Graph of interaction between 2 continuous variables in fixed effects panel model

[JesMHill]

Hello Statalist members,

I am running a fixed effects poisson model using xtpoisson looking at the effect of 2 continuous variables and the interaction between these on a count outcome variable.
My code is as follows:

xtpoisson depvar var1 var2 var1Xvar2, fe

The interaction effect is significant (β=0.145, SE=0.051, p=.005) so in order to understand what direction the effect is in I run marginscontplot to picture the interaction using this code:

marginscontplot var1 var2 , var1(20) var2(3) margopts(predict (iru0))


This is the graph that is produced.

My question is what does the scale on the y-axis refer to? As I am looking at within person changes over time it clearly is not a count, but what it is I don't know.

Many thanks for any help,

Jessica

[Maarten Buis]

If you want to use margins and marginsplot you will have to use the factor variable notation to create the interaction terms. So your command should have been xtpoisson depvar c.var1##c.var2, fe. A useful alternative would be to specify the irr option, and interpret the interaction term directly. See for examples here: http://maartenbuis.nl/publications/interactions.html

[daniel klein]

marginscontplot is a user-written command from SJ, as you are asked to explain.

Concerning your syntax, did you really type

xtpoisson depvar var1 var2 var1Xvar2, fe

If you did, I assume you have before typed something like

generate var1Xvar2 = var1 * var2

If this is the case, i.e. you have created the interaction term yourself, then margins will fail to give the correct answer and so will marginscontplot.

Instead type

xtpoisson depvar var1 var2 c.var1#c.var2 ,fe

and see help fvvarlist for more on factor variable notation.


The y-axis should correspond to whatever you predicted, which is the incidence rate and assumes that the fixed-effects are 0 according to the help file entry on the iru0 option for predict after xtpoisson.


Graphical representation of an interaction effect is very useful and I highly recommend it, but the direction of the effect can also be derived from the coefficients alone.

Best
Daniel

[JesMHill]

Thanks very much for the help. I did indeed create my own interaction variable as you showed. Now using the factor variable notation my graph looks totally different and I also understand the y-axis.

Best
Jessica

[daniel klein]

Note that Maarten's suggestion might lead to problems when one of the variables does not vary within panel-units and read more about this purely technical problem.

Best
Daniel