I didn't see the original either. I don't think much of your experimental strategy, but you can define an effect size for a regression coefficient by relating the range of values for the variable in the data to the minimum difference of interest in the response. If g is the minimum difference of interest in the response, and if r is the range of values for a variable then the minimum detectable value for the regression coefficient is g/r. Since you have standardized the data (a poor practice) you can probably take r to be 5 or 6. All this follows from the observation that the product of the absolute value of a regression coefficient and r is the range of the contribution of that variable to the response.

Donald Burrill wrote:
Hi, Chuck.  Did you ever get a satisfactory reply to your question?
(I've a vague recollection of seeing one or two replies, but seem not to
have saved them.)  Had a thought or two that might be helpful.

In the first place, the b's for both younger and older adults are of the
same sign (not surprisingly), and for both examples that you cited, were
weaker for older persons.  They were significant (i.e., different from
zero, I presume?) for one age group, but not for the other.

That they were not distinguishably different from each other is also not
very surprising:  apparently (guessing from the data you cited) the
threshold for significance (at whatever significance level you were
using - you didn't say) is around 0.2, so that 0.31 and 0.24 are
significant, but 0.15 and 0.19 are not;  for the conditions (including
sample sizes) of your study.

Since the noise level (e.g., standard error) around a difference is
larger than that around a single estimate by about half (actually by a
factor of <square root of 2> for independent samples with the same N),
you'd need _differences_ on the order of 0.3 for them to be significant
and the differences you got (0.16 and 0.05) just don't cut it.

Next question (I can't tell):  are the raw differences you found (laying
aside issues of significance for the moment) in directions that are
consistent with respect to whatever underlying theory(ies) may inform
this research?  If not, that the differences sometimes seem to favor
such a theory and sometimes oppose it would suggest that there may not
be anything going on, at least not in accordance with theory.  However,
if they all (or most of them, for the several measures you took) are in
a consistent direction, your evidence is somewhat stronger with respect
to the theory, only the effect is on the average too small to detect
with a t-test and your sample size.  (As you are undoubtedly aware, ANY
effect can be found to be significant if one throws enough observations
at it:  if N=50 isn't large enough, try N=400 or 2,000 or 40 million.)

But it occurs to me that your categories "younger adults" and "older
adults" aren't very precise, and may well include a range of ages in
each category.  If you have enough data, you might be able to obtain
more than 2 b's as a function of age groups defined more narrowly than
you had done:  in that case, you could plot  b  vs  age (or average age)
and eyeball the plot to see if it appears to wander consistently upward
(or downward) as age increases.  (And if your "younger" and "older"
groups are separated by a range of ages for which you have no Ss, be
alert for a corresponding gap in the apparent progression of b values.)

I wouldn't be surprised if the subgroups so obtained had subsample sizes
so small that none of these b's were significantly different from zero;
but if there are enough of them, and the changes with age are
consistent, you could attack the problem with a sign test, thus:
Suppose you had been able to ten age groups.  Then there are nine
differences from one age group to the next older one.  By the null
hypothesis of no effect due to age, one would expect half of these
differences to be positive and half negative.  If they're ALL in the
same direction, that's impressive evidence that SOMEthing is there, even
if the "something" is too subtle to be detected by a t-test (and given
the rather large standard errors of the regression slopes, arising from
the small subsample sizes).  If not all, but most of the differences are
consistent in this sense, you still may have interesting evidence (and
in any case evidence to which you can attach a significance level, and
for which a larger level (0.10, say) than your usual standard might well
be appropriate);  especially if the number of different age groups is
smallish due to constraints imposed by the original selection process.

Of course, a sign test does have the disadvantage that there isn't
really an "effect size" that one can readily compute.  But sometimes the
logical priority is first "Is there an effect?", and after that one can
be more relaxed in acknowledging that the effect appears to be more
subtle than can readily be estimated with the tools at hand...

HTH -- but I did want to suggest that approach, just in case it might
have been helpful.  Good luck!    -- Don Burrill.

On Sun, 30 Nov 2003, Chuck Robertson wrote:


I have a question about power analysis related to a t-test of some
weights in multiple regression equations. I'm afraid I'm a cognitive
psychologist, so if I give you too much info below about my experiment
please accept my apology.

I've taken memory measures from younger and older adults at several
different encoding times and I also have created constructs of popular
cognitive resources as predictors (perceptual speed, working memory,
vocabulary,..) I am interested in how the regression equations change
in each age group as memory performance increases (as I give people
more time to study). I ran the regression equations for younger and
older adults at a given memory measure and computed a t-test for
difference between the predictor weights for what one might expect to
be different equations.  As an example, at a 1 second encoding time
the older adults have a significant beta weight for speed as a
predictor (b = .31) but the young adults do not (b = .15), in the same
equation young adults have a significant beta weight for vocabulary (b
= .24) but older adults do not (b = .19). This casually looks like a
difference when eyeballing the equations, speed is the only
significant predictor for older adults and vocabulary is the only
significant predictor for young adults at a 1 second encoding time for
the memory material. Interestingly, when I look at the t-test for
difference between the B weights, I fail to reject the null of no
significant difference between the weights. The t-test for difference
between the speed predictors is t = -.51 and t = .14 for the
vocabulary predictors. Of course I do not want to accept the null, I
am really curious as to how much power I had to detect a significant
difference with my sample size of 80 younger and 80 older adults in
each equation.

I need help computing the effect size given my SPSS output from the
regressions. I assume that I need to compute f as an effect size since
this isn't teh typical t-test. Can anyone tell me how to do that given
the weights and the standard errors for the weights from the output? I
must be doing something wrong as my f's are way to large.

Thanks, Chuck

.
.
=================================================================
Instructions for joining and leaving this list, remarks about the
problem of INAPPROPRIATE MESSAGES, and archives are available at:
.                  http://jse.stat.ncsu.edu/                    .
=================================================================



 -----------------------------------------------------------------------
 Donald F. Burrill                                         [EMAIL PROTECTED]
 56 Sebbins Pond Drive, Bedford, NH 03110                 (603) 626-0816
.
.
=================================================================
Instructions for joining and leaving this list, remarks about the
problem of INAPPROPRIATE MESSAGES, and archives are available at:
.                  http://jse.stat.ncsu.edu/                    .
=================================================================


--
Bob Wheeler --- http://www.bobwheeler.com/
        ECHIP, Inc. ---
Randomness comes in bunches.


. . ================================================================= Instructions for joining and leaving this list, remarks about the problem of INAPPROPRIATE MESSAGES, and archives are available at: . http://jse.stat.ncsu.edu/ . =================================================================

Reply via email to