# Does the more complex model necessarily have a lower deviance?

I am aiming to compare the deviances of two models, an ordered probit and a multinomial probit, using the likelihood ratio test (obviously using the same data).

However, I systematically obtain a higher deviance for the most complex model, i.e. the multinomial probit. I don't understand this result.

I am using Matlab to perform the analysis, but as I try and find possible mistakes with the code I would also like to know from a statistical viewpoint if such a result if possible, and if it is, why?