posted 12-13-2007 03:41 PM
Answer to #2 is easy.
Training is generally (always) just add up the numbers for each spot.
I'd like to see that change, and I'll explain more in a bit. Krapohl (1999) demonstrated that doubling EDA scores can reduce inconclusives, without concerns for increased errors. OSS-3, and probably all other computer scoring models weight the EDA more heavily than the other components.
Answer to #1 is not so simple.
First, why is the EDA the only channel with reactions. Health? Medications? Movement or behavior artifacts?
Is your subject really a fit or suitable subject for the polygraph technique? Or, is a marginal candidate, for whom we should be unwilling to provide an unqualified opinion (i.e, without some form of CYA explanation that says "don't even think about arguing the validity of the test results in court.") Just look at the Deskovic mess at anti.
Both hand-scoring and computer scoring models are non-robust against bad data. Bad means not just messy, but unresponsive.
A test of this type might take care of itself, with weak numbers that produce inconclusive results.
Let pretend its meds, and your guy is a garden-variety sex offenders with depression, diabetes, high blood pressure, and allergies. That should give us a good combination of cholinergic, sympathomimetic, and adrenergic effects to goof with our data.
How confident will you be in the results? Would you say he's a fit subject?
Would you test him?
Depending on why, I mikght test him. But I'd also call him a marginal subject, and I'd qualify the results by listing his meds and diagnoses, and describing the quality of the test data.
Hand scoring systems often involve the use of ratios for assigning points, 2:1, 3:1 and so on. Handler and Krapohl showed that there is greater diagnosticity extant in smaller reaction differences, and we all know that the search for larger reaction difference is wishful thinking.
We also have the venerable Bigger-is-better, which I believe DACA describes for both 3 and 7 position systems. (correct me I am an wrong on that.)
One could argue that the 3 position system is theoretically more robust in its application, in that it relies primarily on the bigger is better rule, and makes no assumptions about the linearity of the data - which is important in consideration of the interfering influence of the handfuls of meds some clients take, and is equally important in consideration of the fact that the various polygraph equipment manufacturers have provided us with field equipment that often has unknown linearity.
We still wonder though how much bigger matters, and at what point to small differences become excessively noisy and unreliable.
So I took the time to study the bigger is better situation with some ROC plots.
This is the end result, and the plots illustrate the results with the OSS-3 training set N=292 for each of the components, used alone, using a ratio of 1.1:1. Below that level they get quite noisy.
ROC AOCs are
Pneumo = .75
EDA = .95
Cardio = .77
1.1:1 means that if you can see a difference and could argue that one bump is bigger than another, you can probably score it reliably.
While the cardio and pneumo look close, other experiments reveal roughly twice as much variability in pneumo data compared with cardio data.
That .95 for EDA is impressive and suggests that the EDA alone might be a potent diagnostic indicator. It is tempting to want to discard the other components but our data suggest the polygraph is better with the cardio and the pneumo components. But this strong AOC value helps us understand why it improves the results when we weight the EDA values as more important than the others.
Now wouldn't it be great if there were a very simple way of putting these pieces together in efficient field practices, that is easily understood and provides good interrater reliability.
"Gentlemen, you can't fight in here. This is the war room."
--(Stanley Kubrick/Peter Sellers - Dr. Strangelove, 1964)