8
$\begingroup$

I have 2 factors A and B (5×3) and one covariate X in a within subject design. Here's how I specify my overall model:

lme.out = lme(y~ A*B*X, random=~1|Subject, data=mydata)

My interpretation is that I am looking at a graph y~x, where the slope changes due to the covariate, and the line shifts up or down based on the different levels of A and B (changes in intercepts).

What I want to find out is: if I were to fix factor A (take any of the levels), then looking at the lines (y~x), what is the effect of B? Does the levels of B shift the line up or down (intercepts) or does it alter the slope of the line (X).

Should I be running some sort of contrasts analysis? But I am not sure how contrasts work between factors and covariates.

One way I could think of is to take the subsets of data corresponding to different levels of A and create models such as: lme(y~ B+X, random=~1|Subject, data=mydata[which(mydata$A = A1,]). This way I could compare the resulting intercepts and slopes across these models.

Can anyone tell me if what I am doing make sense? Suggestions of any kind would be greatly appreciated!

$\endgroup$
  • 2
    $\begingroup$ Do you just want to make graphs to look at these relationships? I think that's a good idea. You can do it with lattice or ggplot or base graphics in R depending on what you are most comfortable with. $\endgroup$ – Peter Flom Oct 4 '12 at 21:48
  • $\begingroup$ Thanks Peter. I did look at the graphs. Since I have hypotheses regarding how they would behave I am hoping to be able to formally analyse them. $\endgroup$ – Wynn Oct 4 '12 at 23:05
  • 2
    $\begingroup$ Be aware that "My interpretation is that I am looking at a graph y~x, where the slope changes due to the covariate, and the line shifts up or down based on the different levels of A and B (changes in intercepts)." is a little of: the way your fixed effects are set up (ABX), you're estimating a different intercept AND SLOPE IN X for each combination of the levels of A and B. $\endgroup$ – fabians Feb 20 '14 at 14:11
1
$\begingroup$

For visualizing interaction terms, you may look at the sjPlot-package (see examples here).

Your function call would be

sjp.int(fit, type ="eff")

I'm not sure, however, if this meets your needs?

$\endgroup$
0
$\begingroup$

I personally think that if you want to examine the true relationship between Y and the factors in your model after controlling for X you should be looking at the plotted adjusted rather than raw means computed from your favorite model. For the purposes there are R packages such as lsmeans which are quite handy and user-friendly!

$\endgroup$
0
$\begingroup$

You can investigate the VIFs of your model. VIF stands for Variance Inflation Factor and is a way to measure co-linearity.

https://onlinecourses.science.psu.edu/stat501/node/347

There is a vif function in the car package for R.

$\endgroup$

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.