Wednesday, December 04, 2013

The FDA needs to be more Bayesian

John Wilbanks is one of my favorite open science geeks (I even met him once in person, when he did a presentation for Microsoft Research). In a blog post this week, he writes a well-reasoned explanation for why the “Evil FDA shuts down lone entrepreneur” narrative is wrong.

Tech companies like 23andme, he writes, think in a “Bayesian” way, where the safety or “truth” of a medical claim is a probabilistic concept that depends on the number of data points (i.e. users). This doesn’t sit well with the FDA, to whom Truth is a binary fact: something is either safe or unsafe, period:

That “traditional” submission to the FDA would be of a very specific kind of analysis based on randomized controlled trials. It is designed to keep bad things from happening to people, not to make sure good things happen to people.

He concludes, correctly, that this is a clash of cultures and that if 23andme wants to succeed (and he hopes they do), they need to accept reality. This is how the FDA works. They should have known that:

[U]ntil the FDA learns how to deal with Bayes’s rule and its discomforts - and until DTC companies figure out a business model that isn’t based on massive loss leadership - we’re going to keep coming back to this clash of culture and business models. Both sides need to make some changes if we’re going to avoid doing this over, and over, and over

But why must “both sides” make changes? I’m reminded of similar advice given to Chinese dissidents that they need to “work through the system”, rather than make public their often misleading and “socially irresponsible” opinions.

I am able to make up my own mind about 23andme’s “marketing” claims, and so can you. The FDA, regardless of how understandable their position, is wrong. Shouldn’t those of us who believe in open data just say so?