EDE implements methods for inference about extinction from sighting records and fossil ranges. Some estimate an endpoint directly; others test whether a species could plausibly have persisted to a candidate time. The tests return frequentist p-values, not posterior probabilities that the species is extant; the Bayesian models later in this vignette give those. Their source papers define no point estimate, so EDE reports the one-sided upper confidence bound obtained by inverting each test.
The p-value curves returned with data_out = TRUE have a
column p_value (the former name, chance, is a
deprecated alias).
library(EDE)
years <- c(1900, 1902, 1903, 1905, 1907, 1908, 1910, 1912,
1915, 1918, 1920, 1923, 1925, 1928, 1930, 1933, 1936)
counts <- 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 = counts))The earliest supplied time defines the observation origin. Include an initial zero-count row if observation began before the first sighting. Methods differ in their treatment of counts, following their source papers: OLE, the Weibull test, both Solow (1993) tests and Burgman’s tests count every sighting event; the Robson-Whitlock, Strauss-Sadler, McInerny, Solow-Roberts and Jarić-Ebenhard methods use only distinct sighting times. Strauss-Sadler, Solow-Roberts and Jarić-Ebenhard do not use the observation origin at all.
robson1964() uses the final gap for a Quenouille-type
bias-reduced point estimate and a one-sided confidence interval.
strauss1989() supplies an unbiased endpoint estimate and
one-sided interval under uniform occurrence. ole() fits the
Weibull extreme-value model to the k most recent sighting
events and reports the two-sided interval of Roberts & Solow
(2003).
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]
ole(sd)
#> <OLE (Roberts & Solow 2003)>
#> estimate: 1941.97
#> 95% CI: [1937.48, 1956.75]
ole(sd, k = 10)
#> <OLE (Roberts & Solow 2003)>
#> estimate: 1940.43
#> 95% CI: [1936.35, 1957.28]solow1993(sd, test_year = 2000)
#> <Solow (1993a)>
#> 95% upper confidence bound: 1939.32
solow1993b(sd, test_year = 2000)
#> Warning: p-value never falls to alpha before `test_year`; returning NA.
#> <Solow (1993b)>
#> estimate: 1940.97
#> 95% upper confidence bound: 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: 1980.02solow1993() assumes a stationary Poisson process.solow1993b() allows an exponentially declining sighting
rate. Its upper bound may not exist: the p-value has a positive floor
however late the candidate time.solow2005() uses the Weibull extreme-value p-value
associated with OLE. Its bound is one-sided, so it equals the upper
limit of ole() at twice the significance level.mcinerny2006() conditions on the previous binary
sighting rate.burgman1995() uses discrete periods. By default it
applies equation 2, the test for extinction based on the final run of
empty periods; equation = 4 tests for a change in
conservation status using the longest empty run anywhere in the
record.solow_roberts2003() uses only the two most recent
distinct sightings; its bound coincides with that of
robson1964().jaric2010() uses the average interval and can adjust it
for a trend in consecutive interval lengths. It is an index rather than
a formal test.mccarthy1998() is the Solow (1993a) test with the time
axis replaced by cumulative sighting effort, for records in which effort
changed over time; the effort is supplied by the user, one value per
time unit.The full p-value curve is available with
data_out = TRUE:
curve <- jaric2010(sd, test_year = 2000, data_out = TRUE)
plot(curve$time, curve$p_value, type = "l",
xlab = "candidate time", ylab = "index")
abline(h = 0.05, lty = 2)ole() is the first of a family of estimators that treat
the latest sightings as the upper order statistics of the sighting-time
distribution, whose endpoint is the extinction time. EDE adds
hall_wang1999(), fraga_alves_neves2014(),
dekkers1989() and smith1985(), the choice of
the number of sightings with ede_select_k(), a test of
whether the endpoint is finite (ede_endpoint_test()), and
checks of the sighting model (vogel2009(),
ede_field_test()). They are described in
vignette("extreme-value-methods").
hall_wang1999(sd, k = 8)
#> Warning: the upper confidence limit does not exist at this alpha and record;
#> returning Inf.
#> <Hall & Wang (1999)>
#> estimate: 1938.06
#> 95% CI: [1936, Inf]The effect of the sighting model on all of these methods is
quantified in vignette("choosing-a-method").
Under the constant-rate model, solow1993_bayes() gives
the Bayes factor of Solow (1993a) in favour of persistence to a
candidate time, assuming a uniform prior on the extinction time and a
noninformative prior on the sighting rate. Combined with a prior
probability of persistence, it gives the posterior probability that the
species is extant, which is the quantity the p-values above do not
provide.
jaric_roberts2014() assigns a probability of validity to
every sighting. EDE requires binary counts because reliability belongs
to individual records; this is a package restriction, not one of the
source paper. A sighting at the observation origin fixes the origin and
its reliability is not used.
uncertain <- sighting_data(data.frame(
year = c(1900, 1910, 1920, 1930),
sightings = 1
))
jaric_roberts2014(
uncertain,
reliability = c(1.0, 0.9, 0.6, 0.3)
)
#> <Jaric & Roberts (2014)>
#> estimate: 1930.99
#> 95% upper confidence bound: 2005.22The result includes the effective number of sightings and the
reliability-adjusted endpoint as additional components. Supply
test_year to add the reliability-adjusted p-value of
equations 7-8. Reliabilities should be elicited independently of the
extinction analysis.
Other approaches to uncertain sightings:
thompson2013() handles several observation types (for
example specimens, expert sightings and local reports), each with its
own range of detection and validity probabilities, in discrete
time.brook2019() resamples the record, keeping each sighting
with probability equal to its reliability, and applies any EDE estimator
to every sub-record. Solow (2023) criticised the statistical properties
of the resulting date; the authors present it as pre-processing rather
than an estimator, and its help page discusses both views.rout2009() generalises the Solow (1993a) Bayes factor to
a sighting rate that declines as a power of time, and
rout2009_absences() turns it into a decision rule: the
number of consecutive absent surveys after which declaring eradication
minimises the expected cost. caley_barry2014() models
yearly detections with constant detection and extinction probabilities;
its posterior is computed exactly:
thompson2017() implements the iterative model of
Thompson et al. (2017), which underlies the IUCN Red List tool for
estimating the probability that a taxon is extinct. Each year
contributes either a record, with a probability of being correct, or a
survey that failed to find the species, described by its extent, its
chance of detecting the species and its reliability. Ranges for these
quantities give interval and Monte Carlo bounds.
thompson2019() casts the same data as Bayesian updating, in
which each year’s posterior is the next year’s prior.
The same question arises for the last appearance of a taxon in the
fossil record. strauss1989() assumes fossil horizons are
uniformly distributed; marshall1994() drops that assumption
and bounds the confidence interval using only the distribution of gap
sizes, with marshall1994_trend() to test whether gap size
changes through the section; marshall1997() accepts a
fossil recovery potential that varies through the section; and
weiss_marshall1999() gives a Bayesian estimate for sections
sampled at discrete horizons, with absences recorded. EDE always
estimates the end-point at the largest time, so ages in
millions of years should be entered as time = -age, and
first appearances as negative stratigraphic positions.
ede_compare() applies every method that needs only the
sighting record:
cmp <- ede_compare(sighting_data(monk_seal), test_year = 2050)
cmp
#> method function_name estimate
#> 1 OLE (Roberts & Solow 2003) ole 1963.528
#> 2 Robson & Whitlock (1964) robson1964 1956.000
#> 3 Strauss & Sadler (1989) strauss1989 1961.250
#> 4 Solow (1993a) solow1993 NA
#> 5 Solow (1993b) solow1993b 1957.605
#> 6 Solow (2005) Weibull test solow2005 NA
#> 7 McInerny, Roberts, Davy & Cribb (2006) mcinerny2006 NA
#> 8 Burgman, Grimson & Ferson (1995), equation 2 burgman1995 NA
#> 9 Solow & Roberts (2003) solow_roberts2003 NA
#> 10 Jaric & Ebenhard (2010) jaric2010 NA
#> 11 Hall & Wang (1999) hall_wang1999 1952.630
#> 12 Fraga Alves & Neves (2014) fraga_alves_neves2014 1956.150
#> lower upper interval_type
#> 1 1952.67 2093.799 two-sided
#> 2 1952.00 2028.000 one-sided
#> 3 1952.00 1993.245 one-sided
#> 4 NA 1993.245 one-sided
#> 5 NA NA one-sided
#> 6 NA 2029.444 one-sided
#> 7 NA 1979.000 one-sided
#> 8 NA 1994.000 one-sided
#> 9 NA 2028.000 one-sided
#> 10 NA NA one-sided
#> 11 1952.00 Inf two-sided
#> 12 NA NA two-sided
#> note
#> 1
#> 2
#> 3
#> 4
#> 5 p-value never falls to alpha before `test_year`; returning NA.
#> 6
#> 7
#> 8
#> 9
#> 10 p-value never falls to alpha before `test_year`; returning NA.
#> 11 the upper confidence limit does not exist at this alpha and record; returning Inf.
#> 12plot(cmp) draws the estimates, intervals and upper
bounds as a forest plot.
hall_wang1999() when an extreme-value model is
defensible: its interval is exact, and an infinite upper limit is
informative, not a failure. ole() fits the same model but
under-covers in simulation, most of all with many sightings; limit it
with ede_select_k(). fraga_alves_neves2014()
covers the Gumbel domain.robson1964() for a bound that keeps its coverage
whatever the shape of the sighting rate, accepting that it is wide.mccarthy1998() when sighting effort varied and can
be quantified.solow1993() only when a constant sighting rate is
plausible.solow1993b() for an approximately exponential
decline.jaric2010() when the intervals themselves show a
gradual trend.burgman1995() for discrete frequency data; use
equation = 4 only for questions about a change in
status.jaric_roberts2014() when observations have explicit
reliability assessments, and thompson2013() when they fall
into types with uncertain detection probabilities.thompson2017() when unsuccessful surveys are
documented, as in an IUCN Red List assessment.Running several scientifically defensible methods is useful, but p-values from models whose assumptions are violated should not be combined or interpreted as extinction probabilities.
The assumptions of each method, and a simulation study of their
coverage and bias under constant, declining and varying sighting rates,
are in vignette("choosing-a-method"). In short, the
constant-rate tests are calibrated only when the rate is constant,
robson1964() keeps its coverage under every scenario
simulated but gives wide bounds, and the extreme-value estimators have
bounded intervals only when the record carries strong evidence of a
finite endpoint.
References for the methods introduced after the core estimators (Rout
et al. 2009; Thompson et al. 2013, 2017, 2019; Caley & Barry 2014;
Brook et al. 2019; Marshall 1994, 1997; Weiss & Marshall 1999;
McCarthy 1998) are given on the help page of each function; those of the
extreme-value methods are in
vignette("extreme-value-methods").
Burgman, M. A., Grimson, R. C., & Ferson, S. (1995). Inferring threat from scientific collections. Conservation Biology, 9(4), 923-928.
Cooke, P. (1980). Optimal linear estimation of bounds of random variables. Biometrika, 67, 257-258.
Jarić, I., & Ebenhard, T. (2010). A method for inferring extinction based on sighting records that change in frequency over time. Wildlife Biology, 16(3), 267-275.
Jarić, I., & Roberts, D. L. (2014). Accounting for observation reliability when inferring extinction based on sighting records. Biodiversity and Conservation, 23(11), 2801-2815.
McInerny, G. J., Roberts, D. L., Davy, A. J., & Cribb, P. J. (2006). Significance of sighting rate in inferring extinction and threat. Conservation Biology, 20(2), 562-567.
Roberts, D. L., & Solow, A. R. (2003). When did the dodo become extinct? Nature, 426, 245.
Robson, D. S., & Whitlock, J. H. (1964). Estimation of a truncation point. Biometrika, 51, 33-39.
Solow, A. R. (1993a). Inferring extinction from sighting data. Ecology, 74, 962-964.
Solow, A. R. (1993b). Inferring extinction in a declining population. Journal of Mathematical Biology, 32, 79-82.
Solow, A. R. (2005). Inferring extinction from a sighting record. Mathematical Biosciences, 195, 47-55.
Solow, A. R., & Roberts, D. L. (2003). A nonparametric test for extinction based on a sighting record. Ecology, 84, 1329-1332.
Strauss, D., & Sadler, P. M. (1989). Classical confidence intervals and Bayesian probability estimates for ends of local taxon ranges. Mathematical Geology, 21, 411-427.