{ "cells": [ { "cell_type": "markdown", "id": "35b3177d-c3fc-4c9f-a6a4-e4824c0f16f4", "metadata": {}, "source": [ "# 2. Small Worlds and Large Worlds" ] }, { "cell_type": "markdown", "id": "dbd8a8ba-296a-4092-90db-dff6f28c96b7", "metadata": {}, "source": [ "## 2.1 The garden of forking data" ] }, { "cell_type": "markdown", "id": "fcc2b1c1-c716-4e0c-ba04-42d1aea49252", "metadata": {}, "source": [ "### 2.1.1 Counting possibilities" ] }, { "cell_type": "markdown", "id": "f27a37ae-d826-4273-9602-48c65f257beb", "metadata": {}, "source": [ "The author's footnote refers to [Cox's theorem](https://en.wikipedia.org/wiki/Cox%27s_theorem); other justifications are given in [Bayesian probability § Justification](https://en.wikipedia.org/wiki/Bayesian_probability#Justification).\n", "\n", "As discussed in [Cox's theorem § Interpretation and further discussion](https://en.wikipedia.org/wiki/Cox%27s_theorem#Interpretation_and_further_discussion), there's plenty of reason to doubt this justification.\n", "\n", "An arguably better justification is the [Dutch book theorems](https://en.wikipedia.org/wiki/Dutch_book_theorems); see [Dutch Book Arguments (Stanford Encyclopedia of Philosophy)](https://plato.stanford.edu/entries/dutch-book/) and [Notes on the Dutch Book Argument](https://www.stat.berkeley.edu/~freedman/dutchdef.pdf) (by David A Freedman) for some more rigorous mathematics going back to the original person to make the argument (De Finetti). This justification remains completely finite, which seems desirable, not only if you have prior commitments to finitism but based on the following research (quoting from [David A. Freedman](https://en.wikipedia.org/wiki/David_A._Freedman)):\n", "\n", "> In particular, the 1965 paper with the innocent title \"On the asymptotic behaviour of Bayes estimates in the discrete case II\" finds the rather disappointing answer that when sampling from a countably infinite population the [Bayesian procedure](https://en.wikipedia.org/wiki/Bayesian_inference \"Bayesian inference\") fails almost everywhere, i.e., one does not obtain the true distribution asymptotically. This situation is quite different from the finite case when the (discrete) random variable takes only finite many values and the Bayesian method is consistent in agreement with earlier findings of Doob (1948)." ] }, { "cell_type": "markdown", "id": "74d04ea9-05ad-45f2-a209-3da63cc78420", "metadata": {}, "source": [ "From [Bayesian inference § Alternatives to Bayesian updating](https://en.wikipedia.org/wiki/Bayesian_inference#Alternatives_to_Bayesian_updating):" ] }, { "cell_type": "markdown", "id": "148e3c0e-eaa3-4432-ab83-267e8bcf00f9", "metadata": {}, "source": [ "> [Ian Hacking](https://en.wikipedia.org/wiki/Ian_Hacking \"Ian Hacking\") noted that traditional \"[Dutch book](https://en.wikipedia.org/wiki/Dutch_book \"Dutch book\")\" arguments did not specify Bayesian updating: they left open the possibility that non-Bayesian updating rules could avoid Dutch books. Hacking wrote: \"And neither the Dutch book argument nor any other in the personalist arsenal of proofs of the probability axioms entails the dynamic assumption. Not one entails Bayesianism. So the personalist requires the dynamic assumption to be Bayesian. It is true that in consistency a personalist could abandon the Bayesian model of learning from experience. Salt could lose its savour.\"\n", ">\n", "> Indeed, there are non-Bayesian updating rules that also avoid Dutch books (as discussed in the literature on \"[probability kinematics](https://en.wikipedia.org/wiki/Probability_kinematics \"Probability kinematics\")\") following the publication of [Richard C. Jeffrey](https://en.wikipedia.org/wiki/Richard_C._Jeffrey \"Richard C. Jeffrey\")'s rule, which applies Bayes' rule to the case where the evidence itself is assigned a probability. The additional hypotheses needed to uniquely require Bayesian updating have been deemed to be substantial, complicated, and unsatisfactory." ] }, { "cell_type": "markdown", "id": "c6e2d6e6-06ab-4cd4-8637-f5d03bca03ce", "metadata": {}, "source": [ "### 2.1.2 Combining other information" ] }, { "cell_type": "markdown", "id": "8cc3f9ec-3906-4477-867c-8ac436f94b64", "metadata": {}, "source": [ "### 2.1.3 From counts to probability" ] }, { "cell_type": "markdown", "id": "8317cee9-6e26-4a1f-840e-9eaf5c1350ab", "metadata": {}, "source": [ "The author uses the term \"plausibility\" as a synonym for probability; it's not clear why the word is being introduced." ] }, { "cell_type": "code", "execution_count": 1, "id": "0498e39a-68ff-4bed-966f-4791ff6ebe66", "metadata": {}, "outputs": [ { "data": { "text/html": [ "\n", "
mean | sd | 5.5% | 94.5% | |
---|---|---|---|---|
<dbl> | <dbl> | <dbl> | <dbl> | |
p | 0.6666669 | 0.1571337 | 0.4155369 | 0.917797 |