|
|  |
[Date Prev][Date Next][Thread Prev][Thread Next][Date index][Thread index]
Re: st: Tobit interpretation and post estimation
Elena--
Maarten's answer is quite good, but I want to highlight one important
substantive point: it is not true that "only positive debts are
possible" in household data (look for interest income for
counterexamples). Further, you do not have censored data, I think,
but a nonnegative dep. variable, for which the -tobit- model is
probably ill-suited. A -poisson- regression with the -robust- option
is appropriate for a wide variety of models with nonnegative dep vars,
though given the large number of zeros in your application, you may
prefer -zip- or something else in that family. The relevant
post-estimation commands for the estimation commands are outlined in
their respective help files.
On 4/17/07, Maarten buis <[email protected]> wrote:
--- Elena Giarda <[email protected]> wrote:
> We estimated a tobit model for the household debt (only consumer
> debt, no mortgages) as a function of several explicative variables,
> among which age, age-squared, income, income-squared, wealth,
> wealth-squared and some socio-economic dummy variables.
>
> We decided to estimate a tobit model because of censoring: among our
> households (about 6500), 3373 have positive consumer debt, the
> remaining are censored at zero.
>
> Our problem is how to predict the estimated values. In fact, if we
> use "predict" or "predict, ystar(min,max)" we obtain very small
> values of household debt (both using levels and logs of the
> dependent variable), possibly caused by a large negative estimated
> constant.
After -tobit- a number of different types of predicted values make
sense. -tobit- assumes that everybody has a latent ideal amount of
consumer debt, and that this ideal amount can be either positive or
negative. However, only positive debts are possible, so as soon as
someones ideal debt is negative his realized debt is zero. So we can
predict the ideal debt (using the option xb, or no option at all) or
the realized debt. The difference between the two is easily illustrated
when we assume a linear relationship, like in the example below (where
we assume the minimum value is 17 not 0). If you look at the graph you
can see that the linear relationship is assumed for the latent
variable, and that the observed dependent variable levels off in order
to stay above the minimum value.
*-------------- begin example ------------
sysuse auto, clear
tobit mpg weight, ll(17)
predict xb
predict y , e(17,.)
twoway line xb y weight, sort yline(17)
*---------------- end example ------------
(For more on how to use examples I sent to the Statalist, see:
http://home.fsw.vu.nl/m.buis/stata/exampleFAQ.html )
Hope this helps,
Maarten
*
* 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/
