Bayesian robust transformation and variable selection: A unified approach

The authors consider the problem of simultaneous transformation and variable selection for linear regression. They propose a fully Bayesian solution to the problem, which allows averaging over all models considered including transformations of the response and predictors. The authors use the Box-Cox family of transformations to transform the response and each predictor. To deal with the change of scale induced by the transformations, the authors propose to focus on new quantities rather than the estimated regression coefficients. These quantities, referred to as generalized regression coefficients, have a similar interpretation to the usual regression coefficients on the original scale of the data, but do not depend on the transformations. This allows probabilistic statements about the size of the effect associated with each variable, on the original scale of the data. In addition to variable and transformation selection, there is also uncertainty involved in the identification of outliers in regression. Thus, the authors also propose a more robust model to account for such outliers based on a t-distribution with unknown degrees of freedom. Parameter estimation is carried out using an efficient Markov chain Monte Carlo algorithm, which permits moves around the space of all possible models. Using three real data sets and a simulated study, the authors show that there is considerable uncertainty about variable selection, choice of transformation, and outlier identification, and that there is advantage in dealing with all three simultaneously. The Canadian Journal of Statistics 37: 361–380; 2009 © 2009 Statistical Society of Canada Nous considerons le probleme de la selection de transformations et de variables pour la regression lineaire. Nous proposons une approche Bayesienne a ce probleme qui nous permet de faire la moyenne de tous les modeles consideres y compris les transformations de type Box-Cox de la response et des predicteurs. Pour prendre en consideration le changement d'unite induit par les transformations, nous proposons d'examiner et d'estimer de nouvelles quantites a la place des coefficients de regression. Ces quantites nouvelles, que nous appellons coefficients de regressions generalises, peuvent etre interpretes comme les coefficients de regression dans l'unite originale des donnees, et ne dependent donc pas des transformations selectionees. En particulier, cela nous permet de faire de l'inference sur la taille des effets associes avec chaque variable, et ce, dans l'unite original des donnees. En plus des transformations, nous considerons aussi le probleme de la detection de valeurs aberrantes, ainsi que l'incertitude associee a cette detection. Pour modeliser ces donnees aberrantes, nous utilisons une loi de t avec un nombre de degres de liberte inconnu. L'estimation des parametres est faite en utilisant un algorithm MCMC efficace qui nous permet de traverser l'espace constitue de tous les modeles possibles. En utilisant trois jeux de donnees reelles ainsi que des donnees simulees, nous montrons que l'incertitude associee au choix de variables, de transformations et de donnees aberrantes est considerable, et qu'il est important que les trois selections soient considerees en meme temps. La revue canadienne de statistique 37: 361–380; 2009 © 2009 Societe statistique du Canada

[1]  R. Lockhart,et al.  Box‐Cox transformed linear models: A parameter‐based asymptotic approach , 1997 .

[2]  E. George,et al.  APPROACHES FOR BAYESIAN VARIABLE SELECTION , 1997 .

[3]  R. D. Cook,et al.  Transformations and Influential Cases in Regression , 1983 .

[4]  J. York,et al.  Bayesian Graphical Models for Discrete Data , 1995 .

[5]  D. Madigan,et al.  A method for simultaneous variable selection and outlier identification in linear regression , 1996 .

[6]  D. Ruppert,et al.  Transformations in Regression: A Robust Analysis , 1985 .

[7]  R. Kohn,et al.  Nonparametric regression using Bayesian variable selection , 1996 .

[8]  A. C. Atkinson [Influential Observations, High Leverage Points, and Outliers in Linear Regression]: Comment: Aspects of Diagnostic Regression Analysis , 1986 .

[9]  Michael A. Stephens,et al.  Box‐Cox transformations in linear models: Large sample theory and tests of normality , 2002 .

[10]  Raphael Gottardo,et al.  Markov Chain Monte Carlo With Mixtures of Mutually Singular Distributions , 2008 .

[11]  Luis R. Pericchi,et al.  A Bayesian approach to transformations to normality , 1981 .

[12]  Jun S. Liu,et al.  Bayesian Clustering with Variable and Transformation Selections , 2003 .

[13]  David Ruppert,et al.  Robust Estimation in Heteroscedastic Linear Models. , 1982 .

[14]  James G. Scott,et al.  An exploration of aspects of Bayesian multiple testing , 2006 .

[15]  G. Parisi,et al.  Simulated tempering: a new Monte Carlo scheme , 1992, hep-lat/9205018.

[16]  D. Hinkley,et al.  The Analysis of Transformed Data , 1984 .

[17]  C. Geyer Markov Chain Monte Carlo Maximum Likelihood , 1991 .

[18]  D. Hinkley,et al.  More about transformations and influential cases in regression , 1988 .

[19]  I. Ehrlich Participation in Illegitimate Activities: A Theoretical and Empirical Investigation , 1973, Journal of Political Economy.

[20]  David Ruppert,et al.  On prediction and the power transformation family , 1981 .

[21]  S. Chatterjee,et al.  Influential Observations, High Leverage Points, and Outliers in Linear Regression , 1986 .

[22]  Tsung-Chi Cheng,et al.  Robust regression diagnostics with data transformations , 2005, Comput. Stat. Data Anal..

[23]  Trevor J. Sweeting,et al.  ON THE CHOICE OF PRIOR DISTRIBUTION FOR THE BOX-COX TRANSFORMED LINEAR-MODEL , 1984 .

[24]  P. Royston,et al.  Regression using fractional polynomials of continuous covariates: parsimonious parametric modelling. , 1994 .

[25]  R. Carroll Power Transformations When the Choice of Power is Restricted to a Finite Set. , 1982 .

[26]  Charles Kooperberg,et al.  Spline Adaptation in Extended Linear Models (with comments and a rejoinder by the authors , 2002 .

[27]  Richard F. Gunst,et al.  Applied Regression Analysis , 1999, Technometrics.

[28]  D. Cox,et al.  An Analysis of Transformations , 1964 .

[29]  D. Madigan,et al.  Bayesian Model Averaging for Linear Regression Models , 1997 .

[30]  D. Cox,et al.  An Analysis of Transformations Revisited, Rebutted , 1982 .

[31]  Jeremy MG Taylor,et al.  The Retransformed Mean after a Fitted Power Transformation , 1986 .

[32]  M. Hansen,et al.  Spline Adaptation in Extended Linear Models , 1998 .

[33]  Jean-Marie Dufour,et al.  Pitfalls of Rescalling Regression Models with Box-Cox Transformations , 1994 .

[34]  G. Box,et al.  Transformation of the Independent Variables , 1962 .

[35]  J. Geweke,et al.  Variable selection and model comparison in regression , 1994 .

[36]  David Madigan,et al.  Bayesian Variable and Transformation Selection in Linear Regression , 2002 .

[37]  R. Dennis Cook,et al.  Detection of Influential Observation in Linear Regression , 2000, Technometrics.

[38]  A. C. Atkinson,et al.  Transformations unmasked , 1988 .

[39]  P. Bickel,et al.  An Analysis of Transformations Revisited , 1981 .

[40]  E. George,et al.  Journal of the American Statistical Association is currently published by American Statistical Association. , 2007 .

[41]  P. Green,et al.  Bayesian Variable Selection and the Swendsen-Wang Algorithm , 2004 .

[42]  G. Box An analysis of transformations (with discussion) , 1964 .

[43]  Joseph G. Ibrahim,et al.  Bayesian predictive simultaneous variable and transformation selection in the linear model , 1998 .