data=infert, family=binomial)
logistic.display(model1)
# Having two spontaneous abortions is quite close to being infertile!
# This is actually not a causal relationship

lroc(model1, title=TRUE, auc.coords=c(.5,.1))
# For PowerPoint presentation, the graphic elements should be enhanced as followed 
lroc(model1, title=TRUE, cex.main=2, cex.lab=1.5, col.lab="blue", cex.axis=1.3, 
lwd=3)
lroc1 <- lroc(model1) # The main title and auc text have disappeared
model2 <- glm(case ~ spontaneous, data=infert, family=binomial)
logistic.display(model2)
lroc2 <- lroc(model2, add=TRUE, line.col="brown", lty=2)
legend("bottomright",legend=c(lroc1$model.description, lroc2$model.description),
        lty=1:2, col=c("red","brown"), bg="white")
title(main="Comparison of two logistic regression models")
lrtest(model1, model2) 
# Number of induced abortions is associated with increased risk for infertility

# Various form of logistic regression
# Case by case data
data(ANCdata)
.data <- ANCdata
glm1 <- glm(death ~ anc + clinic, binomial, data=.data) # Note 'calc'
lroc(glm1)

# Frequency format
data(ANCtable)
ANCtable
.data <- ANCtable
attach(.data)
death <- factor (death)
levels (death) <- c("no","yes")
anc <- with(.data, factor (anc))
levels (anc) <- c("old","new")
clinic <- with(.data, factor (clinic))
levels (clinic) <- c("A","B")
.data <- data.frame(death, anc, clinic)
.data
glm2 <- glm(death ~ anc + clinic, binomial, weights=Freq, data=.data)
lroc(glm2)
detach(.data)

# ROC from a diagnostic table
table1 <- as.table(cbind(c(1,27,56,15,1),c(0,0,10,69,21)))
colnames(table1) <- c("Non-diseased", "Diseased")
rownames(table1) <- c("15-29","30-44","45-59","60-89","90+")
table1
roc.from.table(table1)
roc.from.table(table1, title=TRUE, auc.coords=c(.4,.1), cex=1.2)

# Application of the returned list
roc1 <- roc.from.table(table1, graph=FALSE)
cut.points <- rownames(roc1$diagnostic.table)
text(x=roc1$diagnostic.table[,1], y=roc1$diagnostic.table[,2], 
	labels=cut.points, cex=1.2, col="brown")
rm(list=ls())