On Tue, 23 Feb 1999, Mark Goldberg (mark@polair.epi.mcgill.ca) wrote:
>
> I have just discovered an interesting feature of the gam function. In
> fitting complex time series data using quasi-likelihood, I have found that
> the final results of the fit depend on the order of the covariates. For
> example, the model
>
> log(y) = a + lo(day,span=0.05) + lo(year) + lo(x)
>
> will give slightly different estimates of the dispersion parameter,
> residual deviances, residual degrees of freedom, standard errors and the
> like than a model, say,
>
> log(y) = a + lo(year) + lo(x) + lo(day,span=0.05).
>
> The differences all occur in the fifth decimal place, so that the effects
> are extremely minor, but nevertheless disconcerting.
>
> I'm curious to know whether any one else has run into this and whether
> there is a simple explanation.
Try changing the convergence criteria in gam.control. Like glm, gam has
rather loose convergence default criteria: in both cases the lack of
convergence is rarely important.
More seriously, there is a priori no reason why a gam model need have a
single solution, and we have seen examples where equivalent models using
s() that converge to different solutions (with different
(quasi-)likelihoods). When using lo() it is even unclear what problem a
gam is solving.
--
Brian D. Ripley, ripley@stats.ox.ac.uk
Professor of Applied Statistics, http://www.stats.ox.ac.uk/~ripley/
University of Oxford, Tel: +44 1865 272861 (self)
1 South Parks Road, +44 1865 272860 (secr)
Oxford OX1 3TG, UK Fax: +44 1865 272595
-----------------------------------------------------------------------
This message was distributed by s-news@wubios.wustl.edu. To unsubscribe
send e-mail to s-news-request@wubios.wustl.edu with the BODY of the
message: unsubscribe s-news
|