Development and validation of statistically sound criteria for the match of unweathered GC-MS fingerprints in oil spill forensics

Abstract

The investigation of an oil spill's origin frequently relies on determining the equivalence of oil component patterns in samples from the contaminated environment and suspected oil source. This comparison benefits if based on the ratio of the abundance of unweathered characteristic components of the oil product, Diagnostic Ratios, DR. Replicate determinations of DR from one sample are used to set limits for the second sample's DR. The composition equivalence of oil patterns in both samples is indicated if all compared DR are statistically equivalent with a high confidence level. Some studies define DR limits assuming their normality and using Student's t statistics (S-t). However, since the ratio of correlated abundances can be not normally distributed, this criterion can drive to more false comparisons than predicted by the test confidence level. This work developed a computational tool for the reliable description of the non-normal distribution of the DR based on the Monte Carlo Method (MCM), aiming to allow the accurate control of the confidence of DR comparison. This work concluded that S-t defines 95% or 98% confidence limits with probabilities of falsely rejecting samples equivalence, φ, that can be up to 4.3% higher than predicted by the confidence level of the S-t test (i.e., 5% and 2%). The fragilities of the S-t limits significantly reduce the probability (1-θ) of two samples with the same oil producing equivalent values of all compared DR. For the studied 69 DR from unweathered components, the (1-θ) for 98% confidence level limits, set by the MCM and S-t from triplicate injections of one sample, are 94.8% and 91.7%, respectively. These values are below the confidence level (P) defined for each DR because DR are correlated with a correlation coefficient lower than 1. The (1-θ) can be increased to above P by using MCM limits and accepting composition equivalence if at least one of two sample extract injections produces values within limits set from the other sample's replicate injection. The validated user-friendly MS-Excel file used to set and access comparison criteria is made available as Supplementary Material and was checked experimentally. However, it is not feasible to estimate model confidence exclusively from experimentation because it would require too much independent analysis.

Keywords: Compositional match; False result rate; GC-MS; Monte Carlo method; Nonparametric statistics; Oil spill.