ZaZara
AstraZaZara
- Joined
- Jan 31, 2014
- Messages
- 8,425
- Reaction score
- 18,000
These are all excellent points! It is not a simple area and when the press pipes up with ...cos deaths rose a bit.... that will be taken by a lot of people as being code for "she must be guilty". The whole issue surrounding shift patterns against deaths (if it even plays a part, but I bet it will) is that it's potentially a very, very complicated area. It took me a couple of times of reading that part about "statistically, most nurses will experience more deaths on their ward when on shift than off because more of them are onshift when people die" before it dropped as to what they were going on about. It would be quite a thing to have to sit on a jury all day for six months being continually bombarded with information like that.
I certainly think you're correct about what the initial experts were trying to do - consciously or not. The issue with testimony like this is that you need similar experts to refute it and such people I'd guess are somewhat thin on the ground. In the closing part of that paper he says (which I think you quoted) something to the effect of "I hope I don't end up working a Lucy Letby's case". I wonder whether it might an idea for her defence people to give him a call anyway rather than needing him for an appeal?
From an interview with Gerrit Timmer who was Chaiman of the Dutch Association for Statistics at the time when Lucia de Berk was convicted.
Convicting someone on the basis of statistical evidence is in itself pure nonsense, according to Gerrit Timmer even then.
Q. One is inclined to think that every answer that rolls out of an algorithm is true. But people have attached all kinds of variables and goals to such an algorithm beforehand. So as a mathematician, you actually determine the glasses through which I see the world.
A. "As a rule, this colouring is not put into an algorithm. Not consciously, at any rate. Perhaps you discriminate between men and women unconsciously, by the preselection you make.
Q. I think profiling in crime is a nice example: you can imagine that the police organisation wants to know where statistically the most crime will take place, due to undercapacity. In your algorithm, you might select by district, by gender or by ethnicity. And then you end up with profiling that is ethically unjustifiable. As mathematicians, you have simply solved a problem for a client, but unintentionally arrived at an ethically irresponsible solution. Are you aware of the unintended undesirable effects that your algorithms can cause?
A. "In my opinion, the danger lies more in ignorance than in ethics. I studied law myself, and at that time the subject of jurimetrics - statistics - was just coming up. What they did there was mathematically quite wrong. Those lawyers didn't understand enough of it. That led to the examples you are talking about now. They went looking for correlations in a large database at random. Every statistician understands that this is not a good method. If you start comparing billions of data, of course you'll find correlations. Random searches simply aren't allowed.
Q. But how do you work differently?
A. "You have to have a hypothesis first, otherwise you can't test it. We can distinguish a thousand items of clothing nowadays. Even if there is no relationship between clothing and crime, there will still be one type of clothing, perhaps striped jumpers, for which the link with crime is the highest. That says absolutely nothing. It is different if you had reason to suspect striped jumpers beforehand. Then you can use statistics to test that hypothesis. Then you've eliminated the 'coincidence' that you're reasoning towards the answer."
"Random search is the one mistake made out of ignorance. The other mistake is assuming that you can prove things with statistics. At the time of Lucia de Berk, I was chairman of the Dutch Association for Statistics. Lucia de Berk was convicted on statistical evidence. That in itself is nonsense. The chance that it was a coincidence was supposed to be very small in her case. Well, go to a casino and write down which 30 numbers have fallen in a row on the roulette table. The probability of exactly those 30 numbers falling in a row is virtually zero. But that is not proof that it did not happen, is it? Because it did happen. Everything that happens outside our window has a very small chance of happening. So relying on those odds is a close approximation of three times nothing."
Interview AI-expert Gerrit Timmer - 'Bol.com en Ahold zou waarschijnlijk béter zijn dan Amazon'
BBM