This maintenance release adjusts a test of formatDate()
to be compatible with recent changes in R-devel.
residuals.hhh4()
can now also compute type = "pearson"
residuals.
The standard frequency()
and start()
generics can now be used to
extract the respective slots from an "sts"
object.
When indexing a spatial "sts"
object that contains a map by region names,
i.e., <sts>[,<character>]
, the additional argument drop = TRUE
can
now be used to subset the map accordingly.
algo.twins()
implementation of
Held et al. (2006)
has been removed from the package. The source code has been migrated
to a separate R package, twins, which is archived at
https://codeberg.org/EE-hub/twins.update.hhh4(S =)
can now be used for a previously disabled model
component, when it simply uses f = ~1
and adds the requested number
of harmonics (but gives a warning as the specification may need to be
tweaked, e.g., for an offset).The sts()
constructor now also accepts an "sf"
object as map
input;
it is internally converted to "SpatialPolygons"
as required by the
"sts"
class. (Based on a patch by Sophie Reichert.)
as.epidataCS()
is faster in determining potential event sources.
backprojNP(eq3a.method="C")
and as.epidataCS()
) were using it
(inefficiently) and have been replaced by C implementations.
This also reduces the size of the installed package.update.epidataCS()
gained an argument stgrid
to update the
spatio-temporal grid data in an existing "epidataCS"
object.
This enables updates/transformations of endemic variables and/or changes
of the time intervals without needing to do as.epidataCS()
from scratch.Start values for endemic intercepts in twinstim()
are now robust
against non-finite values in offset terms.
intensityplot.twinstim(aggregate = "space")
no longer fails for
endemic-only fits.
pit()
plot could lack some tick marks on the y-axis (for R >= 4.2.0).unionSpatialPolygons()
and polyAtBorder()
now use
sf in place of rgeos.Long unused methods for "gpc.poly"
objects (from package gpclib)
have now been removed and surveillance.options("gpclib")
is obsolete.
Package rgeos is no longer available as a clipper
method for
as.epidataCS()
. The previous default polyclip remains as the only
option.
This is a maintenance release, fixing encoding-related portability issues and increasing test coverage for rarely used functionality.
nbOrder()
has been re-implemented: it is now more efficient and no
longer depends on spdep. Furthermore, it now defaults to
maxlag = Inf
; the historical default maxlag = 1
was barely useful.
It no longer messages (about the range of the detected orders).
Printing "sts"
objects with a map now shows the first row of the
attached data (if present) instead of the object summary.
surveillance no longer relies on the maptools package:
unionSpatialPolygons()
with method = "gpclib"
is deprecated
and now uses the default method with a warning.
Long unused scale.gpc.poly()
and inside.gpc.poly()
are deprecated;
the unused and undocumented diameter.gpc.poly()
method has been removed.
stsplot_spacetime()
is formally deprecated; it has long been
superseded by stsplot_space()
and an animate()
method for "sts"
objects.
vignette("monitoringCounts")
now uses knitr as its engine
to work around Bug 18318.plotHHH4_fitted()
can now produce simple (unformatted) time indexes
if argument xaxis = NA
.Various documentation improvements,
including an example for predict.hhh4()
.
intensityplot.twinstim()
no longer depends on package
maptools.
hhh4()
now warns about interaction terms in model formulae.
These are not implemented and were silently ignored previously.
Fixed a memory leak in algo.twins()
; note that this old experimental
MCMC implementation for a two-component epidemic model may be removed in
future versions of the package.
ks.plot.unif()
: accommodate to NO_S_TYPEDEFS
in R >= 4.3.0.
boda()
with samplingMethod="marginals"
gave all-NA
upperbounds
in INLA >= 21.07.10. boda()
now also works around a scoping issue
(with E
) in recent versions of INLA that led to wrongly scaled
upperbounds.
plotHHH4_season()
gained a period
argument to support harmonics
with periods longer than the frequency of the "sts"
object.
stsplot_space()
now supports passing a col
argument to sp::spplot()
to change the colour of the polygon lines.
plotHHH4_fitted()
can now handle time series with missing values.
If the Nelder-Mead optimizer is used for the variance parameters in
hhh4()
, it is now limited to 500 (not 300) iterations by default
(consistent with the default in optim()
).
Printing an "sts"
object now omits the neighbourhood
component if
that was not set (all-NA
prototype).
simulate.hhh4(..., simplify = TRUE)
now consistently returns a
3d array (nTime x nUnit x nsim), even for nsim = 1
(for which plotting
now works).
The default legend in stsplot_time1()
now only includes plotted elements.
wrap.algo()
no longer prints progress when there is only one area.
summary.hhh4()
now prints the number of excluded observations (due to
missingness), if any.
The print
-method for summary.hhh4()
did not apply the digits
argument to the coefficient matrix. Furthermore, printing of estimated
variance parameters now adheres to significant digits
as documented.
The [
-method for the "hhh4sims"
class was not registered and thus
only available internally. Array-like subsetting of simulated counts
now retains the class.
farringtonFlexible()
with activated populationOffset
(non-default)
always used the population data of the first time series in the fitting step
while iterating over a multivariate "sts"
object.
plotHHH4_ri(..., exp = TRUE)
failed to use a log-scale color axis if
further colorkey
options were passed in a list. The (default) color breaks
could fail to span the range of the data without warning (resulting in
unfilled polygons). This is now checked and the default breaks are now
equally spaced on the log-scale.
stsplot_time1()
did not pass lty
to polygon()
and lwd
to legend()
.
rps()
was wrong for distributions close to a point mass at zero,
e.g., for mu = 1e-3
and x >= 4
. It is now also protected
against wide (quasi-continuous) NegBin distributions that would consume
too much memory with discrete RPS calculation (returning a missing
value with a warning). [both issues spotted by F. Rousseu]
Plots of legacy "disProg"
and "survRes"
objects are now generated
via internal disProg2sts()
conversion and stsplot_time()
.
This fixes their x-axis labels for the default xaxis.years=TRUE
.
The obsolete arguments startyear
and firstweek
are now ignored with
a warning.
The default legend of stsplot_time1()
did not show the fill color
in the non-default case !is.na(col[1])
.
Multivariate hhh4()
with neighbourhood component treated
NA
counts as zero when calculating the weighted sum over units.
A missing count at t-1 in any unit now gives NA
values for the
neighbourhood terms of all units at time t, thus reducing nobs()
.
create.disProg()
is deprecated. Methods for legacy "disProg"
objects
are kept for backwards compatibility, but new projects should use sts()
.
The long-deprecated qlomax()
implementation has been removed.
The CRS
of data(imdepi)
and
data(measlesWeserEms)
have been updated via
sp::rebuild_CRS()
to avoid warnings when rgdal
is loaded with new PROJ and GDAL libraries.
simEpidataCS()
now internally resets the CRS
(temporary), which avoids spurious warnings and also reduces its
runtime by about 25%.
Fix encoding error in vignette("twinstim")
for CRAN's
non-UTF8 Linux test machine.
This version of surveillance (formally) requires the new spatstat umbrella package to avoid collisions of old spatstat and its new sub-packages (we only use spatstat.geom). The spatstat dependence will be dropped in the future.
The epoch<-
replacement method for "sts"
objects now accepts
a "Date"
vector. The standard plots may give nicer x-axis annotation
if indexed by dates. See the xaxis.*
arguments of stsplot_time()
.
tidy.sts()
(and thus autoplot.sts()
) failed for date-indexed
"sts"
objects with non-standard frequencies. [spotted by Junyi Lu]
The nowcast()
function with
method="bayes.trunc.ddcp"
now adds support for negative
binomial response distribution instead of Poisson. Furthermore,
additional components of the design matrix for the discrete time
survival model can be provided, which allows the inclusion of, e.g.,
day of the week effects. Finally, the order of
the polynomial created by the change-points in the discrete time
survival model can now be specified. For further details see the
work of Guenther et al. (2020) about nowcasting the Covid-19
outbreak in Bavaria, Germany.
animate.sts()
can position the timeplot
on
other sides of the map.
The weighted sum in the ne
ighbourhood component of
hhh4()
models is computed more efficiently.
simEpidataCS()
(and thus simulate.twinstim()
)
uses a slightly more efficient location sampler for models with
siaf = siaf.constant()
.
Simulation results will differ from previous package versions even
if the same random seed
is used.
The default main
title for stsplot_space()
now
uses the ISO year-week format for weekly "sts"
data.
Bug fix in the farringtonFlexible()
-function, which for
the argument thresholdMethod=="nbPlugin"
and
thresholdMethod=="muan"
unfortunately computed the limit as
an (1-alpha/2)
prediction interval instead of the documented
(1-alpha)
prediction interval. This affects four threshold
values in Table 2 of vignette("monitoringCounts")
.
The default method "delta"
worked as expected.
In hhh4()
models without AR component,
the matrix of fitted values could lack column names.
Experimental time-varying neighbourhood weights in
hhh4()
were indexed differently in model fitting and in the
simulate()
method (undocumented behaviour). Both now use the
latter variant, where the mean at time t uses products of
weights at time t and observed counts at time t-1.
[reported by Johannes Bracher]
For weekly sts
indexed via start
and freq=52
,
epoch(sts, as.Date=TRUE)
now interprets the start
week
according to ISO 8601. For example, start = c(2020, 5)
corresponds to 2020-01-27, not 2020-02-03.
This affects as.xts.sts()
and the time plot in
animate.sts()
.
stsplot_space()
automatically extends manual color
breaks (at
), if the intervals do not cover the data range.
simEndemicEvents()
and thus
epitest(..., method="simulate")
are no longer slowed down by
intermediate sp::CRS()
computations.
Removed unused rmapshaper from "Suggests" and moved
xts to "Enhances" (used only for as.xts.sts
).
Switched testing framework from (nowadays heavy)
testthat to tinytest. Together with moving
ggplot2 to "Enhances" (used only for autoplot.sts
) ---
and only then --- this switch further reduces the total number of
required packages for a complete check (i.e., installing with
dependencies = TRUE
) in a factory-fresh R environment
from 119 to 94.
spatstat was split into several sub-packages, of which
we only need to import spatstat.geom. This new package
requires R >= 3.5.0
, though.
surveillance now requires R >= 3.6.0
.
New spatial interaction function for twinstim()
:
siaf.exponential()
implements the exponential kernel
f(x) = exp(-x/σ), which is a useful alternative
if the two-parameter power-law kernel is not identifiable.
The plot
-type "maps"
for "hhh4"
fits,
plotHHH4_maps()
, now allows for map-specific color keys
via zmax = NA
(useful for prop = TRUE
).
The nowcast()
-function now also works for
method="bayes.trunc.ddcp"
method when the number of breakpoints
is greater than 1.
The amplitudeShift
transformation for sine-cosine
coefficient pairs in the summary
of multivariate
"hhh4"
models was incorrect in the rare case that the model
used unit-specific seasonal terms (addSeason2formula
with
length(S) > 1
).
algo.hhh()
implementation of the HHH
model has been removed from the package. The function hhh4()
provides an improved and much extended implementation since 2012.The head()
-method for "epidataCS"
objects did
not work with a negative n
argument.
Fix for "matrix"
changes in R-devel.
sts()
now checks for
mismatches in column names of supplied matrices (observed
,
population
, neighbourhood
, ...). This is to catch
input where the units (columns) are ordered differently in
different slots, which would flaw subsequent analyses.simulate.twinSIR()
ignored the atRiskY
indicator
of the underlying "epidata"
, so always assumed a completely
susceptible population. Initially infectious individuals are now
inherited. For the previous behaviour, adjust the supplied
data
via data$atRiskY <- 1
.siaf.powerlaw1()
with fixed sigma = 1
. Useful if sigma
is difficult to
estimate with siaf.powerlaw()
.pit()
's default ylab
was wrong (default are
densities not relative frequencies).
R0()
for "twinstim"
fits with specified
newevents
now handles levels of epidemic factor variables
automatically via the new xlevels
attribute stored in the
fitted model.
Some S3 methods for the "sts"
class are now formally
registered and identical to the established S4 methods.
Minor additions and fixes in the package documentation.
hcl.colors()
, exported since 1.14.0, has been renamed
.hcl.colors()
and is now internal again, to avoid a name
clash with R's own such function introduced in R 3.6.0.W_powerlaw(..., from0 = TRUE)
enables more parsimonious
hhh4
models in that the power-law weights are modified to
include the autoregressive (0-distance) case (see
vignette("hhh4_spacetime")
). The unstructured distance
weights W_np()
gained from0
support as well.
sts()
creation can now handle epoch
arguments of
class Date
directly.
The ranef()
-method for "hhh4"
fits gained a
logical argument intercept
to extract the unit-specific
intercepts of the log-linear predictors instead of the default
zero-mean deviations around the fixed intercepts.
The corresponding plot
method (type="ri"
) gained an
argument exp
: if set to TRUE
random effects are
exp
-transformed and thus show multiplicative effects.
[based on feedback by Tim Pollington]
W_np()
's argument to0
has been renamed to
truncate
. The old name still works but is deprecated.
plotHHH4_ri()
now uses cm.colors(100)
as
col.regions
, and 0-centered color breaks by default.
The help pages of twinSIR()
and related functions now
give examples based on data("hagelloch")
instead of using the
toy dataset data("fooepidata")
. The latter is now obsolete and
will be removed in future versions of the package.
The elements of the control
list stored in the result
of algo.farrington()
are now consistently ordered as in the
default control
argument.
Using negative indices to exclude time points from an
"sts"
object (e.g., x[-1,]
) is now supported and
equivalent to the corresponding subset expression of retained
indexes (x[2:nrow(x),]
) in resetting the start
and
epoch
slots. [reported by Johannes Bracher]
For weekly "sts"
data with epochAsDate=TRUE
,
the as.data.frame()
method computed freq
by
"%Y"
-year instead of by "%G"
-year, which was
inconsistent with the epochInPeriod
variable.
For non-weekly "sts"
data with epochAsDate=TRUE
,
year()
as well as the year
column of the
tidy.sts()
output corresponded to the ISO week-based year.
It now gives the calendar year.
sts_creation()
hard-coded start = c(2006, 1)
.
aggregate()
ing an "sts"
object over time now
recomputes fractions from the cumulated population values if and
only if this is no multinomialTS
and already contains
population fractions. The same rule holds when subsetting units of
an "sts"
object. The aggregate
-method previously
failed to recompute fractions in some cases.
For farringtonFlexible()
with multivariate time series,
only the last unit had stored the additional control items
(exceedence scores, p-values, ...), all others were 0.
[reported by Johannes Bracher]
The supplementary p-values returned by farringtonFlexible()
in control$pvalue
were wrong for the default approach,
where thresholdMethod="delta"
(the original Farrington method)
and a power transformation was applied to the data
(powertrans != "none"
). Similarly, algo.farrington()
returned wrong predictive probabilities in control$pd[,1]
if
a power transformation was used. [reported by Lore Merdrignac]
The control
argument list of algo.farrington()
as stated in the formal function definition was incomplete
(plot
was missing) and partially out of sync with the default
values that were actually set inside the function (b=5
and
alpha=0.05
). This has been fixed. Results of
algo.farrington()
would only be affected if the function was
called without any control
options (which is hardly
possible). So this can be regarded as a documentation error.
The formal control
list of the farrington()
wrapper function has been adjusted accordingly.
The control
argument lists of farringtonFlexible()
and bodaDelay()
as stated in the formal function definitions
were partially out of sync with respect to the following default
values that were actually set inside these functions: b=5
(not 3), alpha=0.05
(not 0.01), pastWeeksNotIncluded=w
(not 26), and, for bodaDelay()
only, delay=FALSE
(not
TRUE
). This has been fixed. Results would only be affected if
the functions were called without any control
options (which
is hardly possible). So this can be regarded as a documentation error.
pairedbinCUSUM()
did not properly subset the sts
object if a range
was specified, and forgot to store the
control
arguments in the result.
wrap.algo()
now aborts if the monitored range is
not supplied as a numeric vector.
In vignette("monitoringCounts")
: several
inconsistencies between code and output have been fixed.
epidataCS2sts()
no longer transfers the
stgrid$BLOCK
indices to the epoch
slot of the
resulting "sts"
object (to avoid epoch[1] != 1
scenarios).
The ranef()
matrix extracted from fitted "hhh4"
models could have wrong column names.
compMatrix.writeTable()
,
makePlot()
, test()
, testSim()
,
readData()
(the raw txt files have been removed as well),
correct53to52()
, enlargeData()
, toFileDisProg()
.autoplot.sts()
gained a width
argument to adjust
the bar width, which now defaults to 7 for weekly time series
(previously was 90% of that so there were gaps between the bars).
"epidataCS"
generation now (again) employs
spatstat's bdist.points()
, which has been
accelerated in version 1.56-0. If you use the
twinstim()
-related modelling part of surveillance, you
are thus advised to update your spatstat installation.
The boda()
examples in
vignette("monitoringCounts")
have been updated to also work
with recent versions of INLA.
Offsets in hhh4
's epidemic components were ignored by
simulate.hhh4()
[spotted by Johannes Bracher] as well as
in dominant eigenvalues ("maxEV").
The color key in fanplot()
is no longer distorted by
log="y"
.
autoplot.sts()
now sets the calling environment as
the plot_env
of the result.
Several twinstim
-related functions finally allow for
prehistory events (long supported by twinstim()
itself):
as.epidataCS()
, glm_epidataCS()
,
as.epidata.epidataCS()
.
The summary()
for SI[R]S-type "epidata"
failed
if there were initially infectious individuals.
qlomax()
,
readData()
, toFileDisProg()
, correct53to52()
,
enlargeData()
, compMatrix.writeTable()
,
test()
, testSim()
, makePlot()
.The as.data.frame()
method for "sts"
objects
gained a tidy
argument, which enables conversion to the
long data format and is also available as function tidy.sts()
.
A ggplot2 variant of stsplot_time()
is now
available via autoplot.sts()
.
as.epidata.data.frame()
gained an argument
max.time
to specify the end of the observation period (which
by default coincides with the last observed event).
The now exported function fanplot()
wraps
fanplot::fan()
. It is used by
plot.oneStepAhead()
and plot.hhh4sims()
, which now
have an option to add the point forecasts to the fan as well.
plotHHH4_fitted()
(and plotHHH4_fitted1()
)
gained an option total
to sum the fitted components over all
units.
Package polyCub is no longer automatically attached (only imported).
scores.oneStepAhead()
no longer reverses the ordering
of the time points by default, as announced in 1.15.0.
Some code in vignette("monitoringCounts")
has been
adjusted to work with the new version of MGLM (0.0.9).
Added a [
-method for the "hhh4sims"
class to
retain the attributes when subsetting simulations.
aggregate(stsObj, by = "unit")
no longer results in
empty colnames (set to "overall"
).
The obsolete map is dropped.
The subset
argument of twinSIR()
was partially
ignored:
If nIntervals = 1
, the model summary()
reported the total number of events.
Automatic knots
, model residuals()
, as well as
the rug in intensityplot()
were computed from the whole set
of event times.
The as.epidata.data.frame()
converter did not actually
allow for latent periods (via tE.col
). This is now possible
but considered experimental (methods for "epidata"
currently
ignore latent periods).
The all.equal()
methods for "hhh4"
and
"twinstim"
objects now first check for the correct classes.
siaf.gaussian()
now also employs a polyCub.iso()
integration routine by default (similar to the powerlaw-type
kernels), instead of adaptive midpoint cubature.
This increases precision and considerably accelerates estimation of
twinstim()
models with a Gaussian spatial interaction
function. Models fitted with the new default
(F.adaptive=FALSE, F.method="iso"
)
will likely differ from previous fits (F.adaptive=TRUE
),
and the numerical difference depends on
the adaptive bandwidth used before (the default adapt=0.1
yielded a rather rough approximation of the integral).
Added quantile()
, confint()
, and plot()
methods for "oneStepAhead"
predictions.
Exported the function simEndemicEvents()
to simulate a
spatio-temporal point pattern from an endemic-only
"twinstim"
; faster than via the general
simulate.twinstim()
method.
twinstim(..., siaf = siaf.gaussian())
uses a larger default initial value for the kernel's standard
deviation (based on the size of the observation region).
Non-default parametrizations of siaf.gaussian()
are
deprecated, i.e., always use logsd=TRUE
and density=FALSE
.
twinstim()
uses a smaller default initial value for the
epidemic intercept, which usually allows for faster convergence.
update.hhh4()
now allows subset.upper
values
beyond the originally fitted time range (but still within the time
range of the underlying "sts"
object).
scores.oneStepAhead()
by default reverses the ordering
of the time points. This awkward behaviour will change in the next
version, so the method now warns if the default reverse=TRUE
is used without explicit specification.
Minor improvements in the documentation and some vignettes: corrected typos, simplified example code, documented some methods.
The C-routines introduced in version 1.14.0 used ==
comparisons on parameter values to choose among case-specific
formulae (e.g., for d==2 in siaf.powerlaw()
).
We now employ an absolute tolerance of 1e-7 (which should fix
the failing tests on Solaris).
Interaction functions for twinstim()
, such as
siaf.powerlaw()
or tiaf.exponential()
, no longer live in
the global environment as this risks using masked base functions.
demo("v77i11")
.
It exemplifies the spatio-temporal endemic-epidemic modelling
frameworks twinstim
, twinSIR
, and hhh4
(see also the corresponding vignettes).Pure C-implementations of integration routines for spatial
interaction functions considerably accelerate the estimation of
twinstim()
models containing
siaf.powerlaw()
, siaf.powerlawL()
, or siaf.student()
.
The color palette generating function used by sts
plots, hcl.colors
, is now exported.
The utility function clapply
(conditional
lapply
) is now exported.
Some utility functions for hhh4
fits are now exported
(update.hhh4
, getNEweights
, coefW
),
as well as several internal functions for use by hhh4
add-on
packages (meanHHH
, sizeHHH
, decompose.hhh4
).
The "fan"
-type plot function for "hhh4sims"
gained a key.args
argument for an automatic color key.
New auxiliary function makeControl()
, which may be
used to specify a hhh4()
model.
twinstim()
now throws an informative error message when
trying to fit a purely epidemic model to data containing endemic
events (i.e., events without ancestors). The help("twinstim")
exemplifies such a model.siaf.powerlaw()$deriv
returned NaN
for the
partial derivative wrt the decay parameter d, if d
was large enough for f to be numerically equal to 0.
It will now return 0 in this case.
twinstim()
could fail (with an error from
duplicated.default
) if the fitted time range was
substantially reduced via the T
argument.
The "simEpidataCSlist"
generated by
simulate.twinstim(..., simplify = TRUE)
was missing the
elements bbox
and control.siaf
.
vignette("twinstim")
, vignette("twinSIR")
, and
vignette("hhh4_spacetime")
, respectively.The calibrationTest()
and pit()
methods for
"oneStepAhead"
forecasts gained an argument units
to allow for unit-specific assessments.
A default scores
-method is now available to compute a
set of proper scoring rules for Poisson or NegBin predictions.
New plot type = "fan"
for simulations from
"hhh4"
models to produce a fan chart using the
fanplot package.
scores.hhh4()
sets rownames for consistency with
scores.oneStepAhead()
."Lambda.const"
matrix returned by
getMaxEV_season()
was wrong for models with asymmetric
neighbourhood weights. [spotted by Johannes Bracher]"maxEV"
) were not affected by this bug.earsC
now has two new arguments thanks to Howard
Burkom: the number of past time units to be used in calculation is
now not always 7, it can be chosen in the baseline
parameter.
Furthermore, the minSigma
parameter allows to get a threshold
in the case of sparse data. When one doesn't give any value for those
two parameters, the algorithm works like it used to.
animate.sts()
gained support for date labels in the
bottom timeplot
.
stsplot_space()
and animate.sts()
can now
generate incidence maps based on the population information
stored in the supplied "sts"
object.
Furthermore, animate.sts()
now supports time-varying
population numbers.
hhh4()
guards against the misuse of
family = factor("Poisson")
for univariate time series.
Previously, this resulted in a negative binomial model by
definition, but is now interpreted as family = "Poisson"
(with a warning).animate.sts()
now supports objects with missing values
(with a warning). Furthermore, the automatic color breaks have been
improved for incidence maps, also in stsplot_space()
.
The as.data.frame
-method for the "sts"
class,
applied to classical time-index-based "sts"
objects
(epochAsDate=FALSE
), ignored a start
epoch different
from 1 when computing the epochInPeriod
indexes.
Furthermore, the returned epochInPeriod
now is a fraction of
freq
, for consistency with the result for objects with
epochAsDate=TRUE
.
simulate.hhh4()
did not handle shared overdispersion
parameters correctly. The different parameters were simply recycled
to the number of units, ignoring the factor specification from the
model's family
. [spotted by Johannes Bracher]
Simulations from endemic-only "hhh4"
models
with unit-specific overdispersion parameters used wrong
variances. [spotted by Johannes Bracher]
oneStepAhead()
predictions of type
"rolling"
(or "first"
) were incorrect for time points
tp
(tp[1]
) beyond the originally fitted time range
(in that they were based on the original time range only).
This usage of oneStepAhead()
was never really supported and
is now caught when checking the tp
argument.
plot.hhh4simslist()
ignored its par.settings
argument if groups=NULL
(default).
The internal auxiliary function, which determines the sets of
potential source events in "epidataCS"
has been implemented
in C++, which accelerates as.epidataCS()
,
permute.epidataCS()
, and therefore epitest()
.
This is only really relevant for "epidataCS"
with a large
number of events (>1000, say).
Negative-binomial hhh4()
models may not converge for
non-overdispersed data (try, e.g.,
set.seed(1); hhh4(sts(rpois(104, 10)), list(family="NegBin1"))
).
The resulting non-convergence warning message now mentions low
overdispersion if this is detected. [suggested by Johannes Bracher]
An additional type="delay"
option was added to the
plot
method of stsNC
objects. Furthermore, an
animate_nowcasts
function allows one to animate a sequence of
nowcasts.
animate
-method for "sts"
objects,
the default top padding of lattice plots is now disabled for the
bottom timeplot
to reduce the space between the panels.
Furthermore, the new option fill
can be used to make the
panel of the timeplot
as large as possible.bodaDelay()
: fixed spurious warnings from rnbinom()
.
vignette("monitoringCounts")
: fixed boda
-related
code and cache to obtain same results as in corresponding JSS paper.
vignette("monitoringCounts")
illustrates the
monitoring of count time series in R with a particular focus on
aberration detection in public health surveillance.
This vignette corresponds to a recently accepted manuscript
for the Journal of Statistical Software
(Salmon, Schumacher, and Höhle, 2016).Non-convergent hhh4()
fits now obey the structure of
standard "hhh4"
objects. In particular, such fits now also
contain the control
and stsObj
elements, allowing
for model update()
s of non-convergent fits.
knox()
warns about symmetric input matrices.
The code of boda()
(with samplingMethod="joint"
)
and bodaDelay()
(with inferenceMethod="INLA"
)
has been adjusted to a change of arguments of INLA's
inla.posterior.sample
function. Accordingly, the minimum
INLA version required to run boda()
and
bodaDelay()
is 0.0-1458166556.
The functions returned by W_powerlaw()
now have the
package namespace as their environment to support situations where
the package is not attached.
Attaching package nlme after surveillance no
longer masks "hhh4"
's ranef
-method. (We now import the
fixef
and ranef
generics from nlme.)
Several new vignettes illustrate endemic-epidemic modeling frameworks for spatio-temporal surveillance data:
vignette("twinstim")
: describes a spatio-temporal point process regression model.
vignette("twinSIR")
: describes a multivariate temporal point process regression model.
vignette("hhh4_spacetime")
: describes an areal time-series model for infectious disease counts.
These vignettes are based on a recently accepted manuscript for the Journal of Statistical Software (Meyer, Held, and Höhle, 2016).
Improved the documentation on various help pages.
The hhh4()
-based analysis of data("fluBYBW")
has been moved to a separate demo script 'fluBYBW.R'. Due to
the abundance of models and the relatively long runtime, we
recommend to open the script in an editor rather than running
all the code at once using demo("fluBYBW")
.
Overhaul of the "sts"
implementation. This mostly
affects package-internal code, which is simpler, cleaner and better
tested now, but requires R >= 3.2.0 (due to callNextMethod()
bugs in older versions of R).
Beyond that, the user-level constructor function sts()
now
has explicit arguments for clarity and convenience.
For instance, its first argument sets the observed
slot and
no longer needs to be named, i.e.,
sts(mycounts, start=c(2016,3), frequency=12)
works just like for the classical ts()
function.
stsplot_time(..., as.one=TRUE)
is now implemented
(yielding a simple matplot
of multiple time series).
plotHHH4_season()
now by default draws a horizontal
reference line at unity if the multiplicative effect of component
seasonality is shown (i.e., if intercept=FALSE
).
Since surveillance 1.8-0, hhh4()
results are of
class "hhh4"
instead of "ah4"
(renamed).
Legacy methods for the old class name "ah4"
have been removed.
The internal model preparation in twinstim()
is more
efficient (the distance matrix of the events is only computed if
event sources actually need to be updated).
stsplot_spacetime()
now recognizes its opts.col
argument.
Conversion from "ts"
to "sts"
using
as(ts, "sts")
could set a wrong start time. For instance,
as(ts(1:10, start=c(1959,2), frequency=4), "sts")@start
was
c(1959,1)
.
algo.twins()
now also accepts "sts"
input and
the automatic legend in the first plot of plot.atwins()
works
again.
The experimental profile
-method for "twinstim"
objects did not work if embedded twinstim()
fits issued warnings.
update.epidata()
can now handle a distance matrix
D
in the form of a classed "Matrix"
.
[suggested by George Wood]
glrnb()
can now handle ret="cases"
for the
generalized likelihood ratio detector based on the negative binomial
distribution. It's based on a brute-force search and hence might be
slow in some situations.
boda()
and bodaDelay()
now support an
alternative method (quantileMethod="MM"
) to compute quantiles
based on the posterior distribution. The new method samples
parameters from the posterior distribution and then computes the
quantile of the mixture distribution using bisectionning, which is
faster and yields similar results compared to the original method
(quantileMethod="MC"
, still the default).
vignette("hhh4")
, updated the package
description as well as some references in the documentation.
Also updated (the cache of) the slightly outdated
vignette("surveillance")
to account for the corrected version
of algo.bayes()
implemented since surveillance 1.10-0.Fixed bug in categoricalCUSUM()
, which ignored alarms
generated for the last time point in range
. Furthermore, the
exact computation in case of returns of the type "value"
for
the binomial are now checked through an attribute.
Fixed bug in the estimateGLRNbHook
function of
algo.glrnb
, which ignored potential fixed alpha
values. If alpha
is fixed this is now taken into consideration
while fitting the negative binomial function. See revised help files
for the details.
Made a hot-fix such that the algo.quality
function now
also works for sts
objects and if the state
or
alarm
slots consists of TRUE/FALSE instead of 0/1.
intensity.twinstim()
did not work for non-endemic models.
A parallelized epitest()
could fail with a strange
error message if some replications were left unassigned.
This seems to happen if forking is used (mclapply
) with
insufficient memory. Incomplete replications are now ignored with a
warning.
This package is now maintained by Sebastian Meyer, who has been an active co-author since version 1.3. We thank Michael Höhle for 10 years of maintenance ever since he created surveillance and published the package on CRAN in November 2005 for R 2.2.0.
Calibration tests for count data (Wei and Held, 2014, Test)
are now implemented and available as calibrationTest()
.
In addition to a default method taking pure counts and predictive
means and dispersion parameters, there are convenient methods for
"hhh4"
and "oneStepAhead"
objects.
Shared overdispersion across units in negative binomial
hhh4()
time series models (by specifying a factor variable
as the family
argument).
scores()
and pit()
are now generic and have convenient
methods for "oneStepAhead"
predictions and "hhh4"
fits.
The initial values used for model updates during the
oneStepAhead()
procedure can now be specified directly
through the which.start
argument (as an alternative to the
previous options "current"
and "final"
).
plotHHH4_fitted()
(and plotHHH4_fitted1()
)
gained an option decompose
to plot the contributions from
each single unit (and the endemic part) instead of the default
endemic + AR + neighbours decomposition.
Furthermore, a formatted time axis similar to stsplot_time1()
can now be enabled via the new argument xaxis
.
The new plot
type
"maps"
for "hhh4"
fits shows maps of the fitted mean components averaged over time.
New plot
-method for simulations from "hhh4"
models (using simulate.hhh4(..., simplify = TRUE)
, which now
has a dedicated class: "hhh4sims"
) to show the final
size distribution or the simulated time series
(possibly stratified by groups of units).
There is also a new scores
-method to compute proper scoring
rules based on such simulations.
The argument idx2Exp
of coef.hhh4()
may now be
conveniently set to TRUE
to exp-transform all coefficients.
Added a coeflist()
-method for "hhh4"
fits.
The generator function sts()
can now be used to
initialize objects of class "sts"
(instead of writing
new("sts", ...)
).
Additional arguments of layout.scalebar()
now allow to
change the style of the labels.
A pre-computed distance matrix D
can now be used as
input for the as.epidata()
converter -- offering an alternative
to the default Euclidean distance based on the individuals coordinates.
(Request of George Wood to support twinSIR
models on networks.)
The first argument of scores()
is now called x
instead of object
(for consistency with calibrationTest()
).
The result of oneStepAhead()
now has the dedicated
class attribute "oneStepAhead"
(previously was just a list).
Changed interpretation of the col
argument of
plotHHH4_fitted()
and plotHHH4_fitted1()
(moved color
of "observed" to separate argument pt.col
and reversed
remaining colors). The old col
specification as a vector of
length 4 still works (caught internally) but is undocumented.
The epoch
slot of class "sts"
is now initialized to
1:nrow(observed)
by default and thus no longer needs to be
explicitly set when creating a new("sts", ...)
for this
standard case.
Initialization of new("sts", ...)
now supports the
argument frequency
(for consistency with ts()
).
Note that freq
still works (via partial argument matching)
and that the corresponding "sts"
slot is still called freq
.
If missing(legend.opts)
in stsplot_time1()
, the
default legend will only be produced if the "sts"
object
contains information on outbreaks, alarms, or upperbounds.
The default summary()
of a "twinstim"
fit is more
concise since it no longer includes the number of log-likelihood and
score function evaluations and the elapsed time during model fitting.
Set the new runtime
argument of summary.twinstim()
to
TRUE
to add this information to the summary as before.
The animate
-method for "sts"
objects gained an
argument draw
(to disable the default instantaneous plotting)
and now invisibly returns the sequential plot objects (of class
"gtable"
or "trellis"
) in a list for post-processing.
The flexible time axis configurations for "sts"
plots
introduced in version 1.8-0 now also work for classical "sts"
objects with integer epochs and standard frequencies
(try plot(..., epochsAsDate = TRUE)
).
stsplot_time()
initiates par
settings only
if the par.list
argument is a list.
The new all.equal()
method for class "hhh4"
compares two fits ignoring their "runtime"
and "call"
elements (at least).
Fixed a bug in algo.bayes
, where an alarm was already
sounded if the current observation was equal to the quantile of the
predictive posterior. This was changed in order to get alarm_t
= I(obs_t > quantile_t) which is consistent with the use in
boda
and bodaDelay
.
Fixed bug in algo.outbreakP
causing a halt in the
computations of value="cases"
when
calc.outbreakP.statistic
returned NaN
. Now, a
NaN
is returned.
wrap.algo
argument control.hook
used
control
argument defined outside it's scope (and not the one
provided to the function). It is now added as additional 2nd argument
to the control.hook
function.
stsplot_time()
did not account for the optional
units
argument for multivariate "sts"
objects
when choosing a suitable value for par("mfrow")
.
hhh4()
could have used a function dpois()
or
dnbinom()
from the global environment instead of the
respective function from package stats.
The default time variable t
created as part of the
data
argument in hhh4()
was incompatible with
"sts"
objects having epochAsDate=TRUE
.
A consistency check in as.epidata.default()
failed for
SI-type data (and, more generally, for all data which ended with an
I-event in the last time block). [spotted by George Wood]
This is a quick patch release to make the test suite run smoothly on CRAN's Windows and Solaris Sparc systems.
The new hhh4()
option to scale neighbourhood weights
did not work for parametric weights with more than one parameter
if normalize=FALSE
.
New functions and data for Bayesian outbreak detection in the
presence of reporting delays (Salmon et al., 2015):
bodaDelay()
, sts_observation()
, and sts_creation()
.
New functions implementing tests for space-time interaction:
knox()
supports both the Poisson approximation and a
Monte Carlo permutation approach to determine the p-value,
stKtest()
wraps space-time K-function methods from
package splancs for use with "epidataCS"
,
and epitest()
for twinstim
models
(makes use of the new auxiliary function simpleR0()
).
New function plapply()
: a parallel and verbose version
of lapply()
wrapping around both mclapply()
and
parLapply()
of package parallel.
New converter as.xts.sts()
to transform "sts"
objects to the quasi standard "xts"
class, e.g., to make use
of package dygraphs for interactive time series plots.
New options for scaling and normalization of neighbourhood
weights in hhh4()
models.
New auxiliary function layout.scalebar()
for use as part
of sp.layout
in sp::spplot()
or in the traditional
graphics system.
"epidataCS"
New argument by
for plot.epidataCS()
, which
defines a stratifying variable for the events (default is the event
type as before). It can also be set to NULL
to make the plot
not distinguish between event types.
The spatial plot of "epidataCS"
gained the arguments
tiles
, pop
and sp.layout
, and can now produce
an sp::spplot()
with the tile-specific population levels behind
the point pattern.
New function permute.epidataCS()
to randomly permute
time points or locations of the events (holding other marks fixed).
twinstim()
New S3-generic coeflist()
to list model coefficients by
component. It currently has a default method and one for
"twinstim"
and "simEpidataCS"
.
New argument newcoef
for simulate.twinstim()
to
customize the model parameters used for the simulation.
New argument epilink
for twinstim()
, offering
experimental support for an identity link for the epidemic
predictor. The default remains epilink = "log"
.
Simulation from "twinstim"
models and generation of
"epidataCS"
is slightly faster now (faster spatstat
functions are used to determine the distance of events to the border).
New option scaled = "standardized"
in iafplot()
to plot f(x) / f(0) or g(t) / g(0), respectively.
Initial data processing in twinstim()
is faster
since event sources are only re-determined if there is effective
need for an update (due to subsetting or a change of
qmatrix
).
formatPval()
disables scientific
notation by default.
The "time"
plot for "epidataCS"
uses the
temporal grid points as the default histogram breaks
.
The special fe()
function which sets up fixed effects
in hhh4()
models gained an argument unitSpecific
as a
convenient shortcut for which = rep(TRUE, nUnits)
.
The convenient plot
option of permutationTest()
uses MASS::truehist()
instead of hist()
and
accepts graphical parameters to customize the histogram.
The bodaFit
function did not draw samples from the
joint posterior. Instead draws were from the respective posterior
marginals. A new argument samplingMethod
is now introduced
defaulting to the proper 'joint'. For backwards compatibility use
the value 'marginal'.
The functions as.epidataCS()
and simEpidataCS()
could throw inappropriate warnings when checking polygon areas
(only if W
or tiles
, respectively, contained holes).
Non-convergent endemic-only twinstim
models
produced an error. [spotted by Bing Zhang]
The "owin"
-method of intersectPolyCircle
could
have returned a rectangle-type "owin"
instead of a polygon.
An error occurred in twinstim()
if finetune=TRUE
or choosing optim()
instead of the default nlminb()
optimizer without supplying a control
list in optim.args
.
The "time"
plot for "epidataCS"
did not
necessarily use the same histogram breaks
for all strata.
Specifying a step function of interaction via a numeric vector
of knots did not work in twinstim()
.
plot.hhh4()
did not support an unnamed type
argument such as plot(x, "season")
.
simEpidataCS()
did not work if t0
was in the
last block of stgrid
(thus it did not work for single-cell
grids), and mislabeled the start
column copied to
events
if there were no covariates in stgrid
.
Evaluating intensity.twinstim()$hFUN()
at time points
before t0
was an error. The function now returns
NA_real_
as for time points beyond T
.
Truncated, normalized power-law weights for hhh4()
models, i.e., W_powerlaw(maxlag = M, normalize = TRUE)
with M < max(neighbourhood(stsObj))
, had wrong derivatives
and thus failed to converge.
update.hhh4(..., use.estimates = TRUE)
did not
use the estimated weight function parameters as initial values for
the new fit. It does so now iff the weight function
ne$weights
is left unchanged.
Accommodate a new note given by R-devel checks, and set the new INLA additional repository in the 'DESCRIPTION' file.
Made linelist2sts()
work for quarters by adding extra
"%q"
formatting in formatDate()
.
In the coefficient vector resulting from a hhh4
fit,
random intercepts are now named.
Parameter start
values in hhh4()
are now
matched by name but need not be complete in that case (default
initial values are used for unspecified parameters).
The update.hhh4()
-method now by default does
use.estimates
from the previous fit. This reduces the
number of iterations during model fitting but may lead to slightly
different parameter estimates (within a tolerance of 1e-5
).
Setting use.estimates = FALSE
means to re-use the previous
start specification.
For univariate "sts"
objects, the (meaningless)
"head of neighbourhood" is no longer show
n.
The "sts"
class now has a dimnames
-method
instead of a colnames
-method. Furthermore, the redundant
nrow
and ncol
methods have been removed
(the dim
-method is sufficient).
If a map
is provided when initialize()
ing an
"sts"
object, it is now verified that all observed
regions are part of the map
(matched by row.names
).
In stsplot_space()
, extra (unobserved) regions of the
map
are no longer dropped but shown with a dashed border by
default.
The R0
-method for "twinstim"
gained an argument
newcoef
to simplify computation of reproduction numbers with
a different parameter vector (also used for Monte Carlo CI's).
New plot type="neweights"
for "hhh4"
fits.
The scores()
function allows the selection of multiple
units
(by index or name) for which to compute (averaged) proper
scores. Furthermore, one can now select which
scores to compute.
Added a formula
-method for "hhh4"
fits to
extract the f
specifications of the three components from the
control list.
The update()
-method for fitted "hhh4"
models
gained an argument S
for convenient modification of component
seasonality using addSeason2formula()
.
The new auxiliary function layout.labels()
generates an
sp.layout
item for sp::spplot()
in order to draw labels.
When generating the pit()
histogram with a single
predictive CDF pdistr
, the ...
arguments can now be
x
-specific and are recycled if necessary using mapply()
.
If pdistr
is a list of CDFs, pit()
no longer requires
the functions to be vectorized.
New method as.epidata.data.frame()
, which constructs the
start/stop SIR event history format from a simple individual-based
data frame (e.g., hagelloch.df
).
New argument w
in as.epidata.default()
to
generate covariate-based weights for the force of infection in
twinSIR
. The f
argument is for distance-based weights.
The result of profile.twinSIR()
gained a class and an
associated plot
-method.
For multivariate oneStepAhead()
predictions,
scores(..., individual=TRUE)
now returns a 3d array instead
of a collapsed matrix. Furthermore, the scores computed by default
are c("logs","rps","dss","ses")
, excluding the normalized
squared error score "nses"
which is improper.
The plot-type="season"
for "hhh4"
fits now by
default plots the multiplicative effect of seasonality on the
respective component (new argument intercept=FALSE
).
The default set of components to plot has also changed.
When as.epidata()
and simEpidata()
calculate
distance-based epidemic weights from the f
functions, they no
longer set the distance of an infectious individual to itself
artificially to Inf
.
This changes the corresponding columns in the "epidata"
in
rows of currently infectious individuals, but the twinSIR
model itself is invariant, since only rows with atRiskY=1
contribute to the likelihood.
Several modifications and corrections in data("hagelloch")
.
Better plotting of stsNC
objects by writing an own plot
method for them. Prediction intervals are now shown jointly with the
point estimate.
Reduced package size by applying tools::resaveRdaFiles
to some large datasets and by building the package with
--compact-vignettes=both
, i.e., using additional GhostScript
compression with ebook quality, see ?tools::compactPDF
.
Added units
argument to stsplot_time
to select
only a subset of the multivariate time series for plotting.
The untie
-method for class "epidataCS"
gained an
argument verbose
which is now FALSE
by default.
"epidataCS"
objects store the clipper
used
during generation as attribute of $events$.influenceRegion
.
In plotHHH4_fitted()
, the argument legend.observed
now defaults to FALSE
.
The default weights for the spatio-temporal component in
hhh4
models now are neighbourhood(stsObj) == 1
.
The previous default neighbourhood(stsObj)
does not make
sense for the newly supported nbOrder
neighbourhood matrices
(shortest-path distances). The new default makes no difference for
(old) models with binary adjacency matrices in the neighbourhood
slot of the stsObj
.
The default for nonparametric weights W_np()
in
hhh4()
is now to assume zero weight for neighbourhood orders
above maxlag
, i.e., W_np()
's argument to0
now
defaults to TRUE
.
Added a verbose
argument to permutationTest()
,
which defaults to FALSE
. The previous behaviour corresponds
to verbose=TRUE
.
simulate.twinstim()
now by default uses the original
data$W
as observation region.
The data("measlesWeserEms")
contain two additional
variables in the @map@data
slot: "vaccdoc.2004"
and
"vacc1.2004"
.
The plot-method for "epidata"
objects now uses colored
lines by default.
The surveillance package now depends on R >= 3.0.2, which, effectively, is the minimum version required since surveillance 1.7-0.
The two diagnostic plots of checkResidualProcess()
are
now by default plotted side by side (mfrow=c(1,2)
) instead of
one below the other.
In farringtonFlexible
alarms are now for
observed>upperbound
and not for observed>=upperbound
which was not correct.
Fixed duplicate "functions"
element resulting from
update.twinstim(*,model=TRUE)
and ensured that
"twinstim"
objects always have the same components (some may
be NULL
).
animate.epidata
works again with the
animation package (ani.options("outdir")
was
removed in version 2.3)
For hhh4
models with random effects, confint()
only worked if argument parm
was specified.
Computing one-sided AIC weights by simulation for twinSIR
models with more than 2 epidemic covariates now is more robust (by
rescaling the objective function for the quadratic programming
solver) and twice as fast (due to code optimization).
simulate.twinstim(..., rmarks=NULL)
can now handle the
case where data
has no events within the simulation period
(by sampling marks from all of data$events
).
The lambda.h
values of simulated events in
"simEpidataCS"
objects were wrong if the model contained an
endemic intercept (which is usually the case).
Automatic choice of color breaks in the animate
-method
for class "sts"
now also works for incidence maps (i.e., with
a population
argument).
hhh4()
did not allow the use of nonparametric
neighbourhood weights W_np()
with maxlag=2
.
scores()
did not work for multivariate oneStepAhead()
predictions if both individual=TRUE
and sign=TRUE
, and
it could not handle a oneStepAhead()
prediction of only one
time point. Furthermore, the "sign"
column of
scores(..., sign=TRUE)
was wrong (reversed).
For "epidataCS"
with only one event,
epidataCSplot_space()
did not draw the point.
The trivial (identity) call
aggregate(stsObj, nfreq=stsObj@freq)
did not work.
Package surveillance now depends on newer versions of
packages sp (>= 1.0-15), polyCub (>= 0.4-2),
and spatstat (>= 1.36-0).
The R packages INLA and runjags are now suggested
to support a new outbreak detection algorithm (boda()
) and
the new nowcast()
ing procedure, respectively.
The R packages for lattice, grid,
gridExtra, and scales are suggested for
added visualization facilities.
More tests have been implemented to ensure package integrity. We now use testthat instead of the outdated package RUnit.
hhh4()
fits now have class "hhh4"
instead of
"ah4"
, for consistency with twinstim()
,
twinSIR()
, and to follow the common convention (cp. lm()
).
Standard S3-methods for the old "ah4"
name are still
available for backwards compatibility but may be removed in the
future.
Plot variants for "sts"
objects have been cleaned up:
The functions implementing the various plot types
(stsplot_*
, previously named plot.sts.*
)
are now exported and documented separately.
The nowcast
procedure has been completely re-written to
handle the inherit right-truncation of reporting data (best
visualized as a reporting triangle). The new code implements the
generalized-Dirichlet and the hierarchical Bayesian approach described in
Höhle and an der Heiden (2014). No backwards
compatibility to the old nowcasting procedure is given.
The package contains a new monitoring function
boda
. This is a first experimental surveillance
implementation of the Bayesian Outbreak Detection Algorithm (BODA)
proposed in Manitz and Höhle (2012). The function
relies on the non-CRAN package INLA, which has to be installed first
in order to use this function. Expect initial problems.
New toLatex
-method for "sts"
objects.
The new function stsplot_space()
provides an improved
map plot of disease incidence for "sts"
objects aggregated
over time. It corresponds to the new type = observed ~ unit
of the stsplot
-method, and supersedes
type = observed ~ 1|unit
(except for alarm shading).
An animate()
-method for the "sts"
class provides
a new implementation for animated maps (superseding the plot
type=observed ~ 1 | unit * time
) with an optional evolving
time series plot below the map.
The plot()
method for "sts"
objects with epochs as
dates is now made more flexible by introducing the arguments
xaxis.tickFreq
, xaxis.labelFreq
and
xaxis.labelFormat
. These allow the specification of
tick-marks and labelling based on strftime
compatible
conversion codes -- independently if data are daily, weekly, monthly,
etc. As a consequence, the old argument xaxis.years
is
removed. See stsplot_time()
for more information.
Inference for neighbourhood weights in hhh4()
models:
W_powerlaw()
and W_np()
both implement weights
depending on the order of neighbourhood between regions, a power-law
decay and nonparametric weights, i.e., unconstrained estimation of
individual weights for each neighbourhood order.
hhh4()
now allows the inclusion of multiplicative
offsets also in the epidemic components "ar"
and "ne"
.
hhh4()
now has support for lag != 1
in the
autoregressive and neighbor-driven components. The applied lags are
stored as component "lags"
of the return value (previously
there was an unused component "lag"
which was always 1 and
has been removed now).
oneStepAhead()
:
Added support for parallel computation of predictions using
parallel::mclapply()
.
New argument type
with a new type
"first"
to base all subsequent one-step-ahead predictions
on a single initial fit.
Nicer interpretation of verbose
levels, and
txtProgressBar()
.
The plot()
-method for fitted hhh4()
objects now
offers three additional types of plots: component seasonality,
seasonal or time course of the dominant eigenvalue, and maps
of estimated random intercepts. It is documented and more customizable.
Note that argument order and some names have changed:
i
-> units
, title
-> names
.
(Deviance) residuals()
-method for fitted hhh4()
models.
Added methods of vcov()
and nobs()
for the "hhh4"
class. For AIC()
and BIC()
, the
default methods work smoothly now (due to changes to
logLik.hhh4()
documented below).
New predefined interaction functions for twinstim()
:
siaf.student()
implements a t-kernel for the distance
decay, and siaf.step()
and tiaf.step()
provide step
function kernels (which may also be invoked by specifying the
vector of knots as the siaf
or tiaf
argument in
twinstim
).
Numerical integration over polygonal domains in the F
and Deriv
components of siaf.powerlaw()
and
siaf.powerlawL()
is much faster and more accurate now since
we use the new polyCub.iso()
instead of polyCub.SV()
from package polyCub.
New as.stepfun()
-method for "epidataCS"
objects.
plot.epidataCS()
:
The spatial plot has new arguments to automatically add
legends to the plot: legend.types
and legend.counts
.
It also gained an add
argument.
The temporal plot now supports type-specific sub-histograms, additional lines for the cumulative number of events, and an automatic legend.
The new function glm_epidataCS()
can be used to fit
an endemic-only twinstim()
via glm()
.
This is mainly provided for testing purposes since wrapping into
glm
usually takes longer.
Fitted hhh4()
objects no longer contain the associated
"sts"
data twice: it is now only stored as $stsObj
component, the hidden duplicate in $control$data$.sts
was
dropped, which makes fitted objects substantially smaller.
logLik.hhh4()
always returns an object of class
"logLik"
now; for random effects models, its "df"
attribute is NA_real_
. Furthermore, for non-convergent fits,
logLik.hhh4()
gives a warning and returns NA_real_
;
previously, an error was thrown in this case.
oneStepAhead()
:
Default of tp[2]
is now the penultimate time point of
the fitted subset (not of the whole stsObj
).
+1
on rownames of $pred
(now the same as for
$observed
).
The optional "twinstim"
result components
fisherinfo
, tau
, and functions
are always
included. They are set to NULL
if they are not applicable
instead of missing completely (as before), such that all
"twinstim"
objects have the same list structure.
iafplot()
...
invisibly returns a matrix containing the plotted
values of the (scaled) interaction function (and the confidence
interval as an attribute).
Previously, nothing (NULL
) was returned.
detects a type-specific interaction function and by default
uses types=1
if it is not type-specific.
has better default axis ranges.
adapts to the new step function kernels (with new arguments
verticals
and do.points
).
supports logarithmic axes (via new log
argument
passed on to plot.default
).
optionally respects eps.s
and eps.t
,
respectively (by the new argument truncated
).
now uses scaled=TRUE
by default.
The argument colTypes
of
plot.epidataCS(,aggregate="space")
is deprecated (use
points.args$col
instead).
The events in an "epidataCS"
object no longer have
a reserved "ID"
column.
hhh4()
now stores the runtime just like twinstim()
.
Take verbose=FALSE
in hhh4()
more seriously.
hhh4()
issues a warning()
if non-convergent.
The following components of a hhh4()
fit now have names:
"se"
, "cov"
, "Sigma"
.
The new default for pit()
is to produce the plot.
The twinstim()
argument cumCIF
now defaults to
FALSE
.
update.twinstim()
no longer uses recursive
modifyList()
for the control.siaf
argument. Instead,
the supplied new list elements ("F"
, "Deriv"
)
completely replace the respective elements from the original
control.siaf
specification.
siaf.lomax()
is now defunct (it has been deprecated
since version 1.5-2); use siaf.powerlaw()
instead.
Allow the default adapt
ive bandwidth to be specified via the
F.adaptive
argument in siaf.gaussian()
.
Unsupported options (logpars=FALSE
,
effRangeProb
) have been dropped from siaf.powerlaw()
and siaf.powerlawL()
.
More rigorous checking of tiles
in
simulate.twinstim()
and intensityplot.twinstim()
.
as.epidataCS()
gained a verbose
argument.
animate.epidataCS()
now by default does not draw
influence regions (col.influence=NULL
), is verbose
if
interactive()
, and ignores sleep
on non-interactive
devices.
The multiplicity
-generic and its default method have
been integrated into spatstat and are imported from there.
The polygon representation of Germany's districts (
system.file("shapes", "districtsD.RData", package="surveillance")
) has been simplified further. The union of districtsD
is
used as observation window W
in data("imdepi")
. The
exemplary twinstim()
fit data("imdepifit")
has been
updated as well. Furthermore, row.names(imdepi$events)
have
been reset (chronological index), and numerical differences
in imdepi$events$.influenceRegion
are due to changes in
polyclip 1.3-0.
The Campylobacteriosis data set campyDE
, where absolute
humidity is used as concurrent covariate to adjust the outbreak
detection is added to the package to exemplify boda()
.
New data("measlesWeserEms")
(of class "sts"
),
a corrected version of data("measles.weser")
(of the old
"disProg"
class).
Fixed a bug in LRCUSUM.runlength
where computations
were erroneously always done under the in-control parameter
mu0
instead of mu
.
Fixed a bug during alarm plots (stsplot_alarm()
),
where the use of alarm.symbol
was ignored.
Fixed a bug in algo.glrnb
where the overdispersion
parameter alpha
from the automatically fitted glm.nb
model (fitted by estimateGLRNbHook
) was incorrectly taken as
mod[[1]]$theta
instead of 1/mod[[1]]$theta
. The error is
due to a different parametrization of the negative binomial
distribution compared to the parametrization in Höhle
and Paul (2008).
The score function of hhh4()
was wrong when fitting
endemic-only models to a subset
including the first time
point. This led to "false convergence".
twinstim()
did not work without an endemic offset if
is.null(optim.args$par)
.
Package gpclib is no longer necessary for the
construction of "epidataCS"
-objects. Instead, we make use of
the new dedicated package polyclip (licensed under the
BSL) for polygon clipping operations (via
spatstat::intersect.owin()
). This results in a slightly
different $events$.influenceRegion
component of
"epidataCS"
objects, one reason being that
polyclip uses integer arithmetic.
Change of twinstim()
estimates for a newly created
"epidataCS"
compared with the same data prepared in earlier
versions should be very small (e.g., for data("imdepifit")
the mean relative difference of coefficients is 3.7e-08, while the
logLik()
is all.equal()
).
As an alternative, rgeos can still be chosen to do the polygon
operations.
The surveillance-internal code now depends on
R >= 2.15.2 (for nlminb()
NA
fix of PR#15052,
consistent rownames(model.matrix)
of PR#14992,
paste0()
, parallel::mcmapply()
). However, the
required recent version of spatstat (1.34-0, for
polyclip) actually needs R >= 3.0.2, which therefore also
applies to surveillance.
Functions unionSpatialPolygons()
and
intersectPolyCircle()
are now exported. Both are wrappers
around functionality from different packages supporting polygon
operations: for determining the union of all subpolygons of a
"SpatialPolygons"
object, and the intersection of a polygonal
and a circular domain, respectively.
discpoly()
moved back from polyCub
to surveillance.
surveillance now Depends on polyCub (>= 0.4-0)
and not only Imports it (which avoids ::
-references in
.GlobalEnv-made functions).
Nicer default axis labels for iafplot()
.
For twinstim()
, the default is now to trace
every iteration instead of every fifth only.
Slightly changed default arguments for plot.epidata()
:
lwd
(1->2), rug.opts
(col
is set according to
which.rug
)
twinstim()
saves the vector of fixed
coefficients as part of the returned optim.args
component,
such that these will again be held fixed upon update()
.
The plot
-method for hhh4()
-fits allows for
region selection by name.
The polyCub
-methods for cubature over polygonal domains
have been moved to the new dedicated package polyCub,
since they are of a rather general use. The discpoly()
function
has also been moved to that package.
As a replacement for the license-restricted gpclib package,
the rgeos package is now used by default
(surveillance.options(gpclib=FALSE)
) in generating
"epidataCS"
(polygon intersections, slightly slower).
Therefore, when installing surveillance version 1.6-0, the
system requirements for rgeos have to be met, i.e., GEOS
must be available on the system. On Linux variants this means
installing 'libgeos' ('libgeos-dev').
The improved Farrington method described in Noufaily et al. (2012) is now available as function farringtonFlexible()
.
New handling of reference dates in algo.farrington()
for
"sts"
objects with epochAsDate=TRUE
. Instead of always
going back in time to the next Date in the "epoch"
slot, the
function now determines the closest Date. Note that this
might lead to slightly different results for the upperbound
compared to previously. Furthermore, the functionality is only
tested for weekly data (monthly data are experimental). The same
functionality applies to farringtonFlexible()
.
To make the different retrospective modelling frameworks of
the surveillance package jointly applicable, it is now possible
to convert (aggregate) "epidataCS"
(continuous-time continuous-space data) into an "sts"
object
(multivariate time series of counts) by the new function
epidataCS2sts
.
Simulation from hhh4
models has
been re-implemented, which fixes a bug and makes it more flexible
and compatible with a wider class of models.
The map
-slot of the "sts"
class now requires
"SpatialPolygons"
(only) instead of
"SpatialPolygonsDataFrame"
.
Re-implementation of oneStepAhead()
for
hhh4
-models with a bug fix, some speed-up and more options.
Slight speed-up for hhh4()
fits,
e.g., by more use of .rowSums()
and .colSums()
.
Crucial speed-up for twinstim()
fits by more efficient
code: mapply
, dropped clumsy for
-loop in
fisherinfo
, new argument cores
for parallel
computing via forking (not available on Windows).
Using tiaf.exponential()
in a twinstim()
now works
with nTypes=1
for multi-type data.
A legend can be added automatically in iafplot()
.
The untie
methods are now able to produce jittered points
with a required minimum separation (minsep
).
simulate.ah4
gained a simplify
argument.
New update
-method for fitted hhh4
-models
(class "ah4"
).
oneStepAhead()
has more options: specify time range
(not only start), choose type of start values, verbose
argument.
pit()
allows for a list of predictive distributions
(pdistr
), one for each observation x
.
New spatial auxiliary function polyAtBorder()
indicating polygons at the border (for a "SpatialPolygons"
object).
animate.epidataCS()
allows for a main
title and
can show a progress bar.
Changed parametrization of zetaweights()
and completed
its documentation (now no longer marked as experimental).
twinstim(...)$converged
is TRUE
if
the optimization routine converged (as before) but contains
the failure message otherwise.
Increased default maxit
for the Nelder-Mead optimizer
in hhh4
from 50 to 300, and removed default artificial lower
bound (-20) on intercepts of epidemic components.
Renamed returned list from oneStepAhead
(mean->pred, x->observed, params->coefficients,
variances->Sigma.orig) for consistency, and
oneStepAhead()$psi
is only non-NULL
if we have a
NegBin model.
Argument order of pit()
has changed, which is also
faster now and got additional arguments relative
and
plot
.
twinstim(...)$runtime
now contains the complete
information from proc.time()
.
Fixed a bug in function
refvalIdxByDate()
which produced empty reference values
(i.e. NA
s) in case the Date entries of epoch
were not
mondays. Note: The function works by subtracting 1:b
years from the
date of the range value and then takes the span -w:w
around this
value. For each value in this set it is determined whether the
closest date in the epoch slot is obtained by going forward or
backward. Note that this behaviour is now slightly changed compared
to previously, where we always went back in time.
algo.farrington()
: Reference values too far back in time
and hence not being in the "epoch"
slot of the "sts"
object are now ignored (previously the resulting NA
s caused the
function to halt). A warning is displayed in this case.
hhh4
: The entry (5,6) of the marginal Fisher
information matrix in models with random intercepts in all three
components was incorrect.
If nlminb
was used as optimizer for the variance parameters
(using the negative marginal Fisher information as Hessian), this
could have caused false convergence (with warning) or minimally
biased convergence (without warning).
As a consequence, the "Sigma.cov"
component of the
hhh4()
result, which is the inverse of the marginal Fisher
information matrix at the MLE, was also wrong.
untie.matrix()
could have produced jittering greater than
the specified amount
.
hhh4
: if there are no random intercepts, the
redundant updateVariance
steps are no longer evaluated.
update.twinstim()
did not work with
optim.args=..1
(e.g., if updating a list of models with lapply).
Furthermore, if adding the model
component only, the
control.siaf
and optim.args
components were lost.
earsC
should now also work with multivariate
sts
time-series objects.
The last week in data(fluBYBW)
(row index 417) has been
removed. It corresponded to week 1 in year 2009 and was wrong
(an artifact, filled with zero counts only).
Furthermore, the regions in @map
are now ordered the same as
in @observed
.
Fixed start value of the overdispersion parameter in
oneStepAhead
(must be on internal log-scale, not
reparametrized as returned by coef()
by default).
When subsetting "sts"
objects in time, @start
was
updated but not @epoch
.
pit
gave NA
results if any x[-1]==0
.
The returned optim.args$par
vector in twinstim()
was missing any fixed parameters.
hhh4()
did not work with time-varying neighbourhood
weights due to an error in the internal checkWeightsArray()
function.
Fixed obsolete .path.package()
calls.
Small corrections in the documentation.
update.twinstim()
performs better in preserving
the original initial values of the parameters.
New pre-defined spatial interaction function
siaf.powerlawL()
, which implements a _L_agged power-law
kernel, i.e. accounts for uniform short-range dispersal.
New method for outbreak detection: earsC
(CUSUM-method described in the CDC Early Aberration Reporting
System, see Hutwagner et al, 2003).
Yet another p-value formatting function formatPval()
is now also part of the surveillance package.
polyCub.SV()
now also accepts objects of classes
"Polygon"
and "Polygons"
for convenience.
New spatial interaction function siaf.powerlaw()
,
a re-parametrization of the now-deprecated siaf.lomax()
.
The temporal plot
-method for class "epidataCS"
now understands the add
parameter to add the histogram to an
existing plot window, and auto-transforms the t0.Date
argument using as.Date()
if necessary.
nobs()
methods for classes "epidataCS"
and
"twinstim"
.
New argument verbose
for twinstim()
which, if
set to FALSE
, disables the printing of information messages
during execution.
New argument start
for twinstim()
, where (some)
initial parameter values may be provided, which overwrite those in
optim.args$par
, which is no longer required (as a naive
default, a crude estimate for the endemic intercept and zeroes for
the other parameters are used).
Implemented a wrapper stepComponent()
for step()
to perform algorithmic component-specific model selection in
"twinstim"
models. This also required the implementation of
suitable terms()
and extractAIC()
methods. The single-step
methods add1()
and drop1()
are also available.
The update.twinstim()
method now by default uses the
parameter estimates from the previous model as initial values for
the new fit (new argument use.estimates = TRUE
).
as.epidataCS()
checks for consistency of the area of
W
and the (now really obligatory) area column in
stgrid
.
simulate.twinstim()
now by default uses the previous
nCircle2Poly
from the data
argument.
direction
argument for untie.epidataCS()
.
The toLatex
-method for "summary.twinstim"
got
different defaults and a new argument eps.Pvalue
.
New xtable
-method for "summary.twinstim"
for
printing the covariate effects as risk ratios (with CI's and p-values).
New argument hide0s
in the plot
-method for class
"ah4"
.
New argument timevar
for addSeason2formula()
,
which now also works for long formulae.
As requested by the CRAN team, examples now run faster. Some
are conditioned on the value of the new package option
"allExamples"
, which usually defaults to TRUE
(but is
set to FALSE
for CRAN checking, if timings are active).
Moved some rarely used package dependencies to "Suggests:", and also removed some unused packages from there.
Dropped strict dependence on gpclib, which has a
restricted license, for the surveillance package to be clearly
GPL-2. Generation of "epidataCS"
objects, which makes use of
gpclib's polygon intersection capabilities, now requires prior
explicit acceptance of the gpclib license via setting
surveillance.options(gpclib = TRUE)
. Otherwise,
as.epidataCS()
and simEpidataCS()
may not be used.
Speed-up by memoisation of the siaf
cubature (using
the memoise package).
Allow for nlm
-optimizer (really not recommended).
Allow for nlminb
-specific control arguments.
Use of the expected Fisher information matrix can be disabled
for nlminb
optimization.
Use of the effRange
-trick can be disabled in
siaf.gaussian()
and siaf.lomax()
. The default
effRangeProb
argument for the latter has been changed from
0.99 to 0.999.
The twinstim()
argument nCub
has been replaced
by the new control.siaf
argument list. The old
nCub.adaptive
indicator became a feature of the
siaf.gaussian()
generator (named F.adaptive
there) and
does no longer depend on the effRange
specification, but uses
the bandwidth adapt*sd
, where the adapt
parameter may be
specified in the control.siaf
list in the twinstim()
call. Accordingly, the components "nCub"
and
"nCub.adaptive"
have been removed from the result
of twinstim()
, and are replaced by "control.siaf"
.
The "method"
component of the twinstim()
result
has been replaced by the whole "optim.args"
.
The new "Deriv"
component of siaf
specifications
integrates the "siaf$deriv" function over a polygonal domain.
siaf.gaussian()
and siaf.lomax()
use polyCub.SV()
(with intelligent alpha
parameters) for this task
(previously: midpoint-rule with naive bandwidth)
scaled
iafplot()
(default FALSE
). The
ngrid
parameter has been renamed to xgrid
and is more
general.
The "simulate"
component of siaf
's takes an
argument ub
(upperbound for distance from the source).
Numerical integration of spatial interaction functions with an
Fcircle
trick is more precise now; this slightly changes
previous results.
New S3-generic untie()
with a method for the
"epidataCS"
class (to randomly break tied event times and/or
locations).
Renamed N
argument of polyCub.SV()
to
nGQ
, and a
to alpha
, which both have new
default values.
The optional polygon rotation proposed by Sommariva &
Vianello is now also implemented (based on the corresponding MATLAB
code) and available as the new rotation
argument.
The scale.poly()
method for "gpc.poly"
is now
available as scale.gpc.poly()
. The default return class of
discpoly()
was changed from "gpc.poly"
to
"Polygon"
.
An intensityplot()
-method is now also implemented for
"simEpidataCS"
.
Significant speed-up (runs about 6 times faster now, amongst others by many code optimizations and by using sparse Matrix operations).
hhh4()
optimization routines can now be customized for
the updates of regression and variance parameters separately, which
for instance enables the use of Nelder-Mead for the variance
updates, which seems to be more stable/robust as it does
not depend on the inverse Fisher info and is usually faster.
The ranef()
extraction function for "ah4"
objects
gained a useful tomatrix
argument, which re-arranges random
effects in a unit x effect matrix (also transforming CAR effects
appropriately).
Generalized hhh4()
to also capture parametric
neighbourhood weights (like a power-law decay). The new function
nbOrder()
determines the neighbourhood order matrix
from a binary adjacency matrix (depends on package spdep).
New argument check.analyticals
(default FALSE
)
mainly for development purposes.
Fixed sign of observed Fisher information matrix in
twinstim
.
Simulation from the Lomax kernel is now correct (via polar coordinates).
Fixed wrong Fisher information entry for the overdispersion
parameter in hhh4
-models.
Fixed wrong entries in penalized Fisher information wrt the combination fixed effects x CAR intercept.
Fixed indexing bug in penalized Fisher calculation in the case of multiple overdispersion parameters and random intercepts.
Fixed bug in Fisher matrix calculation concerning the relation of unit-specific and random effects (did not work previously).
Improved handling of non-convergent / degenerate solutions during
hhh4
optimization. This involves using ginv()
from
package MASS, if the penalized Fisher info is singular.
Correct labeling of overdispersion parameter in
"ah4"
-objects.
Some control arguments of hhh4()
have more clear
defaults.
The result of algo.farrington.fitGLM.fast()
now
additionally inherits from the "lm"
class to avoid warnings
from predict.lm()
about fake object.
Improved 'NAMESPACE' imports.
Some additional tiny bug fixes, see the subversion log on R-Forge for details.
paste0()
in
the package namespace if it is not found in R base at
installation of the surveillance package).Important new twinstim()
-feature: fix parameters
during optimization.
Useful update
-method for "twinstim"
-objects.
New [[
- and plot
-methods for
"simEpidataCSlist"
-objects.
simEpidataCS()
received tiny bug fixes and is now
able to simulate from epidemic-only models.
R0
-method for "simEpidataCS"
-objects (actually
a wrapper for R0.twinstim()
).
Removed dimyx
and eps
arguments from
R0.twinstim()
; now uses nCub
and
nCub.adaptive
from the fitted model and applies the same
(numerical) integration method.
animate.epidata
is now compatible with the
animation package.
More thorough documentation of "twinstim"
-related
functions including many examples.
nlminb
(instead of optim
's "BFGS"
) is
now the default optimizer (as already documented).
The twinstim
-argument nCub
can now be omitted when
using siaf.constant()
(as documented) and is internally set to
NA_real_
in this case. Furthermore, nCub
and
nCub.adaptive
are set to NULL
if there is
no epidemic component in the model.
toLatex.summary.twinstim
now again works for
summary(*, test.iaf=FALSE)
.
print
- and summary
-methods for
"epidataCS"
no longer assume that the BLOCK
index
starts at 1, which may not be the case when using a subset in
simulate.twinstim()
.
The "counter"
step function returned by
summary.epidataCS()
does no longer produce false
numbers of infectives (they were lagged by one timepoint).
plot.epidataCS()
now resolves ... correctly and
the argument colTypes
takes care of a possible
subset
.
simEpidataCS()
now also works for endemic-only models
and is synchronised with twinstim()
regarding the
way how siaf
is numerically integrated (including the
argument nCub.adaptive
).
Fixed problem with simEpidataCS()
related to missing
'NAMESPACE' imports (and re-exports) of marks.ppp
and
markformat.default
from spatstat, which are required
for spatstat::runifpoint()
to work, probably because
spatstat currently does not register its S3-methods.
Improved error handling in simEpidataCS()
. Removed a
browser()
-call and avoid potentially infinite loop.
The .allocate
argument of simEpidata()
has
now a fail-save default.
Simulation without endemic cox()
-terms now works.
Simplified imdepi
data to monthly instead of weekly
intervals in stgrid
for faster examples and reduced package
size.
The environment of all predefined interaction functions for
twinstim()
is now set to the .GlobalEnv
. The previous
behaviour of defining them in the parent.frame()
could have
led to huge save()
's of "twinstim"
objects even with
model=FALSE
.
simulate.twinSIR
only returns a list of epidemics if
nsim > 1
.
simulate.twinstim
uses nCub
and
nCub.adaptive
from fitted object as defaults.
Removed the ...-argument from simEpidataCS()
.
The coefficients returned by simEpidataCS()
are now stored
in a vector rather than a list for compatibility with
"twinstim"
-methods.
Argument cex.fun
of intensityplot.twinstim()
now
defaults to the sqrt
function (as in plot.epidataCS()
.
Support for non-parametric back-projection using the function
backprojNP()
which returns an object of the new
"stsBP"
class which inherits from "sts"
.
Bayesian nowcasting for discrete time count data is implemented in
the function nowcast()
.
Methods for cubature over polygonal domains can now also visualize what
they do. There is also a new quasi-exact method for cubature of the
bivariate normal density over polygonal domains. The
function polyCub()
is a wrapper for the different
methods.
residuals.twinstim()
and residuals.twinSIR()
:
extract the "residual process", see Ogata
(1988). The residuals of "twinSIR"
and
"twinstim"
models may be checked graphically by the new
function checkResidualProcess()
.
Modified arguments of twinstim()
: new ordering, new
argument nCub.adaptive
, removed argument
typeSpecificEndemicIntercept
(which is now specified as part of
the endemic
formula as (1|type)
).
Completely rewrote the R0
-method (calculate "trimmed" and
"untrimmed" R_0 values)
The "trimmed" R0
values are now part of the
result of the model fit, as well as bbox(W)
. The
model evaluation environment is now set as attribute of the
result if model=TRUE
.
New predefined spatial kernel: the Lomax power law kernel
siaf.lomax()
plot
-methods for "twinstim"
(intensityplot()
and iafplot()
)
as.epidataCS()
now auto-generates the stop-column if this is missing
print
-method for class "summary.epidataCS"
[
- and subset-method for "epidataCS"
(subsetting ...$events
)
plot
-method for "epidataCS"
Improved documentation for the new functionalities.
Updated references.
twinSIR
's intensityPlot
is now a method of the
new S3-generic function intensityplot
.
"twinstim"
and the "hhh4"
model.
The "twinSIR"
class of models has been
migrated from package RLadyBug to surveillance.
It may take a while before this version will become available from CRAN.Renamed the "week"
slot of the "sts"
S4 class to "epoch"
.
All saved data objects have accordingly be renamed, but some hassle
is to be expected if one you have old "sts"
objects stored in binary
form. The function convertSTS()
can be used to
convert such "old school" "sts"
objects.
Removed the functions algo.cdc()
and algo.rki()
.
Support for "twinSIR"
models (with associated
"epidata"
objects) as described
in Höhle (2009) has been moved from package
RLadyBug to surveillance.
That means continuous-time discrete-space SIR models.
Support for "twinstim"
models as described in
Meyer et al (2012). That means continuous-time
continuous-space infectious disease models.
Added functionality for non-parametric back projection
(backprojNP()
) and
now-casting (nowcast()
) based on "sts"
objects.
Replaced the deprecated getSpPPolygonsLabptSlots()
calls
by sp::coordinates()
when plotting the map slot.
Minor proof-reading of the documentation.
Added an argument "extraMSMargs"
to the algo.hmm function.
Fixed bug in outbreakP()
when having observations equal to zero
in the beginning. Here, $\hat{\mu}^{C1}$ in (5) of Frisen et al (2008)
is zero and hence the log-based summation in the code failed.
Changed to product as in the original code, which however might be
less numerically stable.
Fixed bug in stcd which added one to the calculated index of idxFA and idxCC. Thanks to Thais Rotsen Correa for pointing this out.
Added algo.outbreakP()
(Frisen & Andersson, 2009) providing a
semiparametric approach for outbreak detection for Poisson
distributed variables.
Added a pure R function for extracting ISO week and year from Date
objects. This function (isoWeekYear) is only called if "%G"
and "%V"
format strings are used on Windows (sessionInfo()[[1]]$os == "mingw32"
)
as this is not implemented for "format.Date"
on Windows.
Thanks to Ashley Ford, University of Warwick, UK for identifying
this Windows specific bug.
For algo.farrington()
a faster fit routine "algo.farrington.fitGLM.fast"
has been provided by Mikko Virtanen, National Institute for Health
and Welfare, Finland. The new function calls glm.fit()
directly, which gives a doubling of speed for long series. However, if one
wants to process the fitted model output some of the GLM routines might
not work on this output. For backwards compatibility the argument
control$fitFun = "algo.farrington.fitGLM"
provides the old (and slow)
behaviour.
A few minor bug fixes
Small improvements in the C-implementation of the twins()
function by Daniel Sabanés Bové fixing the segmentation fault
issue on 64-bit architectures.
Added the functions categoricalCUSUM and LRCUSUM.runlength for the CUSUM monitoring of general categorical time series (binomial, beta-binomial, multinomial, ordered response, Bradley-Terry models).
Added the functions pairedbinCUSUM and pairedbinCUSUM.runlength implementing the CUSUM monitoring and run-length computations for a paired binary outcome as described in Steiner et al. (1999).
Experimental implementation of the prospective space-time cluster detection described in Assuncao and Correa (2009).
Added a demo("biosurvbook")
containing the code of an upcoming
book chapter on how to use the surveillance package. This
contains the description of ISO date use, negative binomial CUSUM,
run-length computation, etc. From an applicational point of view
the methods are illustrated by Danish mortality monitoring.
Fixed a small bug in algo.cdc found by Marian Talbert Allen which resulted in the control$m argument being ignored.
The constructor of the sts class now uses the argument
"epoch"
instead of weeks to make clearer that also
daily, monthly or other data can be handled.
Added additional epochAsDate slot to sts class. Modified plot functions so they can handle ISO weeks.
algo.farrington now also computes quantile and median of the predictive distribution. Furthermore has the computation of reference values been modified so its a) a little bit faster and b) it is also able to handle ISO weeks now. The reference values for date t0 are calculated as follows: For i, i=1,..., b look at date t0 - i*year. From this date on move w months/weeks/days to the left and right. In case of weeks: For each of these determined time points go back in time to the closest Monday
Renamed the functions obsinyear to epochInYear, which now also handles objects of class Date.
Negative Binomial CUSUM or the more general NegBin likelihood ratio detector is now implemented as part of algo.glrnb. This includes the back calculation of the required number of cases before an alarm.
Time varying proportion binomial CUSUM.
Current status: Development version available from http://surveillance.r-forge.r-project.org/
Rewriting of the plot.sts.time.one function to use polygons
instead of lines for the number of observed cases. Due cause
a number of problems were fixed in the plotting of the legend.
Plotting routine now also handles binomial data, where the
number of observed cases y are stored in "observed"
and the
denominator data n are stored in "populationFrac"
.
Problems with the aggregate function not operating correctly for the populationFrac were fixed.
The "rogerson"
wrapper function for algo.rogerson was modified so it
now works better for distribution "binomial"
. Thus a time varying
binomial cusum can be run by calling
rogerson( x, control(..., distribution="binomial"))
An experimental implementation of the twins model documented in Held, L., Hofmann, M., Höhle, M. and Schmid V. (2006). A two-component model for counts of infectious diseases, Biostatistics, 7, pp. 422--437 is now available as algo.twins.
The algo_glrpois function now has an additional "ret"
arguments,
where one specifies the return type. The arguments of the underlying
c functions have been changed to include an additional direction and
return type value arguments.
added restart argument to the algo.glrpois control object, which allows the user to control what happens after the first alarm has been generated
experimental algo.glrnb function is added to the package. All calls to algo.glrpois are now just alpha=0 calls to this function. However, the underlying C functions differentiate between poisson and negative case