Visualizing two regression in one graph

[Mohammed Kasbar]

Dear Statalist respected users,

I ran two regression models and I want to visualize the impact of 2 different interactions (one from each model) on an outcome variable, which is the same on the two models.
For example,
the first model: y = B0 + B1X1 + B2X2 + B3X2*X3 + e

The second model: y= B0 + S1X1 + S2X2 + S3X2*X3 + e

I want to visualize the impact of B3 and S3 on y in one graph.

Thanks a lot in advance.

Kind regards,
Mohammed Kasbar

[Nick Cox]

Tell us more about X1 X2 X3.

What do you imagine would be on the y and x axes of such a graph?

You can plot in a two-dimensional space

(1) observed vs fitted, always

(2) residual vs fitted, always

(3) data and fit for response versus one predictor, if the other predictors are just shifting a fit in that space.

These are evident possibilities, not a complete list.

[Mohammed Kasbar]

Nick Cox

thanks a lot for your reply.
Actually, I am not sure which option of the three which you gave me best fits my case.

I estimated 2 models and I found that the interaction between X2 and X3 in the first model was positive while it was negative in the second model. Please note that X2 is different in the 2 models while X3 is exactly the same. The outcome variable is also the same in the 2 models.

So, I want to show how the change in X2 can change the relationship between X3 and Y.


I am not sure which option of the 3 options you gave me best fit my case.

I wish i well well explained your question.

Thanks a a lot in advance.

Kind regards,
mohammed Kasbar

[Nick Cox]

Thanks for the detail, but we're still short of knowing whether at least two of X1 X2 X3 are indicators [(0, 1) variables] which I think is necessary and sufficient for (c) in #2 to apply.

Superimposing two sets of results is probably going to produce a mess any way. I would put either (a) or (b) side-by-side for the two regressions, and also consider added variable plots.

[Mohammed Kasbar]

Nick Cox

thanks a lot for your reply. Yes, you are right; both X2 and X3 are dichotomous variables.
So, I appreciate if you provide me with more details about the C option you provided in #2.

thanks a lot in advance.

Kind regards,
Mohammed Kasbar

[Nick Cox]

You can plot Y versus X1 and superimpose the predicted values from the regression. That won't be (can't be) a complete rendering of the data.

[Mohammed Kasbar]

Thanks a lot.
This is much appreciated

[Nick Cox]

What I said in #6 was a little too pessimistic as the four distinct values of two indicators can be shown by four distinct markers. This shows some flavour for a regression simpler than yours -- and I would not want to do more than put the graph for another regression alongside.

The veryshortlabel optionof separate is documented at

SJ-5-4 gr0023 . . . . Stata tip 27: Classifying data points on scatter plots
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . N. J. Cox
Q4/05 SJ 5(4):604--606 (no commands)
tips for using separate, gray-scale gradation, and text
characters as class symbols to classify data points on
a scatter plot

namely http://www.stata-journal.com/sjpdf.h...iclenum=gr0023

sysuse auto, clear
set scheme s1color 

gen recmpg = 1000/mpg 

su length, meanonly 
gen hilength = length > r(mean) 

regress recmpg weight i.foreign i.hilength 
predict predicted 

egen group = group(foreign hilength), label 

separate recmpg,  by(group) veryshortlabel 
local response "`r(varlist)'" 

separate predicted, by(group) veryshortlabel 
local predictions "`r(varlist)'" 

scatter `response' weight , ms(Oh + Sh X) mc(orange blue black magenta) ///
|| line `predictions' weight, sort lc(orange blue black magenta) ///
legend(order(1 2 3 4) col(1) pos(3)) ytitle(Gallons per 1000 miles)

[Mohammed Kasbar]

Nick Cox

Dear Nick,

Thanks a lot for your informative reply. I do apologize for replying late but I did not check my page for 2 days as I was busy teaching.
Thanks a lot again for your priceless efforts.

Kind regards,
Mohammed Kasbar