Rich Ulrich <[EMAIL PROTECTED]> wrote in message > I think you have suggested the 
useful way to start - by exploring.  
> By making the naturalistic description, if one is reasonable.
> 
> First, if there are no players who do better/ worse than chance,
> then there is nothing to talk about.  So, how much is there?
> 
> Statistically speaking, you might show a "excess of
> variation"  based on  *many*  cases with  *small*  evidences
> of bias - that could be tough to study.  Maybe interesting, 
> though, as someone's thesis for methodology.
> 
> But I think you have an actual chance if there are a moderate 
> number of cases with *large*  biases in evidence.  Then
> you can say,  Do they have anything in common?  anything
> that differs from the non-biased ones?
>  - Does the *most*  extreme match-failure  have a
> particular feature?

Thanks Rich

another idea I have is that I would extra all 'pairs' from the data
(or possibly only those who have played a certain minimum number of
times), e.g. I get the following from my data:

Tom-Dick (27 wins, 14 losses for Tom, average of ELO ratings
difference at the time of these matches 289)
Dick-Harry (12 wins, 24 losses for Dick, average of ELO ratings
difference at the time of these matches -124)

For each pair listed above I calculate two figures:
1) the posterior (I think its called?) probability based on their
record, e.g. 27/41 for Tom vs Dick and 12/36 for Dick vs Harry
2) the estimated probability based on ELO, e.g.
http://tournaments.tantrix.co.uk/ratings/simple.shtml

I perform a chi-squared test on the two 'series' to determine if they
have the same mean. If they do, then no head to head effect, if they
don't then there is a head to head effect

Am I moving towards something here or barking up the wrong tree
(still/again !!)
.
.
=================================================================
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