2
$\begingroup$

I have a dataset that has two continuous independent variables, months and tenure. The target variable is binary, 1/0. I want to visualize how the events change across months*tenure. How can I do that? I tried proc g3d in SAS:

proc g3d data=ds;
plot months*tenure=target;
run; 

But the results are very large because there are about million data points. What is the best way to visualize such data interaction?

$\endgroup$
2
$\begingroup$

Interestingly enough, we don't already seem to have a question on visualizing interactions with binary dependent data.

The most interesting aspect of your question is that you have binary data. And a lot of it, so I'd recommend plotting the fits from a suitable model, for instance a logistic regression. Let's generate some toy data and fit a model (I'll use R, because I know it better):

set.seed(1)
nn <- 1000
months <- sample(x=0:60,size=nn,replace=TRUE)
tenure <- sample(x=0:10,size=nn,replace=TRUE)
target <- 1/(1+exp(-0.02*months-0.1*tenure-0.01*months*tenure+rnorm(nn,sd=4)))>.95
summary(target)

model <- glm(target~months*tenure,family="binomial")
summary(model)

Alright, now we need to decide over which range of months and tenure we wish to plot the model fit.

mm <- 0:60
tt <- 0:10

I see two ways of plotting this. Probably best to try both and see which one works best.

The first way would be to plot a 3d surface:

foo <- expand.grid(mm,tt)
colnames(foo) <- c("months","tenure")

fit <- cbind(foo,predict(model,newdata=foo,type="response"))
colnames(fit) <- c("months","tenure","target")

fit.matrix <- outer(mm,tt,
  function(months,tenure)predict(model,
    newdata=data.frame(months=months,tenure=tenure),type="response"))

opar <- par(mai=c(0.5,0.5,0,0))
  persp(x=mm,y=tt,z=fit.matrix,
    xlab="Months",ylab="Tenure",zlab="Fitted target",
    ticktype="detailed",theta=-30)
par(opar)

surface

The second possibility would be to plot the response to one variable, say months, for different values of the other variable. I like setting the other variable to the quartiles in the originally observed data:

probs <- c(.25,.5,.75)
plot(range(mm),c(0,1),type="n",xlab="Months",ylab="Fitted probability of target")
for ( ii in seq_along(probs) ) {
  lines(mm,
    predict(model,newdata=data.frame(months=mm,tenure=quantile(tenure,probs[ii])),type="response"),
    lty=ii)
}
legend("bottomright",lty=seq_along(probs),
  legend=paste("Tenure:",quantile(tenure,probs)))

curves 1

Or of course, the other way around:

plot(range(tt),c(0,1),type="n",xlab="Tenure",ylab="Fitted probability of target")
for ( ii in seq_along(probs) ) {
  lines(tt,
    predict(model,newdata=data.frame(months=quantile(months,probs[ii]),tenure=tt),type="response"),
    lty=ii)
}
legend("bottomright",lty=seq_along(probs),
  legend=paste("Months:",quantile(months,probs)))

curves 2

| cite | improve this answer | |
$\endgroup$
0
$\begingroup$

Ok. here is what worked for me. I grouped the targets/total count at each month*tenure combination and was able to plot it.

| cite | improve this answer | |
$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.