Premium
Model averaging in linkage analysis
Author(s) -
Matthysse Steven
Publication year - 2006
Publication title -
american journal of medical genetics part b: neuropsychiatric genetics
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 1.393
H-Index - 126
eISSN - 1552-485X
pISSN - 1552-4841
DOI - 10.1002/ajmg.b.30256
Subject(s) - markov chain monte carlo , mathematics , approximate bayesian computation , bayes' theorem , parametric statistics , markov chain , parameter space , bayes factor , random walk , range (aeronautics) , population , detailed balance , prior probability , monte carlo method , statistics , bayesian probability , computer science , statistical physics , inference , artificial intelligence , materials science , demography , physics , sociology , composite material
Abstract Methods for genetic linkage analysis are traditionally divided into “model‐dependent” and “model‐independent,” but there may be a useful place for an intermediate class, in which a broad range of possible models is considered as a parametric family. It is possible to average over model space with an empirical Bayes prior that weights models according to their goodness of fit to epidemiologic data, such as the frequency of the disease in the population and in first‐degree relatives (and correlations with other traits in the pleiotropic case). For averaging over high‐dimensional spaces, Markov chain Monte Carlo (MCMC) has great appeal, but it has a near‐fatal flaw: it is not possible, in most cases, to provide rigorous sufficient conditions to permit the user safely to conclude that the chain has converged. A way of overcoming the convergence problem, if not of solving it, rests on a simple application of the principle of detailed balance. If the starting point of the chain has the equilibrium distribution, so will every subsequent point. The first point is chosen according to the target distribution by rejection sampling, and subsequent points by an MCMC process that has the target distribution as its equilibrium distribution. Model averaging with an empirical Bayes prior requires rapid estimation of likelihoods at many points in parameter space. Symbolic polynomials are constructed before the random walk over parameter space begins, to make the actual likelihood computations at each step of the random walk very fast. Power analysis in an illustrative case is described. © 2006 Wiley‐Liss, Inc.