Skip to content

Plotting interactions while controlling for other predictors.

4 messages · David Sidhu, Walid Mawass, Guillaume Adeux

Original
David Sidhu
# 
I am interested in creating a plot of an interaction that accounts for other variables present in a model.

I am attaching a reproducible example of my try so far, but I don?t believe this is quite getting at what I?d like.

Essentially, I want to show the interaction between X2 and X3, while controlling for X1.

Thank you very much.
Dave

set.seed(182)
require(data.table)
require(lme4)
require(ggplot2)

X1 = rnorm(1000)
X2 = rnorm(1000)
X3 = sample(1:3, size = 1000, replace = TRUE)
X3 <- as.factor(X3)
DV = rnorm(1000)

Item <- rep(1:10, times = 100)
Subject <- rep(1:100, each = 10)

d <- as.data.frame(cbind(X1, X2, X3, DV, Item, Subject))

m1 <- lmer(DV ~ X1 + X2*X3 + (1|Subject) + (1|Item), data = d)

d$Pred <- predict(m1)
library(ggplot2)
ggplot(d, aes(x = X2, y = Pred, col = factor(X3), group = X3)) + geom_jitter(alpha = .2) + geom_smooth(method = "lm")


---
David M. Sidhu, MSc
PhD Candidate
Department of Psychology
University of Calgary
Walid Mawass
# 
Hi,

The package 'effects' is a good tool for this sort of thing. It let's you
plot the effect of the interaction term from a regression model while
controlling for other predictors by using the mean value for each other
variable by default while plotting the predictions for the interaction term
based on the model.

Hope this helps.
Guillaume Adeux
# 
Hi David,

For plotting interactions between two continuous variables, maybe a gamm
model might be more adapted?

However, by defaut, in the visreg package, you can facet one of your
continuous variables (which is subsetted relative to its distribution) and
plot the other one accordingly.

This would be coded the following way: visreg(m1, X2, by="X3")
#overlay=TRUE to have them all combined in one plot, partial=TRUE to have
the partial residuals (which do not include RE by default), by default X1
will be fixed to its median

This will produce a conditional plot in which all random effects (or X1 in
this case) are fixed at a certain level (you can also fix this at a
particular value with cond=list(X1=XXXX) ).

Nevertheless, this simply calls for predict.lmer(). It is still pretty
convenient IMHO.

Hope this helps.

Sincerely,

GA2

Le mer. 12 d?c. 2018 ? 17:51, David Sidhu <dsidhu at ucalgary.ca> a ?crit :

  
  


  


      

    

    

      
      

        


  

        

      Guillaume Adeux
    

    

      
      #
    

  


  

      
Sorry I hadn't seen X3 was a factor.

The above example will then work how you want it to (one facet per level of
X3) highlithing the effect of X2 on Y (per level of X3) while controlling
for X1 and random effects (set to a particular level).

This will be similar (not identical considering visreg fixes random effects
at a particular level) to creating your own prediction matrix with
new.dat=expand.grid(X1=median(X1),X2=seq(min(X2),max(X2),by=1),X3=c(1:3))
and feeding it to predict.merMod while predicting at the population level
(re.form=~0).

How partial residuals can be extracted however, I am not so sure.

GA2

Le mer. 12 d?c. 2018 ? 18:06, Guillaume Adeux <guillaumesimon.a2 at gmail.com>
a ?crit :