On Wednesday, May 23, 2012 1:01:33 AM UTC+10, Sam Wormley wrote:
> On 5/22/12 9:38 AM, Marvin the Martian wrote:
> > On Mon, 21 May 2012 18:19:34 -0500, Sam Wormley wrote:
> >
> >> On 5/21/12 4:32 PM, Marvin the Martian wrote:
> >>> On Mon, 21 May 2012 10:36:21 -0500, Sam Wormley wrote:
> >>>
> >>>> On 5/21/12 9:50 AM, Marvin the Martian wrote:
> >>>>> What caused the climate change for the 4 billion years before the
> >>>>> last 200 years, Sam?
> >>>>
> >>>> Mostly greenhouse gasses, Marvin!
> >>>
> >>> Ah, circular logic! You are begging the question again, Worm.
> >>>
> >>> It was certainly NOT man made greenhouse gases.
> >>>
> >>>
> >> In the case of this current rapid global warming the evidence is
> >> overwhelming
> >
> > POST HOC.
>
>
http://en.wikipedia.org/wiki/Post-hoc_analysis
>
> In the design and analysis of experiments, post-hoc analysis consists
> of looking at the data for patterns that were not specified a priori.
> It is sometimes called by critics data dredging to evoke the sense
> that the more one looks the more likely something will be found.
In a former age we used to distringuish between what was called
"hypothesis testing" and "statistical exploration".
Both are equally valid and have a sound formal basis.
Hypothesis testing is the basic tool we usually learn from K12 -- how
to decide whether 2 population means are the same, e.g.
Statistical exploration is the basis of data mining. It can be done in a very
unsound manner - i.e. just by "looking for patterns in the numbers".
Or it can be done soundly.
As you say, if you don't do it right you will be misled by the answers.
If you test all kinds of possibilities and one of them is found to be
"significantly" associated with your hypothesis there is really no surprise.
SOME pattern is likely to turn up if you look hard enough.
However, it can usually be made statistically sound by adjusting the degrees of freedom in the problem to count all those different possibilities you tried and rejected before you found one that "matched".
One of the rules of thumb for, e.g., the F test when an "all by all"
is done in certain tests is to replace the df num, den by sqrt(num), den.
This essentially adjusts for the e.g. 2^num possibile subsets you tested
before finding one you "liked". It won't show up as significant using
the adjusted df unless it is very very significant as a stand-alone hypothesis.
It should remind the Q physics people of aspects of Q measurement theory.
Turns out classical prob has the same kinds of funny business, too.
If you just do the 1 hypothesis test you are confident at some level
the null is accepted or rejected; if you did a test BEFORE that one and
gained some knowlege about the same thing, the confidence if something different! Somehow what you knew before somehow "affects" (loosly)
the result of the experiment you are about to do now.
--
He has a long lists of straw man fallacies to support his fallacy fallacy.
Most of them have false premises, tho'.
-- Marvil the Venusian@Boulder, 18 Mar 2012 07:31 -0500
To "protest too much" is to insist so strongly about something not being true that people begin to suspect maybe it is true.
--
http://www.goenglish.com/ProtestTooMuch.asp