A mechanistic approach to Bayesian probability could be seriously misleading. For example, every year on May 19 I have a birthday. This has been happening now, year after year, for 88 years. I do not interpret this as raising the probability of my continuing to have birthdays for the foreseeable future. In fact, I would draw the opposite conclusion.Wilfred Beckerman, “I foresee fewer probable birthdays”, letter to the editor,

Financial Times, 8 October 2013.

Bayesians use historical information to attach some prior probability to a hypothesis, whereas non-Bayesians typically test a hypothesis without assigning a prior probability to it.

UK economist Wilfred Beckerman is Emeritus Fellow of Balliol College, Oxford University. He is commenting on Samuel Brittan’s column “Economics is now an argument about two numbers“, published in FT on October 4th (ungated link).

See also two previous posts, here and here.

**Update**:

I beg to differ with Professor Wilfred Beckerman (Letters, October 8). The model he implicitly uses, in his example, is a Bernoulli trial. This is not appropriate as the trials – having a birthday or not having a birthday – are not independent. The birthdays are just a product of another more fundamental stochastic process: the length of his life, which one hopes will be a long one.

A proper Bayesian analysis would first seek to obtain the probability that he would live to a certain age and from that obtain the probability of his having another birthday.

Bayesian forecasting and analysis is now mainstream and is used very successfully in many fields. Autonomy, owned by Hewlett- Packard, is based on Bayesian logic.Gerry Clarke, “Bayesian analysis, Bernoulli trials and (probable) birthdays“, letter to the editor,

Financial Times, 14 October 2013.