0
$\begingroup$

I have troubles finding an answer to my question, regarding a multinomial logistic regression model, with a single continuous predictor.

Let's say these are the absolute frequencies in the categories that I am predicting:

  • Category A: 32
  • Category B: 11
  • Category C: 85
  • Category D: 44
  • Category E: 10

I ran a multinomial logistic regression with a single continuous predictor and Category C as the baseline. In the model "Category A vs. Category C", the ODDS-Ratio for my predictor was 1.20. A one point increase in the predictor increases the probability of preferring Category A over Category C by 20%.

Now that's nice! However, how do I calculate the "probability of preferring Category A over Category C"?

Intuitively, I would look at how many chose A instead of C, in the subgroup of people that either chose A or C. Something like this:

n(Category A) / ( n(Category A) + n(Category C) ) - That would be 27,4%.

However, I feel that I might do something wrong when I do not take into account that subjects had to choose between 5 categories. And I can not figure out how I would calculate the probability of choosing A over C in this set of 5 given options.

$\endgroup$
  • $\begingroup$ "multinominal" corrected to "multinomial". $\endgroup$ – Nick Cox Feb 18 at 11:15
1
$\begingroup$

If you want to compute the predicted probability of choosing A among {A, ..., E}, then you need to use the following formula:

P(A) = exp(B'X[A]) / sum_j(exp(B'X[j]))

Where "B" corresponds to the vector of model estimates (In your case it seems to have only 2 elements: The constant + the single predictor) and "X" corresponds to the content of the option (In your case this matrix should have only 2 columns: A column of "1's" for the constant + a column corresponding to the value of the predictor) You will find a more detailed explanation of this formula in most textbooks with a chapter on modelling of discrete data (see https://data.princeton.edu/wws509/notes/c3.pdf )

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.