From: Smigrodzki, Rafal (SmigrodzkiR@msx.upmc.edu)
Date: Thu May 30 2002 - 08:37:25 MDT
Harvey Newstrom [mailto:mail@HarveyNewstrom.com] wrote:
On Wednesday, May 29, 2002, at 07:57 pm, CurtAdams@aol.com wrote:
> It doesn't have to be consistent; probabilistic is enough.
Wrong.
### Correct. Absolute consistence, and absolute truth, are philosophical
ephemera. Even genetic testing of blood samples from a crime scene yields
only statistically correct data. It unavoidable to use probabilistic data,
even when judging other people. The essence of the problem cannot be a
choice between absolutely consistent and everything else (like black and
white), but between "consistent enough", versus "a bit too sloppy to be
ethically acceptable".
-----
People argue that blacks are more likely to commit crime. This
does not mean that we can identify criminals by skin color. Some blacks
aren't criminal, and some criminals aren't black. Worse yet, even if
blacks are shown to have a higher probability of criminal behavior, it
is still true that most blacks are not criminal. Any correlation with
"black" meaning "criminal" would be inaccurate more often than it would
be accurate.
### No contest.
-------
But in the first case you have directly measured traits that are
accurate for all members of the group. In the second case you have
unmeasured traits that are not accurate for all members of the group.
The first group defines accurate information. The second group
statistically allows possibilities that might or might not really
exist.
### As I mentioned above, the only accurate information is "accurate enough
for the matter at hand".
So let's say that with the progress of MRI and genetic testing (both
measuring characteristics which are not freely chosen) we will develop a
screening test predicting with 90% specificity that a given person will
commit a violent crime if unsupervised, in the current social setting. The
sensitivity of the test is 70%. The true prevalence of the criminal tendency
is 15 %. The PPV will be 36% and the NPV will be 97 % (check my math, I'm
bad at it).
In other words, you will have a 36%-valid hunch that those who test positive
will do something bad. Also, you will know with 97% certainty that those who
tested negative will stay nice. Since your resources are finite, you have to
optimally apportion them to the surveillance of these two groups. I hope the
security professionals working on my behalf would *not* seriously suggest
that it is ethically imperative and practically right to pay the same amount
of attention to members of both groups.
Rafal
This archive was generated by hypermail 2.1.5 : Sat Nov 02 2002 - 09:14:30 MST