EDE

CRAN status   R-CMD-check   codecov

Extinction date estimation from sighting records.

Given a time-ordered record of sightings, EDE estimates when a species most likely went extinct, or how likely it is to still be extant. It collects methods from the sighting-record and stratigraphic-range literature under one interface: endpoint estimators, frequentist persistence tests, Bayesian models, methods for sightings of uncertain reliability, models that combine records with surveys (including the model behind the IUCN Red List extinction-probability tool), extreme-value end-point estimators with automatic choice of the number of recent sightings, and confidence intervals on the end-points of fossil ranges. Every method is tested against the published examples of its source paper.

Installation

You can install the released version of EDE from CRAN with:

install.packages("EDE")

And the development version from GitHub with:

# install.packages("remotes")
remotes::install_github("rodrigosqrt3/EDE")

Input

Every estimator takes a sighting_data object: a table with one row per time unit and the number of sightings recorded in it.

The earliest supplied time defines the observation origin. If monitoring began before the first sighting, include an initial row with count zero. Methods differ in their treatment of multiple sightings in one time unit, following their source papers: OLE, the Weibull test, both Solow (1993) tests and Burgman’s tests count every sighting event, while the Robson-Whitlock, Strauss-Sadler, McInerny, Solow-Roberts and Jarić-Ebenhard methods use only the distinct sighting times.

library(EDE)

years <- c(1900, 1902, 1903, 1905, 1907, 1908, 1910, 1912, 1915, 1918,
           1920, 1923, 1925, 1928, 1930, 1933, 1936)
sightings <- c(4, 3, 5, 2, 3, 4, 2, 1, 2, 1, 1, 2, 1, 1, 1, 1, 1)

sd <- sighting_data(data.frame(year = years, sightings = sightings))

The last confirmed sighting is 1936. No sightings were recorded afterward.

Estimators

ole(sd)
#> <OLE (Roberts & Solow 2003)>
#>   estimate: 1941.97
#>   95% CI: [1937.48, 1956.75]

robson1964(sd)
#> <Robson & Whitlock (1964)>
#>   estimate: 1939
#>   95% one-sided CI: [1936, 1993]

strauss1989(sd)
#> <Strauss & Sadler (1989)>
#>   estimate: 1938.25
#>   95% one-sided CI: [1936, 1943.41]

solow1993(sd, test_year = 2000)
#> <Solow (1993a)>
#>   95% upper confidence bound: 1939.32

solow1993b(sd, test_year = 2000)
#> <Solow (1993b)>
#>   95% upper confidence bound: NA
#> Warning message:
#> p-value never falls to alpha before `test_year`; returning NA.

solow2005(sd, test_year = 2000)
#> <Solow (2005) Weibull test>
#>   95% upper confidence bound: 1953.99

mcinerny2006(sd, test_year = 2000)
#> <McInerny, Roberts, Davy & Cribb (2006)>
#>   95% upper confidence bound: 1942

burgman1995(sd, test_year = 2000)
#> <Burgman, Grimson & Ferson (1995), equation 2>
#>   95% upper confidence bound: 1940

solow_roberts2003(sd, test_year = 2000)
#> <Solow & Roberts (2003)>
#>   95% upper confidence bound: 1993

jaric2010(sd, test_year = 2000)
#> <Jaric & Ebenhard (2010)>
#>   95% upper confidence bound: 1980.02

solow1993(), solow1993b(), solow2005(), mcinerny2006(), burgman1995(), solow_roberts2003() and jaric2010() are persistence tests. Their source papers define a test, not a point estimate, so they return estimate = NA and the one-sided 1 - alpha upper confidence bound obtained by inverting the test: the earliest time at which persistence is rejected. For the continuous-time methods this bound is computed exactly (for example, equation 17 of Solow 2005); for the discrete-period methods of McInerny et al. and Burgman et al. it is the first rejected time unit. A bound beyond test_year, or one that does not exist, is returned as NA with a warning. Pass data_out = TRUE to get the p-value curve at integer steps up to test_year, in a column named p_value (the former name, chance, is a deprecated alias).

Extreme-value estimators

ole(), hall_wang1999() and fraga_alves_neves2014() treat the last sightings as the extreme order statistics of the sighting-time distribution, whose end-point is the extinction time. All depend on k, the number of recent sightings, which ede_select_k() chooses by ordered goodness-of-fit tests with false-discovery-rate control.

hall_wang1999(sd, k = 8)           # Weibull domain, exact confidence interval
fraga_alves_neves2014(sd)          # Gumbel domain, closed form
set.seed(1)
ede_select_k(sd)$k                 # choose k from the data
ede_endpoint_test(sd, k = 8)       # is a finite end-point supported?
ede_evi(sd)                        # moment estimate of the index over k

Comparing methods

ede_compare() applies every method that needs only the sighting record and returns one table. Methods that cannot be applied stay in the table with the reason in note. plot() draws the table as a forest plot.

cmp <- ede_compare(sighting_data(monk_seal), test_year = 2050)
plot(cmp)

Example data

EDE bundles classic records from the literature, each reproducing its published example: dodo (Roberts & Solow 2003), monk_seal (Solow 1993a), ferret (Solow 1993b; Jarić & Ebenhard 2010), orchid (Solow & Roberts 2003) and extinct_birds, four possibly extinct birds with sightings classified by evidence quality (Roberts et al. 2010; Jarić & Roberts 2014), and alaska_megafauna, radiocarbon ages of Alaskan mammoths and horses with their dating errors (Solow, Roberts & Robbirt 2006), and nz_shells, radiocarbon ages of shells from New Zealand (Solow 1997), and seymour_ammonites, the ten uppermost ammonite taxa of Seymour Island (Solow 1996).

ole(sighting_data(dodo))
#> <OLE (Roberts & Solow 2003)>
#>   estimate: 1690.42
#>   95% CI: [1669.13, 1798.88]

Sighting record and persistence curves

curve_1993 <- solow1993(sd, test_year = 2000, data_out = TRUE)
curve_2005 <- solow2005(sd, test_year = 2000, data_out = TRUE)

plot(curve_1993$time, curve_1993$p_value, type = "l",
     xlab = "candidate extinction year", ylab = "p-value")
lines(curve_2005$time, curve_2005$p_value, col = "firebrick")

The two curves diverge because they encode different models. solow1993() assumes a stationary Poisson process and uses the number of sighting events after the observation origin. solow2005() instead uses the spacing of the most recent sightings under a Weibull extreme-value model, the same model used by ole().

For uncertain records, supply one reliability per sighting (one per occupied row, or one per row):

uncertain <- sighting_data(data.frame(
  year = c(1900, 1910, 1920, 1930),
  sightings = c(1, 1, 1, 1)
))

jaric_roberts2014(uncertain, reliability = c(1, 0.9, 0.6, 0.3))

Pass test_year to jaric_roberts2014() to also get its reliability-adjusted p-value, and data_out = TRUE for the p-value curve.

For a Bayesian view of the constant-rate model, solow1993_bayes() gives the Bayes factor in favour of persistence and the posterior probability of persistence for a chosen prior:

solow1993_bayes(sd, test_year = 2000, prior = 0.5)

Methods

Endpoint estimators

Function Reference Output
ole() Roberts & Solow (2003) point estimate + two-sided CI
robson1964() Robson & Whitlock (1964) point estimate + one-sided CI
strauss1989(), strauss1989_curve() Strauss & Sadler (1989) point estimate + one-sided CI / curve
hall_wang1999() Hall & Wang (1999) minimum-distance end-point + exact two-sided CI
fraga_alves_neves2014() Fraga Alves & Neves (2014) end-point for the Gumbel domain
dekkers1989(), ede_evi() Dekkers, Einmahl & de Haan (1989) moment estimate of the extreme-value index and end-point
smith1985() Smith (1985) likelihood end-point + profile-likelihood CI (regular range only)
ede_select_k() after Bader, Yan & Zhang (2018) number of recent sightings k to use
vogel2009(), ede_field_test() Vogel et al. (2009) PPCC test of a constant sighting rate; field significance over records
mccarthy1998() McCarthy (1998); Roberts & Jaric (2020) Solow test adjusted for sampling effort, upper bound / p-value curve
ede_endpoint_test() Neves, Picek & Fraga Alves (2006); Hasofer & Wang (1992); Fraga Alves, Neves & Rosario (2016) test of a finite end-point (Gumbel null)

Persistence tests

Function Reference Output
solow1993() Solow (1993a) upper bound or p-value curve
solow1993b() Solow (1993b); Solow (2005) unbiased estimate + upper bound, or p-value curve
solow2005() Solow (2005) upper bound or p-value curve
mcinerny2006() McInerny, Roberts, Davy & Cribb (2006) upper bound or p-value curve
burgman1995() Burgman, Grimson & Ferson (1995) upper bound or p-value curve (extinction, eq. 2; change in status, eqs. 3-4)
solow_roberts2003() Solow & Roberts (2003) upper bound or p-value curve
jaric2010() Jarić & Ebenhard (2010) upper bound or index curve

Bayesian models

Function Reference Output
solow1993_bayes() Solow (1993a) Bayes factor + posterior probability of persistence
rout2009(), rout2009_absences() Rout, Salomon & McCarthy (2009) Bayes factor under a declining sighting rate; surveys needed to declare eradication
caley_barry2014() Caley & Barry (2014) posterior probability of persistence from yearly detections (exact constant model, or MCMC)
strauss1989_bayes() Strauss & Sadler (1989) posterior mean + credible interval of the end-point
alroy2014() Alroy (2014) iterative posterior probability of extinction (“shadow values”), best estimate and credible limits

Uncertain sightings

Function Reference Output
jaric_roberts2014() Jarić & Roberts (2014) reliability-adjusted estimate + upper bound, p-value
thompson2013() Thompson et al. (2013) probability of persistence with several observation types
lee2014() Lee et al. (2014) posterior probability of persistence with certain and uncertain (false-positive) sightings
kodikara2020() Kodikara et al. (2020) posterior of the extinction year from yearly certain and uncertain sightings
kodikara2021() Kodikara et al. (2021) posterior of the extinction time under Weibull-rate Poisson processes with a change point
brook2019() Brook, Buettel & Jarić (2019) resampling by reliability around any estimator

Records and surveys

Function Reference Output
thompson2017() Thompson et al. (2017) year-by-year probability extant (IUCN tool), with uncertainty bands
thompson2019() Thompson et al. (2019) Bayesian year-by-year updating

Stratigraphic ranges

Function Reference Output
marshall1994(), marshall1994_trend() Marshall (1994) distribution-free bounds on the CI; test for a trend in gap size
marshall1997() Marshall (1997) one-sided CI under non-uniform recovery potential
weiss_marshall1999() Weiss & Marshall (1999) Bayesian end-point for discretely sampled sections
solow1996() Solow (1996) exact likelihood ratio test of a common end-point of several taxa, and confidence interval for it

Dated remains with dating error

Function Reference Output
solow2006() Solow (1997); Solow, Roberts & Robbirt (2006) maximum likelihood extinction (or settlement) time with normal dating error, profile-likelihood interval
solow2006_region() Solow, Roberts & Robbirt (2006) joint confidence region for the extinction times of two taxa
solow2006_check() Solow, Roberts & Robbirt (2006) simulation check of the uniform-age model
bradshaw2012() Bradshaw et al. (2012) GRIWM: inverse-weighted McInerny estimate over the k latest records, resampled over the dating errors

For fossil data, the end-point estimated is always the one at the largest time: code ages as time = -age, and first appearances as negative positions.

See vignette("EDE") for the statistical background and guidance on choosing a method.

License

GPL (>= 3)