An extreme-value toolkit, methods for radiometrically dated records and several further methods from the literature, and S3 methods for the result classes.
extreme-value-methods, which takes over
the extreme-value material of vignette("EDE") and adds
vogel2009() and mccarthy1998().choosing-a-method: the sighting model
assumed by each method and a simulation study (15 scenario-size
combinations, 1000 replicates) of coverage, bias and the existence of
upper limits. The script is data-raw/simulation.R.hall_wang1999(): the minimum-distance end-point
estimator of Hall & Wang
(1999) for the k most recent sightings, with its exact
confidence interval. It does not depend on the extremal exponent, so it
is defined more often than maximum likelihood and, unlike
ole(), its interval does not rely on a plug-in shape. It
reproduces the 100 m example of the paper (estimate 9.91, extremal
exponent 0.82, 90% interval 9.80 to 9.92).fraga_alves_neves2014(): the closed-form end-point
estimator of Fraga Alves & Neves (2014) for the Gumbel domain of
attraction, which is never earlier than the last sighting.ede_select_k(): automatic choice of the number of
recent sightings k by ordered goodness-of-fit tests with
ForwardStop false-discovery-rate control, adapting the threshold
selection of Bader, Yan & Zhang (2018). It is a package construct,
not a published method for sighting records.dekkers1989() and ede_evi(): the moment
estimator of the extreme-value index (Dekkers, Einmahl & de Haan,
1989) and the end-point estimate that follows from it, for one
k or over a range of k (a stability diagram
with plot()). Unlike the Weibull-domain estimators it also
says whether a finite end-point is plausible. It is not invariant to the
origin of the time axis; origin is an argument and the help
page explains the dependence.smith1985(): maximum likelihood end-point and
profile-likelihood interval for the same model (Smith, 1985). The
likelihood is regular only when the extremal exponent exceeds 2, which
excludes a constant sighting rate; the function warns when the fit is
outside that range, and the help page reports simulated coverage against
hall_wang1999(), which is exact for any shape.hall_wang1999() and ede_endpoint_test(), is
checked against the exact points of Burrows (1979) quoted by Stephens
(1981).vogel2009() and ede_field_test(): the
probability plot correlation test of a constant sighting rate (uniform
sighting times) of Vogel et al. (2009), and the four field-significance
procedures they use to combine the tests of many records. The simulated
null reproduces the published significance levels of the field tests,
and the rejections of their Table 2.ede_endpoint_test(): tests whether the record is
compatible with a finite end-point (the Weibull domain assumed by
ole() and hall_wang1999()) against the Gumbel
null, using the statistics of Neves, Picek & Fraga Alves
(2006) and Hasofer & Wang (1992). The Neves p-value is exact
(Fisher’s distribution of the maximum over the sum of exponentials),
because the Gumbel approximation of the paper is very conservative for
small k. A third test, test = "rosario", is
the statistic G* of Fraga Alves, Neves & Rosario (2016), built on
fraga_alves_neves2014(), which has a Gumbel limit under the
null; its p-value under an exponential tail is simulated exactly as for
the Hasofer & Wang test, and the asymptotic Gumbel p-value of the
paper is given next to it. Its size is checked by simulation and its
power against a finite end-point and a heavy tail is verified; the
paper’s own application (supercentenarian women) is not reproduced
because its data are not published with it.ede_compare() now also reports
hall_wang1999() and
fraga_alves_neves2014().mccarthy1998(): the effort-adjusted Solow test
(“partial Solow equation”) of McCarthy (1998), also used by Roberts
& Jaric (2020), with the sampling effort given per time unit; it
returns the upper bound or the p-value curve.solow2006(), solow2006_region() and
solow2006_check(): inference about the extinction time, or
the settlement time, from dated remains with the dating error taken into
account (Solow, 1997; Solow, Roberts & Robbirt, 2006). The true ages
are uniform between a lower end-point and the endpoint of interest and
each recorded age carries a normal error with known standard deviation;
the endpoint is estimated by maximum likelihood with a
profile-likelihood interval (Solow, 1997),
solow2006_region() gives the joint confidence region of the
extinction times of two taxa, and solow2006_check() the
simulation check of the 2006 paper. The settlement example of Solow
(1997) is reproduced (651 yr BP, interval 617 to 687, and the profile
log-likelihood of -113.26), and so are the maximum likelihood estimates
and the conclusions about the joint region of the mammoth and horse
example of Solow et al. (2006), to the 0.01 ka resolution of the
published figures. The ages can be given as a data frame with one row
per remain, which allows tied dates with different errors.bradshaw2012(): the GRIWM method of Bradshaw et
al. (2012) for dated records with dating error. The sighting-rate model
of McInerny et al. (2006) is applied to the k most recent
records for every k, the results are averaged with weights
inversely proportional to the span of the series, and the dates are
resampled from their normal errors; the estimate is the median and the
interval the quantiles of the resampled values. The estimates for each
k and the weighted estimate of the paper’s Figure 2 are
reproduced for the mammoth and horse ages, and the resampled medians
agree with the extinction dates of the paper’s Figure 4 (12,200 and
11,150 radiocarbon years BP, to 30 years). Without errors it returns the
inverse-weighted McInerny estimate. The Neanderthal example is not
reproduced because its dates are not published in the paper.alroy2014(): the iterative Bayesian method of Alroy
(2014), which turns the frequency of sightings and the length of the
observed range into a posterior probability of extinction for each time
unit after the last sighting (the “creeping-shadow-of-a-doubt” values),
with the best estimate of the extinction time (the shadow value reaches
0.5) and credible limits. The standard and the run-based versions are
included. The worked example of the paper is reproduced (the posterior
is 0.0183 after one unit, 0.072 after three, 0.506 after ten and passes
0.99 after 25), and so is its example of the run-based sampling
probabilities. The real-data examples of the paper use museum records
that are not published with it, and the sampling-control and
maximum-likelihood variants are not implemented.solow1996(): the likelihood ratio test of Solow (1996)
of whether several taxa share an upper (or lower) end-point, with an
exactly known significance level, and the confidence interval for the
common end-point. It reproduces the worked examples of the paper for
two, three and ten ammonite taxa of Seymour Island, and for one taxon it
equals strauss1989().alaska_megafauna (Table 1 of Solow,
Roberts & Robbirt, 2006), nz_shells (Table 1 of Solow,
1997) and seymour_ammonites (Table 1 of Solow, 1996).plot() methods for sighting_data,
ede_estimate, ede_compare (a forest plot) and
ede_select_k objects.as.data.frame() method for ede_estimate
objects.ede_compare() returns an object of class
ede_compare, which still inherits from
data.frame.data_out = TRUE by
solow1993(), solow1993b(),
solow2005(), mcinerny2006(),
burgman1995(), solow_roberts2003(),
jaric2010() and jaric_roberts2014() now have a
column p_value. The former name, chance, is
kept as an alias and will be removed in a future release.hall_wang1999() no longer fails in
uniroot() when Greenwood’s statistic is above the critical
value over the whole search grid and below it only in the limit at
infinity. The interval then starts at the end of the grid and has no
upper limit (Inf, with the usual warning).r = ... examples in three help pages no longer
trigger roxygen’s inline-code parser.A consistency review of every method against its source paper.
Every new method reproduces the published examples of its source paper; where a paper’s own numbers are inconsistent, the help page says so.
solow1993_bayes(): Bayes factor and posterior
probability of persistence of Solow (1993a), equation 3.rout2009() and rout2009_absences(): Bayes
factor under a declining sighting rate and the number of absent surveys
needed to declare eradication (Rout, Salomon & McCarthy 2009).caley_barry2014(): probability of persistence from
yearly detections (Caley & Barry 2014). The constant model is
computed exactly; the non-constant model ports the authors’ MCMC
sampler.strauss1989_bayes(): Bayesian posterior for the end of
a range (Strauss & Sadler 1989, equations 22-27).thompson2013(): probability of persistence with several
observation types of uncertain detection probability (Thompson et
al. 2013), with the annealed and quenched averages computed
exactly.lee2014(): probability of persistence from certain and
uncertain sightings with false positives (Lee et al. 2014), following
the authors’ OpenBUGS model but computed exactly.kodikara2020(): posterior of the extinction year from
yearly certain and uncertain sightings, Models 1-3 of Kodikara et
al. (2020), computed exactly instead of by MCMC.kodikara2021(): extinction time when sightings follow
non-homogeneous Poisson processes with Weibull rates and a change point
(Kodikara et al. 2021), by numerical integration. The default reproduces
the paper with the authors’ likelihood; equation 9 as printed is
available as an option.brook2019(): resampling with sighting reliabilities as
inclusion probabilities, around any estimator (Brook, Buettel &
Jarić 2019).thompson2017(): the iterative model of records and
surveys behind the IUCN Red List extinction-probability tool (Thompson
et al. 2017).thompson2019(): Bayesian year-by-year updating
(Thompson et al. 2019).jaric_roberts2014() gains test_year and
data_out for the reliability-adjusted p-value of equations
7-8.marshall1994() and marshall1994_trend():
distribution-free confidence intervals on a range end-point and the test
for a trend in gap size.marshall1997(): confidence intervals under a
non-uniform fossil recovery potential.weiss_marshall1999() and
weiss_marshall1999_prior(): Bayesian end-point for sections
sampled at discrete horizons.solow1993b() returns the minimum variance unbiased
estimate of the extinction time (Solow 2005, equation 8) in
estimate.burgman1995(equation = 3): the runs test on occupied
and empty periods.ede_compare() applies every method that needs only the
sighting record and returns one table.dodo,
monk_seal, ferret, orchid and
extinct_birds.solow1993(),
solow1993b(), solow2005(),
mcinerny2006(), burgman1995(),
solow_roberts2003() and jaric2010()) no longer
report the year of first rejection as a point estimate, which none of
the source papers define. They return estimate = NA and the
one-sided upper confidence bound from inverting the test in
upper. For continuous-time methods the bound is now exact
rather than rounded up to a whole year: equation 5 of Solow (2005) for
solow1993(), equation 17 of Solow (2005) for
solow2005(), equations 7-8 of Jarić & Ebenhard (2010)
for jaric2010(), the Robson-Whitlock bound for
solow_roberts2003(), and a numerical root for
solow1993b().burgman1995() now applies equation 2 of Burgman et
al. (1995), the test for extinction, by default. The former calculation,
equation 4, tests for a change in conservation status using the longest
empty run anywhere in the record; it remains available as
equation = 4.solow1993() and solow1993b() count every
sighting event, as in Solow (1993a), instead of distinct sighting times
(solow1993()) or dropping all sightings in the origin time
unit (both). Only the sighting that defines the origin is omitted.solow1993() is now labelled Solow (1993a).ole() returned an upper confidence limit earlier than
the last sighting when -log(alpha / 2) >= k. The limit
does not exist in that case and is now returned as Inf with
a warning.mcinerny2006() ignored a zero-count first row. It now
uses it as the known start date and counts every sighting time, as the
source paper allows.burgman1995(equation = 4) returned NaN for
large numbers of sightings. The inclusion-exclusion sum is now computed
on the log scale.test_year when it
is less than one time unit after the last sighting.jaric_roberts2014()
and stated that its binary-count requirement is a package
restriction.ole() and solow2005(), and that
jaric2010() is an index rather than a p-value.k for the most recent sighting events.solow2005() calculation
with the Weibull extreme-value p-value from equation 16 of Solow
(2005).solow_roberts2003() for the nonparametric
two-endpoint persistence test.jaric2010() for average-interval and
sighting-trend inference.jaric_roberts2014() for records with individual
sighting reliabilities.