Re: More Hard Problems Using Bayes' Theorem, Please

From: Dan Fabulich (dfabulich@warpmail.net)
Date: Mon Jul 07 2003 - 19:41:07 MDT

  • Next message: Dossy: "Re: Cutting Taxes to Spur the Economy"

    Mitchell Porter wrote:

    > But what I still don't get is what's wrong with 'orthodox statistics'. I
    > obtained my p1/(p1+p2) through frequentist thinking, which is supposed
    > to be anathema to Bayesians. On the SL4 list, Ben Goertzel implies that
    > "non-Bayesian probabilists" are restricted in their choice of priors:
    > http://sl4.org/archive/0305/6820.html
    >
    > ... but even if what he says is true, I don't see it reflecting on the
    > subjectivist-vs-frequentist debate. Both sides need priors to get
    > anywhere, and there's nothing in frequentism to *compel* the use of
    > Gaussian priors. So, I still don't get the philosophical lesson.

    "Bayesian v. frequentist" philosophical arguments are plagued with
    vagueness; one feels the vagueness more acutely than in many other
    philosophical arguments b/c they both seem to brush up right next to
    what's analytically verifiable, and because the math can be so niggling
    and error-prone.

    (I note, for example, that you happened to misread the problem Eliezer
    posed; you solved a perfectly fine problem instead, but not exactly the
    problem as posed. That problem has some handy features that which allow
    you to simplify and solve the problem just a bit faster. [Though I sure
    as hell wouldn't notice them unless I was working on paper.])

    [There's something else driving this "argument" as well, IMO. There's a
    deep psychological problem with the study of statistics: it's so easy to
    make trivial errors with huge ramifications that, once you get kind of
    good at it, you realize just how absurdly often people make radical
    mistakes in their statistical intuitions; people who do this a lot usually
    leave the path of interpretative charity, assuming that their opponents
    are probably just making a mistake like everyone else, not that they are
    well-informed thinking people with reasonable and possibly serious
    objections.]

    But, let's get back to your question. There is no hard-and-fast
    definition of "frequentist" in this debate, unfortunately. I wish I could
    say that Goertzel was "wrong" in his characterization of frequentists, but
    I'm afraid I just can't guarantee that. One thing I can say, though, is
    that he almost certainly wasn't really talking about the definition of
    frequentist/orthodox that Eliezer was using in his bonus question 3:

    > 3. Using orthodox methods and an expensive statistics program,
    > calculate the probability that, if the hypothesis "I am looking at the
    > red barrel is correct", you will see 7 or more red tokens in ten
    > samples. Explain why this number is actively harmful to genuine
    > statistical understanding in twenty-five words or less.

    IMO, the relevant "frequentist" position at which Eliezer was poking fun
    here was the position that there are some common-sense events that have no
    probability whatsoever; these "singular" events don't have a probability
    of 0 or 1, they just don't have "probabilities" at all... according to
    this type of frequentist, the whole notion is simply not applicable
    sometimes.

    A good example might be to ask what the odds might be that a certain
    couple will get married. In this case, the frequentist feels obligated to
    respond that "probability" is a misnomer here: probability only applies
    when there's a large set of things that can be literally counted. If
    there were a million pairs of nearly identical couples, THEN the
    frequentist would let it count as a "probability", but not otherwise.

    Hence, some kind of frequentist might say of Eliezer's question 1, as
    posed:

    > 1. After repeatedly revising your probability to take each of these
    > observations into account, what is your estimated chance that the barrel
    > is the one containing mostly blue tokens?

    that the question, as stated, doesn't make any sense: there is no such
    thing as an estimated probability or prior chance in this case. Hence,
    nothing to calculate, and nowhere to begin. (It's hard to say what
    exactly a "real" self-described frequentist might say in response to
    Eliezer's example, because it's easy to see how the experiment is, at
    least, repeatable in principle, as opposed to, say, the married couple
    example, which really just can't be repeated with any kind of scientific
    rigor as far as we know: lovers attitudes towards each other are
    irretrievably dependent on history, so we wouldn't call any repeated
    attempts at getting them hitched "repetitions" of the "same" kind of
    event.)

    But that frequentist WOULD be comfortable answering question 3.
    Furthermore, if this frequentist were really hard-core, she might reject
    Bayes' Theorem altogether and try to find the frequency by actually
    repeating the experiment a bunch of times. That, the frequentist might
    insist, would be the only way to guarantee that a meaningful result was
    being calculated.

    So a frequentist like the kind I'M describing, as OPPOSED to what you did,
    would literally write a program (or use an expensive pre-existing one,
    depending on how much money was on hand) and do the experiment over and
    over again, until an answer to question 3 was found to within some
    experimental precision. (Note that talking about "precision" here tends
    to require some background assumptions like those Goertzel mentioned in
    his answer to your question on SL4. What could that "precision" mean
    unless you assumed something Gaussian were going on in the background?)

    Looking back at the problem, I have no idea which barrel is red, (since
    neither barrel was painted,) but, on the presumption that the barrel with
    mostly blue tokens was painted blue, and the other barrel (which you
    failed to notice had an EQUAL number of blue and red tokens) must have
    been painted red. Neither do I find myself able to say much about how the
    frequentist, even my hard-core frequentist, has actively harmed her
    statistical understanding... even frequentists can plug numbers into
    Bayes as well as anybody else, but sometimes they choose not to on
    principle.

    But if I had to give an answer, it would be this: "orthodox statisticians
    cannot explain WHY the given answer is correct; they couldn't have
    predicted the answer, and couldn't therefore have been "surprised" by the
    results." [Hence, they have left the path of science.]

    I appear to have narrowly met the word limit. Well, hopefully my wordy
    answer has served you better than the terse but certainly charming
    response Elizer himself probably would have given.

    -Dan

          -unless you love someone-
        -nothing else makes any sense-
               e.e. cummings



    This archive was generated by hypermail 2.1.5 : Mon Jul 07 2003 - 19:53:17 MDT