Coral ECA Water Quality Assessment Methods Comparison

Coral ECA Water Quality Assessment Methods Comparison

Document: Coral ECA Water Quality Assessment Methods Comparison

Document Type: Report

Author Name: Henry Briceño, Joseph N. Boyer, and Ian L. Drydan

Executive Summary

The Kristin Jacobs Coral Reef Ecosystem Conservation Area (Coral ECA) Water Quality Assessment (WQA) was designed in 2014 by a collaborating body of National Oceanic and Atmospheric Administration (NOAA) scientists, Florida Department of Environmental Protection’s Coral Reef Conservation Program (CRCP) staff, and partners from the Southeast Florida Coral Reef Initiative (SEFCRI). The goal of the WQA was to provide data for managers to assess the status of the Coral ECA, an area which historically did not have a consistent water quality monitoring program.

The goal of this data analysis project is to evaluate and prepare the available water quality data of the Coral ECA for future assessments aimed at identifying both the constituents and the impacts of land-based sources of pollution (LBSP) on the Coral ECA. The main objective of the project is to inform resource managers and decision-makers on the status of water quality in the Coral ECA.

The available data, collected from 2017 to 2021, contain many non-detects (ND) due to high and variable laboratory detection limits. In some instances, censoring reaches 100% of some data sets. High percentages of ND pose a challenge to the use of the database and to the interpretation of results, especially in use for the development of nutrient criteria and in determining compliance to those criteria.

A NOAA Technical Memorandum (Whitall et al. 2019), analyzed the Coral ECA water quality data from September 2016 through December 2018. In that report, the authors used the methodology described by Flynn (2010) to populate a time series by replacing the non-detects with imputed “dummy” values. This technique could be questioned because substitution of non-detects with dummy values creates an artificial dataset which may be biased. To avoid this problem, there are other well-known methods such as survival statistics which are deemed to be more appropriate (Helsel 2006, 2010, 2011).

The tasks of the present project are to: 1) Reformat the available dataset because their current WIN/WQX data format is not easily ingested by standard statistical packages; and 2) Calculate

descriptive statistics for the dataset distributions using both Flynn’s methodology and Survival Statistics and provide a comparison between results of these approaches. The resulting comparison of five statistical methods was carried out for estimating the mean, median, sd, and inter quartile range when censored data are present. A Monte Carlo simulation study was performed to investigate the performance of the estimators under known distributions.

From these simulations we determined that the Censored Maximum Likelihood Estimation using the Weibull distribution (MLE-W) is the most unbiased and most efficient method when the underlying distributional family is correct. The dummy imputation methods can perform poorly in some cases.

In addition, increasing the level of censoring in the dataset resulted in large positive biases in the dummy estimates in some examples while the Weibull MLE remained stable. Therefore, for highly censored datasets, the Weibull MLE was the best overall estimator.

Future analyses on the reported datasets from this project are aimed to answer the following questions:

·        Are there differences in the data between the individual ICA’s, and if possible, also compared to land use coefficients?

·        Are there differences between site types – inlet vs reef vs outfall samples?

·        Is there a significant difference in analyte concentrations between bottom vs surface samples?

·        How do the available concentration data compare to any relevant published thresholds, especially to those of southeast Florida waters?

We recommend that the Weibull MLE be used in performing these future analyses, with suitable checks that the distribution is appropriate.