Improving correlation as a measure of similarity of congener patterns through compositional data analysis

Abstract Persistent organic pollutants such as dioxins and furans (PCDD/F), and other contaminant classes are ubiquitously found in the environment. In this context, identifying potential sources of contamination when new hot spots appear presents a significant challenge. Assigning a hot spot to a specific source (e.g., a nearby small combustion plant) can be difficult, especially when multiple possible sources are involved and the pattern of pollutant load at the hot spot cannot be clearly linked to a single source. This may be due to similarities among the congener patterns of the potential sources or a lack of similarity to any of the known source patterns. Statistical methods such as Pearson correlation—although only indicating statistical associations and not causal relationships—between the congener pattern of a potential source and that of the hot spot can be helpful in this regard. We demonstrate through statistical simulations how the correlation-based quantification of statistical similarity can be improved by fundamentally changing the understanding of the data, thereby increasing the power of the correlation test. This methodology can support statistical similarity analysis by using unbiased statistical estimators—unlike traditional correlation analyses of congener profiles.