under exchangeabilty of the tokens, then a binomial model would seem right (recall that the Poisson is the limiting form of the binomial as n gets large, p gets small, etc)...

maybe an ordinal model, but could be more trouble than it is worth.  

tough to say more without knowing more about the experiment and the plausibility of exchangeability.

Simon Jackman

Sent from my phone

-----Original Message-----

From:  Christopher Zorn <[log in to unmask]>
Subj:  [POLMETH] count (?) query
Date:  Mon Apr 23, 2007 10:33 am
Size:  1K
To:  [log in to unmask]

Two related questions for the list:  I have a student who has some  
data from an experiment in economics.  Respondents were given ten  
"tokens," and allowed to allocate them into two different "pots."   
His response variable is the number of tokens each respondent placed  
in one of the two pots; it varies between zero and ten, inclusive,  
and is always integer-valued.  He is interested in regressing these  
counts on a set of covariates.

1) What's the "right" model for such a response variable?  It is not  
simply a matter of 10 being an exposure/offset, since that is  
constant across observations, and it doesn't strike me that a  
binomial is exactly right either (it's strange to think about the  
probability of "success" for each token).  My initial thought was to  
treat it as an upper-truncated count, but there's precious little  
written on such models, so I am not sure.

2) Assuming it is best thought of as a right-truncated count, is  
anyone aware of a "canned" way to estimate such models, preferably  
using R/Stata/SAS?

Thanks,

  -- CZ

=========================================
Christopher Zorn
Department of Political Science
University of South Carolina
Columbia, SC 29208
Office: 803-777-2207
Mobile: 803-553-4077
E-Mail: [log in to unmask]
http://www.cas.sc.edu/poli/Vitae/zorn.pdf
=========================================