# Predict function - multinomial logistic regression in R [closed]

I'm new to using R. I'm attempting to create a microsimulation of individuals health through time. To do this, I have two survey datasets with the same variables. First, a large base file, second a smaller but more detailed health transition dataset. The outcome variable is self-reported health with three states (1 - good, 2 - fair, 3 - poor), the predictors - Age (continuous), health at time t-1, marital status, highest educational qualification, housing tenure and socio-economic social group.

I have conducted a multinomial logistic regression (test) on the second dataset and now wish to use the predict function to apply this to the larger, fist base dataset. In an ideal world, this will be in the form of predicted category probabilities, that I can then generate random numbers (0,1), and assign new health states.

Currently the best I can come up with is:

test <- multinom(health5 ~ ContAge1 + health4 + marstat1 + highqual1 + tenure1 + socstat1, data = EW5FDR)

newpred <- predict(test, newdata = base, type = "c")


This appears to give me predicted outcome category for the new dataset, my question is: how would I change this to give me predicted category probabilities?

And indeed, is this the correct function to be using in the first place?

• Please explain how it "isn't working" and what kind of advice you are seeking.
– whuber
Commented Oct 6, 2015 at 13:58
• Your response categories are ordered. You should be using ordinal logistic regression, not multinomial LR. Then use the model formula to get the predicted probabilities. You can see an example in my question here: How do you predict a response category given an ordinal logistic regression model? Commented Oct 6, 2015 at 15:06
• I unfortunately can't assume that the categories have equal separation, in that they combine other categories into a compressed measure. I imagine that this would violate the proportional odds assumption? Would this be correct? Thanks for the link though, I'll take a look. Commented Oct 6, 2015 at 15:12
• @Microsim: The proportional odds assumption is checkable, though with only three categories some would prefer to go straight to fitting a multinomial model on a large data-set. BTW type="probs". Commented Oct 6, 2015 at 15:34
• The fact that the categories are not equal interval does not necessarily violate the proportional odds assumption; those are separate issues. You can also fit ordinal logistic regression models without the proportional odds assumption, I believe. Commented Oct 6, 2015 at 19:21