Richard Boylan <[email protected]> asked:
>> I first computed
>> streg y x, dist(weibull) robust cluster(id)
>> mfx compute, predict(mean)
>> Then to check for the imporstance of unobserved effects, I estimated
>> streg y x, dist(weibull) frailty(gamma) shared(id)
>> mfx compute, predict(mean)
>> To my surprise all the results changed drastically.
>> What I find harder to understand is that the overal predicted times change
>> drastically:
>> with the first model:
>> mfx compute, predict(mean)
>> Marginal effects after weibull
>> y = predicted mean _t (predict, mean)
>> = 3.7482296
>> with the second model:
>> mfx compute, predict(mean)
>> Marginal effects after weibullhet
>> y = predicted mean _t (predict, mean)
>> = .50887109
>> Shouldn't I be expecting these numbers to have been similar?
Stephen Jenkins <[email protected]> responded:
> If you told us what the estimate of the Weibull shape parameter ("p") was in
> both specifications, we'd be better placed to comment. If the frailty model
> is the 'true' one, and the hazard is rising at a fast rate, then you could
> get very different predicted means, I suspect.
I agree with Stephen. Basically, there are two levels on which you would
expect different answers (and in some examples, like yours, the difference
can be dramatic).
(1) If the frailty effect is significant (i.e. theta significantly greater
than zero), then you have two different models yielding different
predictions.
(2) Note that with your second model, you are fitting a Weibull/gamma
shared frailty model. When you use the -shared()- option, the prediction
produced by the -mfx- option -predict(mean)- is the estimated mean
time-to-failure for a subject with predictors at their mean levels, AND
conditional on a frailty value of one (the mean frailty).
Thus, when comparing this to a predicted mean time that is unconditional on
the frailty you are really comparing apples and oranges here (or, more
accurately, tangerines and oranges), and when you fit a standard
(non-frailty) model you are in effect producing predictions unconditional on
the frailty.
On the one hand, you have an estimate of the average prediction with
respect to the frailty distribution (the gamma). On the other, you have
a prediction conditional on the mean of the frailty distribution. In
general, these are not one in the same, and since the gamma distribution
can be quite skewed (for large theta) the difference can be substantial.
Note that the use of -robust cluster()- in the first model only effects the
estimated standard errors of your estimates, and thus is not crucial to the
comparison of predicted values between models.
--Bobby
[email protected]
*
* 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/
