[Eason Xu]

Dear Mr Schechter:
I am using xtnbreg to deal with my data. I have tried the first two options you recommended, but it didn't work. As I am using -xtnbreg,fe-, I want to try the fourth option. But how can I feed the result of - xtpoisson,fe- to the xtnbreg as the starting values, it seems I can't find where to specify the starting values. Thank you!

[Joseph Coveney]

. . . how can I feed the result of - xtpoisson,fe- to the xtnbreg as the starting values, it seems I can't find where to specify the starting values.

Try something like the following.

webuse airacc

xtpoisson i_cnt c.inprog, exposure(pmiles) fe
tempname B
matrix define `B' = e(b)
xtnbreg i_cnt c.inprog, exposure(pmiles) fe from(`B')

[John Mullahy]

Re-upping a posting from a couple months ago: https://www.statalist.org/forums/for...59#post1729059

A common confusion with negative binomial regression models is that "overdispersion relative to a Poisson probability model" means Var(y|x)>E(y|x) not Var(y)>E(y). The sample descriptive statistics will show the latter but not the former. (It may be helpful to recall the decomposition Var(y)=Var_xE(y|x)+E_xVar(y|x).)

In my experience it is often the case that nonconvergence with negative binomial models arises when Var(y|x)<E(y|x), which can happen even when Var(y)>E(y) but which cannot in general be accommodated by a negative binomial specification.

Trying to fit a negative binomial model to conditionally-underdispersed outcomes is a round-peg-square-hole problem. The algorithm shouldn't converge in such instances: That's a feature, not a bug.

Of course I don't know whether this is the issue with your data but it may be worth pondering.