## Parsimonious Estimation of Multiplicative Interaction in Analysis of Variance using Kullback-Leibler Information (1999)

Venue: | Journal of Statistical Planning and Inference |

Citations: | 2 - 1 self |

### BibTeX

@ARTICLE{Viele99parsimoniousestimation,

author = {Kert Viele and C. Srinivasan},

title = {Parsimonious Estimation of Multiplicative Interaction in Analysis of Variance using Kullback-Leibler Information},

journal = {Journal of Statistical Planning and Inference},

year = {1999},

volume = {84},

pages = {201--219}

}

### OpenURL

### Abstract

Many standard methods for modeling interaction in two way ANOVA require mn interaction parameters, where m and n are the number of rows and columns in the table. By viewing the interaction parameters as a matrix and performing a singular value decomposition, one arrives at the Additive Main Effects and Multiplicative Interaction (AMMI) model which is commonly used in agriculture. By using only those interaction components with the largest singular values, one can produce an estimate of interaction that requires far fewer than mn parameters while retaining most of the explanatory power of standard methods. The central inference problems of estimating the parameters and determining the number of interaction components has been difficult except in "ideal" situations (equal cell sizes, equal variance, etc.). The Bayesian methodology developed in this paper applies for unequal sample sizes and heteroscedastic data, and may be easily generalized to more complicated data structures...

### Citations

2243 |
Equation of state calculations by fast computing machines
- Metropolis, Rosenbluth, et al.
- 1953
(Show Context)
Citation Context ...rs and defining ff m = \Gamma(ff 1 + : : : + ff m\Gamma1 ) and fi n = \Gamma(fi 1 + : : : + fi n\Gamma1 ). Updating the coordinate axis parameters fl and ffi requires use of the Metropolis algorithm (=-=Metropolis et al 1953-=-). Because the fl vectors are fairly rigidly constrained with respect to one another, care must be taken in updating the set. If a full rank interaction is specified, one cannot update the vectors one... |

1235 |
Information theory and an extension of the maximum likelihood principle
- Akaike
- 1973
(Show Context)
Citation Context ...d calculations for implementing the KL measure. KL information is fundamental to determining the complexity of a model (van Emden 1971) and hence is used is many model selection criteria such as AIC (=-=Akaike 1973-=-) and ICOMP (Bozdogan and Haughton 1996). Two examples are shown in section 4, reanalyzing data on a soybean yield experiment from Zobel, Wright, and Gauch (1988) and a tomato softness study from Oman... |

753 | Markov chains for exploring posterior distributions
- Tierney
- 1994
(Show Context)
Citation Context ... 3 (m\Gamma3) ), and fl 3 = A fl 3 j fl 3 . 2.4 Gibbs Sampler The posterior distribution of the parameters may be estimated through Markov Chain Monte Carlo (MCMC) techniques (Gelfand and Smith 1990, =-=Tierney 1994-=-). The complete conditionals for , ff i , fi j ,st , and oe 2 may be computed using the conjugate property of the prior distributions. As discussed in section 2.1, we avoid singularities in the covari... |

744 |
Sampling-based approaches to calculating marginal densities
- Gelfand, Smith
- 1990
(Show Context)
Citation Context ...; 0;sfl 3 1 ; : : : ;sfl 3 (m\Gamma3) ), and fl 3 = A fl 3 j fl 3 . 2.4 Gibbs Sampler The posterior distribution of the parameters may be estimated through Markov Chain Monte Carlo (MCMC) techniques (=-=Gelfand and Smith 1990-=-, Tierney 1994). The complete conditionals for , ff i , fi j ,st , and oe 2 may be computed using the conjugate property of the prior distributions. As discussed in section 2.1, we avoid singularities... |

533 |
Probability and statistics
- DEGROOT, SCHERVISH
- 2004
(Show Context)
Citation Context ...s similar) is the multivariate normal distribution Nm i 0; oe 2 ff h I \Gamma (1=m)11 T ij ; which is the conditional distribution of m independent N(0; oe 2 ff ) random variables given the sum is 0 (=-=Schervish 1995-=-, section 8.2.2). Although the covariance matrix is singular, it has the advantages that the generated ff i sum to 0 and all the ff i are exchangeable. In practice, we avoid difficulties with the sing... |

58 |
Computing Bayes factors using a generalization of the Savage-Dickey density ratio
- Verdinelli, Wasserman
- 1995
(Show Context)
Citation Context ...power. If one does want a Bayes factor for comparing the point null H 0 : KL = 0 to H 1 : KL ? 0, then a may be decreased to 0 and the Bayes factor may be found using the Savage-Dickey density ratio (=-=Verdinelli and Wasserman 1995-=-). Since the observed Y have normal distributions, it is useful to calibrate the KL information in terms of normal distributions to determine exactly what values are large or small. A N(0; 1) distribu... |

42 |
Limiting behavior of posterior distributions when the model is incorrect
- Berk
- 1966
(Show Context)
Citation Context ...bution of the parameters for the F q model will converge to the parameter values that minimize (mn) \Gamma1 X ij KL(F p (i; j); F q (i; j)); the averaged KL information across the cells in the table (=-=Berk 1966-=-). LetsF q be the minimizing model. The parameters ofsF q are random variables since the F p parameters are unknown. We evaluate fit by looking at the posterior distribution of d(F p ;sF q ) as the F ... |

13 | Diagnostic measures for model criticism - CAROTA, PARMIGIANI, et al. - 1996 |

12 | Informational complexity criteria for regression models - Bozdogan, Haughton - 1998 |

12 | A statistical model which combines features of factor analytic and analysis of variance techniques - Gollob - 1968 |

3 |
The effect of Multiplicative Scatter Correction (MSC) and linearity improvement in NIR spectroscopy
- Isaksson, Næs
- 1988
(Show Context)
Citation Context ...lue decomposition). This reparameterization, originally considered in Gollob (1968), has already been used in a variety of contexts, such as genetics (Elkind and Cahaner 1986), infrared spectroscopy (=-=Isakkson and Naes 1988-=-), latent variable models in education (Muthen and Satorra 1989) and analysis of crop yield data (Zobel, Wright, and Gauch 1989, Cornelius, Crossa and Seyedsadr 1995, hereafter CCS). Some related work... |

3 |
Multiplicative effects in mixed model analysis of variance
- Oman
- 1991
(Show Context)
Citation Context ...ume, without loss of generality, that K = 1. Thus the model under consideration is y ij =s+ ff i + fi j + fl i ffi j + ffl ij This special case is adequate to address heteroscedasticity in the error (=-=Oman 1991-=-). The following discussion extends to the general case with T ? 1 or K ? 1. To insure the parameters are uniquely defined assume P ff i = P fi j = P fl i = P ffi j = 0, and that fl 1 and ffi 1 are no... |

2 |
Bayes Factors.” Journal of the American Statistical Association 90:773-95
- Kass, Raftery
- 1995
(Show Context)
Citation Context .... This section describes a method for quantifying the increase in predictive power from using a particular number of interaction components. An alternative approach would be to compute Bayes Factors (=-=Kass and Raftery 1995-=-) for the interaction components, which we discuss briefly in this section. 7 Carota, Parmigiani, and Polson (1996) (CPP) and Viele (1996) discuss methods of evaluating fit based on model embedding th... |

1 |
A Mixed Model for the Effects of Single Genes
- Elkind, Cahaner
- 1986
(Show Context)
Citation Context ...tors form orthonormal matrices in the singular value decomposition). This reparameterization, originally considered in Gollob (1968), has already been used in a variety of contexts, such as genetics (=-=Elkind and Cahaner 1986-=-), infrared spectroscopy (Isakkson and Naes 1988), latent variable models in education (Muthen and Satorra 1989) and analysis of crop yield data (Zobel, Wright, and Gauch 1989, Cornelius, Crossa and S... |

1 | Stepwise Tests for Multiplicative Components of Interaction - Yochmowitz, Cornell - 1978 |