Statistical Issues in Ecological Risk Assessment

STATISTICAL ISSUES IN ECOLOGICAL RISK ASSESSMENT

Biplob Das

Psyc 802

University of Regina

Student # 200228135

Fall 2008

STATISTICAL ISSUES IN ECOLOGICAL RISK ASSESSMENT

ABSTRACT

This paper discusses some statistical issues arising in the ecological data analysis based on biogeochemical measurements in ecosystems. A brief overview on statistical issues in ecological data analysis has been described in the first part of this discussion. In the later part replication and effect size issues are discussed. Replication of large-scale experiments is desirable, but the numbers of replicates needed are not known. Costs and feasibility of ecosystem experiments depend critically on the numbers of replicates needed because of the high cost per replicate and the scarcity of experimental ecosystems. A partial solution to determine statistical significance of non-replicated ecological variables was tested on two lake data by applying randomization technique. Statistical power plays a critical role in strengthening the interpretation of an experiment or analysis. For effect size issue, the analyses on Canada North Environmental Services (2003) data showed that statistical power is too low to detect any potential cumulative effects due to low sample sizes.

1

INTRODUCTION

Ecological risk assessment is defined as the process of assigning magnitudes and probabilities to the adverse effects of human activities or natural catastrophes (Suter1993). The design and analysis of the vast majority of scientific research has traditionally focused on minimization of false positives (i.e., false positive; or Type I errors) and at the same time maximize power to reject the null hypothesis (i.e., there is an effect). It is rare for an ecological risk assessment to be undertaken to disprove the null hypothesis, since neither industry nor government generally wants to prove that significant risk exists (Holdway 1997). Rather, the norm for ecological risk assessment is to attempt to accept the null hypothesis; that is, there is no risk. Thus, risk assessments are generally fraught with Type II errors (accepting as true a false null hypothesis) while spending the majority of their efforts in minimizing Type I errors. This problem is a very serious statistical one, even when only working with single species toxicity testing, much less when involving the enormous difficulties of measuring and predicting the behaviour of complex ecosystems (Holdway 1997). Thus, to maximize the probability of protecting the environment, it is more appropriate to guard against false negatives (Type II errors) than false positives (Type I errors). This can be done by establishing levels of statistical significance to take into account Type II errors and improving the statistical power of test designs to detect existing effects so that both Type I and Type II error rates are minimized can greatly strengthen the risk assessment (Power and Adams 1993). The process is used to identify and evaluate hazards using measurement, testing and mathematical or statistical models to quantify the relationship between the initiating event and the effect. By expressing results as probabilities, risk assessment acknowledges the inherent uncertainty in predicting future environmental situations, thereby making the assessment more credible. A brief overview on statistical issues in ecological data analysis has been described in the first part of this discussion. The later part is concentrated on randomization and effect size issues with case studies.

2

An overview of issues

UNCERTAINTY

Risk, as mentioned previously, is the probability that a specified negative effect will occur, or in the case of a graded effect, the probability that a particular magnitude of the effect will occur. If it is certain that the undesired event will or will not occur (i.e., the probability of occurrence of a specified event is known to be 1.0 or 0.0) there is de facto no risk.

Uncertainties in risk assessment arise from three sources: 1) the inherent randomness of the world (stochasticity), 2) imperfect or incomplete knowledge of things that could be known (ignorance), 3) mistakes in the execution or measurement of assessment activities (error). Stochasticity refers to uncertainty that can be described and estimated but cannot be reduced because it is an inherent characteristic of the system being studied. Most biological processes such as colonization, reproduction and mortality are stochastic in nature. Moreover, physical drivers of ecosystems such as rainfall, temperature, and wind are also effectively stochastic. Thus, limits on the precision with which variable properties of the environment can be quantified define the upper limit of precision for any risk assessment (Suter 1993). Ignorance refers to lack of knowledge of some aspect of the ecosystem that is potentially knowable. In some cases this is a fundamental ignorance of some scientific issue and can results in undefined uncertainty (the "unknown unknowns") that cannot be described or quantified (Bauer 1992). Mostly though, ignorance is simply a result of the practical limitations on our ability to accurately describe, count or measure everything that pertains to a risk estimate (Suter 1993). Much of the primary literature on uncertainty in risk assessment pertains to identifying, quantifying and reducing this kind of uncertainty. The third source of uncertainty, human error is an inevitable attribute of all human endeavours. Generally these errors can be thought of as a quality assurance problem.

The diversity and complexity of ecological systems ensures that the ignorance component of uncertainty in ecological risk assessments will always be very high (Suter 1993). Moreover, Levins (1966) and others that have followed, suggest that there is an unavoidable trade-off between precision, generality, and realism in any ecological study, including risk assessments. If one envisions each of these descriptors as occupying the three corners of a triangle, it is easy to see that a precise and realistic model (or assessment) will necessarily be limited to a few specific

3

applications. A very general approach applicable to many situations will necessarily be limited in precision. The more realistic the test or model (in terms of inclusion of components), the lower the generality, and so on. This has the following implications for ecological risk assessment: 1) There is no unique