[Date Prev][Date Next][Thread Prev][Thread Next][Date index][Thread index]
Re: st: ordered probit and panels
That's certainly correct. -gllamm- cannot take the time dimension into
account, either. In fact, only -xtregar-, -xtabond- and -xtgee- with
some specific correlation structures, like -corr(ar1)- or -corr(stat)-
would do something related. I don't think people typically bother
modeling this explicitly.
However, if the pattern of declining correlation is there for all
individuals, -cluster- will correct for that, too. Let's put it this
way: it does not bother as to what happens within an individual. If
observations are indeed i.i.d. -- fine, we'll lose efficiency compared
to the naive MLE by going into clustering. If observations are
independent but heteroskedastic -- fine, we can correct for that (at
least in the linear model -- I would suspect that heteroskedasticity
in the error term in the limited dependent variable models like probit
or ordered probit causes all sorts of troubles with normalization of
the regression coefficients). If observations are correlated with what
-xtgee- calls an exchangeable correlation (i.e. the one with a common
random effect, and individual errors conditionally i.i.d. given this
random effect) -- fine, if they follow some time series pattern --
fine... it does not matter for -cluster- at all.
On Tue, 26 Oct 2004 18:06:52 +0200, firstname.lastname@example.org
> Dear Stas,
> thank you for your profound response.
> Just one comment about your question:
> With several panels I would use the cluster option for the single
> inidvidual, assuming that her response of todays (t) is somehow related to
> her response t-1 time-periods before. So she is, certainly, 'similar to
> herself' and has to be grouped. But this 'similarity' might diminish as t-X
> grows bigger (How similar are you to yourself when you were 10 years old?).
> What I meant with 'direc tion of correlation' is that -cluster- does not
> take into account the time structure of this influence, going from the past
> to the present, but not from the present to the past. When I pool panels and
> use the - cluster- option, this time structure is neglected.
* For searches and help try: