Best way to plot interaction effects from a linear model

Jake picture Jake · Sep 8, 2009 · Viewed 33.7k times · Source

In an effort to help populate the R tag here, I am posting a few questions I have often received from students. I have developed my own answers to these over the years, but perhaps there are better ways floating around that I don't know about.

The question: I just ran a regression with continuous y and x but factor f (where levels(f) produces c("level1","level2"))

 thelm <- lm(y~x*f,data=thedata)

Now I would like to plot the predicted values of y by x broken down by groups defined by f. All of the plots I get are ugly and show too many lines.

My answer: Try the predict() function.

##restrict prediction to the valid data 
##from the model by using thelm$model rather than thedata

 thedata$yhat <- predict(thelm,
      newdata=expand.grid(x=range(thelm$model$x),
                          f=levels(thelm$model$f)))

 plot(yhat~x,data=thethedata,subset=f=="level1")
 lines(yhat~x,data=thedata,subset=f=="level2")

Are there other ideas out there that are (1) easier to understand for a newcomer and/or (2) better from some other perspective?

Answer

Ian Fellows picture Ian Fellows · Sep 8, 2009

The effects package has good ploting methods for visualizing the predicted values of regressions.

thedata<-data.frame(x=rnorm(20),f=rep(c("level1","level2"),10))
thedata$y<-rnorm(20,,3)+thedata$x*(as.numeric(thedata$f)-1)

library(effects)
model.lm <- lm(formula=y ~ x*f,data=thedata)
plot(effect(term="x:f",mod=model.lm,default.levels=20),multiline=TRUE)