Thomas S. Dye (tdye@hawaii.edu
)
30 November 2018
The slides for this lecture are publicly available:
http://tsdye.online/bayes/bayesian-inference.html
In plain English: "The posterior probability is proportional to the product of the likelihood and the prior probability."
Bayes' rule is based on a transitivity condition, a sum rule, and a product rule.
Given three propositions, A, B, and C
\(p(X) + p(\bar{X}) = 1\)
\(p(X,Y) = p(X|Y) \times p(Y)\)
Bayes' Rule is a consequence, or corollary, of the sum and product rules and the transitivity condition. Three algebraic steps derive Bayes' Rule. An additional step yields a simplified version.
Transpose \(X\) and \(Y\) in the product rule:
\(p(X,Y) = p(X|Y) \times p(Y)\)
\(p(Y,X) = p(Y|X) \times p(X)\)
Note that \(p(Y,X) = p(X,Y)\):
\(p(X|Y)\ \times\ p(Y) = p(Y|X)\ \times\ p(X)\)
Divide through by \(p(Y)\) to get Bayes' Rule:
\[p(X|Y) = {p(Y|X)\ \times\ p(X) \over p(Y)}\]
Eliminate the denominator on the right hand side and replace the equality sign with a sign of proportionality:
\(p(X|Y) \propto p(Y|X) \times p(X)\)
The 250 year history of Bayesian statistics
Figure 1: Left to right: Rev. Thomas Bayes (1702–1761), Frank P. Ramsey (1903–1930), Bruno de Finetti (1906–1985), Leonard J. Savage (1917–1971).
Figure 2: Left to right: Stuart Geman (1949–), Donald Geman (1943–), Sir Adrian Smith (1946–), Caitlin Buck (1964–)
Figure 3: Buck, C. E., W. G. Cavanagh, and C. D. Litton (1996) Bayesian Approach to Interpreting Archaeological Data. John Wiley & Sons, New York, NY.
Three Bayesian calibration applications are freely available.
Figure 4: https://bcal.shef.ac.uk/top.html
Figure 5: https://c14.arch.ox.ac.uk/oxcal.html
Figure 6: https://chronomodel.com/
This schematic overview of the priors, data, likelihoods, and posteriors glosses over many details that can be found in Buck, Cavanagh, and Litton's book, Bayesian Approach to Interpreting Archaeological Data.
Bayesian inference recognizes that an archaeologist brings knowledge and beliefs to an inquiry. Bayesian calibration organizes this prior information in terms of phases and events. By convention:
A chronological model specifies what is known and believed about the relative ages of the events and phases. Here is a simple model with two phases and two events, where > means "is older than" and = means "is the same age as":
αa > θ1 > βa = αb > θ2 > βb
αa > θ1 > βa = αb > θ2 > βb
Figure 7: Stratigraphic section showing contexts a and b and objects 1 and 2.
Bayesian calibration can model most sources of chronological information that might be used to estimate the age of events.
Figure 8: Age estimates for the events associated with objects 1 and 2. Note the inversion.
Figure 9: Hypothetical Markov Chain Monte Carlo process at the heart of Bayesian calibration.
Figure 10: The inverted age estimates modeled in Chronomodel 2.0.13.
Figure 11: The results of Bayesian calibration are archaeologically interpretable.
Bayesian posteriors can be interpreted as events, occurrences, and processes with the ArchaeoPhases software.
Figure 12: The ArchaeoPhases R
package running as a Shiny application.
Figure 13: Temple construction events in the leeward Kohala field system.
Figure 14: Additions to the system of temples in leeward Kohala.
Figure 15: Tempo of change in the leeward Kohala system of temples.
Figure 16: A processual view of change in old Hawai`i.