Merge pull request #51 from pletschm/develop

Develop

.Rbuildignore

@@ -11,6 +11,7 @@
 ^README\.Rmd$
 ^cran-comments\.md$
 ^prepare_submission\.R$
+^aldvmm_tracker\.txt$
 ^\.travis\.yml$
 ^\.github$
 ^codecov\.yml$

DESCRIPTION

@@ -1,12 +1,12 @@
 Package: aldvmm
 Type: Package
 Title: Adjusted Limited Dependent Variable Mixture Models
-Version: 0.8.7
+Version: 0.8.8
 Authors@R: c(person(given = "Mark", family = "Pletscher", email = "pletscher.mark@gmail.com",
                   role = c("aut", "cre", "cph")),
              person(given = "Achim", family = "Zeileis", email = "Achim.Zeileis@R-project.org", 
                   role = c("ctb"), comment = c(ORCID = "0000-0003-0918-3766")))
-Date: 2023-07-08
+Date: 2023-10-26
 Depends: R (>= 3.5.0)
 Description: The goal of the package 'aldvmm' is to fit adjusted limited 
     dependent variable mixture models of health state utilities. Adjusted 
@@ -21,7 +21,7 @@ License: GPL-3
 Encoding: UTF-8
 LazyData: true
 RoxygenNote: 7.2.3
-Imports: numDeriv, stats, checkmate, optimr, Formula, sandwich, lmtest
+Imports: numDeriv, stats, checkmate, optimx, Formula, sandwich, lmtest
 URL: https://github.com/pletschm/aldvmm/
 BugReports: https://github.com/pletschm/aldvmm/issues
 Suggests: 

NAMESPACE

@@ -43,7 +43,7 @@ export(vcov.aldvmm)
 import(Formula)
 import(checkmate)
 import(numDeriv)
-import(optimr)
+import(optimx)
 import(stats)
 importFrom(lmtest,coefci)
 importFrom(lmtest,coeftest)

NEWS.md

@@ -33,4 +33,7 @@
 * The package "aldvmm" now uses analytical gradients instead of numerical approximations during optimization and in methods used for estimators from the "sandwich" package.
 * New methods for generic functions stats::formula(), stats::residuals() and stats::update(). Objects of class "aldvmm" can now be supplied to sandwich::sandwich(), sandwich::vcovCL(), sandwich::vcovPL(), sandwich::vcovHAC() and sandwich::vcovBS(). sandwich::vcovBS() allows re-estimating the covariance matrix using bootstrapping with and without clustering.
 * Objects of class "aldvmm" now include predicted probabilities of component membership for all observations in the estimation data.
-* New html vignette.
+* New html vignette.
+
+# aldvmm 0.8.8
+* The optimizer package was changed from "optimr" to "optimx". The functionality remains identical.

R/aldvmm.R

@@ -24,7 +24,7 @@
 #' @import numDeriv
 #' @import stats
 #' @import checkmate
-#' @import optimr
+#' @import optimx
 #' @import Formula
 #' @importFrom sandwich estfun
 #' @importFrom lmtest coeftest coefci
@@ -73,19 +73,19 @@ NULL
 #'   \code{"zero"}, \code{"random"}, \code{"constant"} and \code{"sann"}. The
 #'   default value is \code{"zero"}.
 #' @param optim.method an optional character value of one of the following
-#'   \ifelse{html}{\code{\link[optimr]{optimr}}}{\code{optimr::optimr()}}
+#'   \ifelse{html}{\code{\link[optimx]{optimr}}}{\code{optimx::optimr()}}
 #'   methods: \code{"Nelder-Mead"}, \code{"BFGS"}, \code{"CG"},
 #'   \code{"L-BFGS-B"}, \code{"nlminb"}, \code{"Rcgmin"}, \code{"Rvmmin"} and
 #'   \code{"hjn"}. The default method is \code{"BFGS"}. The method
 #'   \code{"L-BFGS-B"} is used when lower and/or upper constraints are set
 #'   using \code{'init.lo'} and \code{'init.hi'}. The method \code{"nlm"}
 #'   cannot be used in the \code{'aldvmm'} package.
 #' @param optim.control an optional list of
-#'   \ifelse{html}{\code{\link[optimr]{optimr}}}{\code{optimr::optimr()}}
+#'   \ifelse{html}{\code{\link[optimx]{optimr}}}{\code{optimx::optimr()}}
 #'   control parameters.
 #' @param optim.grad an optional logical value indicating if an analytical
 #'   gradient should be used in
-#'   \ifelse{html}{\code{\link[optimr]{optimr}}}{\code{optimr::optimr()}}
+#'   \ifelse{html}{\code{\link[optimx]{optimr}}}{\code{optimx::optimr()}}
 #'   methods that can use this information. The default value is \code{TRUE}.
 #'   If \code{'optim.grad'} is set to \code{FALSE}, a finite difference
 #'   approximation is used.
@@ -144,7 +144,7 @@ NULL
 #'   membership.
 #'
 #'   \ifelse{html}{\code{\link[aldvmm]{aldvmm}}}{ \code{aldvmm()}} uses
-#'   \ifelse{html}{\code{\link[optimr]{optimr}}}{\code{optimr::optimr()}} for
+#'   \ifelse{html}{\code{\link[optimx]{optimr}}}{\code{optimx::optimr()}} for
 #'   maximum likelihood estimation of model parameters.  The argument
 #'   \code{'optim.method'} accepts the following methods: \code{"Nelder-Mead"},
 #'   \code{"BFGS"}, \code{"CG"}, \code{"L-BFGS-B"}, \code{"nlminb"},
@@ -153,10 +153,10 @@ NULL
 #'   \ifelse{html}{\code{\link[aldvmm]{aldvmm}}}{ \code{aldvmm()}} because it
 #'   requires a different implementation of the likelihood function. The
 #'   argument \code{'optim.control'} accepts a list of
-#'   \ifelse{html}{\code{\link[optimr]{optimr}}}{\code{optimr::optimr()}}
+#'   \ifelse{html}{\code{\link[optimx]{optimr}}}{\code{optimx::optimr()}}
 #'   control parameters.  If \code{'optim.grad'} is set to \code{TRUE} the
 #'   function
-#'   \ifelse{html}{\code{\link[optimr]{optimr}}}{\code{optimr::optimr()}} uses
+#'   \ifelse{html}{\code{\link[optimx]{optimr}}}{\code{optimx::optimr()}} uses
 #'   analytical gradients during the optimization procedure for all methods 
 #'   that allow for this approach. If \code{'optim.grad'} is set to 
 #'   \code{FALSE} or a method cannot use gradients, a finite difference 
@@ -177,14 +177,17 @@ NULL
 #'   are supplied in \code{'init.est'}, the argument \code{'init.method'} is
 #'   ignored.
 #'
-#'   By default, \ifelse{html}{\code{\link[aldvmm]{aldvmm}}}{ \code{aldvmm()}}
+#'   By default, \ifelse{html}{\code{\link[aldvmm]{aldvmm}}}{\code{aldvmm()}}
 #'   performs unconstrained optimization with upper and lower limits at
 #'   \code{-Inf} and \code{Inf}.  When user-defined lower and upper limits are
-#'   supplied to \code{'init.lo'} and/or \code{'init-hi'}, these default limits
+#'   supplied to \code{'init.lo'} and/or \code{'init.hi'}, these default limits
 #'   are replaced with the user-specified values, and the method
 #'   \code{"L-BFGS-B"} is used for box-constrained optimization instead of the
 #'   user defined \code{'optim.method'}.  It is possible to only set either
-#'   maximum or minimum limits.
+#'   maximum or minimum limits.  When initial values supplied to 
+#'   \code{'init.est'} or from default methods lie outside the limits, the 
+#'   in-feasible values will be set to the limits using the function 
+#'   \ifelse{html}{\code{\link[optimx]{bmchk}}}{\code{optimx::bmchk()}}.
 #'   
 #'   The function \code{aldvmm()} returns the negative log-likelihood, Akaike 
 #'   information criterion and Bayesian information criterion. Smaller values 
@@ -293,7 +296,7 @@ NULL
 #'   (\code{"delta"}).} } }
 #'
 #'   \item{\code{optim.method}}{a character value of the used
-#'   \ifelse{html}{\code{\link[optimr]{optimr}}}{\code{optimr::optimr()}}
+#'   \ifelse{html}{\code{\link[optimx]{optimr}}}{\code{optimx::optimr()}}
 #'   method.}
 #'   \item{\code{level}}{a numeric value of the confidence level used for 
 #'   reporting.}
@@ -490,7 +493,7 @@ aldvmm <- function(formula,
                     lcpar = lcpar,
                     optim.method = optim.method)
 
-  if (!is.finite(test)) {
+  if (!is.finite(test) | (optim.method == "L-BFGS-B" & test >= 1e+20)) {
     stop("Starting values are not feasible.")
   } 
 
@@ -499,19 +502,19 @@ aldvmm <- function(formula,
   # Fit model
   #----------
 
-  fit <- optimr::optimr(fn = aldvmm.ll,
-                        par = init[["est"]],
+  fit <- optimx::optimr(par = init[["est"]],
+                        fn = aldvmm.ll,
                         X = mm,
                         y = y,
-                        lower = init[["lo"]],
-                        upper = init[["hi"]],
                         psi = psi,
-                        dist = dist,
                         ncmp = ncmp,
+                        dist = dist,
                         lcoef = lcoef,
-                        lcmp = lcmp,
                         lcpar = lcpar,
+                        lcmp = lcmp,
                         optim.method = optim.method,
+                        lower = init[["lo"]],
+                        upper = init[["hi"]],
                         method = optim.method,
                         gr = grd,
                         hessian = FALSE,

R/aldvmm.init.R

@@ -5,7 +5,7 @@
 #' creates initial values for the minimization of the negative log-likelihood
 #' returned by
 #' \ifelse{html}{\code{\link[aldvmm]{aldvmm.ll}}}{\code{aldvmm.ll()}} using
-#' \ifelse{html}{\code{\link[optimr]{optimr}}}{\code{optimr::optimr()}}.
+#' \ifelse{html}{\code{\link[optimx]{optimr}}}{\code{optimx::optimr()}}.
 #'
 #' @inheritParams aldvmm
 #' @inheritParams aldvmm.ll
@@ -29,7 +29,10 @@
 #'   are replaced with the user-specified values, and the method
 #'   \code{"L-BFGS-B"} is used for box-constrained optimization instead of the
 #'   user defined \code{'optim.method'}.  It is possible to only set either
-#'   maximum or minimum limits.
+#'   maximum or minimum limits.  When initial values supplied to 
+#'   \code{'init.est'} or from default methods lie outside the limits, the 
+#'   in-feasible values will be set to the limits using the function 
+#'   \ifelse{html}{\code{\link[optimx]{bmchk}}}{\code{optimx::bmchk()}}.
 #'
 #' @return
 #' \ifelse{html}{\code{\link[aldvmm]{aldvmm.init}}}{\code{aldvmm.init()}}
@@ -140,7 +143,7 @@ aldvmm.init <- function(X,
                            dimnames = list(NULL, "(Intercept)")) )
         names(X0) <- lcoef[1]
       }
-
+      
       # Initial values of constant-only model
       tmp <- list()
       for (i in names(zero)) {
@@ -161,7 +164,7 @@ aldvmm.init <- function(X,
       }
 
       # Fit model
-      fit <- optimr::optimr(method       = optim.method,
+      fit <- optimx::optimr(method       = optim.method,
                             fn           = aldvmm.ll,
                             par          = tmp,
                             X            = X0,
@@ -270,6 +273,17 @@ aldvmm.init <- function(X,
   }
   names(init[["hi"]]) <- names(init[["est"]])
 
+  # Set initial values outside constraint to bounds
+  #------------------------------------------------
+
+  bc <- optimx::bmchk(init[["est"]], 
+                      lower = init[["lo"]], 
+                      upper = init[["hi"]])
+
+  if(bc[["parchanged"]] == TRUE) {
+    init[["est"]] <- bc[["bvec"]]
+  }
+
   return(init)
 
 }

R/aldvmm.ll.R

@@ -40,7 +40,7 @@
 #'   distributions) enter the likelihood function as log-transformed values.
 #'
 #'   As the "L-BFGS-B" and "Rcgmin" methods in
-#'   \ifelse{html}{\code{\link[optimr]{optimr}}}{\code{optimr::optimr()}} fail
+#'   \ifelse{html}{\code{\link[optimx]{optimr}}}{\code{optimx::optimr()}} fail
 #'   if they encounter infinite values, the log-likelihood function takes the
 #'   value -1e+20 if it is infinite during these algorithms.
 #'

aldvmm_tracker.txt

@@ -0,0 +1,14 @@
+Planned
+- test.aldvmm.R: check the models really behave in the expected manner wrt messages, warnings and errors.
+- Check number of iterations is returned by optimr/optimx
+- Check all optimx::optimr() return values collected in output object.
+
+Completed
+- Email John C. Nash: Unused parameter error, why only starting value within limits, anyworkaround?
+- aldvmm.init(): Add tests for initial values lying within bounds.
+- Update detailed documentation of use of bounds and the behavior of aldvmm
+- aldvmm(): modify feasibility test of initial values to stop when ll is larger than 1e+20 for optim.method L-BFGS-B
+- Document rule of initial values outside limits in aldvmm.init(). Also mention optimx::bmchk() explicitely.
+- Add test of aldvmm.init() for initial values outside limits behavior.
+
+Discarded

cran-comments.md

@@ -2,10 +2,7 @@
 This is a re-submission. 
 
 It includes the following updates.
-* The package "aldvmm" now uses analytical gradients instead of numerical approximations.
-* New methods for generic functions stats::formula(), stats::residuals() and stats::update().
-* Objects of class "aldvmm" now include predicted probabilities of component membership for all observations in the estimation data.
-* New html vignette.
+* The optimizer package was changed from "optimr" to "optimx". The functionality remains identical.
 
 Maintainer: 'Mark Pletscher <pletscher.mark@gmail.com>'