Thanks a lot Keith, for this most helpful insider’s information!

]]>But seriously, he decided on self controls (comparing earlier periods to the period the accident occurred in) and recruited Rob Tibshirani to do the statistical analysis and be the co-author. I (privately) reviewed the work for Donald (as he was a fellow of my director) and he is aware of the challenges but I think he was always a bit too positive about the approach.

I am not going to read the current work, it seems to be an n’th kick at the same type of design. It won’t be naive though not fully convincing but unlikely others will be able to improve on it. One has to accept unquantified inference risks if you don’t/can’t randomise.

]]>Regardless of whether it sounds or would convince a frequentist, frequentist philosophy of statistics certainly spawned it.

At heart frequentism is the claim that the product rule P(A & B) = P(A|B)P(B) only applies for some well defined A’s and B’s but not others. It applies (supposedly) to something they call “random variables”. Without that restriction, you immediately get full Bayes whether you like it or not.

There is nothing in the mathematics to suggest, let alone require, such a restriction. Moreover, Frequentists have never given an objective operational definition of “random variable” which can be used to decide what’s an RV and what isn’t.

Next week’s stock prices are conventionally treated a “random variables” and frequentists will assign probabilities to them. The outcome of the Presidential election in 2016 is somehow not a random variable and so frequentists wont assign a probability to it like Nate Silver successfully did in the 2012 election.

Neither next week’s stock prices or the outcome of 2016 election are repeatable. Frequentist just seem to have an undefinable intuition, which changes over time, and differs from one to the next, about which variables are RV’s.

The advantage of the multiverse for frequentists is that many formerly “non-repeatable” variables suddenly become “random variables”, and frequentists can magically assign probabilities to them. The variables haven’t physically changed in any way. All that’s happened is that frequentists invented a wild fiction which instantly makes them comfortable using the product rule.

And all of this rigmarole, together with the other millions of man-hours wasted deny the applicability of the product rule, is done for no other purpose than to save Frequentists the trouble of thinking hard about what P(A) and P(B) might mean in general.

What a sad, stupid way to do science.

]]>