If you are interested in this subject, check out "The Sociopath Next Door". It changed my perspective quite a bit.
Finding Sociopaths on Facebook
51–60 of 73 posts
Re: Finding Sociopaths on Facebook
#52Earlier quoted context omitted.
Most likely nothing. It depends on what sort of test it is; if it's something non-intuitive like (say) a habit of writing sentences that always have a prime number of words in them, you'll get your false positives but most those people won't pass any other tests, whereas the actual terrorists will. What Schneier is missing is that while you can't ID people that well from a single test, you can apply a bunch of them.…
Waiting until someone actually commits a crime provides a list that is 100% accurate.
I can just see the Daily Mail headline now:
"Facebook records reveal convicted killer wrote 13-word post 5 years ago - red flag was raised - why was nothing done?"
Re: Finding Sociopaths on Facebook
#53Unfortunately the standard isn't even a lackluster 80%. We've learned that the 'go to' standard is 51%.
Re: Finding Sociopaths on Facebook
#54Two words: Bayes' theorem. I think I learned this in the fourth or fifth week of my first probability course. Thought it was a bit more commonly known.
Re: Finding Sociopaths on Facebook
#55Earlier quoted context omitted.
Most likely nothing. It depends on what sort of test it is; if it's something non-intuitive like (say) a habit of writing sentences that always have a prime number of words in them, you'll get your false positives but most those people won't pass any other tests, whereas the actual terrorists will. What Schneier is missing is that while you can't ID people that well from a single test, you can apply a bunch of them.…
Waiting until someone actually commits a crime provides a list that is 100% accurate.
No, it doesn't. People commit crimes and get away with them al the time, in fact. I'm not proposing that we put people in jail for having criminal potential.
Re: Finding Sociopaths on Facebook
#56An excellent post by Schneier. > The problem isn't just that such a system is wrong, it's that the mathematics of testing makes this sort of thing pretty ineffective in practice. It's called the "base rate fallacy." Suppose you have a test that's 90% accurate in identifying both sociopaths and non-sociopaths. If you assume that 4% of people are sociopaths, then the chance of someone who tests positive actually being…
This is a pretty bad example to illustrate the fallacy because a 25% confidence is actually extremely good. I don't know what good a test for sociopathy is, but if we had a test this good at identifying terrorists, it would be incredibly useful. If signals intelligence could produce a list of people and guarantee that a quarter of the people on that list are terrorists, it would absolutely revolutionize law enforceme…
He got 25% starting from a base rate of 4%. The base rate for terrorists is a little lower than that. I agree that the essay probably ought to emphasize this point so skimmers won't take away the wrong idea.
Re: Finding Sociopaths on Facebook
#57An excellent post by Schneier. > The problem isn't just that such a system is wrong, it's that the mathematics of testing makes this sort of thing pretty ineffective in practice. It's called the "base rate fallacy." Suppose you have a test that's 90% accurate in identifying both sociopaths and non-sociopaths. If you assume that 4% of people are sociopaths, then the chance of someone who tests positive actually being…
I vote for just using real numbers instead of percentages, like "0.4 chance" instead of "40% chance". The reason is that a lot of people get the math of percentages wrong simply because it involves a lot of back and forth mental conversion between percentages and fractional representation. I always found it easier to just use the latter.
Re: Finding Sociopaths on Facebook
#58Earlier quoted context omitted.
This issue has nothing to do with the efficacy of the tests. In a perfect world having a list where 25% of the names are potential trouble-makers (for various definitions of trouble-maker) would be an enormous benefit and allow resources to be targeted more efficiently. The real problem is what happens in the imperfect world where law enforcement, government, self-imposed officers of authority get lazy or down-right…
> In a perfect world having a list where 25% of the names are potential trouble-makers [...] ... would be pointless, as a perfect world would have no concept of "trouble".
"in a perfect world" is an English phrase. It means, in this context, "if this worked perfectly".
It is not a statement about the world in general.
Re: Finding Sociopaths on Facebook
#59Earlier quoted context omitted.
I found his version clearer because "%" distinguishes proportions from quantities.
Yes, but imagine you're taking the test. Suppose you have a test that's 90% accurate in identifying both people with X and people without X. If you assume that 4% of people are people with X and you're told that you test positive for someone who has X. Do you really find it easy to arrive at your actual chance (26%) of having X? Let's not forget that most people on HN are at the smarter end of the bell curve. It'd be…
not sure how to calculate this with percents
true positives .90 * .04 = 0.036 false positives .10 * .96 = 0.096
total positives 0.132
positives that are true positives 0.036 / 0.132 = 0.2727...
i had to think about the calculation as i was doing it wasnt automatic even though it was just multiplication, but I think the difficulty is more to do with the fact that you have to use some relative of bayesian probability not really the fact that you had to deal with percentages
Re: Finding Sociopaths on Facebook
#60Here's the real problem: suppose you have a bunch of data on people and are convinced that somewhere in the data exists the spectrum from sociopath to non-sociopath.
Who is going to label the data with ground truth? Clinicians who "know it when they see it"? What is the ground truth that the classifier is going to train on?
If you're going to do an unsupervised classifier (eg clustering) who is going to label the clusters? What is going to keep the data from turning into uncorrelated mush?