[R] multinomial probabilities with mlogit

[Ingrid Charvet]

Hello,

When fitting a logit multinomial model with "mlogit" I can retrieve the response probabilities using
fit$fitted.values (for a given object "fit")

However, I am trying to calculate those response probabilities myself using the maximum likelihood estimates (i.e. fit$coefficients) given by mlogit.

I have used the model given in Agresti (2002):

Prob_j(x) = exp( linearpredictor_j(x) ) / (1 + sum (linearpredictor(x)))

Which is for a category j the exponential of the linear predictor for category j divided by 1 + the sum of all logits across categories, aside from the reference category.

But I cannot get my fitted probabilities calculated using this equation to match the output of mlogit fit$fitted values.

Can anyone tell me how those fitted values are calculated? Or point me to the corresponding documentation (which I cannot seem to find by googling!)

Many thanks
Ingrid

	[[alternative HTML version deleted]]