CLPEX Chat Board

I've thought enough about the problem of developing a probability model for comparison of two impressions that I wouldn't make any attempt to guess how long it might take to appear. Expressing the features in an impression representing a small portion of skin and doing it in such a way that either distortion effects will not change the expression or working the likelihoods of various distortions into the scheme are significant problems. My feeling is that these problems are so severe that it may never be possible to validate the numerical results. Remember that the problem is not one of spotting candidates. It is instead the problem of telling me how many people might be expected to display, in a small portion of skin, features so similar and so similarly arranged that I couldn't tell them apart. It's not how many people have portions of skin that I can't tell apart. Because we're dealing with latents, it's how many have portions of skin that could be impressed in some way that the impressions can't be discriminated. That means it becomes an extremely complex topological problem, with limitations in the kind of distortions allowed. (For example, the old joke that a topologist can't tell a doughnut from a coffee cup because through continuous deformation they are the same. But in latent examination there is a point in deformation when we can easily tell something's no longer a coffee cup, so to speak.)

Now, it might be proposed that if a latent examiner has decided that a pair of impressions are impressions of skin with exactly the same constellation of features, but that the combined quantity and nature of those features are insufficient to convince the examiner that they must be from the same physical skin, we might be able to use a model that was suited to undistorted representations to tell us how often that less than sufficient constellation is expect6ed to occur. Okay. If that exists, we could say it was insufficient to form a confident belief in the source, but the likelihood of it being the suspect is P. But, if such a likelihood can be calculated for an insufficient constellation, it can be calculated for what would otherwise have been sufficient to induce in the examiner a confident belief in a single source. When that likelihood is known, any examiner saying he has formed a belief in a unique source is then saying the he takes a likelihood of P to be certain belief. His belief then becomes moot. The acceptance of some numerical likelihood as enough to decide reality is not a matter of expertise but merely an arbitrary choice.

So, does our conundrum over whether or not to conclude a probable source change any? Right now, we are, in reality, concluding an identification by stating the fact that what we observe convinces us that the probability of it being anyone but the named source is so low that we believe that person is indeed the source. We do not know what that probability is in numerical terms, but we belief it's very, very small. When we don't find sufficient data, we are really finding the the probability of it being someone else is just too high for us to form a belief in a unique source.

If the kind of model described above were available, we would have numbers. But we would still be deciding that in our examination we had or had not found the probability of duplication so low that we believed. And we would be in the same conundrum over whether to say anything about a case with a probability higher than we were comfortable with.

But is that really how it would be? Once possessed of a numerical probability, can the examiner justify expressing any degree of belief? Where on that scale of numerical probabilities lies his old threshold for belief? Can he say? Will the examiner community find that the cases where the old unspecified threshold determined belief all turn out to have probabilities smaller than some number? Does that number replace the subjective threshold? Has anything been gained, then? Or do conclusions of belief become anachronisms? With no expert belief, now does the fact finder use the evidence? Is that even a problem.

I kind of think it won't be a problem. My reason is the fact that our community threshold has not, absent misinterpretation of features (something that will just as effectively corrupt a mathematical model), produced any demonstrable error of identifying the wrong person. That suggests that if we had the probability calculator, the probabilities of duplication in our Identification conclusions would extremely small, perhaps as small as in a solid DNA case or smaller. And the most significant potential benefit of a valid model may be that cases now deemed inconclusive are shown to have very small likelihoods of duplication and are therefore powerfully probative evidence. The errors, as always, will be almost entirely a result of poor interpretation, since mischaracterizing features in the latent can induce a false congruence with features in the record.

First off I believe that we have to get off the idea that the problem of modeling human friction ridge skin is so difficult that it can never be done. Complex mathmatical models have been built to handle things that are a lot more complex than the little lines on out hands and feet. There is even an algorithim that can generate fingerprint patterns that are hard to distiguish from real life examples. We have AFIS systems that are able to pluck a canidate out of millions of fingers and most of the time the algorithim gets it right.

This is an application of statistics that is not only possible but has been demonstated. The one main hold back to this is of course money! There simply is not very much money going into solving this issue. For years we have said that they way we do things is good enough and that the stat models would be way to hard so why try and why spend the money.

As far as the other issues on how we would incorporate the stat models into our work, I think that Glen has it right. With a basic understanding of what the model is looking at and what its limitations are (currently limited to points) we should be able to justify an identification in the face of low L/R's. Current models only work for the fingers so we would still have to testify the same way for palms. But the model would be very helpful in that you now a method that is not subjective. If the same points are plotted the exact same way the numbers should remain the same. We still will have differences between what points are marked and that will change the numbers from examiner to examiner but you will have some solid statistical data behind your subjective determination. That can only strenghten the science.

David Johnson wrote:But the model would be very helpful in that you now a method that is not subjective. If the same points are plotted the exact same way the numbers should remain the same. We still will have differences between what points are marked and that will change the numbers from examiner to examiner but you will have some solid statistical data behind your subjective determination. That can only strenghten the science.

Firstly, I must state I have the greatest of respect for Glenn and Cedric, they are both highly intelligent and ethical, and; respectively, their dance-floor prowess and taste in shirts can only be described as truly awesome.

I have attended both the Stat course and the probability course run by Glenn and Cedric. Impressive as the work-in-progress is, a major concern of mine was the consistency in the numbers that can be produced both by different examiners working on the same print, and by individual examiners altering the points to improve the Likelihood Ratio. The several hours spent in the lab were nowhere near sufficient to satisfy me that the job was anywhere near finished, but I do believe it is a very good start.

With AFIS, different searchers will achieve different scores for the same print, an individual can improve his/her score on a re-search once they have a known print to plot points more accurately. If the SCRO protagonists were to plot their interpretation of the McKie print, they could be able to use this model to state an impressive LR to support their claim. Imagine going to court with your print and LR, and meeting the defence expert who has used the same model to produce an LR with a magnitude several million short of yours. Perhaps in the future, the identification process will involve three examiners collectively working on a print, deciding where the points are, where to plot them etc in order to come to a mutually agreed LR.

LRs for DNA do not suffer from the interpretive issue that the fingerprint model does, and the similarity between the two fields pretty much ends with them both being tools to solve crimes, so why the big need to treat them both the same way when it comes to interpreting the results. Sure I'd love to be able to tell a detective or a jury, "it's real close, in fact it's probably him/her, actually it's 1 in a brazillion that the print doesn't belong to anyone else", but if I'm not totally convinced it's an ID, I don't know how that's going to sit next to the compartment that holds the warm fuzzy feeling that I get when I make an ID.

I'm sure that both the FSS and IAI have the best interests of the science at heart, and resolution or not, it will be business as usual for us all for quite some time yet.

A while back we were discussing the topic of a numerical point standard and I, among others, was of the opinion that not all points are 'equal' in terms of their discriminatory value. A point in the tip tends to be a bit more generic because it tends to flow along a single plane so to speak. I think alot of examiners, by default, tend to weigh certain points as being more 'active' or more 'passive' (when viewed in conjunction with the surrounding, supporting minutiae)....I speculate that AFIS systems favor numerically scoring a latent higher if point A is present 5 or 6 (or more) ridges away from point B as opposed to 1 or 2 ridges away (in lieu of any discrepancy or another event of course). I wonder how this may or may not play out in a statistical model as well? I'm with Gerald in that I won't make a guess as to how long this may take to appear....that is going to be a long hallway for sure in terms of work and the underlying theory as well.

I agree with Graham in that a % based inconclusive is not a good thing. But I look at it more in terms of what this may or may not do in the context of the justice system. Does the justice system want a 70% or 1 in X number presented to a jury during a trial where there are major ramifications for both sides?? I'm hesitant to embrace that kind of model...

With that said I am all for research and statistical analysis probably has to be examined as an option for latent print examination because we have the means to do so, but I don't think the resolution indicates we have to adopt a model for practical use if its not practical. Plus Research might be going in a totally different direction three or four years from now and who knows where its going to lead. Research away and the models can all be evaluated on their merit and practical use as they come down the pike....

One of the big breakthroughs in using statistics to improve quality in industry was the recognition that every process contains stable factors and unstable factors. Stable factors are present all the time and are predictable whereas unstable factors change from time and are impossible to forecast (but can sometimes be detected).

With DNA the stable factor is the theoretical random match probability of the technique whereas lab errors, contamination, mislabelling of samples etc will be unstable (there may be none of this for a long period then something changes in a lab to make it likely). It is possible that in reality lab errors are more likely to lead to an innocent person being accused than a random DNA match. Even if lab errors are extremely infrequent it may be more significant than some people think because errors tend to “self-select”. Erroneous forensic results will be of interest to the police because they cannot be explained innocently (this is Prosecutor’s Fallacy territory of course). So statements like "only 5 people in the USA would have the same DNA profile as the sample" could be interpreted in a way that is detrimental to the accused.

I think it could be useful for those studying statistics and fingerprints to give some consderation to the concept of stability.

In the meantime I think that prosecutors, juries and others in the justice system should know what it means in practical terms to treat expert opinion as valuable but not infallible.

Outsider wrote:I think it could be useful for those studying statistics and fingerprints to give some consderation to the concept of stability.

I agree, although I suspect the terminology may differ. I think resolving the relationship of 'stable' features in a fingerprint to the 'unstable', or variables, may be at the core of developing a successful statistical model.

In simplistic fashion, we might define those 'stable' features as general ridge flow, pattern type and focal points (if available), and (if limited to 'level 2 detail' as in most current AFIS systems), ridge endings/bifurcations, along with their relationships to one another.

Defining the 'unstable' features is more complicated. Not only are there a multitude of variables that can contribute to distortion, the predictability will of necessity be limited to the specific ridge configuration (including 'stable features'), and the clarity, both overall and specific area.

I see problems with quantifying those 'unstable' features, as well as determining when and how they may affect the 'stable' features, perhaps even turning some of them into 'unstable' features. Perhaps not an impossible task, but daunting nonetheless.
Regards, RLT

That’s not quite what I was meaning by an unstable factor.

Suppose you got some test latent marks / inked prints pairs and got everyone in a fingerprint department to analyze them. The fingerprint experts would not all agree on the number and type of similar features in the image pairs, there will be a variation (a range of results about the average) in the analysis, and this is quite normal and acceptable (this is not about what is the correct analysis, it is just looking at process variation).

Do the exactly the same test with the same prints in a different department or in six months time in the same department, you will find that the analysis is not the same. As long as the long-term average and range of analysis with the same test prints are within certain bounds, the variation is stable (it is quite easy to calculate these limits in manufacturing and some service industries).

Unstable variation is when something makes a significant change to the average or range. You might find that one department consistently reports more points of similarity than others, or there might be a long-term trend increasing or decreasing the number of points reported in a department, or some departments have a very narrow range of analysis between the staff while in others it is wide. Most processes are naturally unstable, you have to "manage" stability into them.

Any statistical analysis of fingerprints that includes the human beings in the loop will encounter variation of analysis. If the idea is to come up with a better way of reporting the closeness of a fingerprint match using human analysis of number and type of points of similarity, then the statistical model can only work with a basically stable process.

The IAI have taken a changed position to allow for the latest developments to be recognised but with a condition that if models don't prove effective and practical as according to the IAI they will not endorse the model. There are already considerable amounts of money and effort spent on this research and these workers needed to see some outcome for encouragement and further financial support. The fingerprint community has been patient awaiting the results as has the wider community in the field of criminalistics. Ofcourse other 'interested' parties will see this Resolution as fence sitting by the science. What do we get if we accept probability as part of our fingerprint conclusions? Does it apply to ten print individualisations? Does it apply to high quality latent individualisations? Or is it applicable only in the examination of latents which are poor quality or have poor quality exemplars? What do we get if we avoid the terms 'insufficient' from 'inconclusive' and substitute these with a Likelihood Ratio? The bottom line is that the term individualisation will be avoided and the door opened wider to allow for a different language for results. What do we get if we support the principles of the science? We have stability in the knowledge that each human develops fingerprint patterns to the exclusion of all others. The Law of Large numbers fully endorses the experience of our science in that no two humans will have the same fingerprints. The Biometric community also endorses this position and relies upon it. How do we know that fingerprints develop uniquely? Research from the relevant scientific community fully supports the principles of unique biological development. We must remember our friction skin is also a sensory organ, so is the biological development priority about our individual arrangement of neurons or the raised ridges? I wrote a paper about the spatial arrangement of features used for individualisation as a means of supporting the quantitative data available for presentation. Using levels one, two and three the model proposed that if only 5 level two features were present they represented a data set of 30 values. If a combination of one Level 1, ten Level 2 and five Level 3 were present these equated to 360 data values. The last example is if three Level 1, fifteen Level 2 and twenty Level 3 details were present they equated to 2109 data values. The point being what we get if the fingerprint data is objectively analysed there will be a large number of data values that can be used to support the conclusion of individualisation. This model is also supported by the data concerning the sequential information and visual appearance of the features used. What we get from a probability model is subjective since the arbritary figures used to assigned value to a selected set of features are conditional and cannot be falsified as the sampling population is currently too small as is the combination of features. Well I think that is what we will get. Research is healthy for any science so this Resolution keeps the modelling concept on the agenda and has breathed interest in our community. regards from oz, les.

Thank you for the clarification, Outsider. It appears that my suspicion regarding differences in terminology (or at least my interpretation) was correct.
Regards, RLT

g. wrote: It was NOT discussed at the business meeting when the resolution was passed by an OVERWHELMING (95+%) majority.

g., I'm wondering how many people were at the business meeting vis-a-vis the total membership. I only mean to suggest that a 95% vote of the members at the business meeting may or may not be a persuasive statistic. The last IAI business meeting I attended had about 100 members.

Langdon Alger wrote: I too fail to worship at the alter of probability models. We spent all this time working away from the image of being point counters just to return to a mathematical model that counts points?

Examiners have always counted points. Any examiner that says otherwise lacks credibility. What we have moved away from is basing our conclusions upon point counting. Furthermore, I submit that most latent print identifications are based on minutae . . . not exclusively . . . but you may say that the minutae do all the heavy lifting when it comes to evidentiary weight. And in many cases the identification can be made based on the minutae alone. Therefore, I do not see a minutae driven statistical model as being at odds with the current practice of latent print identification.

Gerald, I can only echo the previous poster's point. Science routinely builds models of complex systems on a galactic scale. I mean that literally . . . scientists have modeled the universe (that's pretty friggin' big)! The model may not be perfect, but that doesn't mean it isn't useful or that it can't help us get to the next level. This is the real key. Models are not helpful because they answer all our questions. Models are helpful because they tell us more than we knew before.

ian harrison wrote: LRs for DNA do not suffer from the interpretive issue that the fingerprint model does . . .

I disagree. DNA differs in that there is a pre-selected feature set (13 or 16 or whatever loci). I mean they are always looking for the same markers in every DNA sample whereas latent prints exhibit various feature sets. But DNA has its own issues with subjective interpretation. A "clean" DNA sample is like a latent print with exquisite clarity; no one disputes that the features are there. But what about a degraded DNA sample or a mixed sample? What if that DNA peak is right on the edge of the signal detection level? Do you count it or exclude it? Maybe you count it as signal when it supports your identification, and you exclude it when it does not (it's just background noise). Never discount the subjective, interpretative role of the examiner in a DNA comparison.

Outsider wrote: That’s not quite what I was meaning by an unstable factor.

Suppose you got some test latent marks / inked prints pairs and got everyone in a fingerprint department to analyze them. The fingerprint experts would not all agree on the number and type of similar features in the image pairs, there will be a variation (a range of results about the average) in the analysis, and this is quite normal and acceptable (this is not about what is the correct analysis, it is just looking at process variation).

Do the exactly the same test with the same prints in a different department or in six months time in the same department, you will find that the analysis is not the same. As long as the long-term average and range of analysis with the same test prints are within certain bounds, the variation is stable (it is quite easy to calculate these limits in manufacturing and some service industries).

Unstable variation is when something makes a significant change to the average or range. You might find that one department consistently reports more points of similarity than others, or there might be a long-term trend increasing or decreasing the number of points reported in a department, or some departments have a very narrow range of analysis between the staff while in others it is wide. Most processes are naturally unstable, you have to "manage" stability into them.

Any statistical analysis of fingerprints that includes the human beings in the loop will encounter variation of analysis. If the idea is to come up with a better way of reporting the closeness of a fingerprint match using human analysis of number and type of points of similarity, then the statistical model can only work with a basically stable process.

Outsider, the statistical model is not a self contained fingerprint identification machine, no. The L/R will vary depending upon who codes the minutae into the system. So you haven't eliminated the subjectivity of feature selection. What you have done is calculate an objective value for the minutiae you selected. This is still an improvement.

Les Bush wrote: What do we get if we support the principles of the science? We have stability in the knowledge that each human develops fingerprint patterns to the exclusion of all others. The Law of Large numbers fully endorses the experience of our science in that no two humans will have the same fingerprints. The Biometric community also endorses this position and relies upon it. How do we know that fingerprints develop uniquely? Research from the relevant scientific community fully supports the principles of unique biological development.

Les, I'm not sure where you are going with this? Is this the "Fingerprint IDs are absolutely certain because the skin is unique" argument? And that probablities are somehow inferior? It's about time to let go of the "uniqueness" argument in that the uniqueness of skin tells us little to nothing about the uniqueness of skin impressions. Indeed, statistical models are an excellent tool to investigate the discriminating power of latent print impressions.

16_Points wrote:Outsider, the statistical model is not a self contained fingerprint identification machine, no. The L/R will vary depending upon who codes the minutae into the system. So you haven't eliminated the subjectivity of feature selection. What you have done is calculate an objective value for the minutiae you selected. This is still an improvement.

Thanks for the clarification.

That, of course, leads to the observation that you have a very precise statistical model being fed by a feature selection process that is currently of unknown precision and is probably unstable (and therefore unpredictable).

It is clear from the evidence to the Fingerprint Inquiry last year that there has been a wide range in the analysis of Y7, not just between the two sides of the argument, but also within the “match” and “no match” camps. Terry Kent referred to trials he carried out some years earlier which showed the range of the number of features reported from different experts using the same latent mark/exemplar pairs. He mentioned one fingerprint where experts reported between 10 and 30 points of similarity.

It will be important to make sure that people who use the L/R output of the statistical model (which will have the look of objective mathematics) know that the number is just one subjective view of the closeness of the match. Another view from an equally experience expert could give an L/R that is orders of magnitude different.

Another hurdle will be to make sure people using the L/R really understand how to interpret it in the context of the case in hand. This is most important if there is little else to link the suspect to the crime. My experience over the last couple of years trying to make a very simple point about the McKie case to criminal justice professionals has left me wondering if they will ever avoid basic traps like the Prosecutor’s Fallacy.

I hope these hurdles can be overcome because a statistical description of the closeness of a match is potentially much more useful than a simple “match/no match” report.

As something of an aside, I really think the whole issue of probable or "almost" is indeed an issue almost purely because of the traditional position of latent print identification. There was long a fiction that any competent examiner would conclude similarly to another competent examiner. That was largely true but certainly was not often critically tested. When it is generally presumed that if an official examiner concludes an identification, the source has been correctly identified, few adversarial consultations will be sought. And when it is sought, it is likely that the consultant will be a retired official examiner. The result is that those identifications that are just barely sufficient are little more likely to be challenged than the lavishly sufficient identifications. Both the profile of the expert community and the practices of criminal defense tend to mean there are few practitioners to disagree and few identifications exposed to independent review.

It seems to me that this situation could not help but induce official examiners to think seriously about whether those cases that fall just short of sufficient can still be presented to provide some sort of probative value. After all, they seem to be soooo close to being uncontested identifications. But I suspect that if the profile of the examiner community was different, if it was either a reasonably balanced mix of official and private consulting examiners or, as is the case in many fields, private experts served both official and private clients, and if those experts were routinely available to both advocates, it might well be understood that close call identifications would likely be disputed. And the relatively weak conclusion of "almost" an ID would not be very attractive, being even more certain to be vigorously disputed to the point where it had little weight. In effect, in the mind of the fact finder, every identification that is disputed by reasonably credible demonstration and argument becomes a likelihood problem, a matter of judging probability in the face of two credible experts who disagree. A great many more would confront them with the problem, were the professional and legal landscape different.

To put it another way, I submit that if you declare (in whatever way) an "almost" identification, you are inevitably declaring that some examiners would conclude it to be an identification, and others would not. It happens that you're not taking it around to find those who would call it an ID, but it's naive to believe that you wouldn't find them if you looked. The threshold for concluding identification is undefined and therefore variable among examiners, every examiner merely stating a belief that that undefined threshold has been exceeded in this case to form a belief in an identification. In a real sense, no actual threshold exists, except for that one comparison at the instant the belief is formed. That threshold is known only by its effect, not its substance. It's ALL really probabalistic and always has been. But examiners were comfortable, because they weren't confronted with a valid calculated likelihood that they were committing to as being the number that caused them to believe in the identification.

There are purposes within the formation of our friction skin patterns. Often we overlook the sophistication of the sensory component and its role in the resultant pattern design. By doing so we underestimate the power of the pattern to differentiate every human current and past and future. The expression of uniqueness in our skin is an opportunity to provide absolute conclusions about the source of a fingerprint. The closer the clarity of the fingerprint detail is to the actual skin the easier it is to make a conclusion. We know that the use of probability models will not apply to ten print examinations because the design of the model is not robust enough. So for every ten print or high quality latent no probability model can ever exists to replace the certainty of uniqueness in the pattern detail. The combination of skill and experience in becoming an expert in fingerprint pattern recognition means this ability is capable of differentiating the levels of detail in a fingerprint pattern and determine their use in a conclusion. A probability model can only be applied to a known reference configuration of limited number of detail and the power of the calculated confidence represents use of arbitrary figures based on mathematics determined by a mathematician. What added value does this probability estimate give to the conclusion when compared with actual geometrical data values taken from the pattern configuration which provide direct relevant results? To be a fingerprint expert requires knowing the fundamental causes of skin formation and how to use that pattern arrangement in determining source attributions. Taking the expression of uniqueness out of fingerprint science is not on and cannot be acceptable to the relevant scientific community. Regards from oz. Les

Hi There

Steve said 'It will be important to make sure that people who use the L/R output of the statistical model (which will have the look of objective mathematics) know that the number is just one subjective view of the closeness of the match. Another view from an equally experience expert could give an L/R that is orders of magnitude different.

Another hurdle will be to make sure people using the L/R really understand how to interpret it in the context of the case in hand....'

This is my main concern, Steve has clear articulated what I see as being the main issue, and I know I have been stirring with a large wooden spoon for some time now to see if anyone in the fingerprint world can see this. It took some one from the 'OUTSIDE' to point it out. I knew there was no way I could state it so clearly and it would be coming from someone without credibility, so no one would listen.

Thanks Steve you have change my opinion of you, I am sorry I had some negative ones.

Fingerprint Experts are going to have to learn a lot about statistics to be able to use the L/R output of the statistical model in evidence. There are highly qualified people out there who will make mince meat of a fingerprint expert who tries to use objective mathematics in a court room, look no further than Steve.

Being in my 60's I will not ever need to give a statistical model in evidence, but there are heaps of fingerprint experts who are excellent at there job and will possibly not cope with this.

Les I totally agree with all you said yesterday. Do you think this is a training issue in the states.

And I would rather get up in court and say this impression was made by the same finger as the one recorded on the prisoners ten print record. Than get on to the slippery slope of well it maybe his or it may not be I am not sure but I have this printout that shows that there is a high likelihood of it being his finger, no I am sorry I can not explain the print out but I am sure it is accurate as I feed in the subjective data and yes the output may as a result be subjective, and well yes it still maybe someone else's. (Oh and by the way for those who can't recognise this as sarcasm please do not make a fool of yourself responding to this paragraph.)

I am not as it appears to be totally against statistical modelling, I have been known to use it myself, I am just so very pleased that I wasn't interrogated by defence on what I stated. I did say that I was not an expert in such things and it was an answer to a question so I guess the defence was smart enough to leave it at that, maybe he knew as much as I did on the matter so left it alone.

I see it as a band aid over a larger issue, one of training or lack there of. Yes the results may be of interest just like the percentage of whorls in the population or how much weight we place on a spur or trifurcation in an identification, sorry must change my ways, in a high likelihood comparison.

But at the end of the day what makes fingerprint identification unique is the ability to identify some one; unlike DNA we do not have problems with identical twins, bone marrow and blood transfusions, very little issue with shedding and transferring.

A shame we can't put as much effort into bias. That is the biggy in my mind.

g (with a lower case) please explain 'but if you want to debate probabilities, no need to hide behind a pseudonym...we're all reasonable professionals here). So what is 'g' then if not hiding behind a pseudonym? And in my last input into the probabilities debate was your response to me very PROFESSIONAL. I really had the feeling that I had stepped on your foot with my size 11 boots and you took personal issue with it. May I ask what the going price of KIWIs are in the USA or maybe you could offer to pay for my airfare to the states to go to the workshop (you run?), I would love that and we could have a beer, I'll shout you and discuss the issues one to one, you may change my mind. I maybe arguing into a mirror g but at least it is a two way mirror.

Gerald as a point of interest can you give an example of how you would state your findings in a 'brief of evidence' using a likelihood ratio or percentage figure, which ever you prefer. I am hoping for something that the average jury will hear and understand, obviously something that will not introduce a string of questions from defence. This maybe a big issue as the average fingerprint expert will need to feel comfortable with the statement.

I posted the comments below under another topic a few months back, after reading this thread my thoughts on the subject have not changed (other than conceeding that Cedric & Glenn's dance skills must have improved since 2007 based on Ian H's comments, I know mine haven't - I obviously don't get to enough IAI Conferences

(BTW - smilies, quotes etc don't seem to work on our Section PCs so unfortunately some of the fancy bits are lacking)

My concern with this computer generated statistical model - shared with Glenn & Co at a course in 2007 - is this...

We are looking for 'agreement in sequence, having sufficient uniqueness (detail) to individualise'. Where errors such as McKie, Madrid etc. seem to mostly occur is in the 'agreement' of the detail - not in the 'sufficiency' of the detail. (Not that I agree with the constant suggestions that we have a large number of errors lurking in the closet.)

As much as the research being done by the likes of Glenn, Cedric, Christophe & Paul is interesting, to me it does not address the real concern of being confident that we do have agreement when we say that two prints have a common source.

Therefore there seems to be the potential for prints to be plotted on this system and a high likelihood of a match is then generated, yet a competent latent print examiner could clearly find that the two do not have a common source.

The additional suggestion that this system can give us confidence in those marginal matches that have agreement but lack that sufficiency for us to be completely confident is also of concern. Firstly I don't find this to be a regular occurance, in fact it would be extremely infrequently and secondly this model does not (unless things have changed) take into account third level detail.