Hi

On Sun, 20 Feb 2005, Christopher D. Green wrote:
> I see that you did not answer *our* questions either. Many of us 
> obviously understood the trick that was up your sleeve (or at least up 
> that of those you were relying on). As the article you attached says in 
> its very first paragraph: "Even the commonly cited statistic of a higher 
> risk of crash per kilometre among this group..." It then goes on to 
> attempt to minimize this bald fact by continuing, "has been shown to be 
> an artefact: low mileage is intrinsically risky, and studies show that 
> when this is controlled for the older drivers perform as well as, if not 
> better than, younger drivers." The phrase "as well, it not better," in 
> all likelihood translates into "no significant difference," so what we 
> REALLY have, advocacy aside, is that older people are terrible drivers 
> (perhapst in part because they do not drive very far or very often). The 
> implication, however, that their skills would improve if they drove more 
> (as such skills do for younger folks) is patently silly.

Here is a little simulated data set for SPSS to demonstrate the
problem of controlling for covariates that are actually dependent
variables of the underlying predictor.  Run this program and the
effect of age (AGE) on accidents (ACC) is significant when age
alone is in the equation, but becomes non-significant when
kilometres driven (KIL) is added as a second predictor.
(discussion continues below)

SET SEED = 27395137.
INPUT PROGRAM.
LOOP SUBJ = 1 TO 64.
COMP #z1 = NORMAL(1).
END CASE.
END LOOP.
END FILE.
END INPUT PROGRAM.
COMP age = RND(45 + 15*#z1).
COMP kil  = RND(20 +   5*(#z1*-.917 + norm(1)*.4)).
COMP acc = RND(30 +   5*(#z1*.4 + NORMAL(1)*.917)).
regr /vari = age kil acc /dep = acc /enter age /enter.

If you look at the 3 compute statements for the variables of age,
kil, and acc, however, you will see that indeed accidents is a
direct function of age (#z1), along with some noise.  But so is
kil a (negative) function of age.  Controlling for kil is
inappropriate and leads to an incorrect conclusion.

We do not know for certain that the researchers did something
analogous to this, but it certainly sounds like they might have,
and then cast it in terms to make it appear that amount driven
was a causal variable rather than a correlated outcome variable.

Thanks to Stephen for another good example of the dangers of
multiple regression for my Honours stats class.

Best wishes
Jim

============================================================================
James M. Clark                          (204) 786-9757
Department of Psychology                (204) 774-4134 Fax
University of Winnipeg                  4L05D
Winnipeg, Manitoba  R3B 2E9             [EMAIL PROTECTED]
CANADA                                  http://www.uwinnipeg.ca/~clark
============================================================================


---
You are currently subscribed to tips as: [email protected]
To unsubscribe send a blank email to [EMAIL PROTECTED]

Reply via email to