Bayesian variable selection for linear models via non-local priors.
Bayesian model selection for linear, asymmetric linear, median and quantile regression under non-local or Zellner priors. p>>n can be handled.
modelSelection enumerates all models when feasible and uses a Gibbs scheme otherwise. See coef and coefByModel for estimates and posterior intervals of regression coefficients, and rnlp for posterior samples.
modelsearchBlockDiag seeks the highest posterior probability model using an iterative block search.
modelSelection(y, x, data, smoothterms, nknots=9, groups=1:ncol(x), constraints, center=TRUE, scale=TRUE, enumerate, includevars=rep(FALSE,ncol(x)), models, maxvars, niter=5000, thinning=1, burnin=round(niter/10), family='normal', priorCoef, priorGroup, priorDelta=modelbbprior(1,1), priorConstraints, priorVar=igprior(.01,.01), priorSkew=momprior(tau=0.348), phi, deltaini=rep(FALSE,ncol(x)), initSearch='greedy', method='auto', adj.overdisp='intercept', hess='asymp', optimMethod, optim_maxit, initpar='none', B=10^5, XtXprecomp= ifelse(ncol(x)<10^4,TRUE,FALSE), verbose=TRUE) modelsearchBlockDiag(y, x, priorCoef=momprior(tau=0.348), priorDelta=modelbbprior(1,1), priorVar=igprior(0.01,0.01), blocksize=10, maxiter=10, maxvars=100, maxlogmargdrop=20, maxenum=10, verbose=TRUE)
y: Either a formula with the regression equation or a vector with observed responses. The response can be either continuous or of class Surv (survival outcome). If y is a formula then x, groups and constraints are automatically created
x: Design matrix with linear covariates for which we want to assess if they have a linear effect on the response. Ignored if y is a formula
data: If y is a formula then data should be a data frame containing the variables in the model
smoothterms: Formula for non-linear covariates (cubic splines), modelSelection assesses if the variable has no effect, linear or non-linear effect. smoothterms can also be a design matrix or data.frame containing linear terms, for each column modelSelection creates a spline basis and tests no/linear/non-linear effects
nknots: Number of spline knots. For cubic splines the non-linear basis adds knots-4 coefficients for each linear term, we recommend setting nknots to a small/moderate value
groups: If variables in x such be added/dropped in groups, groups indicates the group that each variable corresponds to (by default each variable goes in a separate group)
constraints: Constraints on the model space. List with length equal to the number of groups; if group[[i]]=c(j,k) then group i can only be in the model if groups j and k are also in the model
center: If TRUE, y and x are centered to have zero mean. Dummy variables corresponding to factors are NOT centered
scale: If TRUE, y and columns in x are scaled to have variance=1. Dummy variables corresponding to factors are NOT scaled
enumerate: Default is TRUE if there's less than 15 variable groups. If TRUE all models with up to maxvars are enumerated, else Gibbs sampling is used to explore the model space
includevars: Logical vector of length ncol(x) indicating variables that should always be included in the model, i.e. variable selection is not performed for these variables
models: Optional logical matrix indicating the models to be enumerated with rows equal to the number of desired models and columns to the number of variables in x.
maxvars: When enumerate==TRUE only models with up to maxvars variables enumerated (defaults to all variables). In modelsearchBlockDiag a sequence of models is defined from 1 up to maxvars
niter: Number of Gibbs sampling iterations
thinning: MCMC thinning factor, i.e. only one out of each thinning iterations are reported. Defaults to thinning=1, i.e. no thinning
burnin: Number of burn-in MCMC iterations. Defaults to .1*niter. Set to 0 for no burn-in
family: Family of parametric distribution. Use 'normal' for Normal errors, 'binomial' for logistic regression, 'poisson' for Poisson regression. 'twopiecenormal' for two-piece Normal, 'laplace' for Laplace errors and 'twopiecelaplace' for double exponential. For 'auto' the errors are assumed continuous and their distribution is inferred from the data among 'normal', 'laplace', 'twopiecenormal' and 'twopiecelaplace'. 'laplace' corresponds to median regression and 'twopiecelaplace' to quantile regression. See argument priorSkew
priorCoef: Prior on coefficients, created by momprior, imomprior, emomprior or zellnerprior. Prior dispersion is on coefficients/sqrt(scale) for Normal and two-piece Normal, and on coefficients/sqrt(2*scale) for Laplace and two-piece Laplace.
priorGroup: Prior on grouped coefficients (e.g. categorical predictors with >2 categories, splines). Created by groupmomprior, groupemomprior, groupimomprior or groupzellnerprior
priorDelta: Prior on model space. Use modelbbprior()
for Beta-Binomial prior, modelbinomprior(p) for Binomial prior with prior inclusion probability p, modelcomplexprior for Complexity prior, or modelunifprior() for Uniform prior
priorConstraints: Prior distribution on the number of terms subject to hierarchical constrains that are included in the model
priorVar: Inverse gamma prior on scale parameter. For Normal outcomes variance=scale, for Laplace outcomes variance=2*scale
priorSkew: Either a fixed value for tanh(alpha) where alpha is the asymmetry parameter or a prior on tanh(alpha). For family=='twopiecelaplace' setting alpha=a is equivalent to performing quantile regression for the quantile (1+a)/2. Ignored if family is 'normal' or 'laplace'.
phi: The error variance in Gaussian models, typically this is unknown and is left missing
deltaini: Logical vector of length ncol(x) indicating which coefficients should be initialized to be non-zero. Defaults to all variables being excluded from the model
initSearch: Algorithm to refine deltaini. initSearch=='greedy' uses a greedy Gibbs sampling search. initSearch=='SCAD' sets deltaini to the non-zero elements in a SCAD fit with cross-validated regularization parameter. initSearch=='none' leaves deltaini unmodified
method: Method to compute marginal likelihood. method=='Laplace' for Laplace approx, method=='ALA'
for approximate Laplace approximation. method=='MC' for Importance Sampling, method=='Hybrid'
for Hybrid Laplace-IS (only available for piMOM prior). See Details.
method=='auto'
attempts to use exact calculations when
possible, otherwise ALA if available, otherwise Laplace approx.
adj.overdisp: Only used when method=='ALA'. Over-dispersion adjustment in models with fixed dispersion parameter, as in logistic and Poisson regression. adj.overdisp='none' for no adjustment (not recommended, particularly for Poisson models). adj.overdisp='intercept' to estimate over-dispersion from the intercept-only model, and adj.overdisp='residuals' from the Pearson residuals of each model
hess: Method to estimat the hessian in the Laplace approximation to the integrated likelihood under Laplace or asymmetric Laplace errors. When hess=='asymp' the asymptotic hessian is used, hess=='asympDiagAdj' a diagonal adjustment is applied (see Rossell and Rubio for details).
optimMethod: Algorithm to maximize objective function when method=='Laplace'. Leave unspecified or set optimMethod=='auto' for an automatic choice. optimMethod=='LMA' uses modified Newton-Raphson algorithm, 'CDA' coordinate descent algorithm
optim_maxit: Maximum number of iterations when method=='Laplace'
initpar: Initial regression parameter values when finding the posterior mode to approximate the integrated likelihood. 'none', 'MLE', 'L1', or a numeric vector with initial values. 'auto': if p<n/2 MLE is used, else L1 (regularization parameter set via BIC)
B: Number of samples to use in Importance Sampling scheme. Ignored if method=='Laplace'
XtXprecomp: Set to TRUE to pre-compute the Gram matrix x'x upfront (saves time), to FALSE to compute and store elements only as needed (saves memory)
verbose: Set verbose==TRUE to print iteration progress
blocksize: Maximum number of variables in a block. Careful, the cost of the algorithm is of order 2^blocksize
maxiter: Maximum number of iterations, each iteration includes a screening pass to add and subtract variables
maxlogmargdrop: Stop the sequence of models when the drop in log p(y|model) is greater than maxlogmargdrop. This option avoids spending unnecessary time exploring overly large models
maxenum: If the posterior mode found has less than maxenum
variables then do a full enumeration of all its submodels
Let delta be the vector indicating inclusion/exclusion of each column of x in the model. The Gibbs algorithm sequentially samples from the posterior of each element in delta conditional on all the remaining elements in delta and the data. To do this it is necessary to evaluate the marginal likelihood for any given model. These have closed-form expression for the MOM prior, but for models with >15 variables these are expensive to compute and Laplace approximations are used instead (for the residual variance a log change of variables is used, which improves the approximation). For other priors closed forms are not available, so by default Laplace approximations are used. For the iMOM prior we also implement a Hybrid Laplace-IS which uses a Laplace approximation to evaluate the integral wrt beta and integrates wrt phi (residual variance) numerically.
It should be noted that Laplace approximations tend to under-estimate the marginal densities when the MLE for some parameter is very close to 0. That is, it tends to be conservative in the sense of excluding more variables from the model than an exact calculation would.
Finally, method=='plugin' provides a BIC-type approximation that is faster than exact or Laplace methods, at the expense of some accuracy. In non-sparse situations where models with many variables have large posterior probability method=='plugin' can be substantially faster.
For more details on the methods used to compute marginal densities see Johnson & Rossell (2012).
modelsearchBlockDiag uses the block search method described in Papaspiliopoulos & Rossell. Briefly, spectral clustering is run on X'X to cluster variables into blocks of blocksize and subsequently the Coolblock algorithm is used to define a sequence of models of increasing size. The exact integrated likelihood is evaluated for all models in this path, the best model chosen, and the scheme iteratively repeated to add and drop variables until convergence.
Object of class msfit, which extends a list with elements - postSample: matrix with posterior samples for the model indicator. postSample[i,j]==1
indicates that variable j was included in the model in the MCMC iteration i
postOther: postOther
returns posterior samples for parameters other than the model indicator, i.e. basically hyper-parameters. If hyper-parameters were fixed in the model specification, postOther will be empty.
margpp: Marginal posterior probability for inclusion of each covariate. This is computed by averaging marginal post prob for inclusion in each Gibbs iteration, which is much more accurate than simply taking colMeans(postSample). - postMode: Model with highest posterior probability amongst all those visited
postModeProb: Unnormalized posterior prob of posterior mode (log scale)
postProb: Unnormalized posterior prob of each visited model (log scale)
priors: List with priors specified when calling modelSelection
Johnson V.E., Rossell D. Non-Local Prior Densities for Default Bayesian Hypothesis Tests. Journal of the Royal Statistical Society B, 2010, 72, 143-170.
Johnson V.E., Rossell D. Bayesian model selection in high-dimensional settings. Journal of the American Statistical Association, 2012, 107, 649-660.
Papaspiliopoulos O., Rossell, D. Scalable Bayesian variable selection and model averaging under block orthogonal design. 2016
Rossell D., Rubio F.J. Tractable Bayesian variable selection: beyond normality. 2016
David Rossell
msfit-class for details on the output. postProb to obtain posterior model probabilities. coef.msfit for Bayesian model averaging estimates and intervals. predict.msfit for BMA estimates and intervals for user-supplied covariate values. plot.msfit for an MCMC diagnostic plot showing estimated marginal posterior inclusion probabilities vs. iteration number. rnlp to obtain posterior samples for the coefficients. nlpMarginal to compute marginal densities for a given model.
#Simulate data x <- matrix(rnorm(100*3),nrow=100,ncol=3) theta <- matrix(c(1,1,0),ncol=1) y <- x %*% theta + rnorm(100) #Specify prior parameters priorCoef <- momprior(tau=0.348) priorDelta <- modelunifprior() #Alternative model space prior: 0.5 prior prob for including any covariate priorDelta <- modelbinomprior(p=0.5) #Alternative: Beta-Binomial prior for model space priorDelta <- modelbbprior(alpha.p=1,beta.p=1) #Model selection fit1 <- modelSelection(y=y, x=x, center=FALSE, scale=FALSE, priorCoef=priorCoef, priorDelta=priorDelta) postProb(fit1) #posterior model probabilities fit1$margpp #posterior marginal inclusion prob coef(fit1) #BMA estimates, 95% intervals, marginal post prob