Failed Execution

A criminologist and a computer scientist say they have developed software that can predict whether a U.S. inmate sentenced to death will actually end up getting offed by the state. The Christian Science Monitor reports:

What some observers find alarming about the outcome is that the 19 points of data supplied on each death-row inmate contained no details of the case. Only facts such as age, race, sex, and marital status were included, along with the date and type of offense.

Whole thing here.

Comments

[Douglas Fletcher]

Sounds like it might make a great betting pool, but it probably doesn't move fast enough to interest most gamblers.

[Evan Bick]

A couple of things jumped out at me from the CSM story, which is admittedly pretty short on details:
-If "mindless software" can predict the outcome that accurately, then the death penalty is the opposite of arbitrary. It might be in conflict with our sense of justice, but it's definitely predictable and thus non-arbitrary
-The story doesn't mention what percentage of death row inmates get executed in total. If it's very low or very high, guessing right is not all that impressive.
-The death penalty was, iirc, illegal nationwide until 1977, but the software's learning fodder starts in 1973. It is, presumably, very easy to predict whether or not people who were convicted of capital murder during that interval were executed. They couldn't have been.

[Geoff Nathan]

Yes, but the predictability is based on variables that ought to be irrelevant to whether one dies, such as age, gender and marital status. In a just world the variables would be things like whether the person had killed before, if it was particularly vicious, if the murderer showed no evidence that he or she thought there was anything wrong with what they did, etc. etc.
None of this is to excuse the pitiable lack of detail in the article. Also, neural modelling doesn't always give you a good picture of which variables are most important.

[John]

My problem with this study (which I have not read) is that they are measuring a state system (capital murder is a state crime) using national criteria. I am not a lawyer but as an economist I do work with statistical regressions daily and I am under the impression that each state has different rules regarding trials and sentencing which could cause the results of this study to be questionable. From a purely technical standpoint they should look at each state individually or at least groups of states with similar legal systems rather than aggregating across the country. There will be differences between liberal legal systems in Connecticutt or Massachusetts and conservative ones in the Bible Belt.

[jack]

Wow, an 'artificial neural network.' Sounds impressive. 90% prediction rate; good enough for government work. Why anyone pays attention to this kind of garbage research is amazing to me. I'm just mad I bothered to respond to this nonsense.

[Dan]

Two points:

Couldn't you get pretty close to a 90% success rate if the program *always* returned a value of "the inmate was executed"? I have no idea what percentage of death-penalty verdicts end up being reduced to lesser sentences, but my impression was that the number was low.

Secondly, my reading of this study is that it only examined people who were already on death row -- i.e., people who had already been tried and sentenced. After sentencing the specific details of the crime become largely irrelevant to whether or not the sentence is carried out. A person sentenced to death is, in theory at least, supposed to be executed regardless of what, exactly, he was guilty of. The ability to escape execution is, at that point, a function of how many lawyers you can keep digging for a legal basis for setting aside the verdict. Lawyers are a function of money, and money is strongly correlated to race, gender, age, marital status, et al. So a finding that the details of the case are largely irrelevant and demographic data has high predictive value isn't surprising.

In any event, it is impossible to appraise this study without knowing what all the fields were and what percentage of the 1000 subjects were executed. Anyone have a link to the real study? I didn't find anything on Google.

[thoreau]

Dan raises a good objection. Besides, there are a lot of people on death row who haven't been executed but presumably will be at some point in the future. I wonder whether what the program is really predicting is the length of time on death row. And since the length of time on death row is known to be strongly correlated with the state in which one is sentenced, I wonder what the statistical significance of the other variables is.

Mind you, I'm not trying to dismiss the model, but I do have some questions.

[Maurkov]

Evan, there are definitions of arbitrary that are applicable. You are correct that, being predictible, the results are not random.

I don't think the journalist understood, and shouldn't have quoted the obviously wrong Texas lawyer.

[Blaine]

"no details of the case ... the date and type of offense." You may not call the age, race, sex and marital status generic information, but I would consider them details of the case. The date and offense are difinitely details of the case. I understand the point they are trying to make but it still steps into hyperbole.

[Blaine]

Apologies, I meant to say "You may call the age..." drop the "not"

[Dogzilla]

Take it one step further and let the computer decide who lives and who dies. It can be a complicated algorithm based on objective criteria, or maybe just a (pseudo)random number generator like a big keno game for the damned.

[The Bureau of Corrections]

Citizen- You have been scheduled for termination. Please report to the nearest disintegration booth for administration of your sentence. Failure to comply will result in the issuance of an all points bulletin with shoot on sight provisions. You have 59 minutes to comply. Thank you and have a nice day.

[Bobster]

I wonder if they can create a program for people who want to get married? "Will this marriage work?" Put in the data and find out. The magic 8 ball is out of a job!

[Evan Bick]

Dogzilla, I don't know if you have any musical ability, but please start a band called "keno for the damned"

[Stevo Darkly]

Yes, but the predictability is based on variables that ought to be irrelevant to whether one dies, such as age, gender and marital status. In a just world the variables would be things like whether the person had killed before, if it was particularly vicious, if the murderer showed no evidence that he or she thought there was anything wrong with what they did, etc. etc.

Just to throw this out (possibly having a Larry Summers moment here):

Is it possible that such factors as age, gender, marital status, and race, while not directly linked to the verdict, might have some kind of correlation to such factors as likelihood of having a history of previous killings, showing remorse, having a tendency to be particularly vicious when committing a crime, etc? If so, then the former factors would be, by proxy, reliable indicators of the more relevant latter facts. Especially since most of the latter factors would be much harder to quantify and enter into a computer model. (Was the crime vicious? How vicious, on a scale of 1 to 10? Did the defendant seem to show remorse? How remorseful, on a scale of 1 to 10? Etc.)

For example, probably 20-year-old males are more likely to commit vicious crimes than 45-year-old married women. Take two defendants, a 20-year-old male and a 45-year-old married woman. The model may correctly predict the male will get the death penalty based on such factors as age, gender and marital status alone, without getting into case details that reveal the male was a street thug with a criminal history who beat, raped and murdered somebody, vs. the women had no criminal history and shot an abusive spouse.

[OPUS]

Big deal so someone wrote a program that predicts who actually will die from receiving the death penalty. Lets see...each year about 20,000 murders and each year about 200 are actually executed. I still think the odds are against being oeffed by the feds. Just try walking around Baltimore after the sun goes down.

[joe]

The thing to remember about the criminologist:

That man has no fucking neck!

[David Nieporent]

-If "mindless software" can predict the outcome that accurately, then the death penalty is the opposite of arbitrary. It might be in conflict with our sense of justice, but it's definitely predictable and thus non-arbitrary

Exactly! That was my thought when reading that article: it made no sense as an accusation. And yet the article said it more than once.

Moreover, the entire proposition (that this should be "alarming") is silly. The article isn't looking at who gets sentenced to death; the article looks at, AMONG THOSE SENTENCED TO DEATH, who gets executed. So of course "details of the case" don't matter. We've already factored all relevant details in when we selected our pool.