Anganuzzi. Detecting trends in dolphin abundance indices 



A characteristic of the 

 smoothed test is that, while the 

 smoothing procedure induces a 

 correlation between estimates, 

 the correlation across years be- 

 tween fixed percentiles of the dis- 

 tribution of smoothed estimates 

 is lower. This is a result of the 

 (implicit) sorting of estimates 

 that removes some of the depen- 

 dency. To illustrate this, the 

 average correlation between 

 smoothed estimates obtained be- 

 fore sorting is also shown in 

 Figs. 9 and 10, suggesting that 

 the reduction in correlation due 

 to sorting is -30% for estimates 

 close in time. 



Discussion 



declines rapidly as distance between the estimates in- 

 creases, approaching very low values as the separa- 

 tion between estimates exceeds 4 yr. This suggests that 

 the validity of the test will not be seriously compro- 

 mised by the correlation induced from the smoothing 

 procedure, when the comparison is carried out on esti- 

 mates that are separated by at least 4yr. 



In summary, the two types of 

 tests represent different ways of 

 looking at the data, and a com- 

 parison between them based on 

 the same criteria only partially 

 reflects these differences. The 

 smoothed test provides more in- 

 formation since it allows a 

 graphic comparison between in- 

 dividual estimates. It is based in 

 a more robust assumption about 

 the underlying trend of the popu- 

 lation, since it assumes only that 

 the change has been smooth over 

 a short time-period. It also in- 

 corporates the lack of precision 

 of the estimates in a much more 

 effective way than the linear test. 

 In some cases, however, the 

 linear test might be more suit- 

 able. For example, if the amount 

 of interannual variation is low 

 relative to the precision of the 

 estimates, or if changes are 

 closely approximated by a linear 

 function, the linear test should 

 perform better. Also, if large changes in population 

 size occur over a very short time-period, smoothing 

 the series will tend to underestimate the rate of change. 

 Despite limitations of the comparison, the results 

 from simulations indicated that the smoothed test out- 

 performs the linear test in most situations. The only 

 exception is the tendency to detect spurious trends 



