Whereas I am not much concerned about the general grasp of the Standard Model of particle physics which predicts the existence of the Higgs Boson, I think it quite important that as many people as possible should understand the scientific notion of hypothesis testing. So I'm going to make my own attempt to convey the message.
Suppose we have a hypothesis A which we wish to test. We devise an experiment which will always give result B if A is true, but will sometimes give that result B if A is not true. We can calculate the probability of getting B when A is not true, call it p. We conduct the experiment and get result B. What now is the probability that A is true?
The answer is that there's not enough information. If we had been almost sure that A were true before we did the experiment, we'd be even more sure now. Whereas if we'd thought it fantastically unlikely that A were true, we'd think it only somewhat less unlikely now.
The extra information we need is an estimate before doing the experiment of the probability that A is true. Call that probability q. Now we have the information we need. Once we get result B, we are interested in the relative likelihood of getting the result because A is true - probability q - or because A is not true but the random numbers lined up that way - probability (1-q)p. Our new estimate of the probability that A is true, in the light of the experimental result, is therefore q / (q + (1-q)p ). Note that if we were already certain that A was true - q = 1 - then our new estimate is still 1. And if we were already certain that A was false - q = 0 - then our new estimate is still 0.
The important thing to grasp is that the probability in isolation that we get a particular result from random data is not the same as the probability in a given experiment that gave that result that the data were random. The latter probability depends on what competing, non-random explanations are available and how likely they are to be true.
Let's take a concrete example. We take a penny coin at random from the change we're given in a high-street shop. We wish to test the hypothesis that his has heads on both sides. So we toss it ten times and observe whether we get ten heads. (Yes, I know, it would be easier just to look at both sides, but bear with me). Suppose that we'd estimated the probability of having a double-headed penny as one in ten million. The probability of tossing ten heads and no tails with a standard coin is one in 1024, which is about ten thousand times as probable as the double-headed penny, so our new estimate of the probability of having a double-headed penny is about one in ten thousand - the formula gives one in 9767 to the nearest whole number. Although the likelihood in isolation of getting ten heads was small, we are forced by the result to believe something unlikely has happened, and we prefer the very unlikely explanation - ten heads out of ten by chance - to the extremely unlikely explanation - a double-headed penny.
Let's take another example: a proposed drug. We employ a team of expert scientists to research into a particular biochemical pathway and to devise a drug to interfere with it in some desirable way. We find that the drug works perfectly in vitro. Then we test it in humans, using a carefully devised double-blind trial, and find that it outperforms a placebo to an extent we would expect to equal or exceed with another placebo one time in twenty (i.e a p-value of 0.05). If our prior estimate of the probability of the drug's working was 40%, our new estimate will be 93%.
Alternatively, we might pick a naturally occurring substance by guesswork, dilute it out of existence with water, drip the water onto a sugar tablet, and allow the tablet to dry. We observe no in vitro activity with this tablet beyond the effect of the sugar. But suppose that when we conduct a similar double-blind trial we get a positive result meeting the same p-value of 0.05 . If our prior estimate of the probability of the drug's working was one in a billion, our new estimate is one in 20 million. Homeopathists might see this stark difference in interpretation of the data as unfair: a scientist would merely observe that the experiment wasn't nearly powerful enough to give useful support to such a fantastically unlikely hypothesis. (In practice, drug investigators never publish their own estimates of prior probabilities: instead they give qualitative reasons why they thought the drug worth testing; the reader can form a view of their own.)
Back to the Higgs Boson. Physicists generally tended to think that the Higgs Boson would be there in the energy range they were looking at, say with probability 75%. If we ignore the possibility that they may have discovered a different particle in that range, and accept the announced value of the probability of getting the result from observing noise, then the probability now that the observed results are not due to the Higgs Boson are a bit less than one in 10 million. We'd get a different estimate if we started from something else than 75%: the point is that only if we started from about 50% would we be able to say that the probability that the Higgs Boson has not been found is now one in three and a half million.
____
What I've been writing about is a simple application of Bayes' Theorem. It's less well known than it might be that Thomas Bayes himself is buried at Bunhill Fields in the City of London (as are Daniel Defoe, William Blake, and John Bunyan, among others). If you're a quant in the City and you're finding it difficult to think clearly about some question of probability, I recommend a walk there. Not out of any mystical faith in the powers of the bones of long-dead nonconformists, but because it's there and the walk will do you good.
No comments:
Post a Comment