visualization of interaction effects: using predicted values?

Question

I am currently working on a study in which I am using beta regression (in Stata). My depend variable is bounded between 0 and 1. To interpret the effects of my independent variables I am using the margins command.

However, the problem is that interaction-terms cannot be interpreted using the margins command. My model includes a quadratic term, i.e.: y=a+bX+cX^2. I want to visualize the non-linear connection between the variables by using graphs.

My question might be a bit simple, but I am not sure which values I should choose to visualize the effect. Should I use the original depend variable or the predict values?

Thanks in advance for the help.

I think that in ANOVA-type interaction diagrams the estimated marginal means would be plotted. These are equivalent to the means of the predicted (conditional) means. Not sure if this helps. — tomka, Nov 30 '13 at 14:26
It is certainly possible, & helpful, to plot interactions. My answer here: how-to-visualize-multi-linear-regression may give you some hints to get started. — gung - Reinstate Monica, Nov 30 '13 at 14:40
Thanks for all the answers. From what I have read, I think that plotting the original variables is the right choice. — , Dec 01 '13 at 21:21

score 2 · Answer 1 · answered Nov 30 '13 at 14:42

2

You could have a scatter plot with the data (dependent variable on the Y axis, one independent variable on the X axis) then add lines for different levels of the other independent variable. The lines would be of the predicted dependent variable.

answered Nov 30 '13 at 14:42

Peter Flom

119,535
36
175
383

score 0 · Answer 2 · answered Dec 01 '13 at 23:55

In R you would need to use orthogonal polynomials rather than just regressing against X and X^2 if you wanted to get any valid inference of the impact of adding the X^2 term. I'm not a Stata user, but I read that this command would be effective:

.orthpoly lvl, generate(cntr1 cntr2) degree(2)

You would also want to plot the data (x vs Y as points) and then overlay a regression line for X versus predicted-Y.

visualization of interaction effects: using predicted values?

2 Answers2