Mlogit not converging after changing base outcome

[Ines Simac]

Hello,

I was wondering whether anyone can explain to me why the function "mlogit" does not want to converge anymore when you change the base-outcome.
Usually, mlogit takes the largest category as the baseline - however, when I choose (a more logical) category as the base, it fails to run the model (becomes not concave after a couple of iterations).

[FernandoRios]

Most likely scenario it’s a mixture of numerical stability and the role of initial values.
however there is no meaningful change in results . So I suspect if you could derive the “new “ coefficients and use them as initial values on the alternative mlogit, no issues of convergence would appear

[Ines Simac]

Thanks for the suggestion.
Indeed, there would not be a meaningful change in the results, but in terms of output I would like the results to be expressed relative to another baseline.
Admittedly, I would not know how to go about extracting these coefficients in stata... I assume it would require some matrix algebra?

[Andrew Musau]

Something like this:

webuse sysdsn1, clear
mlogit insure age male nonwhite i.site
mat b=e(b)
mlogit insure age male nonwhite i.site, base(2) from(b, skip)

If you compare to

mlogit insure age male nonwhite i.site, base(2)

you'll notice that log-likelihood values corresponding to iteration 0 differ. Thus, the command uses initial values for the coefficients from the first regression with the default base. This should help with the convergence problems.

[Xuejuan Zh]

Something like this:

webuse sysdsn1, clear
mlogit insure age male nonwhite i.site
mat b=e(b)
mlogit insure age male nonwhite i.site, base(2) from(b, skip)

If you compare to

mlogit insure age male nonwhite i.site, base(2)

you'll notice that log-likelihood values corresponding to iteration 0 differ. Thus, the command uses initial values for the coefficients from the first regression with the default base. This should help with the convergence problems.

Thank you Andrew, im using Hole (2016) 's mixlogit and mixlogitwtp, for the latter it reports "from" not allowed, do you know another form to incorporate the starting values? Thank you so much!

[Andrew Musau]

Thank you Andrew, im using Hole (2016) 's mixlogit and mixlogitwtp, for the latter it reports "from" not allowed, do you know another form to incorporate the starting values? Thank you so much!

I do not use the command mixlogitwtp (from SSC), so I have no idea how one specifies starting values. As this thread concerns mlogit, I recommend that you start a new thread with the title "How to specify starting values in mixlogitwtp", or something similar, so that those who are familiar with this command may be drawn to your question.

[Xuejuan Zh]

I do not use the command mixlogitwtp (from SSC), so I have no idea how one specifies starting values. As this thread concerns mlogit, I recommend that you start a new thread with the title "How to specify starting values in mixlogitwtp", or something similar, so that those who are familiar with this command may be drawn to your question.

Great thank you!