Date: Wed, 10 May 2006 12:00:15 +0200
Reply-To: "adel F." <adel_tangi@YAHOO.FR>
Sender: "SAS(r) Discussion" <SAS-L@LISTSERV.UGA.EDU>
From: "adel F." <adel_tangi@YAHOO.FR>
Subject: Re: comparing linear vs nonlinear model
In-Reply-To: <BAY103-F169B2A54459960D819927FB0A90@phx.gbl>
Content-Type: text/plain; charset=iso-8859-1
Hi,
thanks to all of you for replying my question concerning the comparison between linear vs nonlinear model, very interesting comments so far.
I am sorry that I did not explain my problem.
I have a continuous variable Y which represents Hormone level (H) and I have two continuous Ind variable age (25 -75) weight, two categorical variable smoking (Yes No) Medication (Yes No)
It is weel estiablished in the litterature that the Hormone level (that I am styding) increases with age, en general people looks at linear relationship (hormone vs age)
Until a ceratin age we can assume that we have fairly a linear trend (H = a b*age) and than after an inflexion point (age point) I have a more nonlinear trend (and this is why I have considered X^2 which is age^2 "square")
My question which will be the best model.
The one with the linear relationship (age with other IV) or one with (age + Age^2) and other IV
My sample is large , I have more than 3000 observations
I have looked to the mean of Hormone by age group and I have noticed what I have desctibed as linear trend + (up to a certain age ) a curve trend
It is possible to identify the age where I have this curve change?
Thanks
A lot
David L Cassell <davidlcassell@MSN.COM> a écrit :
Peter Flom postulated:
>An interesting discussion thus far. It prompts me to ask about an idea I
>had, which I haven't seen implemented. This makes me think my idea must
>have something wrong with it, but I don't see what is wrong.....
>and, contrary to my usual demands for context, I am going to phrase this
>very generally.....But I am more interested in models for EXPLANATION than
>'black boxes' that may do a great job of predicting, but aren't so
>explanatory
>
>Suppose you have found (through some sensible method) several models that
>appear to be reasonable. In general, there ARE several models (at least)
>that are sensible for any one data set. Some people suggest various ways
>of averaging across models (e.g. Bayesian model averaging; random forests)
>I have found, though, that these tend to go toward the black box end of
>things. So, suppose you have to choose ONE model, recognizing that it may
>not be THE TRUTH, but that it may be useful (All models are wrong, some
>models are useful)
>
>OK, so, instead of (or in addition to) formal tests of the models (using
>Rsquare, or AIC, or BIC, or whatever) why not compare the predicted values
>of the models to each other and to the actual data values. If the
>predicted values for the various models are all 'close', chosse the simpler
>model. Let 'close' be subject-dependent and rely on the expertise of the
>analyst and substantive experts rather than letting the computer (or any
>preconceived notion of 'good fit' do your thinking.
>
>OK OK, I know this isn't a FORMAL test, that's the point, really. And I
>know that there may be difficulty in getting this method used for
>peer-reviewed articles. But is there anything wrong with it,
>statistically?
>Shouldn't this be what model-development is about? Aren't we data analysts
>supposed to think about what our models mean?
>
>Enough rambling, it isn't even Friday.
This strongly correlates with one of my model-building credos. The
structure
of the model and the meaning of the parameter estimates should not be
left to some cubicle-resident mathematical statistician who looks
depressingly
like Wally.
If several models provide similar fits, then experts in the field ought to
be
able to say why, and to speculate on which model is most useful for the
subject-matter experts. If we do a mixed model with repeated measures
and we end up with several possible covariance structures, but the
subject-matter experts can tell us that the ARH(1) structure has a real
physical meaning which matches what would be expected by the
Morris-Winter model, then why would we want to argue that we can do
marginally better with AIC using a covariance structure which cannot be
interpreted in terms of the theory?
just chiming in late,
David
--
David L. Cassell
mathematical statistician
Design Pathways
3115 NW Norwood Pl.
Corvallis OR 97330
_________________________________________________________________
Is your PC infected? Get a FREE online computer virus scan from McAfee®
Security. http://clinic.mcafee.com/clinic/ibuy/campaign.asp?cid=3963
---------------------------------
Yahoo! Mail réinvente le mail ! Découvrez le nouveau Yahoo! Mail et son interface révolutionnaire.
