Saturday, March 02, 2013

Ministry of Ag & Fish


Nate Silver does not like Sir Ronald Aylmer Fisher. Not his style of dress (ill-kempt & dishevelled), nor his combative, argumentative style; Silver believes there must be a uniquely British mould for producing such intellectuals – after all we also gave Christopher Hitchens to America.

But most of all, Silver detests Fisher’s intellectual legacy, blaming him for the widespread adoption of the purely frequentist approach to statistical inference & the neglect of the theorem of Thomas Bayes.

As if more were needed, the most devastating charge against the man is that he continued to dispute the relationship between smoking & lung cancer, maintaining that researchers who claimed to have established a causal link were confusing correlation with causation.

Although Silver writes passionately about his own belief in the Bayesian approach it is not at all clear to me from his book that he uses formal Bayesian methods, with specified ‘priors’ in his own predictive modelling. Rather he advocates the use of judgement & deep knowledge of the process(es) underlying the forecast to adjust, where necessary, the predictions to take account of uncertainties which can never be reliably & completely captured in a mathematical black box. And in his concluding chapter he states baldly that ‘even common sense can serve as a Bayesian prior’. What is unacceptable, he says, is to believe that mathematical forecasts have a purity entirely free of the forecasters own subjectivity & biases, which do not need to be acknowledged.

I hold no brief for Fisher & am very much in sympathy with the Silver approach, but I am taken aback by the vitriol of his attack upon Fisher, even though I too despair about the cult of significance which he could be said to have inspired. Fisher was more subtle than that &, like Marx, can be said to suffer from the quality of his disciples. It is also important to remember that Fisher –born 1890 – was developing methods which could be used in an age when a computer was a human calculator.

I smiled at the bit where Silver – born 1978 - writes about how a standard SIR model in epidemiology takes only seconds to run on a laptop. Provided of course that somebody else has already done the hard & expensive work of collecting & collating all the data & making it available in electronic form at a click of a button to any one who cares to use it.

I cannot remember ever seeing any of Fisher’s contributions to the debate on smoking & lung cancer, but I was privileged to hear Sir Richard Doll speaking on the subject several times. And he used to point out that, when he started to investigate lung cancer nobody expected tobacco to be implicated.  Some aspect of general air pollution – such as  the Great London Smog of 1952, credited with causing thousands of deaths – seemed much more likely to provide an explanation. They were also of course well aware of the problems of disentangling causation from correlation & were applying the Bradford Hill criteria, with a battery of approaches.

I was once present when someone asked Sir Richard at what point he personally had become convinced that smoking was the culprit. He replied he thought that it was probably when they got the results of a prospective study of patients referred to UCH for ?lung cancer? In every case where the diagnosis was confirmed the victim was a smoker; non-smokers were all found to be suffering from a different problem.

As far as I am concerned – the debate was settled once & for all in March 1962 when it was announced on the BBC 6 o’clock News that it was now ‘beyond doubt’ that smoking causes lung cancer.

Nevertheless even in 1966 an undergraduate dissertation on ‘The Relationship between smoking & health’ would not have got good marks had it failed to mention some of the outstanding doubts or queries.

This was a dissertation for a [service] course in Applied Statistics for those doing degrees in economics & so was, in effect, a discussion of statistical, rather than medical issues. Sample selection (the largest UK study looked only at male doctors), cross-country comparisons, looking at rates of smoking & rates of lung cancer, etc. (As I remember it Australia & Switzerland did not fit the pattern which otherwise showed a positive correlation between rates of smoking & rates of lung cancer; both had rates of the disease which were much lower  than expected.). I also remember – bizarrely as it now seems – concerns about whether people who stopped smoking might gain weight & therefore put themselves at increased risk of a heart attack.

And – most important of all – what could explain the fact that the majority of smokers did not get lung cancer – we were not allowed to consider genetic explanations in those days, so factors such as type of cigarette, the method of inhalation, the precise make up of the tobacco etc were candidates.

Although I don’t think I knew anybody who seriously doubted that the link had been proved, there were many who thought that what was really needed was an answer to that question. The real scientists & medics must grasp the baton handed to them by the statisticians.

I do not know what Fisher’s own reaction had been to the publication of the report by the Royal College of Physicians. He did not have very much time to continue the argument he died unexpectedly on 29 July 1962 in Adelaide from an embolism after a successful operation for bowel cancer.

Links
Smoking and health (1962)
Related post
Pretty girl in crimson rose