Version 3.1-0

DESCRIPTION

@@ -1,7 +1,7 @@
 Package: grpreg
 Title: Regularization Paths for Regression Models with Grouped Covariates
 Version: 3.1-0
-Date: 2017-01-15
+Date: 2017-05-18
 Author: Patrick Breheny [aut, cre], Yaohui Zeng [ctb]
 Maintainer: Patrick Breheny <patrick-breheny@uiowa.edu>
 Depends: R (>= 3.1.0), Matrix

NEWS

@@ -1,16 +1,17 @@
-3.1-0 ()
+3.1-0 (2017-05-18)
   * New: Additional tests and support for coersion of various types with
     respect to both X and y
-  * Internal: new SSR-BEDPP feature screening rule for group lasso
   * Change: Convergence criterion now based on RMSD of linear predictors
   * Change: 'Lung' and 'Birthwt' data sets now use factor representation of
     group, as character vectors are inherently ambiguous with respect to order
   * Change: max.iter now based on total number of iterations for entire path
   * Internal: 'X', 'group', and 'group.multiplier' now bundled together in an
     object called 'XG' to enforce agreement at all times
-  * Fixed:
+  * Internal: new SSR-BEDPP feature screening rule for group lasso
   * Internal: Registration of native routines
   * Internal: Changing PROTECT/UNPROTECT to conform to new coding standards
+  * Fixed: The binding of X and G fixes several potential bugs, including
+    Issue #12 (GitHub)
 
 3.0-2
   *  Fixed bug involving mismatch between group.multiplier and group if

inst/tests/agreement.R

@@ -1,6 +1,3 @@
-##source("~/dev/.grpreg.setup.R")
-#require(grpreg)
-
 #################################
 .test = "gel reproduces lasso" ##
 #################################

inst/tests/basic-functionality.R

@@ -1,5 +1,3 @@
-#source("~/dev/.grpreg.setup.R")
-
 .test = "grpreg() reproduces simple linear regression"
 n <- 5
 p <- 1

inst/tests/coerce.R

@@ -29,4 +29,3 @@ fit3 <- grpreg(X, y, group=g3, family="binomial")
 check(coef(fit1, which=50), coef(fit2, which=50), tol=0.001)
 check(coef(fit2, which=50), coef(fit3, which=50), tol=0.001)
 check(coef(fit1, which=50), coef(fit3, which=50), tol=0.001)
-# check(names(coef(fit1))=="TEST ME")  # Do I have good tests for constant columns?

inst/tests/extra-features.R

@@ -1,5 +1,3 @@
-set.seed(1)
-
 .test = "logLik is correct"
 n <- 50
 group <- rep(0:4,5:1)

inst/tests/grpsurv.R

@@ -1,5 +1,3 @@
-set.seed(1)
-
 .test = "grpsurv works"
 y <- survival::Surv(rexp(50), sample(rep(0:1, c(10,40))))
 X <- matrix(rnorm(50*6), 50, 6)

inst/tests/multitask.R

@@ -1,4 +1,3 @@
-set.seed(1)
 n <- 100
 p <- 5
 

inst/tests/standardization-orthogonalization.R

@@ -1,4 +1,3 @@
-set.seed(1)
 n <- 20
 p <- 5
 l <- 5

man/Lung.Rd

@@ -38,7 +38,7 @@ data(Lung)
     0.}
   }
 }
-\source{\url{http://CRAN.R-project.org/package=survival}}
+\source{\url{https://cran.r-project.org/package=survival}}
 \references{
   \itemize{
     \item Kalbfleisch D and Prentice RL (1980), \emph{The Statistical

man/grpreg.Rd

@@ -36,19 +36,20 @@ tau = 1/3, group.multiplier, warn=TRUE, returnX = FALSE, ...)
     computes a grid of lambda values that ranges uniformly on the log
     scale over the relevant range of lambda values.}
   \item{lambda.min}{The smallest value for \code{lambda}, as a fraction
-    of \code{lambda.max}.  Default is .0001 if the number of observations
-    is larger than the number of covariates and .05 otherwise.}
-  \item{log.lambda}{Whether compute the grid values of lambda on log scale (default) 
-  or linear scale.}
+    of \code{lambda.max}.  Default is .0001 if the number of
+    observations is larger than the number of covariates and .05
+    otherwise.}
+  \item{log.lambda}{Whether compute the grid values of lambda on log
+    scale (default) or linear scale.}
   \item{alpha}{\code{grpreg} allows for both a group penalty and an L2
     (ridge) penalty; \code{alpha} controls the proportional weight of
     the regularization parameters of these two penalties.  The group
     penalties' regularization parameter is \code{lambda*alpha}, while
     the regularization parameter of the ridge penalty is
     \code{lambda*(1-alpha)}.  Default is 1: no ridge penalty.}
   \item{eps}{Convergence threshhold.  The algorithm iterates until the
-    change (on the standardized scale) in any coefficient is less than
-    \code{eps}.  Default is \code{.001}.  See details.}
+    RMSD for the change in linear predictors for each coefficient is
+    less than \code{eps}.  Default is \code{1e-4}.  See details.}
   \item{max.iter}{Maximum number of iterations (total across entire
     path). Default is 10000.  See details.}
   \item{dfmax}{Limit on the number of parameters allowed to be nonzero.
@@ -76,16 +77,16 @@ tau = 1/3, group.multiplier, warn=TRUE, returnX = FALSE, ...)
 \details{
   There are two general classes of methods involving grouped penalties:
   those that carry out bi-level selection and those that carry out group
-  selection.  Bi-level means carrying out variable selection at the group
-  level as well as the level of individual covariates (i.e., selecting
-  important groups as well as important members of those groups).  Group
-  selection selects important groups, and not members within the group --
-  i.e., within a group, coefficients will either all be zero or all
-  nonzero.  The \code{grLasso}, \code{grMCP}, and \code{grSCAD}
-  penalties carry out group selection, while the \code{gel} and
-  \code{cMCP} penalties carry out bi-level selection.  For bi-level
-  selection, see also the \code{\link{gBridge}} function.  For
-  historical reasons and backwards compatibility, some of these
+  selection.  Bi-level means carrying out variable selection at the
+  group level as well as the level of individual covariates (i.e.,
+  selecting important groups as well as important members of those
+  groups).  Group selection selects important groups, and not members
+  within the group -- i.e., within a group, coefficients will either all
+  be zero or all nonzero.  The \code{grLasso}, \code{grMCP}, and
+  \code{grSCAD} penalties carry out group selection, while the
+  \code{gel} and \code{cMCP} penalties carry out bi-level selection.
+  For bi-level selection, see also the \code{\link{gBridge}} function.
+  For historical reasons and backwards compatibility, some of these
   penalties have aliases; e.g., \code{gLasso} will do the same thing as
   \code{grLasso}, but users are encouraged to use \code{grLasso}.
 
@@ -166,8 +167,8 @@ tau = 1/3, group.multiplier, warn=TRUE, returnX = FALSE, ...)
   \item{lambda}{The sequence of \code{lambda} values in the path.}
   \item{alpha}{Same as above.}
   \item{loss}{A vector containing either the residual sum of squares
-  (\code{"gaussian"}) or negative log-likelihood (\code{"binomial"}) of
-  the fitted model at each value of \code{lambda}.}
+    (\code{"gaussian"}) or negative log-likelihood (\code{"binomial"})
+    of the fitted model at each value of \code{lambda}.}
   \item{n}{Number of observations.}
   \item{penalty}{Same as above.}
   \item{df}{A vector of length \code{nlambda} containing estimates of

man/grpsurv.Rd

@@ -44,8 +44,8 @@ group.multiplier, warn=TRUE, returnX=FALSE, ...)
     a sequence of values of length \code{nlambda} is computed
     automatically, equally spaced on the log scale.}
   \item{eps}{Convergence threshhold.  The algorithm iterates until the
-    relative change in any coefficient is less than \code{eps}.  Default
-    is \code{.001}.}
+    RMSD for the change in linear predictors for each coefficient is
+    less than \code{eps}.  Default is \code{0.001}.}
   \item{max.iter}{Maximum number of iterations (total across entire
     path).  Default is 10000.}
   \item{dfmax}{Limit on the number of parameters allowed to be nonzero.

src/gdfit_gaussian.c

@@ -176,10 +176,6 @@ void bedpp_glasso(int *e3, double *yTxxTv1, double *xTv1_sq, double *xTy_sq,
       e3[g] = 0; // reject
     }
 
-    // debug
-    // Rprintf("\t K[%d]: %d;\t Reject: %d;\t LHS_temp[%d]: %f;\t LHS[%d]: %f;\t RHS[%d]: %f;\t yTxxTv1[%d]: %f;\t xTv1_sq[%d]: %f;\t xTy_sq[%d]: %f\n", 
-    //         g, K[g], e3[g], g, LHS_temp, g, LHS, g, RHS, g, yTxxTv1[g], g, xTv1_sq[g], g, xTy_sq[g]);
-
   }
 }