Comments on: Brazil, world centre of terrorism

By: dsquared

dsquared — Wed, 27 Jul 2005 06:31:00 +0000

I see what you mean, but surely if we think that the watch list is reasonably comprehensive then we don't want to set very much weight at all on the "seems like" condition? It looks like a hierarchical model to me; if an identified watch-list member is behaving like he's got a bomb on him, then shoot. Otherwise you're in one of two cases; a watchlist member with no particular reason to suppose he's got a bomb (don't shoot) or someone you've never seen before acting like someone with a bomb. The difficult case is the third one but even here I'm not sure that a checklist (and associated tradeoff) is really the way to model it because the decision is going to be made on the basis of the gestalt; all things considered, do I have a bad (enough) feeling about this guy?

I would also suspect that you would have to be really quite pessimistic about a) the total population of suicide bombers and b) the proportion of that population who were on watchlists before you were a situation where it was the right thing to do to give orders for dealing with case 3's under which you expected to have material numbers of people getting shot[1] on the basis of behavioural scoring.

[1]I suppose that a single police shooting is a material number in this case but even so.

By: Chris Lightfoot

Chris Lightfoot — Wed, 27 Jul 2005 06:13:00 +0000

Daniel -- so far as I know (and I haven't seen the relevant document) the orders are of the form "if you believe that somebody is a suicide bomber and need to stop them from blowing up themselves and others, then shoot them in the head". I hope it's not an order of the form "if you see Billy Bloggs, shoot him"! You might well know that Billy Bloggs has prepared to be a suicide bomber, but that doesn't mean that a police officer should shoot him dead if they see him on the street obviously without a bomb.

So now we've got two conditions: on watch list / not on watch list, and seems like a SB / does not seem like a SB. But we still need to make exactly one decision....

By: Chris Williams

Chris Williams — Tue, 26 Jul 2005 17:50:00 +0000

This is all getting very yellow Mini - you'll remember that the guy who got shot by the Met then, Stephen Waldorf, was hanging around with the bad guy's girlfriend, and even looked like the bad guy. But he wasn't the bad guy.

It looks like the IPCC* are going to be investigating the context of the shooting, which is good.

*Is is just me, or do these initials fill anyone else with a subconscious feeling of safety, derived from long years spent reading Jack Vance's space operas?

By: dsquared

dsquared — Tue, 26 Jul 2005 09:23:00 +0000

Well not necessarily. It strikes me that it’s a triage rather than a separation of distributions. There are people who are definitely suicide bombers and are identifiable specifically (the one-factor model "is he Billy Blogs? yes/no"), and then there are people who might or might not be suicide bombers and need to be separated on the basis of characteristics. If you know the population of SBs is small and you have a functional intelligence-gathering process, then it strikes me that the best use of intelligence would be to work on increasing the comprehensiveness of the first category rather than refining the discrimination in the second. I suppose that this would still leave some discrimination effects because there are cases of mistaken identity but I don’t think that this is well described as a tradeoff; there is clearly a model under which you gain true-positives at the expense of false-positives by changing the orders from "shoot Billy Bloggs" to "shoot people who you are more than 80% certain are Billy Blogs" but I don’t think it’s psychologically plausible. Terrorist hunting looks much more like case-by-case underwriting than actuarial science to me.

By: Chris Lightfoot

Chris Lightfoot — Tue, 26 Jul 2005 09:11:00 +0000

Sorry to keep going on about this, but...

There's lots of things that we can do that would decrease both false positives and false negatives (like get better intelligence) and no guarantee that simply turning up the sensitivity of the half-baked SB-detector we have will increase the number of true positives.

So, in this model, getting more intelligence information will increase the number of factors, separating the two distributions more. There will be places where the marginal effect of changing the decision threshold on the number of true positives is small (at the center of the distribution for the suicide bombers) but that's not necessarily where you'd put it anyway, is it?

By: Matthew

Matthew — Tue, 26 Jul 2005 08:01:00 +0000

Michael – On Sean’s points, I think it’s the last three that must have been crucial. The Police let him get on a bus, so couldn’t have been that concerned about him then. They also (apparently) as he got off the bus approached him with standard warnings, which also must mean they didn’t expect him to explode himself. I suspect they were planning to arrest him at that point still.

I think that’s a good thing, as the first six points I don’t think are that unlikely to be repeated (with an addition of ‘s’ to the day in point 1, and deleting everything after ‘flats’ in point 2 (and yes it does matter), and perhaps generalising 6).

By: dsquared

dsquared — Tue, 26 Jul 2005 07:41:00 +0000

Well … yeh, but only in the sense that there is a Laffer curve existent in some Platonic realm if we had an optimised constant-single-marginal-rate tax system. If we were on the efficient frontier, with a list of characteristics that provided the maximum differentiation between SB and non-SB, then I can see how raising or lowering the number of ticks needed to shoot someone would trade off false positives against false negatives. But as it is, we’re nowhere near that. There’s lots of things that we can do that would decrease both false positives and false negatives (like get better intelligence) and no guarantee that simply turning up the sensitivity of the half-baked SB-detector we have will increase the number of true positives. To me it resembles a credit-scoring problem; if you want to lend to a riskier class of borrower, then dialling down the credit standards on your existing scoring model is usually the wrong thing to do.

By: Chris Lightfoot

Chris Lightfoot — Tue, 26 Jul 2005 07:18:00 +0000

ah, ok, so your statement was based on the assumption that for each ‘indicator of suicice-bomber-ness’ the probability that it would match an innocent person is extremely small, so that the distribution of scores for suicide bombers doesn’t overlap much with the distribution for non-suicide-bombers? But AFAICS that doesn’t eliminate the tradeoff — it just means that the equal error rate can be very small.

By: dsquared

dsquared — Tue, 26 Jul 2005 06:42:00 +0000

Hum but even the Israelis who hand that leaflet out don’t use it as the basis for shooting people five times in the head without asking questions (actually, digging around, the sudden grab-and-shoot approach is much less common in Israel than you’d think). If you’ve reached the point where a) suicide bombers are so prevalent and/or b) intelligence-led operations are so off the menu, then I would have guessed that things would have got so bad that it was probably time to be thinking about how awful it would really be after all for Sevenoaks to be part of the Caliphate.

By: Chris Lightfoot

Chris Lightfoot — Tue, 26 Jul 2005 06:34:00 +0000

I'm going to write a longer post about this.

ditto....

I'm not so sure that such a model exists; you can't increase Y by tolerating a higher X and most policies you can carry out to decrease X won't have the effect of decreasing Y.

OK. I was thinking about the model where we have various suicide-bomber-like behaviours that we watch for (see that Israeli "how to spot a suicide bomber" advice leaflet which I can't find right now, or the thing Michael Brooke quoted earlier for examples) and use to determine who is innocent and who is guilty.

Now, if those suicide-bomber-like behaviours also have possible innocent explanations (i.e., they're things like "is wearing a heavy coat during the summer" rather than "has been observed by police surveillance cameras making a bomb") then we're basically guaranteed that X and Y trade off against one another, aren't we? And in the case where the police are overstretched, don't really know who they're after and fear an attack in the very near future, the tests we user are going to be quite susceptible to false positives.