A plot that illustrates the effect of categorizing continuous covariates in a logistic regression model
The idea behind this problem is the fact that fitting a categorized continuous variable in a logistic regression model can produce very misleading estimates when compared to fitting the original continuous variable. To illustrate this, a plot was generated that plotted the effects of a specific continuous variable from two separate logistic regresison models: (1) fitting the categorized version of the continuous variable in the model; and (2) keeping the continuous variable, continuous.
Let's first read in the data set (
data.csv
). The data set is a simulated data set that was designed for our first Puzzler --- the data was simulated such that there is a nonlinear interaction between sex and age.
library(Design)
x <- csv.get("data.csv", header=TRUE)
Now let's define the logistic regression model in which the continuous variables are kept continuous (and are fitted with restricted cubic splines).
ff <- lrm(dz ~ sex*rcs(age, 4) + rcs(sysbp, 4), data=x)
anova(ff) # Total D.F. = 10
# sex*age interaction is significant (p < 0.05)
Now let's created categorized versions of our continuous variables in order to fit the second model.
x <- upData(x,
agecut = cut(age,
breaks=quantile(subset(age, dz==0),
probs=c(0, 0.25, 0.50, 0.75, 1), na.rm=TRUE)),
sysbpcut = cut(sysbp,
breaks=quantile(subset(sysbp, dz==0),
probs=c(0, 0.25, 0.50, 0.75, 1), na.rm=TRUE))
)
fcut <- lrm(dz ~ sex*agecut + sysbpcut, data=x)
anova(fcut) # TOTAL D.F. = 10
# sex*age interaction no longer significant (p > 0.05)
Just to see the results of each model, let's plot the predicted effect of age across sex in each model.
dd <- datadist(x)
options (datadist = 'dd')
par(mfrow=c(1,2))
plot(ff, age=NA, sex=NA, conf.int=FALSE)
plot(fcut, age=NA, sex=NA, conf.int=FALSE)
We can use the
Hmisc
package's
print.plot.Design()
function to return what was plotted:
print.plot.Design(plot(ff, age=NA, sex=NA, conf.int=FALSE))
# Adjusted to medians: sysbp = 120
print.plot.Design(plot(fcut, age=NA, sex=NA, conf.int=FALSE))
# Adjusted to most frequent level: sysbpcut=(117, 128]
It is important to make sure that both model are adjusted to the same things. For example, if
fcut
wasn't adjusted to the quantile that spanned 120, we would use:
par(mfrow=c(1,2))
plot(ff, age=NA, sex=NA, conf.int=FALSE)
plot(fcut, age=NA, sex=NA, conf.int=FALSE,
sysbpcut="(117,128]")
Now let's put the two plots together.
plff <- plot(ff, age=NA, sex=NA, conf.int=FALSE)
plfcut <- plot(fcut, age=NA, sex=NA, conf.int=FALSE)
par (mfrow =c(1,1), mar=c(5,5,4,2)+0.1, lwd=2)
# First plot the spline regression results: age across sex
plot(ff, age=NA, sex=NA, conf.int=FALSE,
# ylim = range(c(range(plff$x.xbeta"log odds"),
# range(plfcut$x.xbeta"log odds"))),
ylim = c(-2, 2),
label.curves=FALSE, lty=1, col=c("gray", "black"))
# Now plot the categorical regression results: age across sex
Fstep <- stepfun(x=with(x, quantile(subset(age, dz==0),
prob=c(0.25, 0.50, 0.75), na.rm=TRUE)),
y = subset(plfcut$x.xbeta, sex=="female")$"log odds")
par(new=TRUE, col="gray", lty=2)
lines(Fstep, xlim=c(0, 150))
Mstep <- stepfun(x=with(x, quantile(subset(age, dz==0),
prob=c(0.25, 0.50, 0.75), na.rm=TRUE)),
y = subset(plfcut$x.xbeta, sex=="male")$"log odds")
par(new=TRUE, col="black", lty=2)
lines(Mstep, xlim=c(0, 150))
par(new=TRUE, col="black", lty=1, lwd=1)
legend(x=40, y=2,
# y=max(c(range(plff$x.xbeta"log odds"),
# range(plfcut$x.xbeta"log odds"))),
xjust=0, yjust=1,
legend = c("Female - Spline Regression",
"Male - Spline Regression",
"Female - Categorical Regression",
"Male - Categorical Regression"),
lty=rep(1:2, each=2), col = rep(c("gray", "black"), 2), lwd=2)