The Design and Analysis of Experiments for Assessing the Recovery of Shoreline and Avian Species Following the Exxon Valdez Oil Spill

 

E. James Harner

Department of Statistics and Computer Science

West Virginia University

Morgantown, WV 26506

 

 

Shoreline Ecology Program (SEP)

Stratified Random Sample Program

• Randomly chosen sites within strata defined by habitat and oiling
• Sampled only once in 1990
• Sixty-four sites (four per cell initially)

Fixed-site Program

• Subjectively chosen heavily-oiled or unique-habitat sites
• Sampled during two or three successive summers; Allowed time trends to be developed (and extended to SRS sites)
• Twelve fixed sites of which four received no cleanup

 

SRS Design

SRS Matrix

• Four oiling levels by four habitat types

Four oiling levels: Unoiled, Light, Moderate, Heavy

Four habitat types: Exposed bedrock, Sheltered bedrock,

Boulder/Cobble, Pebble/Gravel

• Three sampling transects perpendicular to the water at each site
• Systematic sampling on intertidal elevations and subtidal depths
• Synoptic data on chemistry, toxicology, and biology
• Biology data: upper, middle, and lower intertidal and subtidal
• Chemistry data: intertidal zones, subtidal, and shallow water

Sampling Stations

Measured Variables

Physical variables:

Percent sand, Percent silt/clay, TOC, and Wave-energy

Chemistry variables

PAHs and SHCs analytes

Sediment toxicology variables

Percent mortality of amphipods

Biological variables

Animal species abundances, total abundance, species richness, Shannon diversity, and biomass of algal species

 

Statistical Models

Biology variables (primarily species counts)

• Decision tree for selecting the distributional form
• Decision based on residual and variance function analysis
• Log-normal, Poisson, negative binomial, and binomial distributions
• Oiling effect tested after adjusting for physical variables

Chemistry variables

• Similar to biology variable analyses
• Averaging and nonparametric methods used if necessary

Sediment toxicity variables

• Binomial distribution used

Error Terms and Power

Random terms

• Site-to-site variability
• Transect-to-transect variability

Power criterion

• Signal-to-noise ratio (D/s) 1

Strategies for increasing power

• Increase a (not used in the SEP program)
• Increase the number of sites (constrained by summer sampling window)
• Liberal use of pooled error term (to increase the error degrees of freedom)
• Fit concomitant variables (to reduce the residual variation)

Power Curves

Community Analyses

Univariate--Models for diversity indices

Multivariate--Detrended partial canonical correspondence analysis

• Ordination technique for counts similar to discriminant analysis
• Axes constrained (canonical) to reflect an effect of interest (oiling)
• Axes adjusted (partial) for physical and environmental variables
• Arching effect (detrending) removed
• All species included when present on 20% of the samples

Ordinations

• Superimposed ellipses for oiling levels

 

Recovery Estimates

Based on community parameter analysis

• Individual comparisons for oiling vs. reference in each habitat
• Weighting factors reflect length of oiled shorelines in each cell
• Estimate was 91% in 1990

Based on correspondence analysis

• Based on oiling sites/transects outside the 95% reference ellipse
• Weighting factors reflect length of oiled shorelines in each cell
• Estimate was 73% in 1990

 

 

Wildlife Study

Focus was on marine birds

Bays are the basic experimental unit:

• Sampled along a gradient from unoiled to heavily oiled
• Sampled from 1989 to 1991
• Ten bays in PWS (sampled up to six times per year)
• Fourteen bays along the Kenai (sampled up to two times year)
• Each bay visited one to five times per cruise
• Nearshore and offshore surveys
• 34 habitat variables measured on bays

 

Analyses

Within-year analyses

• Each cruise analyzed for both nearshore and offshore surveys
• Basic model with oil and oil 2
• Models with physical and biological variables
• Adding oil and oil 2 to the "best" habitat model

Year-to-year analyses

• Cruises in the same season compared over years
• Basic model with oil and oil 2 and their interactions with year
• Habitat variables not necessary
• Oil-by-year interaction of principal interest (to assess recovery)

Models for Species Count Data

Normal regression models

• Fit when p (percentage of zeroes) 25%
• Densities computed

Abundances/shoreline length for nearshore surveys

Abundances/transect area for offshore surveys

• Densities log transformed (log 10 (x+1/6))

Poisson regression models

• Fit when 25% < p 75%
• Log link used
• Offset to adjust for shoreline length or transect area
• Adjustments made for over or under dispersion

Error Terms and Power

Random terms

• Bay-to-bay variability
• Visit-to-visit variability

Strategies for increasing power

• Increase a (models successively fit using a = 0.05, 0.1 and 0.2)
• Increase the number of bays (constrained by seasonal sampling window)
• Liberal use of pooled error term for within-year analyses and conservative use of pooled error term for among-year analyses (both to increase the probability of detecting an oiling effect)
• Fit concomitant variables (to reduce the residual variation)

 

Impact and Recovery Assessment

Impact

• Habitat + Oiling models as primary evidence
• Oiling effect after adjusting for habitat variables as compared to Oiling effect without adjustment

Recovery

• Habitat + Oiling models -- absence of an impact detected earlier
• Among-year models -- significance of an oil-by-year interaction with increasingly positive slopes
• Frequency of occurrence of a species over a series of cruises
• Oiling effect classified as strong, moderate, or weak depending on a level
• Interpretation aided by a PCA on species by ecological trait matrix