# Re: st: hetereoskedasticity/serial correlation with RE xtprobit xttobit?

 From "Arne Risa Hole" To statalist@hsphsun2.harvard.edu Subject Re: st: hetereoskedasticity/serial correlation with RE xtprobit xttobit? Date Fri, 6 Jul 2007 20:34:19 +0100

```Thanks Austin for mentioning my -clogithet- module and paper. I'm
afraid, however, that -clogithet- does not estimate a fixed effects
logit with heteroscedasticity but McFadden's choice model with het -
although Stata's -clogit- command can estimate homoscedastic versions
of both of these -clogithet- only does the latter unfortunately.

What you could do is estimate a hetersocedastic binary logit model
with het using -clogithet- with cluster-robust standard errors, along
the lines of AttaUllah Shah's suggestion. You can also use -hetprob-
in this manner if you prefer a probit model.

Arne

On 06/07/07, Austin Nichols <austinnichols@gmail.com> wrote:
```
```Jessica <jholmes@middlebury.edu> :
I don't know about the tests for het or clustering after -xtprobit-
and -xttobit- (perhaps someone else on the list can address these).
-ssc install clogithet- gives you a test for heteroskedasticity in a
fixed-effects logit model, described in

Hole, A.R., 2006. Small-sample properties of tests for
heteroscedasticity in the conditional logit model. Economics Bulletin
3, 1-14.
http://economicsbulletin.vanderbilt.edu/2006/volume3/EB-06C20063A.pdf

However, it seems to me your reviewer missed the main problem, that
-xttobit- is the wrong model. You seem to be estimating ln(real y + 1)
as a function of X with a lower limit of 1. But ln(1)=0 --how many
left-censored observations does Stata report? Is it the same as the
number of obs where y=0?.  Or perhaps you are estimating ln(real y +
exp(1)) as a function of X with a lower limit of 1, where the lower
limit makes more sense for a depvar y>=0.  This still  assumes that
the values observed at zero really should be mostly negative, and the
underlying distribution is normal, but this is unlikely to be the
case.  What you really want is a GLM model or -poisson- type model,
right? See e.g.
http://www.stata.com/statalist/archive/2007-04/msg00549.html
http://www.stata.com/statalist/archive/2006-12/msg00466.html
I would recommend you replace -xttobit- using lnreal1 as the depvar
with -xtpoisson- using real as the depvar.

In general, adding some number to y, taking the log, and running a
-tobit- type model is unjustifiable.  For one, your estimates will
differ depending on what number you specify, and there is rarely a
theoretical justification for one number over another.

On a more substnantive point, are there no fixed time effects in your
model?  To the extent that marginal tax rates and the deductability of
giving varies over time, you would at least want to include year
effects, I think, even if you can't estimate first-dollar marginal tax
rates for individuals (to get a plausibly exogenous tax-price of
giving).

On 7/6/07, Holmes, Jessica <jholmes@middlebury.edu> wrote:
> Hope someone can help! I am using a panel data set that looks at
> charitable giving for 22,000 individuals over 15 years.
>
> I use a random effects Probit to predict whether an individual donates
> in a given year and a random effects Tobit to predict how much he/she
> gives each year:
>
> xtprobit donate male yrsgrad yrsgradsq mar alumnicloserelative
> med_income artsent bankfin commedia comptech consulting education
> environmental govtpp healthcaremed intlang law nonprofitsocialservices
> profservicesbus salesmarketing  ownbusiness mileslt250 reunion soc
> affinity acad devug arts campus sports otheract socsci natsci art
> nongrad loans_z grants_z  finaid_miss djia_pctch campaignyr hock_pc rank
> hock_pc_male hock_pc_yrsgrad hock_pc_sports rank_male rank_yrsgrad if
> stillug==0, i(pidm)
>
> xttobit lnreal1 male yrsgrad yrsgradsq mar alumnicloserelative
> med_income artsent bankfin commedia comptech consulting education
> environmental govtpp healthcaremed intlang law nonprofitsocialservices
> profservicesbus salesmarketing  ownbusiness mileslt250 reunion soc
> affinity acad devug arts campus sports otheract socsci natsci art
> nongrad loans_z grants_z  finaid_miss djia_pctch campaignyr hock_pc rank
> hock_pc_male hock_pc_yrsgrad hock_pc_sports rank_male rank_yrsgrad if
> stillug==0, ll(1) i(pidm)
>
>
> I have been asked by a referee to
>
> 1) test for hetereoskedasticity and correct if necessary
>
> and
>
> 2) test for serial correlaton and correct if necessary.
>
> Stata does not seem to have any commands to detect and/or correct for
> hetereoskedasticity and serial correlation for xtprobit or xttobit. Any
> suggestions would be most appreciated...
>
> Thanks,
>
> Jessica
*
*   For searches and help try:
*   http://www.stata.com/support/faqs/res/findit.html
*   http://www.stata.com/support/statalist/faq
*   http://www.ats.ucla.edu/stat/stata/

```
```*
*   For searches and help try:
*   http://www.stata.com/support/faqs/res/findit.html
*   http://www.stata.com/support/statalist/faq
*   http://www.ats.ucla.edu/stat/stata/
```