The Hessian of the log likelihood function is

$$\frac{\partial^2 \ln(\beta \mid x)}{\partial \beta \partial \beta'} = -\sum_{i=1}^n \underbrace{\Lambda(\beta'x_i)}_{\in(0,1)}\underbrace{\left[1-\Lambda(\beta'x_i)\right]}_{\in(0,1)}\underbrace{x_ix_i'}_{\geq 0}$$

where I have marked my understanding of the ranges of the different factors with underbraces. The Hessian would be zero if $x_i=\mathbf{0}$ for all $i$. Thus, I would conclude that the Hessian was negative semi-definite.

Yet in Greene (p. 691-692)---which specify the Hessian just as I do---it says

Note that the Hessian is always negative definite, so the log-likelihood is globally concave.

What am I missing here?


When they say the Hessian is always negative definite, they are assuming that $X$ is a full rank matrix, which is a very typical assumption for regression models.

So it is true that in your example, the Hessian would not be negative definite. However, they assume that you have informative data: you can't expect to estimate the effect of $x$ if you don't observe different levels of $x$.

| cite | improve this answer | |
  • $\begingroup$ I'll also note that statisticians tend to be a little lazier about stating such things than mathematicians! $\endgroup$ – Cliff AB May 27 '15 at 3:35

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.