[Forwarded on behalf of Mike Palij. -cdg-]
----- Original Message -----
From: "Mike Palij" <[EMAIL PROTECTED]>
To: "Teaching in the Psychological Sciences" <[email protected]>
Cc: "Mike Palij" <[EMAIL PROTECTED]>
Sent: Sunday, June 01, 2008 10:07 AM
Subject: Re: [tips] SAT for selection - Sackett et al. (2008) and Fantasy
Correlations
On Sat, 31 May 2008 23:29:23 -0700, Jim Clark wrote:
Hi
Warning ... simulation that may only be of interest to those "really"
>interested in question of restriction of range and correction.
A. Effect of Selection on SAT on SD of GPA
I initially set out to demonstrate that SD for GPA would shrink
>with selection on SAT, to verify one point of discussion in this thread.
>Following SPSS program generates 100,000 SAT scores (Mu = 500,
>Sigma = 100) and GPAs (Mu = 2.5, Sigma = .5) from a population
with Rho = .71 (i.e., about 50% of GPA predicted by SAT).
I'd like to thank Jim for the time and effort for doing these
analyses. I just want to make a couple of minor points:
(1) Neither SAT or GPA are true normal distributions, they
are truncated normal distributions. For the SAT, scores
below 200 are not possible nor are scores above 800. In
Jim's data, the minimum SAT= 55 and maximum SAT= 937.
Recoding scores below 200 to 200 and above 800 to 800
has a small effect on the SAT statistics but, of course, any
calculations involving SAT >800 cannot occur. We have a
similar situation for GPA but the minimum GPA is 0.34
(an acceptable value) but the maximum is 4.48, so values above
4.00 should be recoded to 4.00.
Applying the above changes for Jim's example of selecting
cases with SAT > 700, we get:
SAT Mean=736.58, SD=29.36
GPA Mea=3.3275, SD=0.3624
Pearson r(SAT,GPA) = 0.337 (original r=.309)
NOTE: the mean and SD for the recoded values are only
slightly smaller than the ones Jim originally reported but
for some reason N=2165 for the recoded data but N=2184
in the original results. There should be no difference in
sample sizes and I don't know why this has happened.
(2) I'm curious as to why the chosen correlation for SAT
and GPA was rho=0.71. Sackett et al report a estimated
value of 0.55 (corrected for range and other "problems"
and based on a sample of N=165,000) on page 218.
Was it just because the r-squared would give a percentage
of variance accounted for of 0.50? Of course, the larger
the value of rho, the greater the effect of selecting on SAT.
-Mike Palij
New York University
[EMAIL PROTECTED]
---
To make changes to your subscription contact:
Bill Southerly ([EMAIL PROTECTED])
