| Title: | Large-Scale Bayesian Variable Selection Using Variational Methods |
|---|---|
| Description: | Fast algorithms for fitting Bayesian variable selection models and computing Bayes factors, in which the outcome (or response variable) is modeled using a linear regression or a logistic regression. The algorithms are based on the variational approximations described in "Scalable variational inference for Bayesian variable selection in regression, and its accuracy in genetic association studies" (P. Carbonetto & M. Stephens, 2012, <DOI:10.1214/12-BA703>). This software has been applied to large data sets with over a million variables and thousands of samples. |
| Authors: | Peter Carbonetto [aut, cre], Matthew Stephens [aut], David Gerard [ctb] |
| Maintainer: | Peter Carbonetto <[email protected]> |
| License: | GPL (>= 3) |
| Version: | 2.6-10 |
| Built: | 2024-11-13 04:44:16 UTC |
| Source: | https://github.com/cran/varbvs |
Fast algorithms for fitting Bayesian variable selection models and computing Bayes factors, in which the outcome (or response variable) is modeled using a linear regression or a logistic regression. The algorithms are based on the variational approximations described in "Scalable variational inference for Bayesian variable selection in regression, and its accuracy in genetic association studies" (P. Carbonetto and M. Stephens, Bayesian Analysis 7, 2012, pages 73-108). This software has been applied to large data sets with over a million variables and thousands of samples.
The main functionality of this package is implemented in function
varbvs. This function selects the most appropriate
algorithm for the data set and selected model (linear or logistic
regression). See help(varbvs) for details. The varbvs interface
is intended to resemble interface for glmnet, the popular package
for fitting genealized linear models.
For more details about the this package, including the license and a
list of available functions, see help(package=varbvs).
Peter Carbonetto [email protected]
P. Carbonetto and M. Stephens (2012). Scalable variational inference for Bayesian variable selection in regression, and its accuracy in genetic association studies. Bayesian Analysis 7, 73–108.
Estimate credible interval from weighted samples.
cred(x, x0, w = NULL, cred.int = 0.95)cred(x, x0, w = NULL, cred.int = 0.95)
x |
Vector of random samples of variable. |
x0 |
Mean of median of variable. |
w |
Weight > 0 assigned to each sample. If w = NULL, all weights are the same. |
cred.int |
Credible interval must contain probability mass of at least this amount. A number between 0 and 1. |
Credible interval [a,b] is defined as smallest interval containing x0
that contains cred.int of the probability mass. Note that the credible
interval is not necessarily symmetric about x0. (Other definitions of the
credible interval are possible.)
The algorithm is quadratic in the length of x (and w), so should not be used for large numbers of samples.
list(a = a,b = b).
Peter Carbonetto [email protected]
P. Carbonetto and M. Stephens (2012). Scalable variational inference for Bayesian variable selection in regression, and its accuracy in genetic association studies. Bayesian Analysis 7, 73–108.
x <- rnorm(100) out <- cred(x,mean(x),cred.int = 0.68)x <- rnorm(100) out <- cred(x,mean(x),cred.int = 0.68)
This gene set was selected in Carbonetto and Stephens (2013) from an interrogation of 3,158 derived from 8 publicly available pathway databases.
data(cytokine)data(cytokine)
cytokine[i] = 1 if SNP i lies within 100 kb of a gene in the "Cytokine
signaling in immune system" gene set, and cytokine[i] = 0 otherwise.
Pathway id 75790 from the Reactome database, or pathway id 366171 from the BioSystems database.
P. Carbonetto and M. Stephens (2013). Integrated enrichment analysis of variants and pathways in genome-wide association studies indicates central role for IL-2 signaling genes in type 1 diabetes, and cytokine signaling genes in Crohn's disease. PLoS Genetics 9, e1003770.
# See demo.cytokine.R vignette.# See demo.cytokine.R vignette.
Expression levels recorded for 3,571 genes in 72 patients with leukemia (Golub et al, 1999). The binary outcome encodes the disease subtype: acute lymphobastic leukemia (ALL) or acute myeloid leukemia (AML).
data(leukemia)data(leukemia)
Data are represented as a 72 x 3,571 matrix x of gene expression
values, and a vector y of 72 binary disease outcomes.
These are the preprocessed data of Dettling (2004) retrieved from the supplementary materials accompanying Friedman et al (2010).
M. Dettling (2004). BagBoosting for tumor classification with gene expression data. Bioinformatics 20, 3583–3593.
J. Friedman, T. Hastie and R. Tibshirani (2010). Regularization paths for generalized linear models via coordinate descent. Journal of Statistical Software 33, 1–22.
T. R. Golub, et al. (1999). Molecular classification of cancer: class discovery and class prediction by gene expression monitoring. Science 286, 531–537.
# See demo.leukemia.R vignette.# See demo.leukemia.R vignette.
Compute normalized probabilities from unnormalized log-probabilities.
normalizelogweights(logw)normalizelogweights(logw)
logw |
Vector of unnormalized log-probabilities. |
Guards against underflow or overflow by adjusting the log-probabilities so that the largest probability is 1.
Normalized probabilities such that the sum is equal to 1.
Peter Carbonetto [email protected]
P. Carbonetto and M. Stephens (2012). Scalable variational inference for Bayesian variable selection in regression, and its accuracy in genetic association studies. Bayesian Analysis 7, 73–108.
logw <- rnorm(6) w <- normalizelogweights(logw)logw <- rnorm(6) w <- normalizelogweights(logw)
Generate a single plot that summarizes the results of fitting the Bayesian variable selection model to the data. When the variables are genetic markers, the groups are chromosomes, and the posterior probabilities are plotted on the vertical axis (typically on the logarithmic scale), the figure resembles a "Manhattan plot" typically used to summarize the results of a genome-wide association study or quantitative trait locus (QTL) mapping study.
## S3 method for class 'varbvs' plot(x, score, groups, vars = NULL, var.labels, draw.threshold = NA, gap = 0,col = "midnightblue", pch = 20, scales = NULL, xlab = "variable", ylab = "posterior probability", main = "fitted varbvs model: variable selection results", abline.args = list(lty = "dotted",col = "orangered"), vars.xyplot.args = list(pch = 20,col = "magenta"), vars.ltext.args = list(col = "black",pos = 4,cex = 0.5), par.settings = list(par.main.text = list(font = 1,cex = 0.8), layout.heights = list(axis.top = 0, axis.bottom = 0)), ...)## S3 method for class 'varbvs' plot(x, score, groups, vars = NULL, var.labels, draw.threshold = NA, gap = 0,col = "midnightblue", pch = 20, scales = NULL, xlab = "variable", ylab = "posterior probability", main = "fitted varbvs model: variable selection results", abline.args = list(lty = "dotted",col = "orangered"), vars.xyplot.args = list(pch = 20,col = "magenta"), vars.ltext.args = list(col = "black",pos = 4,cex = 0.5), par.settings = list(par.main.text = list(font = 1,cex = 0.8), layout.heights = list(axis.top = 0, axis.bottom = 0)), ...)
x |
Output of function |
score |
Value to plot on vertical axis. Must be a numeric vector
with one entry for each variable. If missing, the posterior inclusion
probability for each variable is plotted in the vertical axis, in
which this probability is averaged over hyperparameter settings,
treating |
groups |
Group the variables in the plot according to this argument. This must be a vector with one entry for each variable. If missing, all variables are treated as a single group. This is useful for grouping the genetic markers by chromosome in a genome-wide association study. |
vars |
Indices (type |
var.labels |
Labels to accompany the highlighted variables
only. If missing, labels are retrieved from |
draw.threshold |
Plot a horizontal line at this location on the vertical axis. |
gap |
Amount of space to leave between each group of variables in the plot. |
col |
Argument passed to |
pch |
Argument passed to |
scales |
Argument passed to |
xlab |
Argument passed to |
ylab |
Argument passed to |
main |
Argument passed to |
abline.args |
Additional arguments passed to |
vars.xyplot.args |
Additional arguments passed to |
vars.ltext.args |
Additional arguments passed to |
par.settings |
Argument passed to |
... |
Additional arguments passed to |
Note that plot.varbvs uses function xyplot from the
lattice package, and as.layer from the
latticeExtra package.
An object of class "trellis" generated by functions
xyplot and as.layer.
Peter Carbonetto [email protected]
P. Carbonetto and M. Stephens (2012). Scalable variational inference for Bayesian variable selection in regression, and its accuracy in genetic association studies. Bayesian Analysis 7, 73–108.
xyplot, ltext,
panel.abline, varbvs,
summary.varbvs, varbvsindep
This function predicts outcomes (Y) given the observed
variables (X) and observed covariates (Z), and a model fitted using
varbvs.
## S3 method for class 'varbvs' predict(object, X, Z = NULL, type = c("link","response","class"), averaged = TRUE, ...)## S3 method for class 'varbvs' predict(object, X, Z = NULL, type = c("link","response","class"), averaged = TRUE, ...)
object |
Output of function |
X |
n x p input matrix, in which p is the number of variables, and n is the number of samples for which predictions will be made using the fitted model. X cannot be sparse, and cannot have any missing values (NA). |
Z |
n x m covariate data matrix, where m is the number of
covariates. Do not supply an intercept as a covariate (i.e., a
column of ones), because an intercept is automatically included in
the regression model. For no covariates, set |
type |
Type of prediction to output. The default, "link", gives
the linear predictors for |
averaged |
When |
... |
Other arguments to generic predict function. These extra arguments are not used here. |
Note that the classification probabilities are not guaranteed to be calibrated under the variational
approximation.
When averaged = TRUE, the output is a vector containing the
predicted outcomes for all samples. For family = "binomial",
all vector entries are 0 or 1.
When averaged = FALSE, the return value is a matrix with one
row for each sample, and one column for each hyperparameter setting.
Peter Carbonetto [email protected]
P. Carbonetto and M. Stephens (2012). Scalable variational inference for Bayesian variable selection in regression, and its accuracy in genetic association studies. Bayesian Analysis 7, 73–108.
# See help(varbvs) for examples.# See help(varbvs) for examples.
This function predicts outcomes (Y) given the observed
variables (X) and observed covariates (Z), and a model fitted using
varbvsmix.
## S3 method for class 'varbvsmix' predict(object, X, Z = NULL, ...)## S3 method for class 'varbvsmix' predict(object, X, Z = NULL, ...)
object |
Output of function |
X |
n x p input matrix, in which p is the number of variables, and n is the number of samples for which predictions will be made using the fitted model. X cannot be sparse, and cannot have any missing values (NA). |
Z |
n x m covariate data matrix, where m is the number of
covariates. Do not supply an intercept as a covariate (i.e., a
column of ones), because an intercept is automatically included in
the regression model. For no covariates, set |
... |
Other arguments to generic predict function. These extra arguments are not used here. |
# See help(varbvsmix) for examples.# See help(varbvsmix) for examples.
Generate matrices of pseudorandom values.
rand(m,n) randn(m,n)rand(m,n) randn(m,n)
m |
Number of matrix rows. |
n |
Number of matrix columns. |
Function rand returns a matrix containing pseudorandom values
drawn from the standard uniform distribution (using runif).
Function randn returns a matrix containing pseudorandom values
drawn from the standard normal distribution (using rnorm).
An m x n numeric matrix.
Peter Carbonetto [email protected]
P. Carbonetto and M. Stephens (2012). Scalable variational inference for Bayesian variable selection in regression, and its accuracy in genetic association studies. Bayesian Analysis 7, 73–108.
x <- rand(10,5) y <- randn(10,5)x <- rand(10,5) y <- randn(10,5)
Select a subset of the candidate hyperparameter settings, and return a new varbvs analysis object with these hyperparameter settings only.
## S3 method for class 'varbvs' subset(x, subset, ...)## S3 method for class 'varbvs' subset(x, subset, ...)
x |
Output of function |
subset |
Expression indicating hyperparameter settings to select.
This expression should include one or more of |
... |
Other arguments to generic subset function. These extra arguments are not used here. |
An object with S3 class c("varbvs","list").
Peter Carbonetto [email protected]
P. Carbonetto and M. Stephens (2012). Scalable variational inference for Bayesian variable selection in regression, and its accuracy in genetic association studies. Bayesian Analysis 7, 73–108.
# First run one of the examples in help(varbvs), then try running # this. # # fit.new <- subset(fit,logodds < (-2)) ## First run one of the examples in help(varbvs), then try running # this. # # fit.new <- subset(fit,logodds < (-2)) #
Generate a summary of the Bayesian variable selection model fitted using variational approximation methods.
## S3 method for class 'varbvs' summary(object, cred.int = 0.95, nv, pip.cutoff, ...) ## S3 method for class 'summary.varbvs' print(x, digits = 3, ...) ## S3 method for class 'varbvs' print(x, digits = 3, ...)## S3 method for class 'varbvs' summary(object, cred.int = 0.95, nv, pip.cutoff, ...) ## S3 method for class 'summary.varbvs' print(x, digits = 3, ...) ## S3 method for class 'varbvs' print(x, digits = 3, ...)
object |
Output of function |
cred.int |
Size of credible interval, a number between 0 and 1. |
nv |
Show detailed statistics for top nv variables, ranked
according to their posterior inclusion probabilities. Only one of
|
pip.cutoff |
Show detailed statistics for all variables in which
the posterior inclusion probability (PIP) is at least
|
x |
Output of function |
digits |
Number of digits shown when printing posterior probabilities of top nv variables. |
... |
Additional summary or print arguments. |
The printed summary is divided into three parts. The first part
summarizes the data and optimization settings. It also reports the
hyperparameter setting that yields the largest marginal
likelihood—more precisely, the approximate marginal likelihood
computed using the variational method. For the linear regression only
(family = "gaussian") when no additional covariates (Z)
are included, it reports the estimated proportion of variance in the
outcome explained by the model (PVE), and the credible interval of the
PVE estimate brackets.
The second part summarizes the approximate posterior distribution of
the hyperparameters (sigma, sa, logodds). The "estimate" column is the
value averaged over hyperparameter settings, treating
objectlogw as (unnormalized) log-marginal probabilities. The
next column, labeled "Pr>x", where x = cred.int gives the
credible interval based on these weights (computed using function
cred).
The third part summarizes the variable selection results. This
includes the total number of variables included in the model at
different posterior probability thresholds, and a more detailed
summary of the variables included in the model with highest posterior
probability. For family = "gaussian", the "PVE" column gives
the estimated proportion of variance in the outcome explained by the
variable (conditioned on being included in the model).
An object of class summary.varbvs, to be printed by
print.summary.varbvs.
Peter Carbonetto [email protected]
P. Carbonetto and M. Stephens (2012). Scalable variational inference for Bayesian variable selection in regression, and its accuracy in genetic association studies. Bayesian Analysis 7, 73–108.
# See help(varbvs) for examples.# See help(varbvs) for examples.
Compute fully-factorized variational approximation for Bayesian variable selection in linear (family = gaussian) or logistic regression (family = binomial). More precisely, find the "best" fully-factorized approximation to the posterior distribution of the coefficients, with spike-and-slab priors on the coefficients. By "best", we mean the approximating distribution that locally minimizes the Kullback-Leibler divergence between the approximating distribution and the exact posterior.
varbvs(X, Z, y, family = c("gaussian","binomial"), sigma, sa, logodds, weights, resid.vcov, alpha, mu, eta, update.sigma, update.sa, optimize.eta, initialize.params, update.order, nr = 100, sa0 = 1, n0 = 10, tol = 1e-4, maxiter = 1e4, verbose = TRUE)varbvs(X, Z, y, family = c("gaussian","binomial"), sigma, sa, logodds, weights, resid.vcov, alpha, mu, eta, update.sigma, update.sa, optimize.eta, initialize.params, update.order, nr = 100, sa0 = 1, n0 = 10, tol = 1e-4, maxiter = 1e4, verbose = TRUE)
X |
n x p input matrix, where n is the number of samples, and p is the number of variables. X cannot be sparse, and cannot have any missing values (NA). |
Z |
n x m covariate data matrix, where m is the number of
covariates. Do not supply an intercept as a covariate (i.e.,
a column of ones), because an intercept is automatically
included in the regression model. For no covariates, set
|
y |
Vector of length n containing observations of binary
( |
family |
"gaussian" for linear regression model, or "binomial" for logistic regression model. |
sigma |
Candidate settings for the residual variance
parameter. Must be of the same length as inputs sa and logodds (or
have length equal to the number of columns of logodds). Only used for
linear regression, and will generate an error if |
sa |
Hyperparameter sa is the prior variance of regression
coefficients for variables that are included in the model. This
prior variance is always scaled by sigma (for logistic regression,
we take sigma = 1). Scaling the variance of the coefficients in this
way is necessary to ensure that this prior is invariant to
measurement scale (e.g., switching from grams to kilograms). This
input specifies the candidate settings for sa, of the same length as
inputs sigma and logodds (or have length equal to the number of
columns of logodds). If missing, prior variance is automatically
fitted to data by compute approximate maximum (ML) estimates, or
maximum a posteriori estimates when |
logodds |
Hyperparameter logodds is the prior log-odds that a
variable is included in the regression model; it is defined as
|
weights |
Optional vector of weights for weighted linear
regression; larger weights correspond to smaller variances. It
should be |
resid.vcov |
Optional matrix specifying the covariance of the
residual, scaled by sigma. This is useful to allow for observations
with different variances (similar to |
alpha |
Good initial estimate for the variational parameter alpha for each hyperparameter setting. Either missing, or a p x ns matrix, where p is the number of variables, and ns is the number of hyperparameter settings. |
mu |
Good initial estimate for the variational parameter mu for each hyperparameter setting. Either missing, or a p x ns matrix, where p is the number of variables, and ns is the number of hyperparameter settings. |
eta |
Good initial estimate of the additional free parameters specifying the variational approximation to the logistic regression factors. Either missing, or an n x ns matrix, where n is the number of samples, and ns is the number of hyperparameter settings. |
update.sigma |
Setting this to TRUE ensures that sigma is always fitted to data, in which case input vector sigma is used to provide initial estimates. |
update.sa |
Setting this to TRUE ensures that sa is always fitted to data, in which case input vector sa is used to provide initial estimates. |
optimize.eta |
When |
initialize.params |
If FALSE, the initialization stage of the variational inference algorithm (see below) will be skipped, which saves computation time for large data sets. |
update.order |
Order of the co-ordinate ascent updates for
fitting the variational approximation. The default is
|
nr |
Number of samples of "model PVE" to draw from posterior. |
sa0 |
Scale parameter for a scaled inverse chi-square prior on hyperparameter sa. Must be >= 0. |
n0 |
Number of degrees of freedom for a scaled inverse chi-square
prior on hyperparameter sa. Must be >= 0. Large settings of
|
tol |
Convergence tolerance for inner loop. |
maxiter |
Maximum number of inner loop iterations. |
verbose |
If |
An object with S3 class c("varbvs","list").
family |
Either "gaussian" or "binomial". |
sigma |
Settings for sigma (family = "gaussian" only). |
sa |
Settings for prior variance parameter. |
logodds |
Prior log-odds settings. |
prior.same |
TRUE if prior is identical for all variables. When
logodds is a p x ns matrix, |
sa0 |
Scale parameter for prior on hyperparameter sa. |
n0 |
Degrees of freedom for prior on hyperparameter sa. |
update.sigma |
If TRUE, sigma was fit to data for each setting of
prior logodds ( |
update.sa |
If TRUE, sa was fit to data for each setting of prior logodds. |
logw |
An array with ns elements, in which |
w |
Normalized weights (posterior probabilities) for each of the
hyperparameter settings computed from |
alpha |
Variational estimates of posterior inclusion probabilities for each hyperparameter setting. |
mu |
Variational estimates of posterior mean coefficients for each hyperparameter setting. |
s |
Variational estimates of posterior variances for each hyperparameter setting. |
pip |
"Averaged" posterior inclusion probabilities computed as a
weighted average of the individual PIPs ( |
beta |
"Averaged" posterior mean regression coefficients. |
mu.cov |
Posterior mean regression coefficients for covariates, including intercept, for each hyperparameter setting. |
beta.cov |
"Averaged" posterior mean regression coefficients for covariates, including intercept. |
eta |
Additional variational parameters for |
optimize.eta |
If TRUE, eta was fit to data ( |
fitted.values |
Matrix containing the predicted (or "fitted")
values of the outcome at each hyperparameter setting. For the
logistic regression model ( |
residuals |
For linear regression, this is a matrix containing
the model residuals at each hyperparameter setting. For |
pve |
For each hyperparameter setting, and for each variable,
mean estimate of the proportion of variance in outcome explained
conditioned on variable being included in the model ( |
model.pve |
Samples drawn from posterior distribution giving
estimates of proportion of variance in outcome (y) explained by fitted
variable selection model. This is for |
Two types of outcomes (y) are modeled: (1) a continuous outcome,
also a "quantitative trait" in the genetics literature; or (2) a
binary outcome with possible values 0 and 1. For the former, set
family = "gaussian", in which case, the outcome is
i.i.d. normal with mean and variance sigma, in
which u and b are vectors of regresion coefficients, and u0 is the
intercept. In the second case, we use logistic regression to model
the outcome, in which the probability that y = 1 is equal to
See help(sigmoid) for a
description of the sigmoid function. Note that the regression always
includes an intercept term (u0).
For both regression models, the fitting procedure consists of an inner
loop and an outer loop. The outer loop iterates over each of the
hyperparameter settings (sa, sigma and logodds). Given a setting of
the hyperparameters, the inner loop cycles through coordinate ascent
updates to tighten the lower bound on the marginal likelihood,
. The inner loop coordinate
ascent updates terminate when either (1) the maximum number of inner
loop iterations is reached, as specified by maxiter, or (2)
the maximum difference between the estimated posterior inclusion
probabilities is less than tol.
To provide a more accurate variational approximation of the posterior distribution, by default the fitting procedure has two stages. In the first stage, the entire fitting procedure is run to completion, and the variational parameters (alpha, mu, s, eta) corresponding to the maximum lower bound are then used to initialize the coordinate ascent updates in a second stage. Although this has the effect of doubling the computation time (in the worst case), the final posterior estimates tend to be more accurate with this two-stage fitting procedure.
Outputs alpha, mu and s specify the approximate posterior distribution
of the regression coefficients. Each of these outputs is a p x ns
matrix. For the ith hyperparameter setting, alpha[,i] is the
variational estimate of the posterior inclusion probability (PIP)
for each variable; mu[,i] is the variational estimate of the
posterior mean coefficient given that it is included in the model;
and s[,i] is the estimated posterior variance of the coefficient
given that it is included in the model. These are also the
quantities that are optimized as part of the inner loop coordinate
ascent updates. An additional free parameter, eta, is needed for
fast computation with the logistic regression model (family =
"binomial"). The fitted value of eta is returned as an n x ns
matrix.
The variational estimates should be interpreted carefully, especially when variables are strongly correlated. For example, consider the simple scenario in which 2 candidate variables are closely correlated, and at least one of them explains the outcome with probability close to 1. Under the correct posterior distribution, we would expect that each variable is included with probability ~0.5. However, the variational approximation, due to the conditional independence assumption, will typically get this wrong, and concentrate most of the posterior weight on one variable (the actual variable that is chosen will depend on the starting conditions of the optimization). Although the individual PIPs are incorrect, a statistic summarizing the variable selection for both correlated variables (e.g., the total number of variables included in the model) should be reasonably accurate.
More generally, if variables can be reasonably grouped together based on their correlations, we recommend interpreting the variable selection results at a group level. For example, in genome-wide association studies (see the vignettes) ,a SNP with a high PIP indicates that this SNP is probably associated with the trait, and one or more nearby SNPs within a chromosomal region, or “locus,” may be associated as well. Therefore, we interpreted the GWAS variable selection results at the level of loci, rather than at the level of individual SNPs.
Also note that special care is required for interpreting the results
of the variational approximation with the logistic regression
model. In particular, interpretation of the individual estimates of
the regression coefficients (e.g., the posterior mean estimates
fit$mu) is not straightforward due to the additional
approximation introduced on the individual nonlinear factors in the
likelihood. As a general guideline, only the relative magnitudes of
the coefficients are meaningful.
In many settings, it is good practice to account for uncertainty in the hyperparameters when reporting final posterior quantities. For example, hyperparameter sa is often estimated with a high degree of uncertainty when only a few variables are included in the model. Provided that (1) the hyperparameter settings sigma, sa and logodds adequately represent the space of possible hyperparameter settings with high posterior mass, (2) the hyperparameter settings are drawn from the same distribution as the prior, and (3) the fully-factorized variational approximation closely approximates the true posterior distribution, then final posterior quantities can be calculated by using logw as (unnormalized) log-marginal probabilities.
Even when conditions (1), (2) and/or (3) are not satisfied, this can
approach can still often yield reasonable estimates of averaged
posterior quantities. The examples below demonstrate how final
posterior quantities are reported by function
summary.varbvs (see help(summary.varbvs) for
more details). To account for discrepancies between the prior on
(sigma,sa,logodds) and the sampling density used to draw candidate
settings of the hyperparameters, adjust the log-probabilities by
setting fit$logw <- fit$logw + logp/logq, where logp is the
log-density of the prior distribution, and logq is the log-density
of the sampling distribution. (This is importance sampling; see, for
example, R. M. Neal, Annealed importance sampling, Statistics
and Computing, 2001.)
Specifying the prior variance of the regression coefficients (sa) can
be difficult, which is why we have included the option of fitting this
hyperparameter to the data (see input argument update.sa
above). However, in many settings, especially when a small number of
variables are included in the regression model, it is preferrable to
average over candidate settings of sa instead of fitting sa to the
data. To choose a set of candidate settings for sa, we have advocated
for setting sa indirectly through a prior estimate of the proportion
of variance in the outcome explained by the variables (abbreviated as
PVE), since it is often more natural to specify the PVE rather than
the prior variance (see references below). This is technically only
suitable or the linear regression model (family = "gaussian"),
but could potentially be used for the logistic regression model in an
approximate way.
For example, one could approximate a uniform prior on the PVE by drawing the PVE uniformly between 0 and 1, additionally specifying candidate settings for the prior log-odds, then computing the prior variance (sa) as follows:
sx <- sum(var1.cols(X)) sa <- PVE/(1-PVE)/(sigmoid(log(10)*logodds)*sx)
Note that this calculation will yield sa = 0 when PVE =
0, and sa = Inf when PVE = 1.
Also, bear in mind that if there are additional covariates (Z) included in the linear regression model that explain variance in Y, then it will usually make more sense to first remove the linear effects of these covariates before performing these calculations. The PVE would then represent the prior proportion of variance in the residuals of Y that are explained by the candidate variables. Alternatively, one could include the matrix Z in the calculations above, taking care to ensure that the covariates are included in the model with probability 1.
Finally, we point out that the optimization procedures were
carefully designed so that they can be applied to very large data
sets; to date, this code has been tested on data sets with >500,000
variables and >10,000 samples. An important limiting factor is the
ability to store the data matrix X in memory. To reduce memory
requirements, in the MATLAB interface we require that X be single
precision, but this option is not available in R. Additionally, we
mostly avoid generating intermediate products that are of the same
size as X. Only one such intermediate product is generated when
family = "gaussian", and none for family = "binomial".
Peter Carbonetto [email protected]
P. Carbonetto and M. Stephens (2012). Scalable variational inference for Bayesian variable selection in regression, and its accuracy in genetic association studies. Bayesian Analysis 7, 73–108.
Y. Guan and M. Stephens (2011). Bayesian variable selection regression for genome-wide association studies and other large-scale problems. Annals of Applied Statistics 5, 1780–1815.
X. Zhou, P. Carbonetto and M. Stephens (2013). Polygenic modeling with Bayesian sparse linear mixed models. PLoS Genetics 9, e1003264.
summary.varbvs, varbvs.properties,
varbvspve, varbvsnorm,
varbvsbin, varbvsbinz,
normalizelogweights, varbvs-internal
# LINEAR REGRESSION EXAMPLE # ------------------------- # Data are 200 uncorrelated ("unlinked") single nucleotide polymorphisms # (SNPs) with simulated genotypes, in which the first 20 of them have an # effect on the outcome. Also generate data for 3 covariates. maf <- 0.05 + 0.45*runif(200) X <- (runif(400*200) < maf) + (runif(400*200) < maf) X <- matrix(as.double(X),400,200,byrow = TRUE) Z <- randn(400,3) # Generate the ground-truth regression coefficients for the variables # (X) and additional 3 covariates (Z). Adjust the QTL effects so that # the variables (SNPs) explain 50 percent of the variance in the # outcome. u <- c(-1,2,1) beta <- c(rnorm(20),rep(0,180)) beta <- 1/sd(c(X %*% beta)) * beta # Generate the quantitative trait measurements. y <- c(-2 + Z %*% u + X %*% beta + rnorm(400)) # Fit the variable selection model. fit <- varbvs(X,Z,y,logodds = seq(-3,-1,0.1)) print(summary(fit)) # Compute the posterior mean estimate of hyperparameter sa. sa <- with(fit,sum(sa * w)) # Compare estimated outcomes against observed outcomes. y.fit <- predict(fit,X,Z) print(cor(y,y.fit)) # LOGISTIC REGRESSION EXAMPLE # --------------------------- # Data are 100 uncorrelated ("unlinked") single nucleotide polymorphisms # (SNPs) with simulated genotypes, in which the first 10 of them have an # effect on the outcome. Also generate data for 2 covariates. maf <- 0.05 + 0.45*runif(100) X <- (runif(750*100) < maf) + (runif(750*100) < maf) X <- matrix(as.double(X),750,100,byrow = TRUE) Z <- randn(750,2) # Generate the ground-truth regression coefficients for the variables # (X) and additional 2 covariates (Z). u <- c(-1,1) beta <- c(0.5*rnorm(10),rep(0,90)) # Simulate the binary trait (case-control status) as a coin toss with # success rates given by the logistic regression. sigmoid <- function (x) 1/(1 + exp(-x)) y <- as.double(runif(750) < sigmoid(-1 + Z %*% u + X %*% beta)) # Fit the variable selection model. fit <- varbvs(X,Z,y,"binomial",logodds = seq(-2,-0.5,0.5)) print(summary(fit))# LINEAR REGRESSION EXAMPLE # ------------------------- # Data are 200 uncorrelated ("unlinked") single nucleotide polymorphisms # (SNPs) with simulated genotypes, in which the first 20 of them have an # effect on the outcome. Also generate data for 3 covariates. maf <- 0.05 + 0.45*runif(200) X <- (runif(400*200) < maf) + (runif(400*200) < maf) X <- matrix(as.double(X),400,200,byrow = TRUE) Z <- randn(400,3) # Generate the ground-truth regression coefficients for the variables # (X) and additional 3 covariates (Z). Adjust the QTL effects so that # the variables (SNPs) explain 50 percent of the variance in the # outcome. u <- c(-1,2,1) beta <- c(rnorm(20),rep(0,180)) beta <- 1/sd(c(X %*% beta)) * beta # Generate the quantitative trait measurements. y <- c(-2 + Z %*% u + X %*% beta + rnorm(400)) # Fit the variable selection model. fit <- varbvs(X,Z,y,logodds = seq(-3,-1,0.1)) print(summary(fit)) # Compute the posterior mean estimate of hyperparameter sa. sa <- with(fit,sum(sa * w)) # Compare estimated outcomes against observed outcomes. y.fit <- predict(fit,X,Z) print(cor(y,y.fit)) # LOGISTIC REGRESSION EXAMPLE # --------------------------- # Data are 100 uncorrelated ("unlinked") single nucleotide polymorphisms # (SNPs) with simulated genotypes, in which the first 10 of them have an # effect on the outcome. Also generate data for 2 covariates. maf <- 0.05 + 0.45*runif(100) X <- (runif(750*100) < maf) + (runif(750*100) < maf) X <- matrix(as.double(X),750,100,byrow = TRUE) Z <- randn(750,2) # Generate the ground-truth regression coefficients for the variables # (X) and additional 2 covariates (Z). u <- c(-1,1) beta <- c(0.5*rnorm(10),rep(0,90)) # Simulate the binary trait (case-control status) as a coin toss with # success rates given by the logistic regression. sigmoid <- function (x) 1/(1 + exp(-x)) y <- as.double(runif(750) < sigmoid(-1 + Z %*% u + X %*% beta)) # Fit the variable selection model. fit <- varbvs(X,Z,y,"binomial",logodds = seq(-2,-0.5,0.5)) print(summary(fit))
All these functions are methods for class
"varbvs" objects.
## S3 method for class 'varbvs' nobs(object, ...) ## S3 method for class 'varbvs' case.names(object, ...) ## S3 method for class 'varbvs' variable.names(object, full = FALSE, include.threshold = 0.01, ...) ## S3 method for class 'varbvs' labels(object, ...) ## S3 method for class 'varbvs' coef(object, ...) ## S3 method for class 'varbvsmix' coef(object, ...) ## S3 method for class 'varbvs' confint(object, parm, level = 0.95, ...) ## S3 method for class 'varbvs' fitted(object, ...) ## S3 method for class 'varbvs' resid(object, type = c("deviance","response"), ...) ## S3 method for class 'varbvs' residuals(object, type = c("deviance","response"), ...) ## S3 method for class 'varbvs' deviance(object, ...)## S3 method for class 'varbvs' nobs(object, ...) ## S3 method for class 'varbvs' case.names(object, ...) ## S3 method for class 'varbvs' variable.names(object, full = FALSE, include.threshold = 0.01, ...) ## S3 method for class 'varbvs' labels(object, ...) ## S3 method for class 'varbvs' coef(object, ...) ## S3 method for class 'varbvsmix' coef(object, ...) ## S3 method for class 'varbvs' confint(object, parm, level = 0.95, ...) ## S3 method for class 'varbvs' fitted(object, ...) ## S3 method for class 'varbvs' resid(object, type = c("deviance","response"), ...) ## S3 method for class 'varbvs' residuals(object, type = c("deviance","response"), ...) ## S3 method for class 'varbvs' deviance(object, ...)
object |
An object inheriting from class |
full |
logical; if |
include.threshold |
When |
parm |
Confidence intervals are computed for these selected
variables. These may either be specified as numbers (column indices
of |
level |
Size of confidence level. |
type |
Type of residuals to be returned. This argument is only
relevant for logistic regression models ( |
... |
Further arguments passed to or from other methods. |
The generic accessor functions nobs, case.names,
variable.names and labels can be used to extract various
useful properties of the fitted varbvs model. Method
labels, in particular, returns the names of the candidate
variables (columns of X) which may be used, for example, to
plot posterior inclusion probabilities or effect estimates.
coef returns a matrix containing the posterior estimates of the
regression coefficients at each hyperparameter setting, as well as an
additional column containing "averaged" coefficient estimates.
confint returns confidence intervals (also, equivalently in
this case, "credible intervals") for all selected variables
parm. These are conditional confidence intervals; that
is, conditioned on each variable being included in the regression
model.
The confint return value is different from the usual confidence
interval (e.g., for an lm result) because a
confidence interval is provided for each hyperparameter setting, as
well as an additional "averaged" confidence interval. The confidence
intervals are returned a list, with one list element per selected
variable, and each list element is a matrix with columns giving lower
and upper confidence limits for each hyperparameter setting, as well
as the averaged limits.
Note that confidence intervals cannot currently be requested for
covariates (columns of Z).
fitted returns a matrix containing the predicted (or "fitted")
values of the outcome at each hyperparameter setting. For the logistic
regression model (family = "binomial"), each matrix entry gives
the probability that the binary outcome is equal to 1.
Likewise, resid and residuals each return a matrix
containing the model residuals at each hyperparameter setting.
deviance returns the deviance for the fitted model at each
hyperparameter setting.
varbvs, nobs,
case.names,
variable.names, labels,
coef, coef,
fitted, residuals,
deviance
The Bayes factor is the ratio of the marginal likelihoods
under two different models (see Kass & Raftery, 1995). Function
varbvsbf provides a convenient interface for computing the
Bayes factor comparing the fit of two different varbvs models.
varbvsbf (fit0, fit1) bayesfactor (logw0, logw1)varbvsbf (fit0, fit1) bayesfactor (logw0, logw1)
fit0 |
An output returned from |
fit1 |
Another output returned from |
logw0 |
log-probabilities or log-importance weights under H0. |
logw1 |
log-probabilities or log-importance weights under H1. |
Computes numerical estimate of
the probability of the data given the "alternative" hypothesis (H1) over the probability of the data given the "null" hypothesis (H0). This is also known as a Bayes factor (see Kass & Raftery, 1995). Here we assume that although these probabilities cannot be computed analytically because they involve intractable integrals, we can obtain reasonable estimates of these probabilities with a simple numerical approximation over some latent variable assuming the prior over this latent variable is uniform. The inputs are the log-probabilities
where Pr(Z0 | H0) and Pr(Z1 | H1) are uniform over all Z0 and Z1.
Alternatively, this function can be viewed as computing an importance sampling estimate of the Bayes factor; see, for example, R. M. Neal, "Annealed importance sampling", Statistics and Computing, 2001. This formulation described above is a special case of importance sampling when the settings of the latent variable Z0 and A1 are drawn from the same (uniform) distribution as the prior, Pr(Z0 | H0) and Pr(Z1 | H1), respectively.
The estimated Bayes factor.
Peter Carbonetto [email protected]
P. Carbonetto and M. Stephens (2012). Scalable variational inference for Bayesian variable selection in regression, and its accuracy in genetic association studies. Bayesian Analysis 7, 73–108.
R. E. Kass and A. E. Raftery (1995). Bayes Factors. Journal of the American Statistical Association 90, 773–795.
R. M. Neal (2001). Annealed importance sampling. Statistics and Computing 11, 125–139.
Compute the mean and variance of the coefficients, and the posterior inclusion probabilities (PIPs), ignoring correlations between variables. This is useful for inspecting or visualizing groups of correlated variables (e.g., genetic markers in linkage disequilibrium).
varbvsindep (fit, X, Z, y)varbvsindep (fit, X, Z, y)
fit |
Output of function |
X |
n x p input matrix, where n is the number of samples, and p is the number of variables. X cannot be sparse, and cannot have any missing values (NA). |
Z |
n x m covariate data matrix, where m is the number of
covariates. Do not supply an intercept as a covariate
(i.e., a column of ones), because an intercept is
automatically included in the regression model. For no
covariates, set |
y |
Vector of length n containing observations of binary
( |
For the ith hyperparameter setting, alpha[,i] is the
variational estimate of the posterior inclusion probability (PIP) for
each variable; mu[,i] is the variational estimate of the
posterior mean coefficient given that it is included in the model; and
s[,i] is the estimated posterior variance of the coefficient
given that it is included in the model.
alpha |
Variational estimates of posterior inclusion probabilities for each hyperparameter setting. |
mu |
Variational estimates of posterior mean coefficients for each hyperparameter setting. |
s |
Variational estimates of posterior variances for each hyperparameter setting. |
Peter Carbonetto [email protected]
P. Carbonetto and M. Stephens (2012). Scalable variational inference for Bayesian variable selection in regression, and its accuracy in genetic association studies. Bayesian Analysis 7, 73–108.
Find the "best" fully-factorized approximation to the
posterior distribution of the coefficients, with linear regression
likelihood and mixture-of-normals priors on the coefficients. By
"best", we mean the approximating distribution that locally minimizes
the Kullback-Leibler divergence between the approximating distribution
and the exact posterior. In the original formulation (see
varbvs), each regression coefficient was drawn
identically from a spike-and-slab prior. Here, we instead formulate
the “slab” as a mixture of normals.
varbvsmix(X, Z, y, sa, sigma, w, alpha, mu, update.sigma, update.sa, update.w, w.penalty, drop.threshold = 1e-8, tol = 1e-4, maxiter = 1e4, update.order = 1:ncol(X), verbose = TRUE)varbvsmix(X, Z, y, sa, sigma, w, alpha, mu, update.sigma, update.sa, update.w, w.penalty, drop.threshold = 1e-8, tol = 1e-4, maxiter = 1e4, update.order = 1:ncol(X), verbose = TRUE)
X |
n x p input matrix, where n is the number of samples, and p is the number of variables. X cannot be sparse, and cannot have any missing values (NA). |
Z |
n x m covariate data matrix, where m is the number of
covariates. Do not supply an intercept as a covariate (i.e., a
column of ones), because an intercept is automatically included in
the regression model. For no covariates, set |
y |
Vector of length n containing values of the continuous outcome. |
sa |
Vector specifying the prior variance of the regression
coefficients (scaled by |
sigma |
Residual variance parameter. If missing, it is automatically fitted to the data by computing an approximate maximum-likelihood estimate. |
w |
If missing, it is automatically fitted to the data by computing an approximate maximum-likelihood estimate. |
alpha |
Initial estimates of the approximate posterior mixture assignment probabilities. These should be specified as a p x K matrix, where K is the number of mixture components. Each row must add up to 1. |
mu |
Initial estimates of the approximate regression coefficients conditioned on being drawn from each of the K mixture components. These estimates should be provided as a p x K matrix, where K is the number of mixture components. |
update.sigma |
If |
update.sa |
Currently, estimate of mixture component variances is
not implemented, so this must be set to |
update.w |
If |
w.penalty |
Penalty term for the mixture weights. It is useful
for "regularizing" the estimate of |
drop.threshold |
Posterior probability threshold for dropping
mixture components. Should be a positive number close to zero. If,
at any point during the optimization, all posterior mixture
assignment probabilities for a given mixture component |
tol |
Convergence tolerance for co-ordinate ascent updates. |
maxiter |
Maximum number of co-ordinate ascent iterations. |
update.order |
Order of the co-ordinate ascent updates for
fitting the variational approximation. The default is
|
verbose |
If |
See https://www.overleaf.com/8954189vvpqnwpxhvhq.
An object with S3 class c("varbvsmix","list").
n |
Number of data samples used to fit model. |
mu.cov |
Posterior mean regression coefficients for covariates, including intercept. |
update.sigma |
If |
update.sa |
If |
update.w |
If |
w.penalty |
Penalty used for updating mixture weights. |
drop.threshold |
Posterior probabiltiy threshold used in the optimization procedure for setting mixture weights to zero. |
sigma |
Fitted or user-specified residual variance parameter. |
sa |
User-specified mixture variances. |
w |
Fitted or user-specified mixture weights. |
alpha |
Variational estimates of posterior mixture assignent probabilities. |
mu |
Variational estimates of posterior mean coefficients. |
s |
Variational estimates of posterior variances. |
lfsr |
Local false sign rate (LFSR) for each variable computed from variational estimates of posterior assignment probabilities and posterior means and variances. See Stephens (2017) for a definition of the LFSR. |
logZ |
Variational lower bound to marginal log-likelihood at each iteration of the co-ordinate ascent algorithm. |
err |
Maximum difference in the variational posterior probabilities at each iteration of the co-ordinate ascent algorithm. |
nzw |
Number of nonzero mixture components (including the "spike") at each iteration of the co-ordinate ascent algorithm. |
Peter Carbonetto [email protected]
M. Stephens (2017). False discovery rates: a new deal. Biostatistics 18, 275–294.
# Generate the data set. set.seed(1) n <- 200 p <- 500 X <- randn(n,p) sd <- c(0,0.2,0.5) w <- c(0.9,0.05,0.05) k <- sample(length(w),p,replace = TRUE,prob = w) beta <- sd[k] * rnorm(p) y <- c(X %*% beta + rnorm(n)) # Fit the model to the data, in which the variances of the mixture # prior are automatically selected. fit1 <- varbvsmix(X,NULL,y) # Fit the model, but use only 3 mixture components in the prior # instead of the default of 20. fit2 <- varbvsmix(X,NULL,y,3) # Use the "ground-truth" prior variances (the ones used to simulate # the data). fit3 <- varbvsmix(X,NULL,y,sd^2) # Compare predicted outcomes against observed outcomes. y.fit1 <- predict(fit1,X) print(cor(y,y.fit1)) ## Not run: library(lattice) print(xyplot(beta.est ~ beta.true, data.frame(beta.true = beta, beta.fitted = rowSums(fit$alpha * fit$mu)), pch = 20,col = "royalblue",cex = 1)) ## End(Not run)# Generate the data set. set.seed(1) n <- 200 p <- 500 X <- randn(n,p) sd <- c(0,0.2,0.5) w <- c(0.9,0.05,0.05) k <- sample(length(w),p,replace = TRUE,prob = w) beta <- sd[k] * rnorm(p) y <- c(X %*% beta + rnorm(n)) # Fit the model to the data, in which the variances of the mixture # prior are automatically selected. fit1 <- varbvsmix(X,NULL,y) # Fit the model, but use only 3 mixture components in the prior # instead of the default of 20. fit2 <- varbvsmix(X,NULL,y,3) # Use the "ground-truth" prior variances (the ones used to simulate # the data). fit3 <- varbvsmix(X,NULL,y,sd^2) # Compare predicted outcomes against observed outcomes. y.fit1 <- predict(fit1,X) print(cor(y,y.fit1)) ## Not run: library(lattice) print(xyplot(beta.est ~ beta.true, data.frame(beta.true = beta, beta.fitted = rowSums(fit$alpha * fit$mu)), pch = 20,col = "royalblue",cex = 1)) ## End(Not run)
For each candidate variable j, this function returns a Bayes factor measuring the improvement in fit when variable j is included in the model instead of variable i; that is, a larger Bayes factor indicates a better model fit by swapping variables i and j. From an optimization perspective, this could be viewed as addressing the following question: if you had to update the variational parameters for one variable so as to improve the "fit" of the variational approximation after setting the posterior inclusion probability for variable i to zero, which variable would you choose?
varbvsproxybf(X, Z, y, fit, i, vars)varbvsproxybf(X, Z, y, fit, i, vars)
X |
n x p input matrix, where n is the number of samples, and p is the number of variables. X cannot be sparse, and cannot have any missing values (NA). |
Z |
n x m covariate data matrix, where m is the number of
covariates. Do not supply an intercept as a covariate (i.e., a
column of ones), because an intercept is automatically included in
the regression model. For no covariates, set |
y |
Vector of length n containing values of the continuous outcome. |
fit |
An object inheriting from class |
i |
Variable against will. Typically, will be a variable included in the regression model with high probability, but not always. |
vars |
Set of candidate "proxy" variables. This set may include
|
varbvsproxybf returns a list with the following components:
BF |
Matrix containing Bayes factors for each candidate proxy variable and for each hyperparameter setting. |
mu |
Matrix containing estimated posterior means for each candidate proxy variable and for each hyperparameter setting. |
s |
Matrix containing estimated posterior variances for each candidate proxy variance for each hyperparameter setting. |
Peter Carbonetto [email protected]