How can I Graph the Logistic Regression Plot with Interaction Terms in Stata?

[smith Jason]

Hi, There,
I have a online dataset at
use http://ssc.wisc.edu/sscc/pubs/files/grad.dta,clear

The logit model and marginal effect with interaction term in Stata is
logit grad treat##highSES, or
margins highSES#treat

I don't know how to graph the logistic regression plot with interaction terms in Stata below.


Thank you for your help!

[ericmelse]

Possibly this might be helpful: The Stata Journal, 2005, 5 (1), 64-82, M.N. Mitchell & X. Chen, Visualizing main effects and interactions for binary logit models.

Note that to install the packages discussed in this paper you should type in the command window:

findit viblmcc

[ericmelse]

And here another page with further explanation of this topic.

[Roman Mostazir]

You should be able to see the plot for probability with -marginsplot- command followed by your commands already in your code:

logit grad treat##highSES, or
margins highSES#treat
marginsplot

[smith Jason]

You should be able to see the plot for probability with -marginsplot- command followed by your commands already in your code:

logit grad treat##highSES, or
margins highSES#treat
marginsplot

Thank you!

However, the code you provided can not generate the graph I want.

[smith Jason]

Possibly this might be helpful: The Stata Journal, 2005, 5 (1), 64-82, M.N. Mitchell & X. Chen, Visualizing main effects and interactions for binary logit models.

Note that to install the packages discussed in this paper you should type in the command window:

findit viblmcc

Thank you. However, the code in the resources you provided doesn't work. postgr3 is deprecated for many years.

[smith Jason]

Can someone help me about my question above?
Thank you!

[Weiwen Ng]

In the original post, the author drew a logistic cumulative distribution function using twoway function and the logistic CDF. The x-axis is basically XB from the regression. Then, they're drawing line segments. You're basically drawing from y = 0, x = XB (i.e. the XB for whatever combo you're drawing) to y = marginal probability, x = XB.

You can keep searching for an implementation of this type of margins graph superimposed on the logistic CDF. The thing is, you can already see the marginal effects on the probability scale from marginsplot. What's the marginal utility (in the econ sense) of the plot on the logistic curve? I am not sure it's enormous. Actually, if I were teaching a class on logistic regression, it could be useful.

[smith Jason]

Could you please provide the Stata code to create the plot in the 1st floor?
Thank you!

[Weiwen Ng]

Could you please provide the Stata code to create the plot in the 1st floor?
Thank you!

I believe you actually want to run this code on your margins command to get the XBs. The second line will post the results to memory and give you the symbolic names of each combo.

margins highSES#treat, predict(xb)
margins highSES#treat, predict(xb) post coeflegend

This code plots the logistic curve.

twoway function y = logistic(x), range(-5 5) xtitle("XB") ytitle("Probability")

Then, for each line segment, you'd add the line coordinates. The format is y1 x1 y2 x2.

y1 = 0
x1 = XB
y2 = logistic(XB)
x2 = XB

To explain the above, look at the graph in your first post. It may not be obvious, but everything on that graph has X and Y coordinates. What's the scale? Well, the x-axis runs along the line y = 0. The scale is defined here by the values that the plot takes, and you plotted the logistic function, so inherently the y-axis is going to run from 0 to 1. So, you're typing something like:

twoway function y = logistic(x), range(-5 5) || pci 0 0 0.5 0, label("something") || pci . . . . , label("...") || ... || , xtitle("XB") ytitle("Probability")

You could obviously do this manually. You'd be better off using the symbolic names of each combination in margins to do the calculations. However, it seems like the pci command may only take actual numbers as inputs, i.e. you can't feed it a scalar or a variable as an input. Or at least, I haven't been able to do that.

Anything more detailed than that, and I think unfortunately you will have to ask others for help - remembering that people here aren't obliged to help, we all have our own jobs/dissertations/etc.

[smith Jason]

Thank you!

[smith Jason]

Can anybody help?
Thank you very much!

[Weiwen Ng]

I believe I've given you the individual pieces you need to solve the puzzle, at least manually. I've shown how to plot a logistic curve. I've shown you how to manually plot each line segments (although note that Stata actually applies any point label to the origin, not the destination, so you should reverse the pairs of coordinates I gave you). I've shown how to derive the start and end coordinates for each line segment.

If you don't understand individual pieces, can you say which one is the sticking point?

If you are asking for an automated solution to prepackage all of this, then remember what I said at the end of #10: we all have our own jobs, so it's not a guarantee that someone finds this style of presentation compelling enough to cook up an automated solution.