## Bayesian Selection of Log-Linear Models (1995)

Venue: | Canadian Journal of Statistics |

Citations: | 7 - 2 self |

### BibTeX

@ARTICLE{Albert95bayesianselection,

author = {James H. Albert},

title = {Bayesian Selection of Log-Linear Models},

journal = {Canadian Journal of Statistics},

year = {1995},

volume = {24},

pages = {327--347}

}

### OpenURL

### Abstract

A general methodology is presented for finding suitable Poisson log-linear models with applications to multiway contingency tables. Mixtures of multivariate normal distributions are used to model prior opinion when a subset of the regression vector is believed to be nonzero. This prior distribution is studied for two and three-way contingency tables, in which the regression coefficients are interpretable in terms of odds-ratios in the table. Efficient and accurate schemes are proposed for calculating the posterior model probabilities. The methods are illustrated for a large number of two-way simulated tables and for two three-way tables. These methods appear to be useful in selecting the best log-linear model and in estimating parameters of interest that reflect uncertainty in the true model. Key words and phrases: Bayes factors, Laplace method, Gibbs sampling, Model selection, Odds ratios. AMS subject classifications: Primary 62H17, 62F15, 62J12. 1 Introduction 1.1 Bayesian testing...

### Citations

1404 | Categorical Data Analysis - Agresti - 1990 |

801 | Markov chains for exploring posterior distributions (with discussion
- Tierney
- 1994
(Show Context)
Citation Context ...ion, the Gibbs simulation algorithm of this paper can be time-consuming when the number of possible models is large. Since many alternative Markov chain simulation algorithms are currently available (=-=Tierney, 1991-=-), it seems desirable to look for new methods that are less demanding computationally and can be used to search through large model spaces. These improvements will make the procedure more attractive f... |

744 | Inference from iterative simulation using multiple sequences. Statistical Science 7, 457–511. With Discussion - Gelman, Rubin - 1992 |

433 |
Monte Carlo Methods
- Hammersley, Handscomb
- 1964
(Show Context)
Citation Context ...i fi 1=2 : (3) Generally, for the examples considered in this paper, the Laplace approximation (3) is very accurate. One can check the accuracy of the approximation by the use of importance sampling (=-=Hammersley and Handscomb, 1964-=-, Geweke, 1989). Write the integral R p(y; fi 0 )d k as R [p(y;sk )=h( k )]h( k )d k , where the importance sampling density h() is a multivariate t density with location ~ fi 0 , scale matrix ~ \Sigm... |

394 |
Sampling based approaches to calculating marginal densities
- Gelfand, Smith
- 1990
(Show Context)
Citation Context ...tios of the table. In Section 4 we discuss efficient schemes for computing the posterior model probabilities. The basic tools are the Laplace method (Tierney and Kadane, 1986), and the Gibbs sampler (=-=Gelfand and Smith, 1990-=-, George and McCulloch, 1993) in the comparison of many models. These methods are illustrated in Section 4 for a large number of simulated twoway contingency tables and for two three-way tables where ... |

347 |
Variable Selection via Gibbs Sampling
- George, McCulloch
- 1993
(Show Context)
Citation Context ...tion 4 we discuss efficient schemes for computing the posterior model probabilities. The basic tools are the Laplace method (Tierney and Kadane, 1986), and the Gibbs sampler (Gelfand and Smith, 1990, =-=George and McCulloch, 1993-=-) in the comparison of many models. These methods are illustrated in Section 4 for a large number of simulated twoway contingency tables and for two three-way tables where there are a large number of ... |

343 |
Bayesian Inference in Econometric Models Using Monte-Carlo Integration
- Geweke
- 1989
(Show Context)
Citation Context ...i fi 1=2 : (3) Generally, for the examples considered in this paper, the Laplace approximation (3) is very accurate. One can check the accuracy of the approximation by the use of importance sampling (=-=Geweke, 1989-=-). Write the integral R p(y; fi 0 )dfi 0 as R [p(y; fi 0 )=h(fi 0 )]h(fi 0 )dfi 0 , where the importance sampling density h() is a multivariate t density with location ~ fi 0 , scale matrix ~ \Sigma 0... |

284 |
A Guide to Simulation
- Bratley, Fox, et al.
- 1987
(Show Context)
Citation Context ...f Section 3.2 was run for 5000 iterations and posterior probabilities of each model were approximated by the corresponding relative frequencies in the simulation run. Using the method of batch means (=-=Bratley, et al, 1987-=-), simulation standard errors for these estimated probabilities can be computed. For the two examples of this section, the posterior model probabilities appear to be accurate to the precision that is ... |

281 | Model selection and accounting for model uncertainty in graphical models using Occam’s window - Madigan, Raftery - 1994 |

214 |
Accurate approximations for posterior moments and marginal densities
- Tierney, Kadane
- 1986
(Show Context)
Citation Context ...n parameters are interpretable in terms of odds-ratios of the table. In Section 4 we discuss efficient schemes for computing the posterior model probabilities. The basic tools are the Laplace method (=-=Tierney and Kadane, 1986-=-), and the Gibbs sampler (Gelfand and Smith, 1990, George and McCulloch, 1993) in the comparison of many models. These methods are illustrated in Section 4 for a large number of simulated twoway conti... |

200 | Discrete Multivariate Analysis - Bishop, Fienberg, et al. - 1975 |

126 |
Bayesian statistical inference for psychological research. Psychol Rev 1963; 70
- Edwards, Lindman, et al.
(Show Context)
Citation Context ...he degree of evidence against a null model that is communicated by a p-value. In testing point-null hypotheses, it is well known that p-values typically overstate the evidence against the hypotheses (=-=Edwards et al, 1963-=-, Berger and Sellke, 1987) and so the use of p-values may lead to overfitting, or selection of a model with too many parameters. A second concern regarding this classical methodology relates to the pr... |

109 |
Theory of Probability (3rd ed
- Jeffreys
- 1961
(Show Context)
Citation Context ...his statement would be matched to the value b = 1.5. In the examples of Section 4, we investigate the effect of choosing alternative values of the scale parameter on the values of the Bayes factors. (=-=Jeffreys, 1961-=-, advocated the use of a Cauchy prior in Bayesian testing problems and Crook and Good, 1990, use a log-Cauchy hyperprior on the Dirichlet precision parameter.) In the above two-way contingency table, ... |

104 |
Testing a Point Null Hypothesis: The Irreconcilability of P Values and Evidence
- Berger, Selke
- 1987
(Show Context)
Citation Context ... against a null model that is communicated by a p-value. In testing point-null hypotheses, it is well known that p-values typically overstate the evidence against the hypotheses (Edwards et al, 1963, =-=Berger and Sellke, 1987-=-) and so the use of p-values may lead to overfitting, or selection of a model with too many parameters. A second concern regarding this classical methodology relates to the procedure of first choosing... |

104 | Approximate Bayes factors and accounting for model uncertainty in generalised linear models. Biometrika 83:251266
- Raftery
- 1996
(Show Context)
Citation Context ...ted N(0; P \Gamma1 j ). However, this choice of P j must be made with some care, since the size of the Bayes factor can depend significantly on this value. For this reason, some authors (for example, =-=Raftery, 1993-=-) recommend tabulating values of the Bayes factor for a range of values of P j . An alternative method, which will be used here, is to place a prior distribution on P j (Good, 1976). To motivate this ... |

93 | Bayes factors and model uncertainty - Kass, Raftery - 1995 |

65 |
Testing precise hypotheses
- Berger, Delampady
- 1987
(Show Context)
Citation Context ..., values of fi j close to zero are more likely apriori than values away from zero. This assumption is also made in many other Bayesian applications of testing point null hypotheses (see, for example, =-=Berger and Delampady, 1987-=-). The significant aspects of the prior distribution are the precision parameters in the matrix \Sigma \Gamma1 . First note that the components of j have precisions equal to 0, reflecting vague prior ... |

42 | Bayes Factors for linear and log-linear models with vague prior information - Smith, Spiegelhalter - 1982 |

38 | Applications of a Method for the Efficient Computation of Posterior Distribution - Naylor, Smith - 1982 |

33 | Simple conditions for the convergence of the Gibbs sampler and Metropolis-Hastings algorithms - Roberts, Smith - 1992 |

28 | On the application of symmetric Dirichlet distributions and their mixtures to contingency tables - Good - 1976 |

25 | A note on Bayes factors for log-linear contingency tables with vague prior information - Raftery - 1986 |

11 | Bayes factors for independence in contingency tables - Gunel, Dickey |

6 | A Bayesian test for a two-way contingency table using independence priors. The Canadian - Albert - 1990 |

5 |
Statistical Decision Theory and Bayesian Inference
- BERGER
- 1985
(Show Context)
Citation Context ... full set of interaction parameters described in the previous paragraph. There are some concerns about the use of this informative normal prior distribution from the viewpoint of Bayesian robustness (=-=Berger, 1985-=-). First, although the user may be able to specify the central portion of the prior distribution of the local odds-ratio under the alternative hypothesis H, it is unlikely that he/she has significant ... |

4 | The robustness and sensitivity of the mixed-Dirichlet Bayesian test for “independence” in contingency tables - Good, Crook - 1987 |

3 | Oral-contraceptive use in relation to myorcardial infarction. Lancet i - SHAPIRO, SLONE, et al. - 1979 |

2 | Using loglinear models to analyze cross-classified categorical data - Fienberg - 1980 |

1 | Bayesian estimation of models for cross-classified data - Sedransk, Yang - 1993 |