version 0.6

DESCRIPTION

@@ -1,13 +1,13 @@
 Package: mclogit
 Type: Package
 Title: Mixed Conditional Logit Models
-Version: 0.5.1
-Date: 2017-07-16
+Version: 0.6
+Date: 2018-04-25
 Author: Martin Elff
 Maintainer: Martin Elff <mclogit@elff.eu>
 Description: Specification and estimation of conditional logit models of binary
-    responses and multinomial counts is provided, with or without alternative-
-    specific random effects. The current implementation of the estimator for random
+    responses and multinomial counts is provided, with or without
+    random effects. The current implementation of the estimator for random
     effects variances uses a Laplace approximation (or PQL) approach and thus should
     be used only if groups sizes are large.
 License: GPL-2
@@ -19,6 +19,6 @@ URL:
 BugReports: http://github.com/melff/mclogit/issues
 RoxygenNote: 5.0.1
 NeedsCompilation: no
-Packaged: 2017-07-15 23:56:48 UTC; elff
+Packaged: 2018-05-10 09:48:14 UTC; elff
 Repository: CRAN
-Date/Publication: 2017-07-17 09:46:59 UTC
+Date/Publication: 2018-05-10 12:16:38 UTC

MD5

@@ -1,25 +1,23 @@
-5b79bb423fd1a848c875aa6eff03c133 *DESCRIPTION
+bc526bce97fcc5ab31252ecf1274de1c *DESCRIPTION
 8e2f6305165fc3d69a294d09227ba690 *NAMESPACE
 5f46b25a049e81e86774e78cd294ff89 *R/AIC-mclogit.R
 fda85116896df5934a2ce2b9e3118dee *R/anova-mclogit.R
 825e92e98bee4d5cba8ebbb5bad2f448 *R/getSummary-mblogit.R
 174a746177e5e0ffa74047959c4b59ba *R/getSummary-mclogit.R
 8443552bb939bf83bafc6255125dd6a9 *R/mblogit.R
 e59d1445b41da59347a1d181b9e44025 *R/mclogit-fit.R
-373e59f4f1e50208db954430b959aabc *R/mclogit.R
+4391129b68ab0c0b4b4f6c6920d82f9a *R/mclogit.R
 202e8dd6dd5ee19a281df6277484e46c *R/zzz.R
-39d40b69dd3cffb9bf8c3d6fbd638fae *data/Transport.rda
-030cc52f5962d7f76f12370c2c108ff9 *data/electors.rda
+f07c5b8a1b9a311ce9f14c89a3501be5 *data/Transport.rda
+170a2686691e2ec1bc0d197ec11a8397 *data/electors.rda
 de4c63b49c62998527d72c185ceab0d3 *demo/00Index
 6ba2e10f92f3743f465cda6cc059f5a1 *demo/mclogit.test.R
-198cfecbf3ad7e32f75624a87953e05f *inst/ChangeLog
-b98c84290f3b8c0921224cde137b44c8 *inst/ChangeLog.sync-conflict-20170309-201336
-055faf0212b89aec35ac26ea67dc7f75 *inst/NEWS.Rd
-85a099ef2942ee748e7c3baef305479f *inst/NEWS.sync-conflict-20170309-201318.Rd
+285f7ffde7690a50a68c9e2d79922631 *inst/ChangeLog
+3108d598ac3ef8fec366d0563fc95d72 *inst/NEWS.Rd
 c05ff442f7a2312ff8f3d6b5aa96ddeb *man/Transport.Rd
 6384c0365cb1b46ea7230227dd9c8392 *man/electors.Rd
 5bb2193c1f179dc06510bfd2b8a46dfe *man/getSummary-mclogit.Rd
 215c7ba13f2a1ca831b87c00adfb6e4a *man/mblogit.Rd
-14524331f76dc7dd591490b6a448865f *man/mclogit.Rd
+c1493568f138301373475f881ff48877 *man/mclogit.Rd
 2592f185bc66478485d812cd6eaa59d6 *man/mclogit.fit.Rd
 89ef120edb709000ad26477de8194dd7 *man/mclogit_control.Rd

R/mclogit.R

@@ -24,16 +24,25 @@ quickInteraction <- function(by){
   return(structure(f,unique=uf))
 }
 
-constInSets <- function(X,sets){
-    ans <- integer(0)
+matConstInSets <- function(X,sets){
+    ans <- logical(ncol(X))
     for(i in 1:ncol(X)){
         v <- tapply(X[,i],sets,var)
-        if(all(v[is.finite(v)]==0)) ans <- c(ans,i)
+        ans[i] <- all(v[is.finite(v)]==0) 
     }
-    names(ans) <- colnames(X)[ans]
     ans
 }
 
+listConstInSets <- function(X,sets){
+    ans <- logical(length(X))
+    for(i in 1:length(X)){
+        v <- tapply(X[[i]],sets,var)
+        ans[i] <- all(v[is.finite(v)]==0) 
+    }
+    ans
+}
+
+
 mclogit <- function(
                 formula,
                 data=parent.frame(),
@@ -102,10 +111,12 @@ mclogit <- function(
     xlevels <- .getXlevels(mt,mf)
     icpt <- match("(Intercept)",colnames(X),nomatch=0)
     if(icpt) X <- X[,-icpt,drop=FALSE]
-    const <- constInSets(X,sets)
-    if(length(const)){
-        warning("removing ",paste(names(const),collapse=",")," from model")
-        X <- X[,-const,drop=FALSE]
+    const <- matConstInSets(X,sets)
+    if(any(const)){
+        warning("removing ",
+                gsub("(Intercept)","intercept",paste(colnames(X)[const],collapse=","),fixed=TRUE),
+                " from model due to insufficient within-choice set variance")
+        X <- X[,!const,drop=FALSE]
     }
     if(!length(start)){
       drop.coefs <- check.mclogit.drop.coefs(Y,sets,weights,X,
@@ -115,14 +126,21 @@ mclogit <- function(
         X <- X[,!drop.coefs,drop=FALSE]
       }
     }
+    if(ncol(X)<1)
+        stop("No predictor variable remains in model")
 
+    if(length(random) && control$trace)
+        cat("Fitting plain conditional logit to obtain starting values")
 
     fit <- mclogit.fit(Y,sets,weights,X,
-                        control=control,
-                        start = start,
-                        offset = offset)
+                       control=control,
+                       start = start,
+                       offset = offset)
 
     if(length(random)){ ## random effects
+
+        if(control$trace)
+            cat("Fitting random effects/random coefficients model")
 
         null.dev <- fit$null.deviance
 
@@ -131,7 +149,6 @@ mclogit <- function(
 
         groups <- random$groups
         rX <- model.matrix(rt,mf,contrasts)
-
         groups <- mf[groups]
 
         nlev <- length(groups)
@@ -141,6 +158,21 @@ mclogit <- function(
             for(i in 2:nlev)
                 groups[[i]] <- quickInteraction(groups[c(i-1,i)])
         }
+
+        gconst <- listConstInSets(groups,sets)
+        if(any(gconst)){
+            rconst <- matConstInSets(rX,sets)
+            if(any(rconst)){
+                cat("\n")
+                warning("removing ",
+                        gsub("(Intercept)","intercept",paste(colnames(rX)[rconst],collapse=","),fixed=TRUE),
+                        " from random part of the model\n because of insufficient within-choice set variance")
+                rX <- rX[,!rconst,drop=FALSE]
+            }
+            if(ncol(rX)<1)
+                stop("No predictor variable remains in random part of the model.\nPlease reconsider your model specification.")
+        }
+
 
         Z <- lapply(groups,mkZ,rX=rX)
         G <- mkG(rX)
@@ -370,8 +402,10 @@ predict.mclogit <- function(object, newdata=NULL,type=c("link","response"),se.fi
                       contasts.arg=object$contrasts,
                       xlev=object$xlevels
                       )
-    drop <- match("(Intercept)",colnames(X))
-    X <- X[,-drop,drop=FALSE]
+
+    cf <- coef(object)
+    X <- X[,names(cf)]
+
     eta <- c(X %*% coef(object))
     if(se.fit){
         V <- vcov(object)

inst/ChangeLog

@@ -1,3 +1,10 @@
+2018-04-25:
+  - Improved handling of with misspecified random effect structures.
+  - Added documentation about new handling of misspecified models.
+
+2017-10-25:
+  - Fixed handling of dropped predictors in `predict.mclogit`.
+
 2017-01-26:
   - Fixed some bugs in predict models for `mclogit` objects.
   - Made sure that dummy coding is used for response factors even if they are ordinal

inst/NEWS.Rd

@@ -2,6 +2,34 @@
 \title{\emph{mclogit} News}
 \encoding{UTF-8}
 
+\section{Version 0.6}{
+  \subsection{NEW FEATURES}{
+    \itemize{
+      \item \code{mclogit} now complains about (i.e. throws an error
+      exception) when the random effects structure cannot be estimated,
+      e.g. because random effects are constant within choice sets
+      and therefore drop out by the formation of conditional logits.
+    }
+  }
+}
+
+\section{Version 0.5}{
+  \subsection{NEW FEATURES}{
+    \itemize{
+      \item \code{mclogit} now supports conditional logit models with
+      random slopes.
+      \item \code{mblogit} now supports multinomial baseline models with
+      random intercept and random slopes.
+    }
+  }
+  \subsection{BUGFIXES}{
+  \itemize{
+    \item \code{predict} methods of objects created by \code{mclogit}
+    and \code{mblogit} are better in handling missing data.
+    }
+  }
+}
+
 \section{Version 0.4}{
   \subsection{NEW FEATURES}{
   \itemize{

man/mclogit.Rd

@@ -22,9 +22,9 @@
   where the choice set may vary across choice occasions.
 
   Conditional logit models without random effects are fitted by
-  Fisher-scoring/IWLS.
-  The implementation of mixed conditional logit currently is limited
-  to PQL and random intercepts.
+  Fisher-scoring/IWLS. Models with random effects 
+  (mixed conditional logit models) are estimated via maximum likelihood
+  with a simple Laplace aproximation (aka PQL).
 }
 \usage{
 
@@ -105,6 +105,27 @@ mclogit(formula, data=parent.frame(), random=NULL,
   same structure as an object of class "\link[stats]{glm}".
 }
 
+\note{
+  Covariates that are constant within choice sets are automatically
+  dropped from the model formula specified by the \code{formula}
+  argument of \code{mclogit}.
+
+  If the model contains random effects, these should
+  \itemize{
+    \item either vary within choice sets (e.g. the levels of a factor
+    that defines the choice sets should not be nested within the levels
+    of factor)
+    \item or be random coefficients of covariates that vary within
+    choice sets. 
+  }
+
+  In earlier versions of the package (prior to 0.6) it will lead to a
+  failure of the model fitting algorithm if these conditions are not
+  satisfied. Since version 0.6 of the package, the function
+  \code{mclogit} will complain about such model a misspecification
+  explicitely.
+}
+
 \examples{
 data(Transport)