The Skin Cream Testing Paradox

Over at Bad Science, in a discussion of the recent BBC Horizon program on cosmetics, Ben Goldacre describes what I shall call the Skin Cream Testing Paradox:

But this is the astonishing thing. Other manufacturers don’t submit their creams to a university dermatology lab, or publish academic papers: so these aren’t so much the best results, they’re the only ones. Why? The tests cost just £15,000. Cellular Radiance Cream by La Prairie, for example, costs £340 a pot. Fifty pots would fund one study.

But those tests could bite any company right back: because if they show no effect, then your business is trash; but if they do show an effect, then busybodies wade in to regulate your pharmaceutically-active product. If it was my cosmetics company, I’d stick with the sciencey diagrams and hope for the best.

That is, if testing shows the creams to be effective then the manufacturers are subject to costly regulations, and if they are shown to be ineffective then the manufacturers have to admit that they are conning their customers.  Thus the manufacturers have no incentive to fund the testing of the efficacy of their own products.  They chose to keep their heads firmly in the sand.

However, what has not yet been pointed out is that each manufacturer has a strong incentive to fund tests on its competitor's products.  I wonder if there is a simple way to trigger an outbreak of such competitive cross-testing?

Corruption and Parking Tickets

Statistics can be a wonderfully subversive tool.  Take this discussion by Andrew Gelman of a paper by Ray Fisman and Ted Miguel of the correlation between unpaid parking tickets of United Nations delegates and the level of corruption in their countries.  I think someone should publish league tables!

Building Large Systems in C (and C++)

From Steve Yegge's old Amazon blog:

If you're trying to build a large system in C, you'll inevitably wind up making a choice between two alternatives:

1. Use the AlternateHardAndSoftLayers design pattern, in which you write most of your code in a high-level language, and the performance-critical parts in optimized C.
2. Try to make C a high-level language itself, which either leads to the madness of choosing C++, or the madness of implementing your own high-level language in C, without any actual syntactic support.

Most companies choose the second option, usually because they haven't learned the hard way that 90% of your code never needs to be optimized, and doing so won't produce any human-observable difference in runtime performance.

Indeed. The use of C++ for the whole of a large system can be seen as a 'premature optimization' in the sense of Donald Knuth.

What happens when you generalize Probability Theory to allow Minus Signs ...

Over at Shtetl-Optimized, Scott Aaronson has a nice introduction to the field of quantum computing disguised as a series of answers to questions he was asked on his recent job interview tour.  I particularly like the following characterization of quantum computing as:

what happens if we generalize probability theory to allow minus signs, and base it on the L₂ norm instead of the L₁ norm

I now feel I know enough to be able to talk authoritatively on the subject over coffee at work.

Owls at Ascot Station

On Friday afternoon I was waiting for my train at Ascot when I noticed this owl pellet on the platform, just beside a lamp-post.  During the previous night, an owl must have alighted on the lamp-post to regurgitate the hard, indigestible parts of the animals it had eaten.  The pellet appears to mostly consist of beetle wing-cases, and the hair and bones of a small shrew or mouse.  You can see the shoulder blades of the latter in the lower left corner of the photo.