delta = 7 argument to
ssd_plot_cdf().ssdtools version 1.0.0 is the first major release of
ssdtools with some important improvements and breaking
changes.
An important change to the functionality of
ssd_fit_dists() was to switch from model fitting using fitdistrplus
to TMB which
has resulted in improved handling of censored data. Although it was
hoped that model fitting would be faster this is currently not the
case.
As a result of the change the fitdists objects returned
by ssd_fit_dists() from previous versions of
ssdtools are not compatible with the major release and
should be regenerated.
As a result of an international collaboration British Columbia and Canada and Australia and New Zealand selected a set of recommended distributions for model averaging and settings when generating final guidelines.
The distributions are
> ssd_dists_bcanz()
[1] "gamma" "lgumbel" "llogis" "lnorm" "lnorm_lnorm" "weibull" The ssd_fit_bcanz() and ssd_hc_bcanz()
functions were added to the package to facilitate the fitting of these
distributions and estimation of hazard concentrations using the
recommended settings.
In the previous version of ssdtools a distribution was
considered to have converged if the following condition was met
stats::optim() returns a code of 0 (indicating
successful completion).In the new version an additional two conditions must also be met
at_boundary_ok = TRUE or the user can
specify different boundary values - see below)computable = FALSE)Censoring can now be specified by providing a data set with one or more rows that have
It is currently not possible to fit distributions to data sets that have
Rows that have a zero or missing value for the left column and an infinite or missing value for the right column (fully censored) are uninformative and will result in an error.
For uncensored data, Akaike Weights are calculated using AICc (which corrects for small sample size). In the case of censored data, Akaike Weights are calculated using AIC (as the sample size cannot be estimated) but only if all the distributions have the same number of parameters (to ensure the weights are valid).
Weighting must be positive with values <= 1000.
Previously the density functions for the available distributions were
exported as R functions to make them accessible to
fitdistrplus. This meant that ssdtools had to
be loaded to fit distributions. The density functions are now defined in
C++ as TMB templates and are no longer exported.
The distribution, quantile and random generation functions are more
generally useful and are still exported but are now prefixed by
ssd_ to prevent clashes with existing functions in other
packages. Thus for example plnorm(), qlnorm()
and rlnorm() have been renamed ssd_plnorm(),
ssd_qlnorm() and ssd_rlnorm().
The following distributions were added (or in the case of
burrIII3 readded) to the new version
burrIII3 - burrIII three parameter distributioninvpareto - inverse pareto (with bias correction in
scale order statistic)lnorm_lnorm log-normal/log-normal mixture
distributionllogis_llogis log-logistic/log-logistic mixture
distributionThe following arguments were added to
ssd_fit_dists()
rescale (by default FALSE) to specify
whether to rescale concentrations values by dividing by the largest
(finite) value. This alters the parameter estimates, which can help some
distributions converge, but not the estimates of the hazard
concentrations/protections.reweight (by default FALSE) to specify
whether to reweight data points by dividing by the largest weight.at_boundary_ok (by default FALSE) to
specifying whether a distribution with one or more parameters at a
boundary has converged.min_pmix (by default 0) to specify the boundary for the
minimum proportion for a mixture distribution.range_shape1 (by default c(0.05, 20)) to
specify the lower and upper boundaries for the shape1 parameter of the
burrIII3 distribution.range_shape2 (by default the same as
range_shape2) to specify the lower and upper boundaries for
the shape2 parameter of the burrIII3 distribution.control (by default an empty list) to pass a list of
control parameters to stats::optim().It also worth noting that the default value of
computable argument was switched from
FALSE to TRUE to enforce stricter requirements
on convergence (see above).The following were added to handle multiple distributions
ssd_dists() to specify subsets of the available
distributions.delta argument (by default 7) to the
subset() generic to only keep those distributions within
the specified AIC(c) difference of the best supported distribution.The function ssd_fit_burrlioz() was added to approximate
the behaviour of Burrlioz.
Hazard concentration estimation is performed by ssd_hc()
(which is wrapped by predict()) and hazard protection
estimation by ssd_hp(). By default confidence intervals are
estimated by parametric bootstrapping.
To reduce the time required for bootstrapping, parallelization was implemented using the future package.
The following arguments were added to ssd_hc() and
ssd_hp()
delta (by default 7) to only keep those distributions
within the specified AIC difference of the best supported
distribution.min_pboot (by default 0.90) to specify minimum
proportion of bootstrap samples that must successfully fit.parametric (by default TRUE) to allow
non-parametric bootstrapping.control (by default an empty list) to pass a list of
control parameters to stats::optim().and the following columns were added to the output data frame
wt to specify the Akaike weight.method to indicate whether parametric or non-parametric
bootstrap was used.nboot to indicate how many bootstrap samples were
used.pboot to indicate the proportion of bootstrap samples
which fitted.It also worth noting that the
dist column was moved from the last to the first
position in the output data frame.Confidence intervals cannot be estimated for interval censored data.
Confidence intervals cannot be estimated for unequally weighted data.
The pvalue argument (by default FALSE) was
added to ssd_gof() to specify whether to return p-values
for the test statistics as opposed to the test statistics
themselves.
There have also been some substantive changes to the plotting functionality.
Added following functions
ssd_plot_data() to plot censored and uncensored data by
calling geom_ssdpoint() for the left and for the right
column (alpha parameter values should be adjusted accordingly)geom_ssdsegment() to allow plotting of the range of a
censored data points using segments.scale_colour_ssd() (and scale_color_ssd())
to provide an 8 color-blind scale.Made the following changes to ssd_plot()
bounds (by default
c(left = 1, right = 1)) argument specify how many orders of
magnitude to extend the plot beyond the minimum and maximum
(non-missing) values.linetype (by default NULL) argument
to specify line type.linecolor (by default NULL) argument
to specify line color.ylab from “Percent of Species
Affected” to “Species Affected”.Renamed - GeomSsd to GeomSsdpoint. -
StatSsd to StatSsdpoint
Soft-deprecated - geom_ssd() for
geom_ssdpoint(). - stat_ssd(). -
ssd_plot_cf() for
fitdistrplus::descdist().
ssddataThe dataset boron_data was renamed
ccme_boron and moved to the ssddata R
package together with the other CCME datasets.
The ssddata package provides a suite of datasets for
testing and comparing species sensitivity distribution fitting
software.
Added
ssd_data() to return original data for a
fitdists object.ssd_ecd_data() to get empirical cumulative density for
data.ssd_sort_data() to sort data by empirical cumulative
density.npars() now orders by distribution name.Implemented the following generics for fitdists
objects
glance() to get the model likelihoods,
information-theoretic criteria etc.augment() to return original data set.logLik() to return the log-likelihood.summary.fitdists() to summarize.wt (Akaike weight) column to
predict(), ssd_hc() and
ssd_hp()ic to predict(),
ssd_hc() and ssd_hp() because unused.ssd_fit_dists().actuar
package.comma_signif() so that now rounds to 3
significant digits by default and only applies
scales::comma() to values >= 1000.... argument to
comma_signif().rdist() functions now use length of n if
length(n) > 1.slnorm() to get starting values for ‘dlnorm’
distribution.rllog() that was causing error.ssd_hc() and
predict() where ci = TRUE to explicit
ssd_hc(ci = FALSE) and
predict(ci = FALSE).shape and scale arguments to
llog() with lshape and
lscale.location and scale arguments to
lgumbel() with llocation and
lscale.burrIII2).ssd_hp() to calculate hazard percent at specific
concentrations.ssd_exposure() to calculate proportion exposed
based on distribution of concentrations.predict() and added parallel argument.ssd_fit_dists() now checks if standard errors
computable.burrIII3).sdist(x) functionality to set starting values for
distributions.ssd_plot_cdf() to plot cumulative distribution
function (equivalent to autoplot())nobs() for censored data now returns a missing
value.ssd_fit_dists() distributions now ordered
alphabetically.ssd_hc() argument hc = 5L for
percent = 5L.dllog() etc for dllogis().ssd_cfplot() for
ssd_plot_cf().llog distribution with small concentrations.