Re: Is Racial Profiling a Type of Bayesian Inference?

On Jul 25, 8:43 pm, aruzinsky <aruzin...@xxxxxxxxxxxxxxxxxxxx> wrote:
On Jul 25, 10:40 am, Russell <russell.mar...@xxxxxxx> wrote:

On Jul 25, 11:34 am, aruzinsky <aruzin...@xxxxxxxxxxxxxxxxxxxx> wrote:

On Jul 24, 6:12 pm, Russell <russell.mar...@xxxxxxx> wrote:

On Jul 24, 6:51 pm, aruzinsky <aruzin...@xxxxxxxxxxxxxxxxxxxx> wrote:

On Jul 24, 10:44 am, Russell <russell.mar...@xxxxxxx> wrote:

On Jul 23, 8:41 pm, aruzinsky <aruzin...@xxxxxxxxxxxxxxxxxxxx> wrote:

I'm certain that racial profiling is an informal application of
statistics, but I am uncertain of the kind of statistics.

I shouldn't touch this topic with a 10 foot pole, but...

It's a Yes or No question!

Please note that the body of your post, which was what I
addressed, was "I'm certain that racial profiling is an
informal application of statistics, but I am uncertain
of the kind of statistics.", which in fact is not a
question at all.

My motto is all too often the study of data requires cares.
A corollary of that might be there is no such thing as
informal statistics.  Either the methodology is applied
carefully or it isn't really statistics, just quackery.

A euphemism for "fudge" is "model" as in "Gaussian white noise
model."  Name one area of applied statistics without fudge.

A famous statistician once said, "All models are wrong.  Some
are useful."

I'm sure many officers try hard, I doubt if the people in
law enforcement, especially on patrol, have the time for
careful analysis, even if they have the desire and skills.
I'm more inclined to believe that cognitive illusions tend
to take over, driven by the heuristics hardwired into our
brains by millions of years of evolution (see the book
_Inevitable Illusions: How Mistakes of Reason Rule Our
Minds_, or several others that deal with these ideas).
Good training helps condition officers to control their
reactions to these hueristics, and in some cases the
hueristics may even invoke the correct response and
should be followed.  
But in any case, it's not statistics,
at least as I use the word.  Just my opinion...


I am leaning toward informal statistical decision theory.  You do not
have to know the exact number of bullet chambers in a gun to quickly
decide not to play Russian roulette.

That's a classic a risk analysis.

Russell- Hide quoted text -

- Show quoted text -

My point was that racial profiling typically attempts to optimize
costs (maximizing expected utility) whereas Bayesian Inference does
not therefore calling it an application of statistical decision theory
is more appropriate.


As statistical decision theory, racial profiling can be done more
rigorously to set police policy.

It is possible to get crime statistics, e.g., here

and guesstimate the utility costs.  An advantage of this procedure is
that points of contention (probably utility costs) can easily be
recognized compared to recognizing points of contention in endless
hours of tradition debate using connotative words and denotatively
ambiguous homonyms such as "racist."- Hide quoted text -

- Show quoted text -

My point is that the results don't have to be perfect,
the assumptions of the models will never be perfectly
met in real life data (if they were, I'd think the data
suspect), but these things need to be done with care
and interpreted with care to be truly useful.  One can
download all the data one can find, but how good is the
data?  What else are potential sources of statistical
bias in the data?  What covariates are highly colinear
with race and crime?  Correlation is not causation.
Such things need to be studied before one even decides
whether to use technique A or B for the analysis.  (You
might be right in your choice of technique.  I don't
know.)  One will almost certainly find the data wanting
in some respect.  What might that do to the results?

I agree that endless hours of debate might be illuminated
by actual facts.  But I don't want armed police making
decisions based on a guesstimate of the utility costs or
guesstimates produced in whatever statistical framework
one thinks is most applicable.  It's not that such
results are necessarily fundamentally wrong in all
cases, but they might be wrong in too many cases and they
might provide a cloak of scientific respectability for
decisions that is not justified.

Russell- Hide quoted text -

- Show quoted text -

But, you are already a voluntary user of grossly inaccurate
guesstimation.  You are making an important decision based on "might",
i.e., nonzero probability, instead of more accurate probability

The use of the word "might" in this case is rhetorical
rather than statistical, and even if it wasn't, I'm not
going to draw a gun based on the resulting probabilities.