[R-sig-eco] Poisson regression vs. linear model on transformed data

[Philip Dixon]

Theory suggests that the Poisson, overdispersed Poisson, Negative Binomial, 
Zero inflated Poisson, or some other appropriate distribution should be the 
best method.  Practical implementation issues often suggest otherwise.

Linda Young did a comparative study of LM, Poisson and Negative Binomial 
regression for overdispersed data in a 1 way layout.  She found the LM often 
worked best and identified those situations.  The paper is

Linda J. Young, Nancy L. Campbell and George A. Capuano, 1999
Analysis of Overdispersed Count Data from Single-Factor Experiments: A 
Comparative Study
JABES 4:258-275.

Philip Dixon