Hi

James M. Clark
Professor of Psychology
204-786-9757
204-774-4134 Fax
[email protected]
 
Department of Psychology
University of Winnipeg
Winnipeg, Manitoba
R3B 2E9
CANADA


>>> "Mike Palij" <[email protected]> 09-Jan-09 9:53 AM >>>
On Thu, 08 Jan 2009 20:04:23 -0800, Karl Wuensch wrote:
>        I'm even less conservative than Stephen.  I would not apply the
>Bonferroni adjustment.  After all, these are PLANNED comparisons, eh?

This is a curious point:

Why should the state of knowledge (i.e., able to predict the size
of difference, the direction of a difference, etc.) affect the probability
of making an error of inference?

JC:
I think of it as quasi-Bayesian kind of thinking and use the analogy of 
perception.  If you walk into a crowded room EXPECTING to see a certain person 
(the prediction), then you require less perceptual information (the data) in 
order to identify that the person is indeed present.  In essence, you have a 
lower threshold for detection because of the expectation.  Lacking the 
expectation, your threshold for detection is higher and you will require 
greater evidence.  In Bayesian terms, your a priori probability is higher given 
the theoretical/empirical expectation, therefore you require less current 
evidence to achieve a certain level of a posteriori confidence than if you 
lacked that expectation.  It is important to emphasize that your expectations 
in research need to be well-founded, either on the basis of theory or past 
research.

A second point with respect to the specific case Karl is commenting on (i.e., 
simple effects following an interaction) is that the omnibus test of the 
interaction is quite insensitive except when a cross-over interaction occurs.  
Otherwise, the interaction is diluted into main effects and interaction.  In 
the typical pretest-posttest by treatment/control design, for example, one 
expects (ideally) no change for the control group and a change for the 
treatment group.  But this specific pattern gives rise to differences for the 
two main effects and the interaction, whereas the simple effects analysis 
correctly (?) allocates all the variability to the difference between pre and 
posttest for the treatment group.  Illustrative numbers below.

Factorial Interaction (brackets are interaction effect)
                          Pre          Post              Row effect

        Control      20            20                 20     -5
                             (+5)         (-5)
        Treatment 20            40                 30     +5
                             (-5)          (+5)
Col effect          20            30                 25 Mg
                          -5            +5


Simple Effects (brackets are simple effects)

                          Pre          Post              Row

        Control      20            20                 20
                             (0)         (0)
        Treatment 20            40                 30
                             (-10)          (+10)

In many cases, latter simple effect for treatment will be significant absent 
significant interaction.

Take care
Jim



---
To make changes to your subscription contact:

Bill Southerly ([email protected])

Reply via email to