Page 1 of 1

mlogit ink with fixed parameters

Posted: Tue Mar 01, 2022 1:52 pm
I have fixed a problem that occurs when you fix a parameter that is part of an mlogit link set. For example, suppose you have 4 states in a multistate analysis, meaning you want to estimate 3 of these transitions and obtain the 4th one by subtraction because the mlogit link forces the sum of the 3 estimated parameters to be <= 1. So suppose you are estimating A to B, A to C, and A to D, so that A to A is obtained by subtraction. Previously if you fixed A to B equal to 0.2, this information was not included in the estimation of A to C and A to D, with these 2 parameters still constrained sum to <= 1. When the fixed value of A to B was included in the sum, you might end up with the sum > 1. This would not usually create a problem because MARK uses a penalized likelihood and usually would still get a valid analysis, although the optimization might hang and not complete without the user providing good initial values. The more robust approach is to scale the mlogit link set to include the information that A to B = 0.2, and force the sum to be <= 0.8. This is now done with the new release.