A new direction for large-scale experimental design and analysis

Abstract

Increasing awareness of the effects of large-scale, natural, and anthropogenic perturbations (eg global climate change, aquatic eutrophication, dam removal, etc) has initiated a groundswell of public, political, and scientific support for actions that will preserve and restore ecosystems and the services they provide. Consequently, ecosystem ecology and restoration are developing rapidly to meet the demand for urgent action to restore degraded habitats. This has revealed the inadequacy of traditional statistical methods in addressing large-scale, unreplicated research and unplanned events, because these methods rely on a statistical paradigm of replicates, homogeneity, randomness, normal distributions, and controlled experiments. However, the subjects of ecosystem studies are often complex, non-random, non-normal, not replicable, and generally violate traditional statistical assumptions. Even experimental treatments, designed for homogeneous application within and between replicates, may not follow these rules. For example, fire is important in maintaining structure and function in terrestrial and wetland ecosystems, yet fire intensity – and therefore its effects – varies across the landscape due to variations in attributes such as topography, soil moisture, and species composition. Fire events are unique in spatial extent and temporal characteristics. Without proven statistical methods to address these issues, the rigor and validity of emerging research findings have been questioned. We have been involved in the design of a large-scale ecosystem study in the Florida Everglades for the past 2 years. Its main objective is to assess alternative approaches to accelerate recovery of nutrient-enriched areas where homogeneous stands of cattails have replaced sawgrass communities. We are attempting to integrate ecology and management to advance the field of ecosystem restoration, particularly its statistical rigor. Facing experiments with unprecedented spatial extent, enormous spatiotemporal variations, as well as financial constraints and the risk of further system degradation through nutrient mobilization, we have looked for improved methodologies for design and analysis to avoid the use of replicates. In December 1990, in a Special Feature in Ecology (71[6]: 2037–68), Carpenter and others challenged the applicability of traditional replicated design to large-scale, natural, and human perturbations, emphasizing the importance of conducting unreplicated studies to further the development of ecological theory. This special feature introduced new tools to bring statistical rigor to the analysis of unreplicated studies. The authors argued that (1) “insufficient replication may be worse than no replication at all, if the experimenter does not consider the power of statistical tests when interpreting results”, and (2) “resources that could be invested in duplicates might be better spent on more detailed mechanistic analyses conducted within the context of the large-scale experiment”. Numerous statistical tools have emerged over the past 15 years, including Bayesian statistics, likelihood methods, information theoretic approaches, intervention analyses, and before-after control-impact paired series, which provide a unified framework for predicting landscape-level ecosystem responses to perturbations and management strategies, and offer new means of gaining insights into ecological processes. So far, we have found no publication where a large-scale ecosystem restoration study was deliberately designed using an unreplicated approach. We attribute this to concerns that these methods are not “proven”, and so manuscripts based on research conducted in this manner may be rejected by peer-reviewed journals. Clearly, there is a gap between theory and practice and an urgent need to educate the ecological community regarding such approaches. Ecosystem restoration requires new statistical tools to investigate the consequences of large-scale conservation and restoration actions, such as wetland mitigation, canal filling, and nutrient fluxes. It is neither sound science nor common sense to replicate large-scale human activities that degrade ecosystems, so statistical methods to utilize unreplicated sites are needed. Over the past three decades, research on global change (eg elevated atmospheric CO2) increased from small-scale mechanistic studies to field studies to landscape simulation models at large temporospatial scales. With increasing scale, uncertainties related to heterogeneity, multiple interactions, and other factors arise; financial costs increase as well. Both restoration ecology and traditional ecology need to integrate design and analytical approaches into appropriate goals and theories. With the support of some authors of the 1990 Special Feature, we have organized a symposium for the 2006 ESA Annual Meeting to describe our experiences with integrating new statistical approaches into design and analysis of large-scale, unreplicated studies. Discussions will focus on applications in a variety of ecosystems, including Florida's Everglades, the San Francisco Bay estuary, grasslands, and coral reefs. While genuine replication is a powerful tool and should be used whenever possible, ecologists, with the help of statisticians, need to develop statistical techniques that meet the special requirements of large and diverse ecosystems. Large-scale experimentation in ecology will be greatly enhanced by adding analytical techniques appropriate to the scales of landscape processes and ecosystem perturbations. ShiLi Miao, Senior Supervising Scientist, Everglades Division, South Florida Water Management District, FL Susan Carstenn, Assistant Professor of Environmental Science, Hawaii Pacific University, HI