MCMC Convergence Slides

MCMC Convergence Slides - Convergence of MCMC Algorithms in...

Info icon This preview shows pages 1–8. Sign up to view the full content.

View Full Document Right Arrow Icon
Convergence of MCMC Algorithms in Finite Samples Anna Kormilitsina and Denis Nekipelov SMU and UC Berkeley September 2009 Kormilitsina, Nekipelov Divergence of MCMC September 2009
Image of page 1

Info iconThis preview has intentionally blurred sections. Sign up to view the full version.

View Full Document Right Arrow Icon
Introduction Motivation MCMC widely used Bayesian method in frequentist context Due to simplicity of coding and promise to converge to global extremum, popular in structural estimation In general requires verification of set of “regularity conditions”: practitioners rarely consider These assumptions can be violated in very common structural models Violation can lead to divergence of algorithm We use example of macro DSGE model: erroneous inference can lead to misinterpretation of policy parameters Kormilitsina, Nekipelov Divergence of MCMC September 2009
Image of page 2
Introduction Our approach MCMC chain: complex dynamic system; in general, stability in such systems can be an issue We use continuous-time approximation for Markov chain; allows us to use results on Lyapunov stability Lyapunov in -stability implies divergence of MCMC chain (with probability 1) We formulate requirements for objective function to guarantee stability If stability is local, convergence will not occur from some regions of parameter space Kormilitsina, Nekipelov Divergence of MCMC September 2009
Image of page 3

Info iconThis preview has intentionally blurred sections. Sign up to view the full version.

View Full Document Right Arrow Icon
Introduction Our results: preview MCMC can diverge even when structural model is identified Create test for stability of chain initialized in particular subset of parameter space: based on Lyapunov stability of its continuous-time approximation Test creates potential for “automatic” choice of support of prior distribution Results are illustrated using commonly used model of (Christiano, Eichenbaum, Evans, 2005) Find that even in simple case MCMC does not have global convergence Kormilitsina, Nekipelov Divergence of MCMC September 2009
Image of page 4
Theory Our definition of “MCMC” MCMC - very large class of algorithms We analyze narrow class of quasi-Bayesian procedures in (Chernozhukov, Hong, 2003) Based on using objective for M-estimation to form a quasi-density (Laplace quasi-posterior) Idea: convergence of statistics of quasi-posterior to extremum estimates (Bernstein-von Mises theorem) leads to convergence of quasi-posterior moment to the M-estimator Study this problem in the context of sampling based on Metropolis-Hastings Kormilitsina, Nekipelov Divergence of MCMC September 2009
Image of page 5

Info iconThis preview has intentionally blurred sections. Sign up to view the full version.

View Full Document Right Arrow Icon
Theory Characterization of Markov chains Create sample of parameter draws from quasi-posterior Procedure can be treated as dynamic system Elements: proposal density, objective function + tuning parameters; output { θ t } Usually have large samples, proposals can be chosen normal/truncated normal Sequence of draws can be approximated by diffusion-based stochastic process Kormilitsina, Nekipelov Divergence of MCMC September 2009
Image of page 6
Theory Characterization of Markov chains Result from theory of SDE: form stochastic differential equation for Langevin diffusion process L t dL t = 1 2 log f ( L t ) dt + dW t where W t standard Brownian motion f will be the stationary distribution of the solution to Langevin equation
Image of page 7

Info iconThis preview has intentionally blurred sections. Sign up to view the full version.

View Full Document Right Arrow Icon
Image of page 8
This is the end of the preview. Sign up to access the rest of the document.

{[ snackBarMessage ]}

What students are saying

  • Left Quote Icon

    As a current student on this bumpy collegiate pathway, I stumbled upon Course Hero, where I can find study resources for nearly all my courses, get online help from tutors 24/7, and even share my old projects, papers, and lecture notes with other students.

    Student Picture

    Kiran Temple University Fox School of Business ‘17, Course Hero Intern

  • Left Quote Icon

    I cannot even describe how much Course Hero helped me this summer. It’s truly become something I can always rely on and help me. In the end, I was not only able to survive summer classes, but I was able to thrive thanks to Course Hero.

    Student Picture

    Dana University of Pennsylvania ‘17, Course Hero Intern

  • Left Quote Icon

    The ability to access any university’s resources through Course Hero proved invaluable in my case. I was behind on Tulane coursework and actually used UCLA’s materials to help me move forward and get everything together on time.

    Student Picture

    Jill Tulane University ‘16, Course Hero Intern