| Title: | High-Dimensional Regression and CAR Score Variable Selection |
|---|---|
| Description: | Implements the regression approach of Zuber and Strimmer (2011) "High-dimensional regression and variable selection using CAR scores" SAGMB 10: 34, <DOI:10.2202/1544-6115.1730>. CAR scores measure the correlation between the response and the Mahalanobis-decorrelated predictors. The squared CAR score is a natural measure of variable importance and provides a canonical ordering of variables. This package provides functions for estimating CAR scores, for variable selection using CAR scores, and for estimating corresponding regression coefficients. Both shrinkage as well as empirical estimators are available. |
| Authors: | Verena Zuber and Korbinian Strimmer. |
| Maintainer: | Korbinian Strimmer <[email protected]> |
| License: | GPL (>= 3) |
| Version: | 1.1.11 |
| Built: | 2024-12-22 06:20:55 UTC |
| Source: | CRAN |
The "care" package implements the CAR regression approach described in Zuber and Strimmer (2011). CAR scores measure the correlations between the response and the Mahalanobis-decorrelated predictors. The squared CAR score is a natural measure of variable importance and provides a canonical ordering of variables - see Zuber and Strimmer (2011) for details.
This package provides functions for estimating CAR scores, for variable selection using CAR scores, and for estimating corresponding regression coefficients. Both shrinkage as well as empirical estimators are available.
The name of the package refers to CAR estimation and CAR regression.
Verena Zuber and Korbinian Strimmer (https://strimmerlab.github.io/)
Zuber, V., and K. Strimmer. 2011. High-dimensional regression and variable selection using CAR scores. Statist. Appl. Genet. Mol. Biol. 10: 34. <DOI:10.2202/1544-6115.1730>
Website: https://strimmerlab.github.io/software/care/
carscore, slm,
efron2004, lu2004.
carscore estimates the vector of CAR scores, either using the standard
empirical estimator of the correlation matrix, or a shrinkage estimator.
carscore(Xtrain, Ytrain, lambda, diagonal=FALSE, verbose=TRUE)carscore(Xtrain, Ytrain, lambda, diagonal=FALSE, verbose=TRUE)
Xtrain |
Matrix of predictors (columns correspond to variables). |
Ytrain |
Univariate response variable. |
lambda |
The correlation shrinkage intensity (range 0-1).
If not specified (the default) it is estimated using an
analytic formula from Sch\"afer and Strimmer (2005). For |
diagonal |
For |
verbose |
If |
The CAR scores are the correlations between the response and the Mahalanobis-decorrelated predictors. CAR score is an abbreviation for Correlation-Adjusted (marginal) coRelation, where the first correlation matrix refers dependencies among predictors.
In Zuber and Strimmer (2011) it is argued that squared CAR scores are a natural measure for variable importance and it is shown that variable selection based on CAR scores is highly efficient compared to competing approaches such as elastic net lasso, or boosting.
If the response is binary (or descrete) the corresponding quantity
are CAT scores (see catscore).
carscore returns
a vector containing the CAR scores (or marginal correlations for diagonal=TRUE).
Verena Zuber and Korbinian Strimmer (https://strimmerlab.github.io).
Zuber, V., and K. Strimmer. 2011. High-dimensional regression and variable selection using CAR scores. Statist. Appl. Genet. Mol. Biol. 10: 34. <DOI:10.2202/1544-6115.1730>
# load care library library("care") ###### # empirical CAR scores for diabetes data data(efron2004) xnames = colnames(efron2004$x) n = dim(efron2004$x)[1] car = carscore(efron2004$x, efron2004$y, lambda=0) car # compare orderings # variables ordered by squared CAR scores xnames[order(car^2, decreasing=TRUE)] # "bmi" "s5" "bp" "s3" "s4" "s6" "sex" "age" "s2" "s1" # compare with ordering by t-scores / partial correlations pcor = pcor.shrink(cbind(efron2004$y,efron2004$x), lambda=0, verbose=FALSE)[-1,1] xnames[order(pcor^2, decreasing=TRUE)] # "bmi" "bp" "s5" "sex" "s1" "s2" "s4" "s6" "s3" "age" # compare with ordering by marginal correlations mcor = cor(efron2004$y,efron2004$x) #mcor = carscore(efron2004$x, efron2004$y, diagonal=TRUE, lambda=0) xnames[order(mcor^2, decreasing=TRUE)] # "bmi" "s5" "bp" "s4" "s3" "s6" "s1" "age" "s2" "sex" # decomposition of R^2 sum(car^2) slm(efron2004$x, efron2004$y, lambda=0, lambda.var=0)$R2 # pvalues for empirical CAR scores pval = 1-pbeta(car^2, shape1=1/2, shape2=(n-2)/2) pval <= 0.05 ###### # shrinkage CAR scores for Lu et al. (2004) data data(lu2004) dim(lu2004$x) # 30 403 # compute shrinkage car scores car = carscore(lu2004$x, lu2004$y) # most important genes order(car^2, decreasing=TRUE)[1:10] # compare with empirical marginal correlations mcor = cor(lu2004$y, lu2004$x) order(mcor^2, decreasing=TRUE)[1:10] # decomposition of R^2 sum(car^2) slm(lu2004$x, lu2004$y)$R2# load care library library("care") ###### # empirical CAR scores for diabetes data data(efron2004) xnames = colnames(efron2004$x) n = dim(efron2004$x)[1] car = carscore(efron2004$x, efron2004$y, lambda=0) car # compare orderings # variables ordered by squared CAR scores xnames[order(car^2, decreasing=TRUE)] # "bmi" "s5" "bp" "s3" "s4" "s6" "sex" "age" "s2" "s1" # compare with ordering by t-scores / partial correlations pcor = pcor.shrink(cbind(efron2004$y,efron2004$x), lambda=0, verbose=FALSE)[-1,1] xnames[order(pcor^2, decreasing=TRUE)] # "bmi" "bp" "s5" "sex" "s1" "s2" "s4" "s6" "s3" "age" # compare with ordering by marginal correlations mcor = cor(efron2004$y,efron2004$x) #mcor = carscore(efron2004$x, efron2004$y, diagonal=TRUE, lambda=0) xnames[order(mcor^2, decreasing=TRUE)] # "bmi" "s5" "bp" "s4" "s3" "s6" "s1" "age" "s2" "sex" # decomposition of R^2 sum(car^2) slm(efron2004$x, efron2004$y, lambda=0, lambda.var=0)$R2 # pvalues for empirical CAR scores pval = 1-pbeta(car^2, shape1=1/2, shape2=(n-2)/2) pval <= 0.05 ###### # shrinkage CAR scores for Lu et al. (2004) data data(lu2004) dim(lu2004$x) # 30 403 # compute shrinkage car scores car = carscore(lu2004$x, lu2004$y) # most important genes order(car^2, decreasing=TRUE)[1:10] # compare with empirical marginal correlations mcor = cor(lu2004$y, lu2004$x) order(mcor^2, decreasing=TRUE)[1:10] # decomposition of R^2 sum(car^2) slm(lu2004$x, lu2004$y)$R2
Diabetes data (10 variables, 442 measurements) as used in the study of Efron et al. (2004). The data is standardized such that the means of all variables are zero, and all variances are equal to one.
data(efron2004)data(efron2004)
efron2004$x is a 422 x 10 matrix containing the measurements
of the explanatory variables (age, sex, body mass, etc.).
The rows contain the samples and the columns the variables.
efron2004$y contains the response.
The orginal data are available in the lars R package, see https://cran.r-project.org/package=lars. Note that this uses a slightly different standardization.
Efron, B., et al. 2004. Least angle regression (with discussion). Ann. Statist. 32:407–499. <DOI:10.1214/009053604000000067>
# load care library library("care") # load Efron et al. (2004) diabetes data set data(efron2004) dim(efron2004$x) # 442 10 colnames(efron2004$x) length(efron2004$y) # 442# load care library library("care") # load Efron et al. (2004) diabetes data set data(efron2004) dim(efron2004$x) # 442 10 colnames(efron2004$x) length(efron2004$y) # 442
Gene expression data (403 genes for 30 samples) from the microarray study of Lu et al. (2004).
data(lu2004)data(lu2004)
lu2004$x is a 30 x 403 matrix containing the log expression levels.
The rows contain the samples and the columns the genes.
lu2004$y is the age of for each sample.
This data set contains measurements of the gene expression of 403 genes from 30 human brain samples. In addition, the age of each patient is provided.
The original data are available from the GEO public functional genomics database at URL https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE1572 and are described in Lu et al. (2004). The selected 403 genes result from prescreening and preprocessing as described in Zuber and Strimmer (2011).
Lu, T., et al. 2004. Gene regulation and DNA damage in the ageing human brain. Nature 429:883–891. <DOI:10.1038/nature02661>
Zuber, V., and K. Strimmer. 2011. High-dimensional regression and variable selection using CAR scores. Statist. Appl. Genet. Mol. Biol. 10: 34. <DOI:10.2202/1544-6115.1730>
# load care library library("care") # load Lu et al. (2004) data set data(lu2004) dim(lu2004$x) # 30 403 hist(lu2004$x) length(lu2004$y) # 30 lu2004$y # age# load care library library("care") # load Lu et al. (2004) data set data(lu2004) dim(lu2004$x) # 30 403 hist(lu2004$x) length(lu2004$y) # 30 lu2004$y # age
slm fits a linear model and computes
(standardized) regression coefficients by plugin of shrinkage estimates of correlations and variances.
Using the argument predlist several models can be fitted on the same data set.
make.predlist constructs a predlist argument for use with slm.
slm(Xtrain, Ytrain, predlist, lambda, lambda.var, diagonal=FALSE, verbose=TRUE) ## S3 method for class 'slm' predict(object, Xtest, verbose=TRUE, ...) make.predlist(ordering, numpred, name="SIZE")slm(Xtrain, Ytrain, predlist, lambda, lambda.var, diagonal=FALSE, verbose=TRUE) ## S3 method for class 'slm' predict(object, Xtest, verbose=TRUE, ...) make.predlist(ordering, numpred, name="SIZE")
Xtrain |
Matrix of predictors (columns correspond to variables). |
Ytrain |
Univariate continous response variable. |
predlist |
A list specifying the predictors to be included when fitting the linear regression. Each entry in the list is a vector containing the indices of variables used per model. If left unspecified single full-sized model using all variables in Xtrain is assumed. For a given ordering of covariables a suitable |
lambda |
The correlation shrinkage intensity (range 0-1).
If not specified (the default) it is estimated using an
analytic formula from Sch\"afer and Strimmer (2005). For |
lambda.var |
The variance shrinkage intensity (range 0-1). If
not specified (the default) it is estimated
using an analytic formula from Opgen-Rhein and Strimmer
(2007). For |
diagonal |
If |
verbose |
If |
object |
An |
Xtest |
A matrix containing the test data set. Note that the rows correspond to observations and the columns to variables. |
... |
Additional arguments for generic predict. |
ordering |
The ordering of the predictors (most important predictors are first). |
numpred |
The number of included predictors (may be a scalar or a vector). The predictors
are included in the order specified by |
name |
The name assigned to each model is |
The regression coefficients are obtained by estimating the joint joint covariance matrix of the response and the predictors, and subsequently computing the the regression coefficients by inversion of this matrix - see Opgen-Rhein and Strimmer (2007). As estimators for the covariance matrix either the standard empirical estimator or a Stein-type shrinkage estimator is employed. The use of the empirical covariance leads to the OLS estimates of the regression coefficients, whereas otherwise shrinkage estimates are obtained.
slm returns a list with the following components:
regularization: The shrinkage intensities used for estimating correlations and variances.
std.coefficients: The standardized regression coefficients, i.e. the regression coefficients
computed from centered and standardized input data. Thus, by construction the intercept is zero.
Furthermore, for diagonal=TRUE the standardized regression coefficient for each predictor is
identical to the respective marginal correlation.
coefficients: Regression coefficients.
numpred: The number of predictors used in each investigated model.
R2: For diagonal=TRUE this is the multiple correlation coefficient
between the response and the predictor, or the proportion of explained variance, with range
from 0 to 1.
For diagonal=TRUE this equals the sum of squared marginal
correlations. Note that this sum may be larger than 1!
sd.resid: The residual unexplained error.
predict.slm returns the means predicted for each sample and model as well as the corresponding
predictive standard deviations (attached as attribute "sd").
Korbinian Strimmer (https://strimmerlab.github.io).
Opgen-Rhein, R., and K. Strimmer. 2007. From correlation to causation networks: a simple approximate learning algorithm and its application to high-dimensional plant gene expression data. BMC Syst. Biol. 1: 37. <DOI:10.1186/1752-0509-1-37>
Sch\"afer, J., and K. Strimmer. 2005. A shrinkage approach to large-scale covariance estimation and implications for functional genomics. Statist. Appl. Genet. Mol. Biol. 4: 32. <DOI:10.2202/1544-6115.1175>
# load care library library("care") ## example with large number of samples and small dimension ## (using empirical estimates of regression coefficients) # diabetes data data(efron2004) x = efron2004$x y = efron2004$y n = dim(x)[1] d = dim(x)[2] xnames = colnames(x) # empirical regression coefficients fit = slm(x, y, lambda=0, lambda.var=0) fit # note that in this example the regression coefficients # and the standardized regression coefficients are identical # as the input data have been standardized to mean zero and variance one # compute corresponding t scores / partial correlations df = n-d-1 pcor = pcor.shrink(cbind(y,x), lambda=0)[-1,1] t = pcor * sqrt(df/(1-pcor^2)) t.pval = 2 - 2 * pt(abs(t), df) b = fit$coefficients[1,-1] cbind(b, pcor, t, t.pval) # compare results with those from lm function lm.out = lm(y ~ x) summary(lm.out) # prediction of fitted values at the position of the training data lm.out$fitted.values mu.hat = predict(fit, x) # precticted means mu.hat attr(mu.hat, "sd") # predictive error sd(y-mu.hat) # ordering of the variables using squared empirical CAR score car = carscore(x, y, lambda=0) ocar = order(car^2, decreasing=TRUE) xnames[ocar] # CAR regression models with 5, 7, 9 included predictors car.predlist = make.predlist(ocar, numpred = c(5,7,9), name="CAR") car.predlist slm(x, y, car.predlist, lambda=0, lambda.var=0) # plot regression coefficients for all possible CAR models p=ncol(x) car.predlist = make.predlist(ocar, numpred = 1:p, name="CAR") cm = slm(x, y, car.predlist, lambda=0, lambda.var=0) bmat = cm$coefficients[,-1] bmat par(mfrow=c(2,1)) plot(1:p, bmat[,1], type="l", ylab="estimated regression coefficients", xlab="number of included predictors", main="CAR Regression Models for Diabetes Data", xlim=c(1,p+1), ylim=c(min(bmat), max(bmat))) for (i in 2:p) lines(1:p, bmat[,i], col=i, lty=i) for (i in 1:p) points(1:p, bmat[,i], col=i) for (i in 1:p) text(p+0.5, bmat[p,i], xnames[i]) plot(1:p, cm$R2, type="l", ylab="estimated R2", xlab="number of included predictors", main="Proportion of Explained Variance", ylim=c(0,0.6)) R2max = max(cm$R2) lines(c(1,p), c(R2max, R2max), col=2) par(mfrow=c(1,1)) ## example with small number of samples and large dimension ## (using shrinkage estimates of regression coefficients) data(lu2004) dim(lu2004$x) # 30 403 fit = slm(lu2004$x, lu2004$y) fit# load care library library("care") ## example with large number of samples and small dimension ## (using empirical estimates of regression coefficients) # diabetes data data(efron2004) x = efron2004$x y = efron2004$y n = dim(x)[1] d = dim(x)[2] xnames = colnames(x) # empirical regression coefficients fit = slm(x, y, lambda=0, lambda.var=0) fit # note that in this example the regression coefficients # and the standardized regression coefficients are identical # as the input data have been standardized to mean zero and variance one # compute corresponding t scores / partial correlations df = n-d-1 pcor = pcor.shrink(cbind(y,x), lambda=0)[-1,1] t = pcor * sqrt(df/(1-pcor^2)) t.pval = 2 - 2 * pt(abs(t), df) b = fit$coefficients[1,-1] cbind(b, pcor, t, t.pval) # compare results with those from lm function lm.out = lm(y ~ x) summary(lm.out) # prediction of fitted values at the position of the training data lm.out$fitted.values mu.hat = predict(fit, x) # precticted means mu.hat attr(mu.hat, "sd") # predictive error sd(y-mu.hat) # ordering of the variables using squared empirical CAR score car = carscore(x, y, lambda=0) ocar = order(car^2, decreasing=TRUE) xnames[ocar] # CAR regression models with 5, 7, 9 included predictors car.predlist = make.predlist(ocar, numpred = c(5,7,9), name="CAR") car.predlist slm(x, y, car.predlist, lambda=0, lambda.var=0) # plot regression coefficients for all possible CAR models p=ncol(x) car.predlist = make.predlist(ocar, numpred = 1:p, name="CAR") cm = slm(x, y, car.predlist, lambda=0, lambda.var=0) bmat = cm$coefficients[,-1] bmat par(mfrow=c(2,1)) plot(1:p, bmat[,1], type="l", ylab="estimated regression coefficients", xlab="number of included predictors", main="CAR Regression Models for Diabetes Data", xlim=c(1,p+1), ylim=c(min(bmat), max(bmat))) for (i in 2:p) lines(1:p, bmat[,i], col=i, lty=i) for (i in 1:p) points(1:p, bmat[,i], col=i) for (i in 1:p) text(p+0.5, bmat[p,i], xnames[i]) plot(1:p, cm$R2, type="l", ylab="estimated R2", xlab="number of included predictors", main="Proportion of Explained Variance", ylim=c(0,0.6)) R2max = max(cm$R2) lines(c(1,p), c(R2max, R2max), col=2) par(mfrow=c(1,1)) ## example with small number of samples and large dimension ## (using shrinkage estimates of regression coefficients) data(lu2004) dim(lu2004$x) # 30 403 fit = slm(lu2004$x, lu2004$y) fit