Hurdle estimation, the meaning of lnsigma

[joanna tyrowicz]

Hi,
I am thinking of estimating a hurdle model, but cannot find decent documentation on the meaning of output. First, what is sigma (or lnsigma). The user guide informs rather vaguely "Information about the estimated standard deviation of the error term in the outcome model appears under lnsigma and /sigma.", which simply does not make much sense. Second, how can I test the validity of the linear model relative to the exponential one? And what is exactly exponential about the exponential specification?

I will appreciate all help.
Best,
JT/

[joanna tyrowicz]

And one more thing: for some reason it does not compute standard errors after vce(clustrer varname). The varname is good (works with vce(cluster) for other estimation commands and brings the expected direction of the size in the standard errors).

[Enrique Pinzon (StataCorp)]

Hello Joanna,

In many Stata commands, there is output that refers to sigma and lnsigma. You can think of them as two ways of presenting a parameter, in levels and in natural logarithm form. For churdle these are variance parameters that correspond to the distribution of the error term for the interior solution of the hurdle model. For instance, in the case of churdle linear,the error term comes from a truncated normal distribution and the sigma is the standard deviation parameter for this distribution.

With regard to which model to select, one alternative is to type estat ic after estimation and see which model gives you the smaller information criteria. Note, however, that this is not my field of expertise and perhaps someone else in the forum might have a better suggestion regarding model selection which is an active field research.

With regard to the exponential mean, it means that the conditional mean of the interior part of the hurdle model has an exponential mean, i.e. exp(xbeta + sigma^2/2).

Finally, with regard to the vce(cluster varname) could you please give me more information. Is the command exiting with an error message? Could you please tell me exactly what is going on.

For a nice discussion of hurdle models look at chapter 17 of Wooldridge 2010 Econometric Analysis of Cross Section and Panel Data, in particular section 17.6.1.

[joanna tyrowicz]

Dear Enrique,
thank you for your reply and explanations. It is not computing any standard errors after the vce(cluster varname) command. I paste the output below. Do you think it is a bug that can be fixed?
JT/

. churdle lin share cond_others, ll(0) vce(cluster sub_ID) select(cond_others $demo)

Iteration 0: log pseudolikelihood = -2782.5142
Iteration 1: log pseudolikelihood = -2715.2666
Iteration 2: log pseudolikelihood = -2712.4692
Iteration 3: log pseudolikelihood = -2712.4305
Iteration 4: log pseudolikelihood = -2712.4305

Cragg hurdle regression Number of obs = 1,272
Wald chi2(1) = .
Prob > chi2 = .
Log pseudolikelihood = -2712.4305 Pseudo R2 = 0.0338

(Std. Err. adjusted for 212 clusters in sub_ID)

Robust
share Coef. Std. Err. z P>z [95% Conf. Interval]

share
cond_others .5641421 . . . . .
_cons -5.164374 . . . . .

selection_ll
cond_others .0440762 . . . . .
_cons -.1918973 . . . . .

lnsigma
_cons 1.68583 . . . . .

/sigma 5.396928 . . .

[joanna tyrowicz]

Dear Enrique,
can you help on the lack standard errors?
Best,
JT/

[Enrique Pinzon (StataCorp)]

Hello Joanna,

Sorry for the late reply. I was out of the office for a couple of weeks. Do you mind emailing [email protected] with your data and do-file for me to take a closer look at the output. Thanks.

[joanna tyrowicz]

Sorry, I should have written, your advice helped! This must be some kind of a bug, because in other comands the cluster var may be a string.