Changes in Version 1.5-4-1.0.0
o Synchronised code with upstream strucchange.
o Fix for the 'breaks' option not being fully honoured in breakpoints() if
the option is numeric.
Changes in Version 1.5-3-1.0.4
o Replaced links in citations with DOIs (sync with upstream strucchange)
o Debug symbols are now retained on macOS, increasing the binary size. Use
R CMD INSTALL --strip to decrease the size if you need to keep size low.
Changes in Version 1.5-3-1.0.3
o Fix for deprecation of digits = 0 in latest R version
Changes in Version 1.5-3-1.0.2
o Lowered the required precision for tests of R.sq in summary.breakpoints().
Changes in Version 1.5-3-1.0.1
o Fix for build failures on Solaris and other architectures.
Changes in Version 1.5-3-1.0.0
o Initial release of the C++-optimised version of strucchange, strucchangeRcpp.
The optimisations are enabled by default, use
options(strucchange.use_armadillo = FALSE) to disable them.
o Arguments of mefp.matrix and breakpoints.matrix now use `obj` as the name for
X, to be consistent with the generics. Avoid positional arguments, as they
differ per method.
o breakpoints(): Added the ability to select the optimal number of breakpoints
by LWZ or RSS via the argument `breaks`.
o Added a function magnitude() for getting informaton about breakpoint
magnitude using a variety of statistics.
o Added the R² statistic of the model fit to summary.breakpoints() and
summary.breakpointsfull(), available as an element `r.squared`.
o breakpoints() called on a response with no variance returns no break instead of an error
instead of an error.
o Fixed a bug where efp() with h = 1 would create an invalid time series and
fail with an error.
Changes in Version 1.5-3
o Fixups for R CMD check on different platforms.
Changes in Version 1.5-2
o Added lrvar = FALSE argument to efp() so that optionally a long-run variance
estimator (Andrews or Newey-West) instead of the standard OLS estimator can
be used for the error variance.
o Added vcov. = NULL argument to efp() so that optionally other covariance
matrix estimators can be plugged into RE and ME tests (e.g., sandwich(),
kernHAC(), NeweyWest(), etc.).
o The default recresid() gained an argument qr.tol = 1e-7 which allows to
pass another tolerance to lm.fit() for detecting linear dependencies in
(small) subsamples. Furthermore, an argument engine = c("R", "C") has been
added along with an alternative faster C implementation (by Nikolaus
Umlauf).
o The formula method of breakpoints() now passes ... to recresid(), e.g., for
the qr.tol and engine arguments above.
o The breaks argument of the breakpoints() formula argument is now checked
to be at least 1.
o breakpoints(..., hpc = "foreach") now also works if the "foreach" package
is not attached.
o Improved support for formulas like 'y ~ .' in efp(), Fstats(), and
breakpoints(). (Suggested by Matthieu Stigler.)
o Bug fix in gefp(..., decorrelate = FALSE). Scaling is done with the square
root of the diagnoal of the variance - as opposed to the diagonal of the
square root of the variance. (Reported by Dries Debeer.)
Changes in Version 1.5-1
o ordL2BB() now uses a direct simulation method based on mvtnorm::rmvnorm()
which is much faster, making the computation of p values and critical values
for the ordinal maxLM statistic much faster and feasible "on the fly".
o Reduced number of significant digits in breakpoints summary to
getOption("digits") - 3.
o Reference output updated for recent versions of R.
Changes in Version 1.5-0
o Added new efpFunctional generators for conducting various types of
structural change tests based on empirical fluctuation processes of class
"gefp". In particular a (maximum) MOSUM functional was added as well
as several functionals suitable for aggregation along categorical
variables. The documentation for previously available functionals
such as supLM() was also enhanced.
o The new functionals mentioned above for assessing parameter instability
along (ordered) categorical variables are catL2BB (unordered),
ordL2BB and ordwmax (ordered). These are discussed in more detail in
Merkle, Fan, and Zeileis (2013, Psychometrika).
o Added a new default method for sctest(). This essentially just calls
gefp(object, fit = NULL) and then (optionally) calls plot() and sctest()
using the specified functional. However, several convenience options have
been added, e.g., using the maximum likelihood information (rather than
the outer product of gradients) for the covariance matrix or specifying
the functional via a character string.
o Documentation of the sctest() generic and its methods have been enhanced.
Methods for "formula", "efp", and "Fstats" are suitable for assessing
structural changes in linear regression models while the "default" and
"gefp" methods (see above) are suitable for general parametric models.
o Improved plot() method for gefp/efpFunctional to allow for more
flexibility in boundary display. Rather than only boundary = TRUE) or
FALSE one can now specify a list of graphical parameters, e.g.,
boundary = list(col = "slategray", lty = 2).
o Updated Depends/Imports in DESCRIPTION/NAMESPACE with new R CMD check
requirements.
Changes in Version 1.4-7
o plotProcess() function in "efpFunctional" objects now takes a
boundary = TRUE argument by default which can be set to FALSE to
suppress plotting of the boundary function.
o Added a check (and a more intelligible warning) in the "formula"
method of breakpoints() whether the 'breaks' argument supplied by
the user is too large.
Changes in Version 1.4-6
o Default recresid() can now also deal with regressors that do not
vary across (small) subsamples.
Changes in Version 1.4-5
o Further improvements in new recresid() default method.
Now also works correctly if some coefficients are not identified
on the initial subsamples in the recursion.
o Resaved datasets to reduce storage requirements.
o Fixed bug in breakpoints() for time series that contain NAs.
Changes in Version 1.4-4
o Default recresid() method now tries to choose adaptively between
using the faster updating formula and the slower full
QR decomposition to yield numerically more stable results.
In previous versions of the function the QR decomposition was
used only in the first iteration.
o Improvement in breakdates() computations.
Changes in Version 1.4-3
o Speed-up in breakpoints() for the intercept-only case,
i.e., breakpoints(y ~ 1).
Changes in Version 1.4-2
o Improved time index computations in gefp().
o Added replication notes in ?durab.
Changes in Version 1.4-1
o efp(), Fstats(), and breakpoints() are now more cautios about using
time series properties from the data and try to check whether any
NAs were removed. In general, the functions will yield best results
if all NA processing is done before calling them.
o Better handling of time series properties for the boundaries in
the examples of SP2001.
Changes in Version 1.4-0
o added optional high performance computing support by means of the
"foreach" package for the breakpoints() formula method. This can
be leveraged to alleviate the computational burden in the dynamic
programming approach. Simply register a parallel backend (e.g.,
by means of "doMC" or "doSNOW") and call breakpoints() with
additional argument hpc = "foreach".
Changes in Version 1.3-7
o added optional start end end arguments to recresid().
Changes in Version 1.3-6
o enhanced documentation for new Rd parser.
Changes in Version 1.3-5
o added some further references to the vignette,
and provide the associated .bib file in ~/inst/doc/.
o removed \itemize in .Rd files for new R-devel
Changes in Version 1.3-4
o fixed CITATION encoding
o removed Z.sty dependency in vignette
Changes in Version 1.3-3
o enhanced references in the vignette, CITATION and
man pages
o fixed some outdated information in the vignette
Changes in Version 1.3-2
o added new data set with bibliographic information about
structural change publications
Changes in Version 1.3-1
o renamed SP500 to SP2001 to avoid conflicts with MASS
Changes in Version 1.3-0
o added NAMESPACE
o improved dependency declaration in DESCRIPTION