Barlow and Berkson: Evaluating methods for estimating rare events with zero heavy data 
355 
A 
B 
o 
<D 
<D 
> 
< 1) 
cl 
o 
Q> 
<D 
> 
TO 
CD 
CL 
D-s D-p P-p NB-p 
C 
0 5 
o 
-f- 
8 a 
0 0 
L ' J. 
— ' — l 
[ ! 
1 11 1 
, 
r ir i 
-0 5 
~ L ~ 
1 o - 
D-s 
D-p 
P-p NB-p 
D 
u 
o — 
Dl 
3- 
E 
D-p 
Comparison of bycatch estimates to the total amount of bycatch simulated to 
evaluate performance of estimation methods. The stratum-level delta-lognormal 
method (D-s), delta-lognormal method for all sets pooled (D-p), generalized linear 
model with Poisson error distribution for all sets pooled (P-p), and generalized 
linear model with negative binomial error distribution for all sets pooled (NB-p) 
were evaluated. Each of the 5 panels corresponds to one of the spatial scenarios: 
(A) = co-occurrence clumping ( Turtles dump , Sefs ciu turtles ), (B) = sets-only clumping 
(Turtles uni{orm , Sets clump . sels ), (C) = independent clumping (Turtles dump , Sets dump . 
sets ), (D) = turt!es-only clumping (Turtles dump , Sefs uniform ), and (E)=fully uniform 
(Turtles uniform , ‘Sefs uniform ). Each of the plots within a panel corresponds to an 
estimation method. The scale of the y-axes varies by rows of panels for display 
purposes. The horizontal line at a relative error of zero marks where the median 
of an unbiased estimation method should fall. Notches are placed around the 
medians, and if the notches of 2 plots do not overlap, there is strong evidence 
that those medians differ. The box of each plot includes the first through third 
quartile. Whiskers extend to the most extreme data point that is no more than 
1.5 times the interquartile range from the box. Small circles represent outli- 
ers. For purposes of display, in the panel for the sets-only clumping scenario 
(Turtles uniform , ■5ets clump . sets ), one outlier was removed from each of the P-p and 
NB-p box plots. 
