Zoology - Theses
Permanent URI for this collection
Search Results
1 - 1 of 1
-
ItemAn investigation of hypothesis testing and power analysis in impact assessment, using case studies of marine infauna( 2002)Statistical analysis is now widely used in impact assessment, and a common approach is to use a BACI design which considers changes Before and After a disturbance occurs, in conjunction with differences between the potentially Impacted location and one or more Control locations. The statistical power of an assessment is the confidence with which a nominated difference may be detected by a formal hypothesis test, and power analysis provides a means of quantifying this confidence. Despite the effort that goes into many impact assessments, results may be inconclusive or even wrong. Retrospective examination of completed assessments is one means of identifying factors which contribute to their success or failure. In this thesis, several BACI-style statistical tests for impact and their statistical power were examined, using case studies of marine infaunal assemblages in the vicinity of coastal wastewater discharges. The infauna is an assemblage often used in impact assessment, but is also known for large and often unpredictable natural changes in abundance which can make the identification of disturbance-related changes particularly difficult. Estimates of variance of infaunal abundance generated from the case study data were themselves extremely variable, with 43% differing from a best estimate by 50% or more. Even estimates for the same taxon, from studies in the same region and using identical sampling methods, differed by an order of magnitude or more in 25% of cases. The worst estimates of variance were usually obtained from single surveys, which had no component of large-scale temporal variation. Such variability in estimates of variance suggests that those based on single sampling times may be particularly unreliable, and that it may be desirable to allow for a larger than expected error variance by initially sampling more replicates than are expected to be needed. The specification of an alternate hypothesis, and subsequently an effect size, is an essential step in calculating the power of an hypothesis test, but it may be difficult to specify an effect size that is meaningful for the particular situation. The effect sizes actually observed in the infaunal data were very variable, with the most extreme being in excess of 10000% change. However, many of the observed effects, including some of the most extreme cases, appeared to be natural events rather than the result of a disturbance. The main alternative to the Control versus Impact designs is a gradient approach, which identifies changes in an ecological variable along a disturbance-related environmental gradient. This is usually done at a single time, resulting in possible confounding of a naturally-existing gradient with one imposed on the fauna by a disturbance. Including Before/After comparisons of gradients from multiple times within each period in a BAG design reduces this confounding. In contrast to the more common BACI-style assessments, it also provides an indication of the spatial extent of an impact, although comparison with a more conventional BACIP design suggested it might be more costly to undertake. The use of soft sediment infauna in impact assessment is often an expensive undertaking, because manual sorting of macrofauna from large quantities of sediment can be very time consuming. A process of compositing and subsampling may be effective for infaunal samples in assessments where random samples are taken at a range of spatial scales. When hierarchical analysis of variance is used, an arithmetic averaging of data from lower levels in the hierarchy occurs for tests which use the variance among higher levels as their error term. Subsampling of composite samples for each location in an MBACI design would produce physically averaged values which reflect the overall state of each location. Using these physical means would result in an identical test statistic to that resulting from the arithmetic averaging of lower level data within the hierarchical anova, provided that the compositing and subsampling process itself did not bias the estimates of the location means. Data from infaunal samples collected specifically for this compositing exercise established that the process did not appear to generate bias. Using hypothetical monitoring programs, cost savings of approximately 40% were possible with the use of compositing and subsampling in one scenario, while in another, additional sampling could be undertaken to increase power from 0.47 to 0.81 while still achieving an overall cost saving of 17%. The patterns of variability seen in the case studies, particularly in the space x time interactions, resulted in some large and unpredictable error variances and observed effect sizes. These patterns were not dissimilar to those seen in other published studies, suggesting that impact assessments in general could be prone to the problems identified in this study. Thus, responses to those problems, such as not relying on single sampling times for estimates of error variance in long-term programs, or considering compositing and subsampling when laboratory processing time is a concern, may also be applicable.