186 



Fishery Bulletin 102(1) 



To maintain a focus on the effects of soak time, the models 

 were limited to simple combinations of fixed effects and 

 interaction terms. Dawn and dusk were added to various 

 models of each species in each fishery. To reduce complex- 

 ity, year and season were limited to models of seven spe- 

 cies (bigeye tuna, oilfish, swordfish, blue shark, albacore, 

 southern bluefin tuna, long-nosed lancetfish) in the two 

 South Pacific fisheries. The seven species represented 

 four taxonomic groups and the full range of responses 

 observed in preliminary analyses of the soak-time-catch- 

 rate relationship. 



Random effects We added random effects to all models to 

 allow catch rates of segments within each longline opera- 

 tion to be related. The random effects model assumes that 

 there is an underlying distribution from which the true 

 values of the probability of catching the species, jt, are 

 drawn. The distribution is the among-operation varia- 

 tion or "random effects distribution." The operations are 

 assumed to be drawn from a random sample of all opera- 

 tions, so that the random effects (0 ( ) in the relationship 

 between catch rate and soak time for each operation (i) are 



independent and normally distributed with 0~N(Q, a 2 ). 

 The random effects and various combinations of the fixed 

 effects were added to the linear predictor presented in 

 Equation 5. 



For each species in the South Pacific yellowfin tuna 

 data set we compared the performance of models under 

 an equal correlation structure with that of models under 

 an autoregressive correlation structure. Under an au- 

 toregressive structure, catch rates in the different hourly 

 segments within the operations are not equally correlated. 

 For example, the correlation between segments might be 

 expected to decline with increased time between seg- 

 ments. However, we used an equal correlation structure 

 for all models because the Akaike's information criterion 

 (AIC) and Sawa's Bayesian information criterion (BIO 

 indicated that there was no clear advantage in using the 

 autoregressive structure rather than an equal correlation 

 structure. 



Implementation We implemented the models in SAS 

 (version 8.0) using GLIMMIX, a SAS macro that uses 

 iteratively reweighted likelihoods to fit generalized linear 



