15
Seminar Series in Quantitative Studies in Consumer Behavior 11 April 2006 George & Foster (2000) George & Foster (2000) Calibration and Empirical Bayes Calibration and Empirical Bayes Variable Selection Variable Selection Discussion of

George & Foster (2000) “Calibration and Empirical … · Seminar Series in Quantitative Studies in Consumer Behavior 11 April 2006 George & Foster (2000) “Calibration and Empirical

Embed Size (px)

Citation preview

Page 1: George & Foster (2000) “Calibration and Empirical … · Seminar Series in Quantitative Studies in Consumer Behavior 11 April 2006 George & Foster (2000) “Calibration and Empirical

Seminar Series in Quantitative Studies in Consumer Behavior

11 April 2006

George & Foster (2000)George & Foster (2000)

““Calibration and Empirical BayesCalibration and Empirical BayesVariable SelectionVariable Selection””

Discussion of

Page 2: George & Foster (2000) “Calibration and Empirical … · Seminar Series in Quantitative Studies in Consumer Behavior 11 April 2006 George & Foster (2000) “Calibration and Empirical

11 April 2006 2

Overview

Calibration Empirical Bayes

Bayes

Cp / AIC

BIC

RIC

Data based methods for prior specification

Page 3: George & Foster (2000) “Calibration and Empirical … · Seminar Series in Quantitative Studies in Consumer Behavior 11 April 2006 George & Foster (2000) “Calibration and Empirical

11 April 2006 3

Common Approaches

Choose model that maximizes:

dimension of model γ

“dimension penalty”

Page 4: George & Foster (2000) “Calibration and Empirical … · Seminar Series in Quantitative Studies in Consumer Behavior 11 April 2006 George & Foster (2000) “Calibration and Empirical

11 April 2006 4

Common Approaches

F Criterion

2Cp

(AIC)Mallows (1973)Akaike (1973)

log n BIC Schwartz (1978)

2 log p RIC Foster & George (1994)

Page 5: George & Foster (2000) “Calibration and Empirical … · Seminar Series in Quantitative Studies in Consumer Behavior 11 April 2006 George & Foster (2000) “Calibration and Empirical

11 April 2006 5

Bayes Approach

Model Space Posterior

Marginal Likelihood

Prior Distributions

Calibrate a particular class of priors to F

Page 6: George & Foster (2000) “Calibration and Empirical … · Seminar Series in Quantitative Studies in Consumer Behavior 11 April 2006 George & Foster (2000) “Calibration and Empirical

11 April 2006 6

Priors

Bernoulli / g-priorHyperparameters

Note:

c controls the expected size of nonzero coefficients

w controls the number of nonzero coefficients

Page 7: George & Foster (2000) “Calibration and Empirical … · Seminar Series in Quantitative Studies in Consumer Behavior 11 April 2006 George & Foster (2000) “Calibration and Empirical

11 April 2006 7

g-priors are…

…conditionally compatible

…a limiting case of a latent factor regression model (West, 2001)

…computationally tractable

...can be problematic for model selection (Jeffreys, 1961; Berger & Pericchi, 2001):

Page 8: George & Foster (2000) “Calibration and Empirical … · Seminar Series in Quantitative Studies in Consumer Behavior 11 April 2006 George & Foster (2000) “Calibration and Empirical

11 April 2006 8

Calibration

Page 9: George & Foster (2000) “Calibration and Empirical … · Seminar Series in Quantitative Studies in Consumer Behavior 11 April 2006 George & Foster (2000) “Calibration and Empirical

11 April 2006 9

F(c,w)

0

-5

-10

5

10

very negative

very positive

Page 10: George & Foster (2000) “Calibration and Empirical … · Seminar Series in Quantitative Studies in Consumer Behavior 11 April 2006 George & Foster (2000) “Calibration and Empirical

11 April 2006 10

Calibration

F(c,w) Criterion c w

2Cp

(AIC)3.92 0.5

0.5

0.5

log n BIC n

2 log p RIC p2

Page 11: George & Foster (2000) “Calibration and Empirical … · Seminar Series in Quantitative Studies in Consumer Behavior 11 April 2006 George & Foster (2000) “Calibration and Empirical

11 April 2006 11

Empirical Bayes Estimation of c and w

• “Default” choices can bias posterior away from “true” underlying model

• Fully Bayes approach: prior p(c,w)

• Stochastic search methods only visit a small area of the model space

Page 12: George & Foster (2000) “Calibration and Empirical … · Seminar Series in Quantitative Studies in Consumer Behavior 11 April 2006 George & Foster (2000) “Calibration and Empirical

11 April 2006 12

Note:

Maximize (numerically) to get

Not computationally feasible for large p

Empirical Bayes Estimation of c and w

Marginal Likelihood

Page 13: George & Foster (2000) “Calibration and Empirical … · Seminar Series in Quantitative Studies in Consumer Behavior 11 April 2006 George & Foster (2000) “Calibration and Empirical

11 April 2006 13

Note:

Only p distinct terms - maximization is (computationally) feasible

MAP model maximizes

Empirical Bayes Estimation of c and w

Orthogonal Design

Page 14: George & Foster (2000) “Calibration and Empirical … · Seminar Series in Quantitative Studies in Consumer Behavior 11 April 2006 George & Foster (2000) “Calibration and Empirical

11 April 2006 14

Conditional EBayes Approach

conditional on a model γ

maximize rather than marginalize:

Note:

Under the “usual” estimate

Page 15: George & Foster (2000) “Calibration and Empirical … · Seminar Series in Quantitative Studies in Consumer Behavior 11 April 2006 George & Foster (2000) “Calibration and Empirical

11 April 2006 15

Conditional EBayes Approach

Notes:

Computation easy, but still have to search model space

When p(y , γ | c, w) dominated by single model, CMML and CCML are similar