Dave Fournier1, James Ianelli1,2 and Steve Martell1,3
1ADMB Foundation; 2Alaska Fisheries Science Center; 3International Pacific Halibut Commission
April 21, 2015 9:00 (PST): FSH 203
A proposal for rehabilitating the multinomial likelihood for compositional data
The multinomial likelihood is widely used in many stock assessment models for evaluating compositional data. However, this simple model does not account for over-dispersion and does not allow for correlation among adjacent categories. In response to recent publications proposing alternatives for the multinomial, we propose to extend and modify the ideas put forward by Hranfnkelsson and Stefánsson (2004) to better accommodate over-dispersion and correlation in compositional data. Specifically, to develop a self-scaling multinomial type estimation procedure and optionally incorporate positively autocorrelated errors. Models that incorporate composition data need to take into account over-dispersion and correlation if uncertainty in model parameters, and the risks associated with management actions, is desired.