-
Dependency changes:
- Occupational prestige data for the examples is now drawn from
carData
package (formerlycar
)
- Occupational prestige data for the examples is now drawn from
-
Other minor changes to namespace calls to pass CRAN checks
-
predVals()
has been rewritten to compute fitted values according to the "observed value" approach advocated in the following article:Hanmer, M. J. and Ozan Kalkan, K. (2013), Behind the Curve: Clarifying the Best Approach to Calculating Predicted Probabilities and Marginal Effects from Limited Dependent Variable Models. American Journal of Political Science, 57: 263--277. doi: 10.1111/j.1540-5907.2012.00602.x
For more details, see
?predVals
. -
k-fold cross-validation can now be performed in parallel for adaptive LASSO models, controlled via the
.parallel
argument ofpolywog()
. Parallel computation of bootstrap iterations is now handled via the.parallel
argument ofbootPolywog()
orcontrol.bp()
. -
Adds
model.matrix
andmodel.frame
methods for objects of class"polywog"
-
Polynomial expansions of the design matrix are now handled in C++, and obsolete functions
polym2()
andrawpoly()
have been removed -
New arguments of
polywog()
:lambda
,nlambda
, andlambda.min.ratio
for finer control of the sequence of penalization factor values examinedfoldid
for direct specification of cross-validation folds (only available when fitting via the adaptive LASSO)thresh
andmaxit
for finer control of the convergence criterion, replacing old argumentscad.maxit
-
Dependency changes:
- All dependencies imported instead of attached to the search path, except
miscTools
which must be attached to provide themargEff
generic glmnet
1.9-5 required (for parallel cross-validation)ncvreg
2.4-0 required (for bug fix incv.ncvreg
)iterators
andRcpp
requiredcar
no longer required (but still suggested)matrixStats
andgames
no longer required
- All dependencies imported instead of attached to the search path, except
-
polywog()
now has argumentunpenalized
to exclude some terms from the adaptive LASSO penalty -
bootPolywog()
now has argumentmaxtries
to control failure when a non-collinear bootstrap model matrix cannot be found -
bootPolywog()
now has argumentmin.prop
to ensure a minimum amount of variation in the bootstrapped response variable in binary models -
The
fitted.values
element of"polywog"
objects is now on the response scale instead of the link scale (i.e., transformed to probabilities whenfamily = "binomial"
) -
Fixed bug where the
polywog.fit
element ofcv.polywog()
output would not contain fitted values -
Fixed bug that sometimes caused
predVals()
to fail unexpectedly
-
New function
cv.polywog()
to select both the polynomial degree and the penalization parameter by cross-validation -
New method
margEff.polywog()
to compute observation-wise and average marginal effects from a fitted model -
varNames
element of a"polywog"
object is now a character vector rather than a list (and is generated more safely) -
"polyTerms"
attribute of matrix returned bypolym2()
is now a matrix rather than a data frame -
predict.polywog()
now works correctly whennewdata
is a model frame
- Initial release