[PREV - QUESTIONING_DEBIASING]    [TOP]

SCATTER_THE_NOISE

                                             August    7, 2023
                                             October  24, 2023
                                             December 11, 2023

"Noise" (2021) by Kahneman, Sibony and Sunstein

The subject at hand is improving
organizational decision-making.  The
claim is that while assessing the
accuracy of decisions can be difficult,
it's easier to observe when multiple
judgments are scattered widely.  The
authors argue that simply trying to
reduce this inconsistency without            Identifying bias in a series
worrying quite so much about correctness     of judgments would require
is a quick way to improve the process.       some knowledge of the right
                                             answer in order to determine
                                             how far the average judgment
                                             deviates from a correct one.

                                             But if, for example, you have a
                                             number of experts on staff
                                             trying to make a prediction, you
                                             can look at the variations
                                             between predictions even if you
                                             don't yet know who's right.




Now really, they don't mean "noise" they mean
"scatter": the way your hits spread out around      The familiar distinction
an average when you've been aiming at a target.     is between "precision"
                                                    and "accuracy"--
    None of your associations with the word
    "noise" suggest what they actually want                Though that bit of
    to talk about-- this is not, for                       technical jargon has
    example, a book about signal-to-noise                  it's own issues:
    ratios in a communications channel.                    an average English
                                                           speaker would presume
                                                           those are synonyms.

                                                              PRECISELY_ACCURATE
Their primary
example is legal         This is what they
decisions:               began with and tried
                         to generalize from.

You would hope that questions of guilt and
sentencing would be determined by the quality
of the evidence and the circumstances of the
case.  In practice, it's well understood that
the individual characteristics of the judge
matter tremendously.  We want a system that
approximates justice but it often seems that
what we have is a crap shoot about which judge
you get assigned.

The authors consistently treat reducing
scatter (which they insist on calling "noise")
as the golden road to reducing overall error,
with only an occasional admission that in the
process you might sometimes inadvertanly
increase bias.

But consider that one way you could fix the
problem of scatter in legal judgements is to
make every judge a hanging judge.  If every
judge invariably went "thumbs down", then
scatter would be reduced to zero-- a great
improvement, right?

    They may very well be correct that large
    amounts of "noise" are often symptomatic
    of a problem, but it doesn't necessarily
    follow that reducing that "noise" is the
    way to fix the problem.

    Really, there's no particular reason to assume
    scatter and bias are independent of each other.

    Myself, I would expect that if an
    organization put emphasis on reducing
    scatter, then reduced scatter is what            For any general
    they would get even though accuracy              principle you can
    would quite likely go to hell.                   formulate, human beings
                                                     will find the wrong way
        E.g. the independent experts you've          to apply it.
        been examining might start consulting
        each other to make sure they're
        clustered closer together.


    Now the authors are no doubt correct that
    variation is often easier to identify than
    correctness, and large amounts of variation
    can be a strong indication of a systemic
    problem...

       But the actual goal is *hitting
       the target*, so you should keep
       your eye on the target.

                                             Somewhere in this book there may
                                             be some verbiage that addresses
                                             this point, but even if so it's
                                             lost in the roar of the headline
                                             message.

                                             On nearly every page they
                                             regard it as a given that any
                                             reduction in noise is an
                                             improvement in decision making,
                                             and any improvement in decision
                                             making can be regarded as a way
                                             of reducing noise.

                                             It's competely shot-through (and
                                             for me, completely shot down)
                                             with this confusion of symptom
                                             and disease.


                                             Maybe I shouldn't have been
                                             surprised: this confusion is
                                             evident right there on the cover,
                                             in the book's subtitle:

                                                NOISE 
                                                A Flaw in Human Judgement



--------
[NEXT - PRECISELY_ACCURATE]