Multinomial Model plot in R

Hey there,

I try to plot a multinomial regression. However, something does no work the way I wished...

The formula is: mod.fak_diff_1996.7 <- multinom(factor(fac.num)~ factor(revenu), data=voxit_destill_1996)

fac.num is a nominal variable with 3 values, revenu is a ordinal scale which is a measure for income. As I would like to visualise the probability of being part of fac.num(1) having an income value 1, 2, 3...
I tried with this plot, but there is something wrong/missing...

pred.probs <- predict(mod.fak_diff_1996.7, type = "probs")

plot.probs.1996 <- ggplot(pred.probs, aes(x = revenu, color = level)) + eom_point(aes(y=mean)) + geom_errorbar(aes(ymin = lower, ymax = upper)) + theme_minimal() + ylab("Vorausgesagte Wahrscheinlichkeit") + xlab("Einkommen") + scale_y_continuous(labels = percent)

Would be very greatful if someone could help me :slight_smile:

What is the error message? Can you provide some test data? Is the error due to the typo in eom_point()?


Unfortunately, the typo is just in my message here...
The error message is: data must be a data frame, or other object coercible by fortify(), not a numeric vector Have you any idea what I did wrong here?


Well that suggests that predict only returned a vector (or matrix) of probabilities, and not a data frame. You can visualize the content of pred.probs to check that, and see if you need some additional computation to obtain the variables mean, lower, upper and level that you want to plot. You have to format them as a data.frame and then give them to ggplot().

hm, how would I do that? :thinking:

To display the content of pred.probs, just type it in the console and execute it.

Well then I get a probability for each case and each of my three groups... This is what I need, no? So, I don't get it, why I cannot create the graph...

What does it look like? Which multinom function are you using, the one from nnet? What are the variables level, mean, upper and lower that you want to plot (the mean of what?)?

yes, the one from nnet. It should look like here at the end:

Are you specifically referring to the post Multinomiale Logistische Regression in R from May 16th? Then there is a trap: he doesn't use the predict() function that is provided by nnet, but he uses the function predicts() that is implemented in his package glm.predict. That function seems to compute automatically a bunch of other statistics that are ready for plotting:

test_data <- data.frame(fac.num = sample(LETTERS[1:3], 5, replace=TRUE),
                                 revenu = sample(letters[1:3], 5, replace=TRUE))

mod <- nnet::multinom(factor(fac.num)~ factor(revenu), data=test_data)
predict(mod, type = "probs")
#            A            B            C
# 1  0.9998222 0.0001139447 6.387267e-05
# 2  0.5000312 0.2499988295 2.499700e-01
# 3  0.5000312 0.2499988295 2.499700e-01
# 4  0.2500355 0.5000046104 2.499599e-01
# 5  0.2500355 0.5000046104 2.499599e-01
# 6  0.5000312 0.2499988295 2.499700e-01
# 7  0.5000312 0.2499988295 2.499700e-01
# 8  0.2500355 0.5000046104 2.499599e-01
# 9  0.2500355 0.5000046104 2.499599e-01
# 10 0.9998222 0.0001139447 6.387267e-05
glm.predict::predicts(mod, values = "F")
#        mean        lower     upper factor(revenu level
# 1 0.2968459 1.843498e-71 1.0000000             a     A
# 2 0.3254653 8.661279e-89 1.0000000             a     B
# 3 0.3776888 6.733555e-97 1.0000000             a     C
# 4 0.4433916 9.687225e-02 0.8230188             b     A
# 5 0.2714595 2.324654e-02 0.7715126             b     B
# 6 0.2851489 2.135547e-02 0.7699831             b     C
# 7 0.2261655 3.432994e-02 0.6102105             c     A
# 8 0.4740600 3.614687e-02 0.9152489             c     B
# 9 0.2997745 1.472718e-02 0.8785041             c     C
1 Like

And most importantly:

class(predict(mod, type = "probs"))
# [1] "matrix" "array" 
class(glm.predict::predicts(mod, values = "F"))
# [1] "data.frame"

Yes, I'm confuse because I worked with the glm.predict package and it did not work...

Again, without error messages, test data, and exact commands, I can't tell. It seems to work pretty well on my dummy data.

I'd suggest taking a good look at the documentation of glm.predict to understand the important parameters. If you still can't get it to work, post a new question with a clear description of what you tried (a reprex) and what the error message is.

1 Like

the output from predict contains only the probabilities of being in the various classes, given the input variiable values. You don't have those values because they were in the original dataframe, not in pred.probs.

One thing to try is to use cbind to glue together voxit_destill_1996 and pred.probs (not bind_cols because pred.probs is a matrix rather than a dataframe) and then make your plot, since then you'll have your inputs and your predictions in one place.

Is the plot on page 190 of anything like what you're trying to make? I'm not sure you can do confidence intervals on these very easily.

This topic was automatically closed 21 days after the last reply. New replies are no longer allowed.

If you have a query related to it or one of the replies, start a new topic and refer back with a link.