bmonomvn {monomvn}

Bayesian Estimation for Multivariate Normal Data with Monotone Missingness

Description

Bayesian estimation via sampling from the posterior distribution of the of the mean and covariance matrix of multivariate normal (MVN) distributed data with a monotone missingness pattern, via Gibbs Sampling. Through the use of parsimonious/shrinkage regressions (lasso/NG & ridge), where standard regressions fail, this function can handle an (almost) arbitrary amount of missing data

Usage

bmonomvn(y, pre = TRUE, p = 0.9, B = 100, T = 200, thin = 1,
         economy = FALSE, method = c("lasso", "ridge", "lsr", "factor",
         "hs", "ng"), RJ = c("p", "bpsn", "none"), capm = TRUE,
         start = NULL, mprior = 0, rd = NULL, theta = 0, rao.s2 = TRUE,
         QP = NULL, verb = 1, trace = FALSE)

Arguments

y

data matrix were each row is interpreted as a random sample from a MVN distribution with missing values indicated by NA

pre

logical indicating whether pre-processing of the y is to be performed. This sorts the columns so that the number of NAs is non-decreasing with the column index

p

when performing regressions, p is the proportion of the number of columns to rows in the design matrix before an alternative regression (lasso, ridge, or RJ) is performed as if least-squares regression has “failed”. Least-squares regression is known to fail when the number of columns equals the number of rows, hence a default of p = 0.9 <= 1. Alternatively, setting p = 0 forces a parsimonious method to be used for every regression. Intermediate settings of p allow the user to control when least-squares regressions stop and the parsimonious ones start; When method = "factor" the p argument represents an integer (positive) number of initial columns of y to treat as known factors

B

number of Burn-In MCMC sampling rounds, during which samples are discarded

T

total number of MCMC sampling rounds to take place after burn-in, during which samples are saved

thin

multiplicative thinning in the MCMC. Each Bayesian (lasso) regression will discard thin*M MCMC rounds, where M is the number of columns in its design matrix, before a sample is saved as a draw from the posterior distribution; Likewise if theta != 0 a further thin*N, for N responses will be discarded

economy

indicates whether memory should be economized at the expense of speed. When TRUE the individual Bayesian (lasso) regressions are cleaned between uses so that only one of them has a large footprint at any time during sampling from the Markov chain. When FALSE (default) all regressions are pre-allocated and the full memory footprint is realized at the outset, saving dynamic allocations

method

indicates the Bayesian parsimonious regression specification to be used, choosing between the lasso (default) of Park & Casella, the NG extension, the horseshoe, a ridge regression special case, and least-squares. The "factor" method treats the first p columns of y as known factors

RJ

indicates the Reversible Jump strategy to be employed. The default argument of "p" method uses RJ whenever a parsimonious regression is used; "bpsn" only uses RJ for regressions with p >= n, and "none" never uses RJ

capm

when TRUE this argument indicates that the number of components of \beta should not exceed n, the number of response variables in a particular regression

start

a list depicting starting values for the parameters that are use to initialize the Markov chain. Usually this will be a "monomvn"-class object depicting maximum likelihood estimates output from the monomvn function. The relevant fields are the mean vector $mu, covariance matrix $S, monotone ordering $o (for sanity checking with input y), component vector $ncomp and penalty parameter vector $lambda; see note below

mprior

prior on the number of non-zero regression coefficients (and therefore covariates) m in the model. The default (mprior = 0) encodes the uniform prior on 0 < m < M. A scalar value 0 <= mprior <= 1 implies a Binomial prior Bin(m|n=M,p=mprior). A 2-vector mprior=c(g,h) of positive values g and h represents gives Bin(m|n=M,p) prior where p~Beta(g,h)

rd

=c(r,delta); a 2-vector of prior parameters for \lambda^2 which depends on the regression method. When method = "lasso" then the components are the \alpha (shape) and \beta (rate) parameters to the a gamma distribution G(r,delta); when method = "ridge" the components are the \alpha (shape) and \beta (scale) parameters to an inverse-gamma distribution IG(r/2,delta/2)

theta

the rate parameter (> 0) to the exponential prior on the degrees of freedom paramter nu for each regression model implementing Student-t errors (for each column of Y marginally) by a scale-mixture prior. See blasso for more details. The default setting of theta = 0 turns off this prior, defaulting to a normal errors prior. A negative setting triggers a pooling of the degrees of freedom parameter across all columns of Y. I.e., Y is modeled as multivariate-t. In this case abs{theta} is used as the prior parameterization

rao.s2

indicates whether to Rao-Blackwellized samples for \sigma^2 should be used (default TRUE); see the details section of blasso for more information

QP

if non-NULL this argument should either be TRUE, a positive integer, or contain a list specifying a Quadratic Program to solve as a function of the samples of mu = dvec and Sigma = Dmat in the notation of solve.QP; see default.QP for a default specification that is used when QP = TRUE or a positive integer is is given; more details are below

verb

verbosity level; currently only verb = 0 and verb = 1 are supported

trace

if TRUE then samples from all parameters are saved to files in the CWD, and then read back into the "monomvn"-class object upon return

Details

If pre = TRUE then bmonomvn first re-arranges the columns of y into nondecreasing order with respect to the number of missing (NA) entries. Then (at least) the first column should be completely observed.

Samples from the posterior distribution of the MVN mean vector and covariance matrix are obtained sampling from the posterior distribution of Bayesian regression models. The methodology for converting these to samples from the mean vector and covariance matrix is outlined in the monomvn documentation, detailing a similarly structured maximum likelihood approach. Also see the references below.

Whenever the regression model is ill–posed (i.e., when there are more covariates than responses, or a “big p small n” problem) then Bayesian lasso or ridge regressions – possibly augmented with Reversible Jump (RJ) for model selection – are used instead. See the Park & Casella reference below, and the blasso documentation. To guarantee each regression is well posed the combination setting of method="lsr" and RJ="none" is not allowed. As in monomvn the p argument can be used to turn on lasso or ridge regressions (possibly with RJ) at other times. The exception is the "factor" method which always involves an OLS regression on (a subset of) the first p columns of y.

Samples from a function of samples of mu and Sigma can be obtained by specifying a Quadratic program via the argument QP. The idea is to allow for the calculation of the distribution of minimum variance and mean–variance portfolios, although the interface is quite general. See default.QP for more details, as default.QP(ncol(y)) is used when the argument QP = TRUE is given. When a positive integer is given, then the first QP columns of y are treated as factors by using

default.QP(ncol(y) - QP)

instead. The result is that the corresponding components of (samples of) mu and rows/cols of S are not factored into the specification of the resulting Quadratic Program

Value

bmonomvn returns an object of class "monomvn", which is a list containing the inputs above and a subset of the components below.

call

a copy of the function call as used

mu

estimated mean vector with columns corresponding to the columns of y

S

estimated covariance matrix with rows and columns corresponding to the columns of y

mu.var

estimated variance of the mean vector with columns corresponding to the columns of y

mu.cov

estimated covariance matrix of the mean vector with columns corresponding to the columns of y

S.var

estimated variance of the individual components of the covariance matrix with columns and rows corresponding to the columns of y

mu.map

estimated maximum a' posteriori (MAP) of the mean vector with columns corresponding to the columns of y

S.map

estimated MAP of the individual components of the covariance matrix with columns and rows corresponding to the columns of y

S.nz

posterior probability that the individual entries of the covariance matrix are non–zero

Si.nz

posterior probability that the individual entries of the inverse of the covariance matrix are non–zero

nu

when theta < 0 this field provides a trace of the pooled nu parameter to the multivariate-t distribution

lpost.map

log posterior probability of the MAP estimate

which.map

gives the time index of the sample corresponding to the MAP estimate

llik

a trace of the log likelihood of the data

llik.norm

a trace of the log likelihood under the Normal errors model when sampling under the Student-t model; i.e., it is not present unless theta > 0. Used for calculating Bayes Factors

na

when pre = TRUE this is a vector containing number of NA entries in each column of y

o

when pre = TRUE this is a vector containing the index of each column in the sorting of the columns of y obtained by o <- order(na)

method

method of regression used on each column, or "bcomplete" indicating that no regression was used

thin

the (actual) number of thinning rounds used for the regression (method) in each column

lambda2

records the mean \lambda^2 value found in the trace of the Bayesian Lasso regressions. Zero-values result when the column corresponds to a complete case or an ordinary least squares regression (these would be NA entries from monomvn)

ncomp

records the mean number of components (columns of the design matrix) used in the regression model for each column of y. If input RJ = FALSE then this simply corresponds to the monotone ordering (these would correspond to the NA entries from monomvn). When RJ = TRUE the monotone ordering is an upper bound (on each entry)

trace

if input trace = TRUE then this field contains traces of the samples of \mu in the field $mu and of \Sigma in the field $S, and of all regression parameters for each of the m = length(mu) columns in the field $reg. This $reg field is a stripped-down "blasso"-class object so that the methods of that object may be used for analysis. If data augmentation is required to complete the monotone missingness pattern, then samples from these entries of Y are contained in $DA where the column names indicate the i-j entry of Y sampled; see the R output below

R

gives a matrix version of the missingness pattern used: 0-entries mean observed; 1-entries indicate missing values conforming to a monotone pattern; 2-entries indicate missing values that require data augmentation to complete a monotone missingness pattern

B

from inputs: number of Burn-In MCMC sampling rounds, during which samples are discarded

T

from inputs: total number of MCMC sampling rounds to take place after burn-in, during which samples are saved

r

from inputs: alpha (shape) parameter to the gamma distribution prior for the lasso parameter lambda

delta

from inputs: beta (rate) parameter to the gamma distribution prior for the lasso parameter lambda

QP

if a valid (non–FALSE or NULL) QP argument is given, then this field contains the specification of a Quadratic Program in the form of a list with entries including $dvec, $Amat, $b0, and $meq, similar to the usage in solve.QP, and some others; see default.QP for more details

W

when input QP = TRUE is given, then this field contains a T*ncol(y) matrix of samples from the posterior distribution of the solution to the Quadratic Program, which can be visualized via plot.monomvn using the argument which = "QP"

Note

Whenever the bmonomvn algorithm requires a regression where p >= n, i.e., if any of the columns in the y matrix have fewer non–NA elements than the number of columns with more non–NA elements, then it is helpful to employ both lasso/ridge and the RJ method.

It is important that any starting values provided in the start be compatible with the regression model specified by inputs RJ and method. Any incompatibilities will result with a warning that (alternative) default action was taken and may result in an undesired (possibly inferior) model being fit

Author(s)

Robert B. Gramacy [e-mail]

References

R.B. Gramacy and E. Pantaleo (2010). Shrinkage regression for multivariate inference with missing data, and an application to portfolio balancing. Bayesian Analysis. 5(1), 237-262. doi:10.1214/10-BA602 Preprint available on arXiv:0710.5837 https://arxiv.org/abs/0907.2135

Roderick J.A. Little and Donald B. Rubin (2002). Statistical Analysis with Missing Data, Second Edition. Wilely.

https://bobby.gramacy.com/r_packages/monomvn/

See Also

blasso, monomvn, default.QP, em.norm in the now defunct norm and mvnmle packages, and returns

Examples

## standard usage, duplicating the results in
## Little and Rubin, section 7.4.3
data(cement.miss)
out <- bmonomvn(cement.miss)
out
out$mu
out$S

##
## A bigger example, comparing the various
## parsimonious methods
##

## generate N=100 samples from a 10-d random MVN
xmuS <- randmvn(100, 20)

## randomly impose monotone missingness
xmiss <- rmono(xmuS$x)

## using least squares only when necessary,
obl <- bmonomvn(xmiss)
obl

## look at the posterior variability
par(mfrow=c(1,2))
plot(obl)
plot(obl, "S")

## compare to maximum likelihood
Ellik.norm(obl$mu, obl$S, xmuS$mu, xmuS$S)
oml <- monomvn(xmiss, method="lasso")
Ellik.norm(oml$mu, oml$S, xmuS$mu, xmuS$S)


##
## a min-variance portfolio allocation example
##

## get the returns data, and use 20 random cols
data(returns)
train <- returns[,sample(1:ncol(returns), 20)]

## missingness pattern requires DA; also gather
## samples from the solution to a QP
obl.da <- bmonomvn(train, p=0, QP=TRUE)

## plot the QP weights distribution
plot(obl.da, "QP", xaxis="index")

## get ML solution: will warn about monotone violations
suppressWarnings(oml.da <- monomvn(train, method="lasso"))

## add mean and MLE comparison, requires the
## quadprog library for the solve.QP function
add.pe.QP(obl.da, oml.da)

## now consider adding in the market as a factor
data(market)
mtrain <- cbind(market, train)

## fit the model using only factor regressions
obl.daf <- bmonomvn(mtrain, method="factor", p=1, QP=1)
plot(obl.daf, "QP", xaxis="index", main="using only factors")
suppressWarnings(oml.daf <- monomvn(mtrain, method="factor"))
add.pe.QP(obl.daf, oml.daf)


##
## a Bayes/MLE comparison using least squares sparingly
##

## fit Bayesian and classical lasso
p <- 0.25
obls <- bmonomvn(xmiss, p=p)
Ellik.norm(obls$mu, obls$S, xmuS$mu, xmuS$S)
omls <- monomvn(xmiss, p=p, method="lasso")
Ellik.norm(omls$mu, omls$S, xmuS$mu, xmuS$S)

## compare to ridge regression
obrs <- bmonomvn(xmiss, p=p, method="ridge")
Ellik.norm(obrs$mu, obrs$S, xmuS$mu, xmuS$S)
omrs <- monomvn(xmiss, p=p, method="ridge")
Ellik.norm(omrs$mu, omrs$S, xmuS$mu, xmuS$S)