Compound-Specific Isotope Analysis Coupled with Multivariate Statistics to Source-Apportion Hydrocarbon Mixtures

Compound Specific Isotope Analysis (CSIA) has been
shown to be a useful tool for assessing biodegradation,
volatilization, and hydrocarbon degradation. One major
advantage of this technique is that it does not rely on
determining absolute or relative abundances of individual
components of a hydrocarbon mixture which may change
considerably during weathering processes. However, attempts
to use isotopic values for linking sources to spilled or
otherwise unknown hydrocarbons have been hampered
by the lack of a robust and rigorous statistical method for
testing the hypothesis that two samples are or are not
the same. Univariate tests are prone to Type I and Type
II error, and current means of correcting error make hypothesis
testing of CSIA source-apportionment data problematic.
Multivariate statistical tests are more appropriate for use
in CSIA data. However, many multivariate statistical
tests require high numbers of replicate measurements.
Due to the high precision of IRMS instruments and the high
cost of CSIA analysis, it is impractical, and often
unnecessary, to perform many replicate analyses. In this
paper, a method is presented whereby triplicate CSIA
information can be projected in a simplified data-space,
enabling multivariate analysis of variance (MANOVA) and
highly precise testing of hypotheses between unknowns
and putative sources. The method relies on performing
pairwise principal components analysis (PCA), then performing
a MANOVA upon the principal component variables (for
instance, three, using triplicate analyses) which capture
most of the variability in the original data set. A probability
value is obtained allowing the investigator to state
whether there is a statistical difference between two
individual samples. A protocol is also presented whereby
results of the coupled pairwise PCA−MANOVA analysis
are used to down-select putative sources for other analysis
of variance methods (i.e., PCA on a subset of the original
data) and hierarchical clustering to look for relationships
among samples which are not significantly different. A Monte
Carlo simulation of a 10 variable data set; tanks used to
store, distribute, and offload fuels from Navy vessels; and
a series of spilled oil samples and local tug boats from
Norfolk, VA (U.S.A.) were subjected to CSIA and the statistical
analyses described in this manuscript, and the results
are presented. The analysis techniques described herein
combined with traditional forensic analyses provide a
collection of tools suitable for source-apportionment of
hydrocarbons and any organic compound amenable to GC-combustion-IRMS.