Date: Wed, 17 Dec 2008 09:28:20 -0800
Reply-To: Shawn Haskell <shawn.haskell@STATE.VT.US>
Sender: "SAS(r) Discussion" <SAS-L@LISTSERV.UGA.EDU>
From: Shawn Haskell <shawn.haskell@STATE.VT.US>
Subject: Re: Regression: do you always need main effects with interactions?
Content-Type: text/plain; charset=ISO-8859-1
On Dec 17, 10:33 am, peterflomconsult...@mindspring.com (Peter Flom)
> Kevin Viel <citam.s...@GMAIL.COM> wrote
> >>Prof Drews-Botsch is funny. If we did what he says, all the time, we
> >>would make almost no progress at all. Scientific advance requires
> >>exploration as well as confirmation.
> >Actually, she :)
> Oopps, my bad. To use some of what you say, sex is one of the more significant ways in which humans vary
> >Perhaps we should stipulate a dichotomy? Exploration and hypothesis-
> >generating work is essential, but even then I recommend discipline in the
> >form of a very limited number of alterations.
> I don't think it's a strict dichotomy. I don't think any data analysis is purely exploratory, and very few are
> purely confirmatory. We almost always have *some* idea what we are expecting, and we almost never have a very exact idea.
> Also, if we only look in places where we expect to find something, then we won't make any big discoveries.
> At the same time, yes, hypothesis generation can lead to hypothesis confirmation.
> >This process is unfortunately expensive. In a recent analysis I noticed an
> >apparent association between contrast-induce nephropathy and platelet
> >count. My colleagues did not tell me to look for it. How do we treat it?
> >It deserves to be mentioned, but as the second model, in my opinion. Since
> >we now have knowledge of this association, it can be the (a) main exposure
> >variable in the next investigation, if there is one...
> Yes, see my middle sentence just above. I am not sure what you mean by 'second model', but
> I agree that we must make clear, when we write things up, whether they are a priori or post hoc models (oddly,
> one grad student who I am working with is referring to some things as a priori post hoc tests .... I have pointed
> out to her that a test can't be both .... :-).
> [I had written]
> >>In a data set I am working with now, there are about 1000 variables, and
> >>it took about 10 years to collect all the data. Shall we test each of
> >>1000 models separately, getting new data each time?
> [Kevin replied]
> >Potentially, yes. It depends how refined your models are. As an analogy,
> >consider a stick figure, a mannequin, and a fashion model on the catwalk.
> >I can use each as a model for a patient, but may not reflect the situation
> >so well.
> Again I agree. One way to look at the progress of a particular field is to see
> how refined the model is. But there are times for models of all degree of refinement.
> >The age of using the general term "patient" is over, or should be. Any
> >relationship you find with those 1000 variables may need to be refined. To
> >illustrate, no longer should doctors say "patients given aspirin experience
> >fewer acute coronary events." Patients differ in too many meaningful
> >variables, especially genetic, which we are revealing every day.
> This is a pet peeve of mine, that we often see statements that ignore human variation.
> Peter L. Flom, PhD
> Statistical Consultant
> www DOT peterflom DOT com
we seem to have strayed a bit but in a good way. i agree with Peter's
philosphy of research and scientific endeavor. I can only recall ever
running 1 strictly "confirmatory" type of analysis - i only had 2
predictor variables to consider realistically. In wildlife science,
and probably most other disciplines, situations and general effects
are almost "never" identical in space or time - i can imagine
exceptions. it is this variability that is most interesting and leads
to real discovery in my field. Thus, I may have 12 potential
predictor variables and appropriate interactions a priori but
understand that all won't be important for conceivable (hypothesis
testing) and inconceivable (exploratory) reasons. In this context, a
"best" model within a somehow-defined model set often leads to
simultaneous hypothesis testing and hypothesis formation - induction-
deduction-induction - the scientific method is a big beautiful rolling
stone that keeps on truckin along.........SH