GiveWell have an interesting post arguing against what they call the “explicit expected value formula” approach to giving: http://blog.givewell.org/2011/08/18/why-we-cant-take-expected-value-estimates-literally-even-when-theyre-unbiased/

It covers many subjects of interest to people here: Giving What We Can's recommendations, existential risk charities like SIAI, problems like Pascal's mugging etc. It's also been crossposted on LessWrong where it's generated a long comment thread. Perhaps people'd prefer to comment there or at GiveWell, but either way I'd love to hear people's thoughts - I'll aim to post mine later.

It seems to focus too much on normal distribution. Charities in particular seem to have something closer to a log-normal distribution. The effectiveness of charities can easily vary by orders of magnitude. Thus has a much larger upper tail than a normal distribution, and thus would allow much higher answers.

DanielLC wrote:It seems to focus too much on normal distribution. Charities in particular seem to have something closer to a log-normal distribution. The effectiveness of charities can easily vary by orders of magnitude. Thus has a much larger upper tail than a normal distribution, and thus would allow much higher answers.

That seems plausible - it'd be interesting to see an application of this Bayesian approach to a real charity, like SCI. I'm a little unclear on how I'd set probably distributions for my 'prior' view of all charities' cost-effectiveness and my 'estimate' of SCI's cost-effectiveness. My existing view of SCI's cost-effectiveness makes me think the prior distribution has a large upper tail, but how would I factor this in?

Another thing I'm unclear about is how the figure of X/(X^2+1) is derived as the value of giving into a Pascalian mugger.