Thursday, December 27, 2012

"The best predictor of future behavior is …


… past behavior."


Past as prelude. So neat, so clean. So full of certitude. Like a fortune cookie Confucianism. Something you might hear on CSI: Special Victims Unit. A maxim cited by pop psychologist "Dr. Phil" McGraw, in one of his many self-help books.

I'm sure you have heard the mantra. It's creeping into risk assessment reports and court testimony by forensic psychologists. Sometimes, it's augmented with incendiary metaphors: The subject is "a ticking time bomb"; he is "carrying a hand grenade and it's just a matter of when he pulls the pin."

One current case of mine involves a guy with a cluster of several violent offenses a few years ago, when he was in his 20s. He was using drugs back then, and hanging around with a bad crowd. Plus, he is chronically psychotic. Not a good combination.

But if you predict future violence based on a set of risk factors like his, you will be wrong more often than not. Only about four out of ten of those individuals judged to be at moderate to high risk of future violence go on to reoffend violently, according to research. The low base rates of violent recidivism will be working against you.

Birth of a legend

So where does this idea that "the best predictor of future behavior is past behavior" come from, and does it hold any water?

Perusing psychology texts, it appears that the principle has circulated for decades. But as it gained traction, some boiled it down into a simpler, one-size-fits-all mantra. So, for example, the 2003 Complete Idiot's Guide to Psychology claims as an established "psychological fact of life" that, "when it comes to human beings, the best predictor of future behavior is past behavior." Period. End of story.

But this is a gross oversimplification. Psychological scientists who study human behavior agree that past behavior is a useful marker for future behavior. But only under certain specific conditions:
  1. High-frequency, habitual behaviors are more predictive than infrequent behaviors.
  2. Predictions work best over short time intervals.
  3. The anticipated situation must be essentially the same as the past situation that activated the behavior.
  4. The behavior must not have been extinguished by corrective or negative feedback. 
  5. The person must remain essentially unchanged.
  6. The person must be fairly consistent in his or her behaviors.
Here, by way of illustration, is a typical study of the phenomenon, involving college students' class attendance habits:

In a semester-long course, researcher Icek Ajzen found that a student's attendance rate for the first eight sessions correlated 0.46 with his or her attendance rate for the second eight sessions. As you can see, all the conditions are in place: Class attendance is a habitual and routinized behavior, the prediction span is very short, and there is little likelihood of meaningful changes in either the situation or the person. Yet still, the correlation is far from perfect.

Other examples from the classic studies: Frequency of exercise during a given time period is a pretty good indicator of exercise habits in the near future. Ditto for cigarette smoking and drug use.

But over longer time periods, even high frequency, habitual behaviors may undergo dramatic change. A smoker or heavy drinker might successfully quit the habit. A chronic thief might land a decent job, start a family and settle down.

As this last example suggests, researchers have also determined that the situation plays a critical role in behavior. The situation is often more determinative than individual character traits. Personality theorist Walter Mischel - frequently cited in connection with the "best predictor" maxim - suggests that behavioral consistency is best described through if-then relationships between situations and behaviors, as in: "She does A when X, but B when Y." So, a person may engage in heavy drug use when in the company of drug-using peers, but may stop using when she gets a fulfilling job and moves to the suburbs, or when she is staying with her strict grandmother.

Forensic psychologists jump aboard

It’s one thing to find a simplistic maxim where one would expect to - in an "Idiot’s Guide" or on Dr. Phil. But it is troubling to see it incorporated in forensic contexts, where the stakes are much higher.

Confusion creeps in when a risk marker is mistaken for an inevitability. It is true that people with a history of violence have a higher likelihood of committing violence in the future than do people who habitually turn the other cheek. Risk is especially acute for those with very extensive histories of violence across a range of situations. But this does not mean that everyone who has committed past acts of violence will continue to aggress forever (any more than someone with no prior violence is guaranteed to remain peaceable forever).

It's like claiming to know that because your teenage neighbor had a fender bender (or two) when he was first learning to drive, he will definitely crash his car again. He is probably at a higher risk of another collision than is his middle-aged mother, with her clean driving record. But he may or may not crash again. There are many intervening variables - whether he learned from his mistakes, the frequency and locations and times of day of his future driving, his choice of companions, the actions of other drivers on the road, the weather conditions, and so on.

The maxim also conflates all types of violence, and all types of offenders. For example, with detected recidivism among sex offenders falling somewhere between a low of about 3 percent and a high of no more than 15 percent, it's pretty hard to argue past as prelude. And if we apply the mantra to murderers, as did "Dr. Death" in Texas, we will be even further off the mark. In California over the past two decades, about 1,000 people have been paroled from prison after serving time for first- or second-degree murder. Their recidivism rate for murder?

Precisely zero, according to Nancy Mullane's Life After Murder.

The best-predictor axiom ignores such base rates, which are essential to accurate prediction. If we know the base rate of the criminal behavior we are trying to predict - whether murder or sex offending or general violence - and we know the frequency with which a person has engaged in that behavior, we can use a mathematical formula called Bayes's theorem to calculate a rough likelihood of the behavior's reoccurrence. (I recommend Nate Silver's The Signal and the Noise for great examples of the applications of this theory across a range of contexts, from poker to climatology.)

The maxim also snubs its nose at the age-crime curve, perhaps the most universal finding of a century of criminology research. As they reach their mid-30s or so, criminal offenders begin to slow down. Some mature naturally, some go through successful mentorship or treatment programs, some settle down and have families, some form mellower friendships, some simply burn out. Whatever the reasons, as research by Shadd Maruna and Sampson and Laub drives home, desistance is a virtual inevitability for all but the most die-hard minority of offenders.

This is not to say that the maxim is entirely useless. It may work fairly well under certain limited circumstances, if all of the following hold true:
  1. We are predicting over a relatively short time frame.
  2. The individual has a high frequency of violence.
  3. The violence occurs in a variety of situations.
  4. The person is faced with the same or similar situations.
  5. He or she has not been deterred by negative feedback.
  6. He or she has not changed in any other significant way.
But given lengthier time frames of prediction, our subject and his circumstances both undergo inevitable and often unpredictable changes, and we lose fidelity.

A ticking time bomb fails to ignite

In the case of the report I was reading this week, the mantra was a complete bust. The guy got out of jail and did great. He voluntarily sought treatment and cooperated with all terms of his supervision. By the time I saw him, he was leading a life as peaceable as a newborn lamb's. In his spare time, he even volunteered to help the needy at his local church.

If the evaluator had heeded the literature on criminal desistance, she might have seen this coming. The fellow had reached the age at which desistance becomes more the rule than the exception. He no longer associated with his old criminal peers. Perhaps most importantly, he had stopped using the drugs that had exacerbated his psychosis.

The past-as-prelude mantra fits with today's dominant, dark view of offenders as a bundle of perpetual risk factors, ticking time bombs just waiting to explode.

What it doesn't fit so well with is reality. 

1 comment:

  1. "Yet still, the correlation is far from perfect. " We don't need "perfect" - all we need is that it is more useful than all the other predictions available in order to meet the criterion of "best single predictor." So, enough with your straw man bashing. A correlation of .46 is pretty healthy in the world of Psychological research.

    If you don't have anything else to go on, it is the best bet available - if you have to bet...and sometimes we have to bet...it is the nature of life to have to rely on heuristics and shortcuts, when algorithmic solutions are less available.

    Furthermore, I don't really care about data for the entire further 8 weeks of attendance...if the student merely misses ONE furher class then the prediction has earned its keep.

    And the murder recidivism rate is definitely not zero in the rest of the country...although it is much lower than most would expect. Keep cherry-picking, though...

    ReplyDelete

 
Real Time Web Analytics