• Home
  • Calculating the reliability of likelihood ratios: Addressing modelling problems related to small n and tails

Calculating the reliability of likelihood ratios: Addressing modelling problems related to small n and tails

Sourcetitle: 
Proceedings of 14th Australasian International Conference on Speech Science and Technology
Year of publication: 
2012
PublicationType: 
Conference paper - peer reviewed

In forensic speech science we are often faced with the problem of having a relatively small amount of data which is alsomultivariate and distributionally complex. This results in a serious problem exactly in the scenario where potentially largestrengths of evidence could be obtained, i.e., when the trace data are on a tail of the distribution which models either theprosecution or defence hypothesis and a large magnitude log likelihood ratio is calculated. By definition the sampling ofa distribution is sparse on its tails and this problem is compounded if the model is trained on a small amount of data – smallfluctuations in the training data can lead to large changes in the calculated likelihoods on the tails and thus large changesin the calculated likelihood ratios for trace data on the tails. Large-magnitude calculated log likelihood ratios are thereforeinherently unreliable.

http://clas.mq.edu.au/sst2012/

To the top

Page updated: 2013-01-07 10:24

Send as email
Print page
Show as pdf

X
Loading