The name is absent



50

(a) Given the previous values for θ*, K and φ, we generate a new configuration by
sampling new values for the indicator variables using (3.21), replacing
θ↑,θ⅛,....
When sampling φi = 0 we associate the observation i, xi, with a new draw from
gi0 given in (3.20) and update the configuration accordingly.

Note that using a base measure G0 conjugate to fi is mathematically convenient.
It makes sampling from
gi0 and computation of hi easier. For computational reasons
conjugate families should be considered when appropriate. In the model introduced
in Section 3.4 this is the case.

Repeated use of step (a) defines a Markov chain with limiting distribution equal
to the posterior distribution
p(θ*,K,φ ∣ rrn). But, this Markov chain mixes poorly.
Since it shifts one value of
θi at a time, rarely does a value of in a clusters change.
This change needs the chain to pass trough a middle state of low probability in which
all indices in the cluster
Sk are moved to other clusters. Convergence is accelerated
by avoiding this phenomenon. Once the values of
θ*, K and φ have been obtained in
step (a), the vector
θ* is resampled conditional on the values of K and φ. That is, a
second step is added to the Gibbs sampling scheme (after each step (a)),

(b) Given K and φ, we draw a new set of parameters Θ* by sampling new values
from the posterior distribution (3.14) .

Sequential iterations over (a) and (b) defines a Markov chain that converges to
the posterior distribution
p(β*,K,φ xn). For details about this convergence see
MacEachern and Müller (1998) . The posterior distribution of any function of the
parameters can be estimated based on the posterior sample. For example the aver-
age of (3.17), with respect to the simulated values of flɪ,...,
θ*κ, and θn+1 estimates
the predictive distribution
p(xn ∣ ʃ") = G. This average is used for the density
estimation. See the example given in the next subsection.

Now we extend the Gibbs sampling scheme to include the common parameter
σ. Let p(σ) denote the prior for σ. We draw this parameter from the appropriate



More intriguing information

1. The name is absent
2. Can we design a market for competitive health insurance? CHERE Discussion Paper No 53
3. Naïve Bayes vs. Decision Trees vs. Neural Networks in the Classification of Training Web Pages
4. Investment and Interest Rate Policy in the Open Economy
5. The Shepherd Sinfonia
6. Tobacco and Alcohol: Complements or Substitutes? - A Statistical Guinea Pig Approach
7. The name is absent
8. Natural hazard mitigation in Southern California
9. Education as a Moral Concept
10. The Importance of Global Shocks for National Policymakers: Rising Challenges for Central Banks
11. The name is absent
12. Studying How E-Markets Evaluation Can Enhance Trust in Virtual Business Communities
13. Midwest prospects and the new economy
14. Direct observations of the kinetics of migrating T-cells suggest active retention by endothelial cells with continual bidirectional migration
15. Second Order Filter Distribution Approximations for Financial Time Series with Extreme Outlier
16. TRADE NEGOTIATIONS AND THE FUTURE OF AMERICAN AGRICULTURE
17. The name is absent
18. Correlation Analysis of Financial Contagion: What One Should Know Before Running a Test
19. Flatliners: Ideology and Rational Learning in the Diffusion of the Flat Tax
20. The name is absent