On correlation neglect

In the Times, Danny Finkelstein expresses disquiet about Operation Yewtree:

"My word against some bloke after more than 20 years is good enough for you?" [Liz Kershaw] asked. The police wouldn't need other evidence to charge?
"Well" replied the officer. "If it was just one girl obviously the Crown Prosecution Serivce would probably throw it out. But if more than one girl came forward, well..."
The clearest statement I have ever seen of the dubious policy of using legally weak individual allegations to support each other.

He's touching upon a widespread cognitive error here - correlation neglect.

If lots of women were to independently make allegations against someone, those allegations might be credible - though whether credible enough to overcome the reasonable doubt hurdle is another matter. The problem comes when the allegations might be correlated. If one allegation against a public figure encourages others to make allegations, the subsequent allegations might not have as much weight as they would if they were independent. The latter allegations might be cases of misremembering or mistaken identity. If so, there's a danger of an information cascade; several allegations might be due not to independent pieces of information which support each other but to a common mistake.

Now, I've used the word "might" a lot in that paragraph. But we do have a much clearer example of the danger of correlation neglect in criminal law. In 1999 Sally Clark was convicted of murdering her two babies on the evidence of a pathologist who claimed that the chances of two babies dying of cot death as Ms Clark claimed were vanishingly small. His statement was an example of correlation neglect; it might have been true if the chance of one baby suffering cot death were independent of the chance of his brother dying. But if there are common environmental risks of cot death, then the chances are correlated, so the chances of two deaths are much higher than the pathologist claimed. Ms Clark was eventually released on appeal.

It's not just prosecutors who are prone to correlation neglect. It also happens in investing. Benjamin Enke and Florian Zimmerman show that people fail to discount correlated signals about future returns, and so over-invest in assets which look good. And Erik Eyster and Georg Weizsacker show that investors treat (pdf) correlated assets as uncorrelated and so fail to diversify risk properly.

And, of course, it also happens in politics. We fail to discount the political opinions of our friends because they are correlated: our friends tend to have similar backgrounds and outlooks on life. We are, say (pdf) Glaeser and Sunstein, "credulous Bayesians", who over-react to weak information.

Such correlation neglect contributes to the bubblethink I discussed yesterday. It also generates political polarization (pdf) and overconfidence; Enke and Zimmerman show that it can be a source of stock market bubbles because investors become overconfident about future returns. In the legal context, such overconfidence can lead to false convictions.

These, of course, can have terrible effects: after she was released, Ms Clark drank herself to death. In this sense, correlation neglect can be deadly. Whether this has any relevance to the UK's decision to go to war in Iraq is a question I shall leave to others.

Comments

[JR]

Well I was with you there until - "These, of course, can have terrible effects: after she was released, Ms Clark drank herself to death." That's a pretty big leap of attribution!

[Xerxes]

another weakness in the sally clark trial was the prosecutors fallacy

the jury should have compared the probability of joint cot death with the probability of being a murderer

instead they equated the low chance of joint cot death with the probability of innocence

fail

[politicallycorrect]

["My word against some bloke after more than 20 years is good enough for you?" [Liz Kershaw] asked. The police wouldn't need other evidence to charge?]

The law in many countries is that no other evidence is needed to *convict*, because campaigners against crimes committed by men on women have always argued that rapists of course lie and victims of rape are not accusers but witnesses, and a witness statement by the victim is as such proof of the crime. Also because they cite statistics that 80% of the abusers are men who are known to the women victim, so there is no issue of identification of the culprit.

[The clearest statement I have ever seen of the dubious policy of using legally weak individual allegations to support each other.]

In many countries they are not legally weak, where is no legal need for corroboration by any other evidence for witness statements in at least some types of crimes where women are victims and men are abusers. Since the witness is the victim herself, and her testimony in itself is sufficient as proof of guilt in many jurisdictions, the rapist then needs to prove beyond reasonable doubt that the witness is unreliable.

Campaigners have always complained that this means that most trials of men who abuse women then become in effect trials of the victim as she gets accused to be an unreliable witness. The fix that many campaigners propose to protect women victims of men's abuses is to forbid questioning the reliability of the witness victim in court in cases of abuse of women by men, leaving open to the rapist or abuser the option to find some other evidence to prove beyond reasonable doubt that they are innocent despite there being proof of their guilt from the testimony of the victim herself. The goal is to spare the women victims of men's crimes further abuse as their abusers attack them in court with character assassination.

Other campaigners reckon that a statement by a witness who is also the victim of a crime committed by a man against a woman should be final proof of the crime, under the prudential principle that "better safe than sorry", and that securing convictions of men abusers to protect women victims is essential for a civilized society, and if mistakes happen, that's something that is the the fault of rapists: if men did not abuse women, or protected women from abuse by other men, there would be no mistakes.

The reason why the CPS and the police are reluctant to bring cases where there is only a witness victim who is a victim of a man, seems to be that many juries, especially those containing men, will return verdicts of "innocent" when they simply don't believe the statements of the witness victim, as juries seem to follow the principle of "benefit of doubt" even when the law does not allow that.

The CPS and police seem to think that juries however usually convict men who abuse women when there are more than one witness statements by women victims, under the prudential principle that "there is no smoke without fire". which then seems to override the "benefit of doubt".

Some campaigners argue that this shows that the legal system created by men considers the statements of women victims as witnesses as worth less than the character assassination of the men who abuse them, so that it takes some or several to get a conviction.

This has meant that for example an unpopular male teacher can be convicted with very high probability if even a small number of his pupils give witness statements as victims, with no other corroboration, in many jurisdictions, which has led many potential abusers to avoid getting jobs in schools as the risk is very high; this has led many schools to have almost exclusively female staff. Campaigners would probably argue that this has made pupils safer from predatory men, thanks to "better safe than sorry".

[He's touching upon a widespread cognitive error here - correlation neglect.]

I would rather say "better safe than sorry" mixed with "there is no smoke without fire".

Campaigners for women victims of the abuses of men of course could consider arguments about "correlation neglect" as mere sophistry and "lawyering up" to protect men who abuse many women, and cite statistics that show that only 98% of rapists are ever convicted, and only 2% of those convicted were mistakenly convicted.

[We are, say (pdf) Glaeser and Sunstein, "credulous Bayesians", who over-react to weak information.]

The argument of those campaigners is often that every man is a potential abuser of women, abusers only fear getting caught, because men are ruthless predators, and therefore aiming to convict all men who abuse women is an essential priority in deterring the abuse of women by men.

[Paul]

This is nothing new ; it dates back to the 1990s and is called "Similar Fact" law.

The idea is that claims from multiple complainants are deemed to prop each other up even if they are only vaguely related. In the past, DLT's et als persecutors would have had to show much closer types of similarity (it goes back to "Brides in the Bath".

In the 1990s, not sure when, a judgement went down that they only had to be vaguely connected.

This encourages the Police to simply collect and encourage allegations irrespective of whether there is any consistency - it's effectively mudslinging - and ignoring exactly how they are collected.

This has been going on in social care for years ; effectively the Police ask the criminally minded to tell lies, usually in collusion with compensation solicitors, resulting in a huge morasse of random (or copied - sometimes word for word) allegations which are driven by cash, which are then just thrown at defendants who convict on no smoke without fire, ignoring the Police going round dumping petrol over everything.

CPJ and Perjury never happen, sometimes complaints are literally impossible - a 'victim' and a 'defendant' never worked together - this is ignored because the route by which the information is passed over "Mr Jones bu**ered you, didn't he" is the Police themselves.

[Dipper]

I thought that the issue in the Clark case was that probability of two cot deaths was very unlikely - about 1 in a million. But in a country of 50 million, things with a probability of 1 in a million happen to about 50 people, so whilst it is unlikey that a person at random would have two cot deaths, it is quite likely that a large country would be able to find someone who had had an unlikely combination and drag them (erroneously) in front of the courts.

[Dipper]

Paul - can you tell us more?

[Donald Smith]

Correlation of accusations only, in itself, tells us that the probability of guilt is less relative to the probability of guilt with independent accusations (all other things equal).

Correlated accusations could still be strong evidence.

[An Alien Visitor]

"cite statistics that show that only 98% of rapists are ever convicted, and only 2% of those convicted were mistakenly convicted"

The better safe than sorry approach would be to lock all men up once they reach the age of 16 I guess. When you say only 98% of rapists are ever convicted what you are actually saying is that either:

A number of women who are raped do not report it.

or

According to the jury, some women are claiming they are raped when they are not.

Either way there is documented fact that women have lied about being raped.

So from the other side of the politically correct coin we could say the better safe than sorry bet would be to assume all women who claim they were raped are lying and probably after some form of compensation, i.e. money.

[magistra]

Chris,

I normally enjoy reading your posts, but this one is poorly argued. The reasons you suggest for spurious correlation aren’t plausible. First of all, the whole point about Yewtree was that it was dealing with celebrities. Therefore, in the vast majority of the cases the women and girls who have come forward as witnesses knew exactly who their attacker was before the attack: indeed they’d often wanted to meet the celebrity concerned (as with many of Stuart Hall’s victims).

Secondly, although the exact details of a previous encounter may have been forgotten over twenty years, why should the emotional content of it have changed? In other words, why should an encounter vaguely remembered as positive, even if sexually charged, suddenly be recalled as abusive? And why should people in such a situation, even if they now feel negative towards the celebrity concerned, choose to come forward and be subjected to cross-examination? The process of appearing as a witness in a sex case can in itself be so gruelling that some people have killed themselves as a result (such as the violinist Frances Andrade after the Chetham’s School of Music trial).

Thirdly, the credibility of new witnesses who come forward is checked at the trial (and presumably before). In many of the recent cases, they’re both more accurate as to times and places than the defendants, they describe a similar modus operandi to other witnesses (without having had the chance to hear their testimony) and they complained about the offence at the time, only to be told to keep quiet.

And that brings me to my final point. If you disallow multiple testimonies, unless the witnesses come forward independently and at the same time, you tip the balance of justice in the favour of these repeat sexual predators, of whom there are a considerable number. This doesn’t seem to be a behaviour they grow out of as they mature, as with many offences of violence or drug-taking. That is because such men deliberately target people who are less powerful than themselves and who are therefore less likely to be believed. That’s particularly clear in the case of Jimmy Saville, with threats against those who went against him, but it also happens in non-celebrity cases. Why did the recent grooming cases target children’s homes? Because they were the girls that nobody cared about, whom no one would listen to because they were “bad” and untrustworthy. If you really want to stop allegations from twenty years having to be dealt with, then the police and other authorities need to treat current allegations seriously. But instead, what we hear of is police who repeatedly refuse to record sex crimes as such and prosecutors who decide that witnesses aren’t credible, because they don’t meet the criterion of being 100% sexually pure.

[Xerxes]

@Dipper

youre conflating the wrong numbers first the uk population is closer to 60 million

second the probability of cot death refers to the number of baby deaths which is a lot lower than the uk population

thirdly and as i said previously the chance of double murder is very low and you should be comparing against that