12.11.2012 Views

Examination of Firearms Review: 2007 to 2010 - Interpol

Examination of Firearms Review: 2007 to 2010 - Interpol

Examination of Firearms Review: 2007 to 2010 - Interpol

SHOW MORE
SHOW LESS

You also want an ePaper? Increase the reach of your titles

YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.

Continuing the work <strong>of</strong> Curran et al. (94), Campbell and Curran (95) introduce a<br />

“permutation testing approach” <strong>to</strong> the interpretation <strong>of</strong> elemental analysis data<br />

from glass evidence. In previous work, the authors suggested using Hotelling’s<br />

T 2 <strong>to</strong> evaluate multivariate data that may be used in a Bayesian calculation.<br />

Criticisms <strong>of</strong> this approach are addressed; namely, the assumption <strong>of</strong> normality,<br />

and the relatively large sample size not necessarily available <strong>to</strong> the forensic<br />

scientist. The authors had some success removing these constraints using a<br />

shrinkage estima<strong>to</strong>r. The reader is directed <strong>to</strong> regard this as a “pre-screening”<br />

<strong>to</strong>ol and not for interpretation for the courts.<br />

Zadora and Neocleous (96) and Zadora (97) employ a Bayesian approach for<br />

classification <strong>of</strong> glass fragments in<strong>to</strong> use-type categories using RI and SEM-EDS<br />

data. Support vec<strong>to</strong>r machines, naïve Bayes classifiers, and Bayesian networks<br />

were again assessed as well as the performance <strong>of</strong> the likelihood ratio models<br />

when considering between- and within-object variability. A model that takes both<br />

in<strong>to</strong> consideration performs better than a model using between-object variability<br />

only. The accuracy <strong>of</strong> the classification is highly dependent upon glass type, but<br />

in general, the combination <strong>of</strong> SEM-EDS and change in RI values with annealing<br />

increased anywhere from 3% <strong>to</strong> greater than 10% over SEM-EDS and preannealing<br />

RI data alone. Caution should be taken when using this type <strong>of</strong> data<br />

alone for classification purposes as the values for several glass types (e.g.<br />

architectural and au<strong>to</strong>motive) tend <strong>to</strong> overlap. Classification based on preannealed<br />

RI values alone should not be attempted. Likelihood ratios from these<br />

data are usually close <strong>to</strong> one, giving very limited support <strong>to</strong> either the<br />

prosecution’s or defense’s hypothesis.<br />

Jensen and Shen (98) explore the use <strong>of</strong> statistical feature selection as an aid <strong>to</strong><br />

Bayesian analysis <strong>of</strong> glass. In fact, they aim <strong>to</strong> remove the need for expert<br />

knowledge in feature selection and advocate univariate data analysis for<br />

likelihood ratio calculations. While the use <strong>of</strong> au<strong>to</strong>mated feature techniques may<br />

be <strong>of</strong> use in evidence evaluation, it is unadvisable <strong>to</strong> rely on statistical modeling<br />

<strong>of</strong> a likelihood ratio without sufficient knowledge <strong>of</strong> the underlying analytical data.<br />

Zadora and Ramos (99) study the influence <strong>of</strong> adequate population databases<br />

for likelihood ratio calculations. Databases assessed are based on RI and SEM-<br />

EDS data. This study was also presented at the 2009 EFS meeting in Glasgow,<br />

Scotland (100) and the 15 th Annual meeting <strong>of</strong> the EPG in St. Gallen, Switzerland<br />

(101).<br />

This concludes the Glass Report for <strong>2010</strong>. Future papers and presentations will<br />

be summarized in the next report.<br />

104

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!