Statistically sound comparison of standardized and simulation methods for oil spill source identification in real...

The comparison of oil patterns of a spill (Sp) and suspected spill source (SS) samples is based on ratios between correlated GC-MS signals of oil-discriminating compounds, i.e., diagnostic ratios (DR). The Student's t statistics (S-t) and a maximum relative difference (SC), proposed in standard methods, have been used for DR comparison due to their simplicity. An alternative methodology based on Monte Carlo Method (MCM) simulations of correlated signals, capable of accurately defining DR comparison criteria, proved that S-t and SC assumptions regarding DR normality and precision are frequently not valid, affecting comparison reliability. The performance of the approaches was accurately compared from independent signals of the same oil sample from a perfect match between Sp and SS. The present study describes the comparison of the approaches in real oil spill scenarios reproduced in International Round Robin Tests. Since as the number of compared DR increases, also rises the probability of not all equivalent DR being actually considered equivalent, the decision of oil pattern equivalence was based on two comparisons of independent sets of Sp and SS signals. The risk of true oil standard equivalency claims is compared for the three oil spill scenarios studied, which are different considering oil types, DR sets and spill weathering. The ability of the approaches to distinguish the Sp sample from an oil sample known not to be the source of the spill was also assessed. The MCM based on two independent DR comparison trials was the only one consistently producing fingerprint comparison risks of correct equivalence claims larger than 98 %. MCM also performed better in distinguishing different oil patterns. It was concluded that comparing >22 DR does not change the risk of correct oil pattern equivalence assessment significantly. The complexity of the MCM approach is overcome by using user-friendly and validated software.