Less Wrong is a community blog devoted to refining the art of human rationality. Please visit our About page for more information.

bigjeff5 comments on The Amanda Knox Test: How an Hour on the Internet Beats a Year in the Courtroom - Less Wrong

42 Post author: komponisto 13 December 2009 04:16AM

You are viewing a comment permalink. View the original post to see all comments and the full post content.

Comments (632)

You are viewing a single comment's thread.

Comment author: bigjeff5 02 March 2011 06:33:25PM 3 points [-]

I'm ashamed to say I failed this one, and badly.

I'm new to the Way of Bayes, and I've been reading the sequences regularly (that's why I'm on this post, actually) and thinking I've been doing pretty well, but this illustrates that I have a long, long way to go.

I did do a whole lot better than the jury in this case, such that I would never have convicted Knox and her boyfriend, but I still came away with about a 60% probability of their having been guilty, with Guede at >95%.

I started with the pro-guilty website, and thinking back I clearly privileged their hypothesis. In fact I had only assigned about an 80% probability that they were guilty after reading their evidence, which would not have been enough to convict. They made scant mention of how little DNA evidence was actually used to convict them (even calling the braw-clasp significant, damning evidence), and they certainly didn't compare this to the massive amounts of DNA evidence found for Guede. In fact they made it sound as though the evidence was equivalent.

I then updated based on the pro-defense website, down to 60%, a clear case of anchoring. Had I considered the defense's evidence from a fresh perspective there is no way I could have come away with the estimate I did. The overwhelming DNA evidence against Guide should have relegated all other evidence to almost negligible importance. Everything else was unreliable, and I did not treat it as such even though in my mind I even thought that it was unreliable.

A good analogy for the DNA evidence compared to everything else would be like trying to star-gaze at high noon - the light from the sun overwhelms the starlight because the starlight is far, far too weak. Likewise the circumstantial evidence was far too weak to be of any importance when compared to the DNA evidence.

Comment author: komponisto 04 March 2011 03:57:58AM 1 point [-]

I'm ashamed to say I failed this one, and badly...this illustrates that I have a long, long way to go.

On the other hand, you'd be surprised at how difficult it is for people to understand this even after seeing the explanation. While of course it's always preferable to get things right on the first try, the fact that you saw it so clearly after reading this post places you in a high percentile for learning speed.

(I know whereof I speak here, because I've spent some time arguing about this in other parts of the internet, only to be dismayed by the true extent of people's ability to clack. Heck, even here in the comments section of this post there are some pretty confused comments with depressingly high scores.)

I think your analysis is excellent, and is an admirable example of "cognitive debugging". This part in particular is worth dwelling on:

Everything else was unreliable, and I did not treat it as such even though in my mind I even thought that it was unreliable.

It is really difficult for people to discount information, and to actually update on evidence of the form "that information may not be reliable", rather than treating it as an "opposing argument" to be rationalized away, or simply forgetting it altogether. Something like the bra clasp provides an example of this: people hear about it, and decide that Sollecito is guilty. Then, when they hear the defense argue that it's unreliable, they mentally dismiss it as "the defense trying to poke holes in the prosecution's case". They may respect the social obligation to provide some sort of answer to the defense's arguments, but they don't consider that the defense arguments should actually impact their belief about whether Sollecito is guilty; the belief-formation process was over and done with at the first stage!

It's not that they explicitly, consciously think like this, of course; rather, they have just failed to fully incorporate the "rules" into their unconscious belief-generating system. The third and fourth virtues are hard.

I like the star-gazing-at-noon analogy, and may use it myself in the future.

Comment author: bigjeff5 04 March 2011 05:07:27AM 2 points [-]

I think your analysis is excellent, and is an admirable example of "cognitive debugging". This part in particular is worth dwelling on:

Everything else was unreliable, and I did not treat it as such even though in my mind I even thought that it was unreliable.

That's what upset me the most. I made a special note of it, in order to give it a very low weighting, yet it was still a major factor in my assessment. It had to be, because the prosecution's case hinged on it.

I think next time I notice something like that I will have to assign it a specific weight as soon as I decide that it should be given a low weight. Instead I just labeled it as "weak" in my head and moved on.

So when I think "that's likely true" I should stop and consider what the actual number is that I would assign to it. This would help considering in spotting cases where I assign more than 100% total probability to an outcome. If I am thinking it is 70% likely that they are guilty and 60% likely that they aren't, there is an obvious problem. If I don't assign any numbers until I'm finished looking at the problem then I'll never spot this error, as I'll simply calculate not-guilty likelihood from the guilty likelihood.

Much to think about and much to learn still.