Identification using palmar flexion creases

[Raul]

Hi all - this weeks story was interesting - at Wolverhampton, UK, we have developed algorithms for conducting flexion crease matching that show individualisation potential if good exemplar prints are found (error rate of about 3/100 000). I am interested in the case. So that I can refer to it - could you let me have details, please? Thanks.

[Shane Turnidge]

Hi Raul,
If you can be a bit more specific about the details you need I should be able to help you out. Feel free to e-mail me @ shane.turnidge@peelpolice.on.ca
I would be interested to hear about the statistical model used in Wolverhampton.

Shane Turnidge

[clpexco]

The following e-mail discussion is posted here with permission from the authors to further the concept:

I read through the chain and there are a few things that stand out. I love what you are attempting to do and the idea is completely sound. I think the issues that might come up (as almost always occurs here) is in the expression of the probabilities and how that is communicated. DNA, protein markers, serology suffered from similar issues too...

So yes, keep collecting the data and write a paper. The more frequency data that are available the better for expressing the weight of evidence.

My criticisms would be focused on the following (and not necessarily trivial issues):

1)

<<And finally, the empirical data gathered from the 500 palm prints in our collection suggested that the probability that these two images were from two different people was less than 1% of the general population. >>

This is a common error in reporting probabilities. The more accurate way to say it is "Based on this sampling, less than 1% of the population would be expected to exhibit these characteristics. The defendant cannot be excluded from this subset (or "and the defendant is one of X% that would be expected to exhibit).

2) The frequency approach (above) assumes in the test, the range of appearance (i.e. the numerator in the likelihood ratio, is "1"). In other words, he has the feature OR he DOES NOT have the feature. In your test, were there instances where a feature could be in one category? or another? or it wasn't entirely clear if it fell in the category or not? Or could it be distorted in a way that the feature could have been present in the skin, but unrecorded in the exemplar? If yes, than your numerator would change...meaning that a higher percent of people could exhibit those traits but they aren't necessarily being considered...

3) This would directly impact an LR approach. If we do assume a numerator of "1", then the LR and 1/X are equivalent to the inverse of each other. So less than 1/100 would equal a LR of at least 100.

So the way to say this would be:

It is at least 100 times more likely to observe these corresponding features if the defendant left the mark THAN if some other (randomly selected, unrelated) individual left the mark.

[for analogy sake, this is equivalent to flipping a coin and getting about 7 heads in a row. Impressive if you say "watch this. I will now flip a coin and get 7 heads in a row..." (i.e. you generated the suspect on some other unrelated evidence and his palm creases JUST HAPPEN to match)....but NOT impressive if you flip the coin a million times and see if you ever get 7 in a row (i.e. you searched a database of 1 million and found this guy in the database of palm print creases that matched the crime scene mark).

4) Because of the way you reported it, how many others matched at ALL 7 or so creases...you are ok. You would have serious problems of independence if you attempted to MULTIPLY all those probabilities together. Pr [A] x Pr , etc. They are most likely NOT independent and I can already from your data that a couple appear to be highly correlated...

So as long as you compare the set of features and how many individuals shared features ABFG together, then you are OK. Independence is not being invoked.

Those are the things that popped out to me.

I am no statitician though and don't take my word for it. I'd go to the likes of Ian Evett, Champod, or Cedric Neumann, maybe even Nicole Egli or Anil Jain, etc. They would know a lot more about the best way to approach it.

Good luck, but keep generating these data and publish the method for classification/scoring (that will be key for the robustness of the data).

g.

*************************************

Hi Glenn and thank you for your response. I by no means am a math guy and your input is really valuable.

I’m in the lessons learned phase of this case and I know I would do things slightly different today than I did in 2008.

I get what you’re saying in item 1) but the evaluation of the evidence in this case was based on a combination of my knowledge and experience with palm prints and the data from the sample.

After the initial analysis I identified the seven features that I believed warranted further investigation, based on my previous knowledge and experience with palms. These by no means were the only features available, just the ones that gained my attention. Also, I was careful to only note secondary palm print creases that were related to major creases. The major creases were treated like class characteristics for the purposes of this analysis even though I could have put extra weight on the path of the proximal transverse crease because of its more random path. So in hindsight, I guess it would have been better to report that; “Based on a combination of my knowledge and experience with palm prints and the data gathered from this sampling, less than 1% of the population would be expected to exhibit these characteristics.”

I also reported the discrepancies between the two images. One image was taken with a point and shoot camera while the other was taken with a Nikon D200.

The frequency approach may have skewed the data somewhat. There were instances where some secondary creases were not recorded in the sample, particularly the data associated to the finger creases. But the creases in the palms were far less susceptible to those problems.

Interestingly, there were colleagues who were suggesting that I multiply the probabilities together à la DNA probability determinations. I stayed away from that approach because I felt the approach was far too aggressive. I also felt it was unnecessary given that the court would only be interested in being reasonably convinced. The difference between 99% and 99.89% in a courtroom is almost nothing.

I certainly can see the need for a statistician in a future case unless someone comes up with a wonderful probability model or a searchable AFIS type database from which we could generate more incriminating probabilities. I may write a paper on this process if I can find the time but I doubt I’ll be able to continue generating data by analyzing additional samples. I see the future of this type of process as evolutionary rather than something that works great right out of the box. I think it is going to take a lot of effort from a fair number of people if it is to evolve.

[Neville]

would it be better to talk of rolling a dice than flipping a coin