mimrgns and marginsplot; plotting exponentiated predictions?

[Kira Riehm]

Hello, I am currently trying to make a graph using marginsplot after the mimgrns command. I am having trouble getting marginsplot to use the exponentiated predictions on the y-axis. My code is as follows:

mi estimate, eform: ///
    logit ALC_ c.MED_ c.ADHD_ c.DEP_ AGE i.WAVE i.SEX i.ETHNIC i.FAMHIST_ALC i.FAMHIST_DRUG i.WAVE i.LUNCH, or vce(cluster SCHOOL)
mimrgns, at(MED_=(0(1)42)) atmeans eform cmdmargins
marginsplot, ///
    recast(line) ///
    recastci(rarea) ///
    ciopts(fintensity(30) ///
    lwidth(none)) ///
    yla(-4(1)4, ang(h)) ///
    xtitle("Social Media Score", size(small)) ///
    ytitle("Probability of Initiating Alcohol Use", size(small)) ///
    title("Probability of Initiating Alcohol Use, by Social Media Score") ///
    subtitle("Overall Sample") ///
    saving(GRAPH_overall, replace)
graph export GRAPH_overall.pdf, replace

I've attached a PDF of the graph I get. It looks good, except that I want the y-axis to be predicted probabilities, and right now I think these are on the log scale. I tried adding the eform option to mimrgns but marginsplot afterward still plots the un-exponentiated values. Is there anyway around this, either with another command, or some way of exporting these values to a separate dataset to plot independently of marginsplot? Thank you in advance for any help that can be offered.

(Also, I have read the online help file for mimrgns and understand that the CIs plotted with marginsplot are not correct.)

[daniel klein]

mimrgns is community-contributed software from SSC.

I find it a bit hard to figure out what exactly you want. You say

I am having trouble getting marginsplot to use the exponentiated predictions

but then

I want the y-axis to be predicted probabilities

Those are not the same thing! Exponentiated log-odds are odds (ratios), not probabilities. If you want the latter, you will have to explicitly request that, As the help file for mimrgns explains, it defaults to predicting xb. You may overwrite that

mimrgns ... , predict(pr)

Best
Daniel

[Kira Riehm]

Thanks for your help Daniel - this clarifies multiple issues on my end, and I got the graph to work!

Best,

Kira

[Luke Aldridge]

Hi daniel klein or others,

I'm interested in Dr. Riehm's original question, which is how to have marginsplot plot exponentiated predicted margins after mimrgns. I'm working with negative binomial regression and log-linear regression, both of which I would like to plot in exponentiated form. The option eform works with mimrgns to display exponentiated values (as does expression(exp(predict(xb))), however marginsplot still plots the non-transformed predicted values.

My work around has been to exponentiate the values and plot in R, though would appreciate a more direct route.

Thanks,
Luke

[daniel klein]

The option eform works with mimrgns to display exponentiated values

Yes, it does; the eform option, however, only affects how coefficients are displayed in the table; it is probably of little use here.

[...] as does expression(exp(predict(xb))), however marginsplot still plots the non-transformed predicted values.

Can you provide a reproducible example? Specifying expression() with mimrgns clearly produces a different graph from the default predictions in the following example:

// example dataset
webuse mheart1s20

mi estimate : logit attack bmi i.female

// marginsplot with non-transfomed prediction
mimrgns , cmdmargins
marginsplot

// now with transfomred predictions
mimrgns , expression(exp(predict(xb))) cmdmargins
marginsplot

Whether that is what you want, is a different question.