Correlates of rediscovery and the detectability of extinction in mammals

Abstract

Extinction is difficult to detect, even in well-known taxa such as mammals. Species with long gaps in their sighting records, which might be considered possibly extinct, are often rediscovered. We used data on rediscovery rates of missing mammals to test whether extinction from different causes is equally detectable and to find which traits affect the probability of rediscovery. We find that species affected by habitat loss were much more likely to be misclassified as extinct or to remain missing than those affected by introduced predators and diseases, or overkill, unless they had very restricted distributions. We conclude that extinctions owing to habitat loss are most difficult to detect; hence, impacts of habitat loss on extinction have probably been overestimated, especially relative to introduced species. It is most likely that the highest rates of rediscovery will come from searching for species that have gone missing during the 20th century and have relatively large ranges threatened by habitat loss, rather than from additional effort focused on charismatic missing species.

Full text

doi: 10.1098/rspb.2010.1579
, 1090-1097 first published online 29 September 2010278 2011 Proc. R. Soc. B
Diana O. Fisher and Simon P. Blomberg
mammals
Correlates of rediscovery and the detectability of extinction in
Supplementary data
tml
http://rspb.royalsocietypublishing.org/content/suppl/2010/09/29/rspb.2010.1579.DC1.h
"Data Supplement"
References http://rspb.royalsocietypublishing.org/content/278/1708/1090.full.html#ref-list-1
This article cites 28 articles, 9 of which can be accessed free
Subject collections (2172 articles)ecology 
Articles on similar topics can be found in the following collections
Email alerting service hereright-hand corner of the article or click
Receive free email alerts when new articles cite this article - sign up in the box at the top
http://rspb.royalsocietypublishing.org/subscriptions go to: Proc. R. Soc. BTo subscribe to
This journal is © 2011 The Royal Society
on February 24, 2011rspb.royalsocietypublishing.orgDownloaded from

Correlates of rediscovery and the
detectability of extinction in mammals
Diana O. Fisher*and Simon P. Blomberg
School of Biological Sciences, Goddard Building (8), The University of Queensland, St Lucia,
Queensland 4072, Australia
Extinction is difficult to detect, even in well-known taxa such as mammals. Species with long gaps in their
sighting records, which might be considered possibly extinct, are often rediscovered. We used data on
rediscovery rates of missing mammals to test whether extinction from different causes is equally detect-
able and to find which traits affect the probability of rediscovery. We find that species affected by habitat
loss were much more likely to be misclassified as extinct or to remain missing than those affected by intro-
duced predators and diseases, or overkill, unless they had very restricted distributions. We conclude that
extinctions owing to habitat loss are most difficult to detect; hence, impacts of habitat loss on extinction
have probably been overestimated, especially relative to introduced species. It is most likely that the high-
est rates of rediscovery will come from searching for species that have gone missing during the 20th
century and have relatively large ranges threatened by habitat loss, rather than from additional effort
focused on charismatic missing species.
Keywords: rediscovery; mammal extinction; habitat loss; introduced predators; overkill; extinction rate
1. INTRODUCTION
Species presumed to be extinct are often rediscovered.
For example, 89 Australian vascular plants were rediscov-
ered between 1981 and 2001, and rediscovery was the
reason for disqualification of 13 mammals from a list of
144 candidate extinct species analysed in 1999 [1,2].
Conservation resources are wasted searching for species
that have no chance of rediscovery, while most missing
species receive no attention. Public understanding of the
extent of the current extinction crisis may be compro-
mised if lists of extinct species are too conservative [3].
On the other hand, premature designation of species
extinction will lead to the withdrawal of conservation
effort [4]. Quantitative predictions of which missing
species are most likely to be extinct and which might be
rediscovered are therefore needed.
Modern-era species extinctions (since 1500) have been
assessed based on two criteria. Provided that a species is tax-
onomically accepted [2], the species is considered extinct if
it remains missing either for a prescribed waiting period (the
pre-1995 IUCN threshold was 50 years), or after a specified
search effort [5]. The current IUCN criteria are phrased in
terms of search effort: ‘exhaustive surveys in known and/or
expected habitat, at appropriate times, throughout its his-
toric range have failed to record an individual... over a
time frame appropriate to the taxon’s life cycle and life
form’ [5]. The emphasis is now on demonstrating adequate
search effort, because extinction is difficult to detect, even in
the best-known taxa. The current IUCN criteria acknowl-
edge this by including a status category of ‘Critically
Endangered’ with a tag of ‘(Possibly Extinct)’. A decade
ago, a comprehensive review found only 36 per cent of
purported mammal extinctions to be resolved. The rest
had insufficient evidence of extirpation or taxonomic val-
idity, or had been rediscovered [2]. MacPhee & Flemming
[2] argued that although the 50-year threshold was not
inherently meaningful, some substantial, unspecified wait-
ing period is necessary because of frequent rediscoveries
of supposedly extinct mammals. Lack of detection for 50
years might not constitute good evidence of extinction,
depending on the distribution of previous sightings [6].
For species with five or more dated sightings, quantitative
techniques have recently been developed to estimate the
probability of extinction, based on the distribution of sight-
ing records [6,7]. However, 70 per cent of purportedly
extinct mammal species are known from fewer than five
sightings (D. Fisher & S. Blomberg 2010, unpublished
data). We do not know which traits are correlated with
longer gaps in species detection records, or the frequency
of species rediscovery.
We might expect the rediscovery rate of missing species
to depend on factors including search effort [8] and the
size of the area that investigators need to search. We pre-
dicted that the probability of species rediscovery would
also depend on species traits that predict extinction risk
in mammals, such as small geographical range size, eco-
logical specialization, large body size and slow
reproduction [9,10]. We might further expect to find
interactions between the cause of decline in missing
species and traits that predict elevated extinction risk,
because small range size and specialization are expected
to disproportionately increase extinction risk in species
affected by habitat loss [11,12], and large body size to
increase extinction risk particularly in harvested and per-
secuted species, and those subject to overkill by
introduced predators [13,14]. The aims of this study
were to quantify the frequency of species rediscovery,
and to identify traits that predict the probability of
rediscovery in missing mammals.
*Author for correspondence (d.fisher@uq.edu.au).
Electronic supplementary material is available at http://dx.doi.org/
10.1098/rspb.2010.1579 or via http://rspb.royalsocietypublishing.org.
Proc. R. Soc. B (2011) 278, 1090–1097
doi:10.1098/rspb.2010.1579
Published online 29 September 2010
Received 23 July 2010
Accepted 9 September 2010 1090 This journal is q2010 The Royal Society
on February 24, 2011rspb.royalsocietypublishing.orgDownloaded from

2. MATERIAL AND METHODS
(a)Data and definitions
In order to compare traits and detection rates of currently
detected (extant) and undetected (missing) species, we first
needed to define an appropriate dataset. We compiled a
global database of species with long gaps in their sighting
records: (i) species that have been reported extinct (includ-
ing ‘Extinct’ (EX), ‘Critically Endangered (Possibly
Extinct)’ (CR (PE)) and ‘Extinct in the Wild’ (EW)
species in the Red List) or flagged as missing in the litera-
ture (so have the most recent detection date reported); and
(ii) species that have published accounts of rediscovery,
including records of detection history (see electronic sup-
plementary material, dataset S1). To minimize the
possibility that our choice of definition of ‘missing’ influ-
enced our results, we analysed both the full dataset and
subsets of the data with a stricter definition of assumed
extinction, omitting species classed as data-deficient (but
with species account information indicating that they are
very likely extinct), now extinct in the wild, or subject to
taxonomic disagreement, and including only missing
species currently designated as EX or CR (PE). We also
analysed a subset of the data including only the most
recent two centuries of records (see below). We established
the status and detection history of these species after they
were reported missing using past and present IUCN Red
Lists and related publications, primary literature, books
and the Committee on Recently Extinct Organisms
mammal database [2]. We classed each species as currently
missing/extinct or rediscovered, and recorded the date of
last sighting and rediscovery.
In order to calculate rediscovery rates, we included the
dates of all rediscoveries (i.e. species that were rediscovered
but are now again missing were classified as rediscovered).
Only one rediscovered species is classified as Extinct by the
IUCN: the desert rat kangaroo (Caloprymnus campestris),
which was found in 1931 after having been missing for 90
years [15], and disappeared again in 1935, shortly after the
red fox (Vulpes vulpes) reached the lake Eyre basin of
South Australia, where it occurred [16]. Two other rediscov-
ered Australian species (the central rock rat, Zyzomys
pedunculatus and Christmas island shrew, Crocidura
trichura), a Cuban species (Garrido’s hutia, Mysateles
garridoi) and a Solomon Islands species (the Vanikoro
flying fox, Pteropus tuberculatus) are classified as Critically
Endangered (Possibly Extinct). We omitted species that
have been ‘rediscovered’ through taxonomic revision (e.g.
‘splitting’), and only included mammals that are named as
taxonomically accepted full species (not subspecies) in the
IUCN Red List, which lists assessments of most current
taxonomically accepted species of mammals [17].
For each species, we recorded available data on the cause
of extinction or the threat associated most with decline. A
main threat was assigned if one threat was reported to be
the major one, whether it be habitat loss (deforestation, agri-
cultural clearing, fragmentation, degradation, overgrazing),
overkill (harvesting, hunting, exploitation, persecution or
bycatch) or introduced species (invasive predators or
diseases, predominantly comprising: black rat, Rattus rattus;
domestic cat, Felis catus; red fox, Vulpes vulpes; Indian mon-
goose, Herpestes javanicus; domestic dog, Canis lupus
familiaris; pig, Sus scrofa; brown rat, Rattus norvegicus;
kiore, Rattus exulans; and a trypanosome contracted from
black rats). We also recorded additional threats interacting
with the major extinction driver, such as loss of vegetation
cover exacerbating predation. If no threat was reported to
be the major one (two or three of the three main classes of
threat were responsible, or the likely cause was not known),
we assigned no threat to that species (i.e. ‘threat’ was treated
as missing data). We recorded geographic range rank (range
with a precision of one order of magnitude: 1 ¼up to
1km
2
;2¼1–10 km
2
;7¼100 000–1 000 000 km
2
). We
used a rank because range was often not known with suffi-
cient precision to treat it as a continuous variable. Range
estimates (e.g. [18]) are from different sources, so we
assume that estimates of range are unlikely to vary by more
than an order of magnitude between authors. We defined
search effort as the number of reported search expeditions
targeted to the species after it was reported missing and
before it was rediscovered (if rediscovered). We divided the
number of reported searches into three ranked intervals:
low (0– 2), medium (3– 6) and high (.10). Reported
searches are likely to be underestimates, but we assume
that the three broad categories reflect meaningful relative
ranks of recent search effort, because except in the case of
undisputed pre-20th century extinctions (which we ranked
as low effort; e.g. Steller’s sea cow, Hydrodamalis gigas),
authors of the IUCN Red List always noted search effort in
their accounts of extinct and possibly extinct species, and
publications that report rediscoveries invariably discussed
the frequency of previous sightings and unsuccessful search
expeditions.
We assessed the original reported density of each species
under three categories (sparse, 1; locally common, 0.5;
dense, 0), if we found unambiguous published statements
about abundance. For example, the Christmas Island
shrew, C. trichura, was assessed as dense because it was
‘once extremely common all over the island and its distinctive
shrill squeaks could be heard all around as one stood quietly
in the rainforest’ [19]; the desert bandicoot, Perameles
eremiana, was assessed as locally common because ‘according
to Finlayson, it was common in northwestern South Austra-
lia, the southwest of the Northern territory and adjacent
parts of western Australia in the 1930s. Its range extended
as far north as the Tanami desert’ [20]; and the Jamaican
rice rat, Oryzomys antillarum, was assessed as sparse because
‘doubtless this is the field mouse described by P. H. Gosse in
his Naturalist’s sojourn in Jamaica as ...far from numerous’
[21]. We assessed research effort as the mean number of
Web of Science citations in 2009 that had the topic keywords
‘taxonomy’ or ‘conservation’, and an address keyword as the
countries in which the species occurred. We also recorded
discovery date (the date that the type specimen was collected,
not necessarily the date of description); elevation (coastal, up
to 50 m.a.s.l.; mid, 50–1000 m.a.s.l.; high, over
1000 m.a.s.l.); habitat openness (closed: forest or swamp;
open: grassland, desert, rocky coast or shrubland); island
status (island versus continent); body mass (g); last sighting
date; century (19th, 20th or pre-19th); colour (cryptically
coloured, black, grey, brown or white versus conspicuously
coloured, spotted or striped); arboreality (arboreal or terres-
trial); diurnality (diurnal versus nocturnal or crepuscular);
gregariousness (group-living versus solitary); and current
human density rank (mean human density in the geographi-
cal range in categories of 50 km
22
: from 0 ¼up to 1
person km
22
to 15 ¼more than 651 km
22
). Human density
rank was obtained from the most recent data in [18] for most
species, and from the United Nations Population Division
Rediscovery bias in mammals D. O. Fisher & S. P. Blomberg 1091
Proc. R. Soc. B (2011)
on February 24, 2011rspb.royalsocietypublishing.orgDownloaded from

database (http://esa.un.org/unpp/) for some island species,
assuming that human density on the island was representa-
tive of the species range, in mammals restricted to small
islands.
Clavero & Garcia-Berthou [22] noted that the searchable
classification system of threats in the IUCN Red List does
not cover all species, so researchers need to assess detailed
information about the causes of extinction provided in
other fields of the database. Therefore, we assessed all
extended species accounts, in order to extract information
on threatening processes and detection history. We include
our data on missing and rediscovered mammals and sources
as electronic supplementary material (dataset S1 and text S1).
(b)Statistical analysis
We tested for associations between species characteristics and
probability of rediscovery using Cox proportional hazards
regression [23]. We modelled the annual ‘survival’ of miss-
ing/extinct status with respect to potential covariates (i.e.
rediscovery was analogous to death in a survival model, in
which mean survival is compared between groups with differ-
ent traits, and the result expressed as a mean survival curve
for each group). In species missing before the 19th centur y,
there was no variation in search effort, and no species in
which extinction was attributed mainly to habitat loss.
Therefore, we did two tests: (i) using a dataset including
species from all centuries and search effort, but excluding
the ‘century’ variable; and (ii) with a reduced dataset includ-
ing century (19th and 20th), but excluding search effort and
species missing before the 19th century.
Case-wise deletion of species without complete data is
likely to give misleading results in a multiple regression [24],
so we dealt with missing data using multiple imputation (MI
[25]). MI provides unbiased estimates of parameters from
the data, together with standard errors that take account of
the imputation procedure [26]. Five datasets were generated
after 10 000 iterations of a Gibbs sampler using the ‘mice’
package for R [27,28]. Convergence was assessed by visual
examination of the trace plots for each variable. Missing
data were never greater than 6 per cent for any of the variables
in our dataset. Data were imputed for the continuous variables
range rank, log mass, human rank and inverse density using a
Bayesian multiple regression, with the conventional improper
flat prior [29]. The three binary variables colour, diurnality
and gregariousness were imputed using logistic regression.
Threat, an ordinal variable, was imputed using polytomous
regression. Combined inference for the multiply imputed
datasets was conducted using the ‘mitools’ package for R [30].
For analyses of the ‘all centuries’ and the ‘19th and 20th
century’ datasets, we tested the assumptions of Cox
regression (the hazard for any species is a fixed proportion
of the hazard for any other species, risk is multiplicative) by
calculating weighted residuals using the cox.zph function in
R[31]. We tested for multi-collinearity using variance
inflation factors, calculated using the ‘car’ package in R
[32]; no values were .2 (indicating no problematic
multi-collinearity for either dataset [33]).
We did not test for correlates of rediscovery rate using a
method to control for phylogenetic non-independence,
because rate data are strongly skewed (non-normally distri-
buted), and no method currently exists to account for
phylogenetically correlated data in a Cox regression. How-
ever, rediscovered species were dispersed evenly across the
mammal phylogeny, and showed no apparent phylogenetic
signal (see electronic supplementary material, figure S1).
Using the ‘all centuries’ dataset, we tested for correlations
between species characteristics and discovery date, using
both generalized least-squares models to account for phylo-
genetic non-independence and linear models of raw species
data. We report the results of all models using raw species
data, because there was no significant difference between
any of the phylogenetic and non-phylogenetic linear model
results (p.0.99 in each case). Rediscoveries are scattered
across families and genera of mammals, rather than being
clumped in certain phylogenetic lineages (for details of the
phylogeny and tree-building methods, see electronic sup-
plementary material, figure S1).
3. RESULTS
(a)Detectability of extinction in all centuries
We identified 187 mammal species that have been missing
(claimed or suspected to be extinct) since 1500. This
number includes all such mammals for which we were
able to find key variables for analysis. In the complete
dataset, 67 species that were once missing have been
rediscovered. When species from all centuries were
included, mammals that declined mainly due to habitat
loss were much more likely to have been wrongly sus-
pected to be extinct than species affected by overkill or
introduced species. The rate of rediscovery in species
affected by loss of habitat was 3.4 times, and in species
affected by persecution 1.8 times as high as in species
affected by introduced predators and diseases (table 1
and figure 1). Mammals with larger geographical ranges
and lower original population densities were also more
likely to be wrongly suspected to be extinct (table 1).
To check that our results did not depend on interpretation
of the major threat and multiple threat interactions, we
also repeated this analysis using only species listed as
affected by only one threat. The results were qualitatively
unchanged, with the same significant variables and
Table 1. The probability of rediscovery versus predictive variables in the final model of a Cox proportional hazards regression
with multiple imputation, using data from all centuries, excluding the effect of century and including search effort (s.e. is
standard error, pis the p-value for the final imputed dataset). Results from the broad dataset are non-italicized. If different,
results from the restricted dataset (with a narrower definition of ‘missing’) are italicized.
variable coefficient exp(coef) s.e. p
range rank 0.457 0.395 1.58 1.48 0.009 0.101 ,0.0001
threat (habitat loss) 1.217 1.379 3.38 3.97 0.034 0.363 ,0.001
threat (overkill) 0.612 0.841 1.84 2.32 0.390 0.402 0.04
density category 0.762 0.898 2.14 2.46 0.401 0.433 ,0.01
search category (medium) 1.068 0.812 2.91 2.25 0.302 0.326 ,0.001 0.01
1092 D. O. Fisher & S. P. Blomberg Rediscovery bias in mammals
Proc. R. Soc. B (2011)
on February 24, 2011rspb.royalsocietypublishing.orgDownloaded from

relative effect sizes (electronic supplementary material,
table S1). In the reduced dataset with a narrower defi-
nition of ‘missing’, 56 species were rediscovered and 98
were missing. The conclusions based on this reduced
dataset were also unchanged, with the same significant
variables, and even larger effect sizes for the main
correlates (table 1).
The distribution of search effort (the number of
reported search expeditions targeting each missing
species) was highly skewed. Most species were subject
to two or fewer expeditions, but six species in the broad
dataset that were exterminated in the 20th century were
the targets of more than 11 reported searches each
(figure 2). Search effort affected the detection rate of
missing species. Most species with up to two targeted
searches (low effort) remain missing, but most species
with three to six searches (intermediate effort) have
been rediscovered (figure 2). Species with intermediate
effort were rediscovered at a rate 2.9 times higher than
species with low effort (table 1). However, the association
between effort and detectability was nonlinear. None of
the missing species subject to more than 11 rediscovery
attempts have been found. No other modelled variables
(including body size, life history, habitat, appearance,
cryptic habits or density of overlapping human popu-
lations) were significantly associated with the probability
of rediscovery (table 1).
Harvested, persecuted or exploited mammals elicited
more attention than species affected by other threats.
They were targeted by more than twice as many reported
searches as species that declined from habitat loss or
introduced species (F
2,174
¼4.6, p¼0.01; 3.4 +0.93
(s.e.) searches versus 1.5 +0.18 for habitat loss and
1.7 +0.28 for introduced species).
(b)Detectability of extinction in the 19th and 20th
centuries
Habitat loss was not considered a main cause of extinc-
tion until after 1800 (electronic supplementary material,
dataset S1). In the model including only 19th and
20th century extinctions and effects of century, an inter-
action between threat and range size was the only
significant effect. In mammals that declined from habitat
loss, species with larger geographical ranges were much
more likely to be wrongly thought of as extinct, and, con-
versely, claims of extinction in species with very small
ranges have nearly always been confirmed (table 2 and
figure 3). In these species, each order of magnitude
increase in range increased the odds of rediscovery by a
factor of 1.6. Species with the largest ranges (100 000 –
1 000 000 km
2
, rank of 7) therefore had odds of
rediscovery 26.84 times higher than species with the smal-
lest ranges that declined from habitat loss (up to 1 km
2
,
rank of 1; table 2 and figure 3). There was no interaction
effect between range size and overkill, or range size and
introduced species. The results based on the reduced data-
set were unchanged, with the same significant variables
and even larger effect sizes (table 2).
Mammals missing in the 20th century were nearly
three times as likely to be rediscovered as those that dis-
appeared in the 19th century. There was also an
interaction between century and threat; mammals
affected by overkill were 6.62 times as likely to be redis-
covered in the 20th century as in the 19th century.
However, these effects of century were marginally non-
significant (table 2). No other modelled variables
(including body size, life history, habitat, appearance,
cryptic habits, or density of overlapping human popu-
lations) were associated with probability of rediscovery
(table 2).
4. DISCUSSION
A substantial proportion (more than a third) of mammal
species that have been classified as extinct or possibly
extinct, or flagged as missing, have been rediscovered.
Searching for missing species takes substantial effort
and funding, and many missing species have a high scien-
tific or public profile and high potential conservation
importance if found [3,7,8]. It is therefore important
that investigators prioritize their effort to missing species
that are most likely to be detected [7].
The missing species most likely to be rediscovered are
those with large ranges that declined from habitat loss.
Mammal extinctions have been attributed to habitat loss
only in the last two centuries, and our analysis of this
time period showed that larger range size predicted
higher probability of rediscovery only in species affected
by habitat loss. This is consistent with most evidence in
birds, showing that habitat loss causes disproportionate
global and local extinction of restricted range endemics
in comparison with other threats [13], and with models
showing that endemics– area relationships predict
50 100 150 200 250
y
ears
probability of rediscovery
300 350 400 450 500
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1.0
0
Figure 1. Cox proportional hazards regression showing the
mean difference in the probability of rediscovery between
species affected by different threats, in all centuries. Habitat
loss (including fragmentation and degradation), dotted line;
introduced species (including invasive diseases and pre-
dators), dashed line; overkill (including harvesting,
persecution and exploitation), solid line. The y-axis shows
the probability of rediscovery for each category per year.
Time zero is when all species in the category are missing
(0 ¼zero rate of rediscovery, 1 ¼100% rediscovered). Cox
proportional hazards regression does not extrapolate
beyond the data: the curve for species affected by habitat
loss ends at 180 years because the earliest suspected extinc-
tion attributed to habitat loss was 180 years ago (the
Bahian tree rat, Phyllomys unicolor). Several species affected
by introduced species or overkill disappeared in the 16th
century (n¼187).
Rediscovery bias in mammals D. O. Fisher & S. P. Blomberg 1093
Proc. R. Soc. B (2011)
on February 24, 2011rspb.royalsocietypublishing.orgDownloaded from

extinction rates from habitat loss better than species–area
relationships [34].
Our finding that mammals affected by habitat loss are
most likely to be rediscovered suggests that the current
number of species considered extinct owing to habitat
loss is likely to be overestimated. Because small range is
a variable used to both ascertain extinction risk and
assign the cause (habitat loss), circularity might lead to
overestimation of the proportion of extinctions that are
due to habitat loss [11]. Severely declined mammals are
number of searches
number of species
Peromyscus pembertoni
Peromyscus guardia
Lipotes vexillifer
Equus ferus
Bos sauveli
Thylacinus cynocephalus
1 2 3 4 5 6 7 8 9 101112131415161718192021222324 250
5
10
15
20
25
30
35
40
Figure 2. Frequency distribution of the number of reported searches targeting particular missing species, in all centuries. Redis-
covered species are white bars, missing species are black bars. Species with high numbers of searches (12, all unsuccessful) are
Pemberton’s deer mouse (Peromyscus pembertoni) and the Angel Island deer mouse (Peromyscus guardia) from Mexico; the baiji
(Lipotes vexillifer) from China; wild horse (Equus ferus) from China, Mongolia, Belarus, Germany, Kazakhstan, Lithuania,
Poland, the Russian Federation and Ukraine; kouprey (Bos sauveli) from Thailand, Laos and Cambodia; and the thylacine
(Thylacinus cynocephalus) from Australia (for an explanation of sources for search effort, see electronic supplementary material,
text S1; n¼178).
50 100 150 200 250
years
probability of rediscovery
300 350 400 450 5000
0.1
0
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1.0
(a)(b)
500 100 150 200 250
years
300 350 400 450 500
Figure 3. Cox proportional hazards regression showing mean probability of rediscovery for species affected by habitat loss in
seven categories of range size. All species were last recorded in the 19th or 20th century. (a) Small geographical ranges
(thick solid line, ,1km
2
; thick dotted line, 1 –10 km
2
; thin solid line, 11 – 100 km
2
; thin dotted line, 101 –1000 km
2
);
(b) larger geographical ranges (thick solid line, 1001–10 000 km
2
; dotted line, 10 001– 100 000 km
2
; thin solid line,
100 001–1 000 000 km
2
). Time zero is when all species in the category are missing (0 ¼zero rate of rediscovery, 1 ¼100%
rediscovered; n¼164).
Table 2. The probability of rediscovery versus predictive variables in the final model of a Cox proportional hazards regression
with multiple imputation, using data from the 19th and 20th centuries, including the effect of century and excluding search
effort (s.e. is standard error, pis the p-value for the final imputed dataset). Results from the broad dataset are non-italicized.
If different, results from the restricted dataset (with a narrower definition of ‘missing’) are italicized.
variable coefficient exp(coef) s.e. p
century 1.07 0.93 2.92 2.53 0.57 0.59 0.061 0.12
threat (overkill) : century (20th) 1.89 2.45 6.62 11.59 0.99 1.23 0.058 0.046
threat (habitat loss) : range rank 0.47 0.51 1.59 1.67 0.19 0.21 0.016
1094 D. O. Fisher & S. P. Blomberg Rediscovery bias in mammals
Proc. R. Soc. B (2011)
on February 24, 2011rspb.royalsocietypublishing.orgDownloaded from

likely to be considered as specialists on their last detected
habitat. Now being restricted to a small range, they will be
categorized as threatened by habitat loss, even if the cause
of previous decline was different. This will inflate per-
ceived extinction risk owing to habitat loss if some of
these species actually persist undetected in other habitats
or distant sites. Fisher [35] found that species affected by
habitat loss are more likely to be rediscovered at the per-
iphery than the centre of their former range, suggesting
that spreading habitat change has pushed them to the
range edge, and that high human population pressure
was associated with rediscovered species changing habitat
from previous records in primary forest, to rediscovery in
marginal habitat such as regrowth, cropland and planta-
tions. Both of these effects are likely to make mammals
that have declined from habitat loss particularly hard to
detect. We found no significant effect of human popu-
lation density on the probability of rediscovery, although
increased frequency of extinction from habitat loss and
overkill might be expected in more populated regions. It
is possible that this effect was cancelled out because
there were also more opportunities for rediscovery in
populated areas, because of increased encounter rates
and number of people with identification skills.
Across all centuries, range size was strongly correlated
with the probability of rediscovery of missing mammals,
and species with very small ranges were unlikely to be
rediscovered. This might simply be due to the elevated
extinction risk associated with small ranges. All recent
analyses have concluded that small range size and the clo-
sely correlated trait of small population size are the most
important indicators of extinction risk and declines of
threatened mammals [10,11,36]. The rediscovery rate
of species with large ranges might also be higher because
scattered remnant populations are more likely to escape
detection, an explanation reinforced by the finding that
species originally occurring at lower population density
were rediscovered at higher rates, despite the fact that
low population density predicts extinction risk in mam-
mals generally [10]. This interpretation also seems
inconsistent with previous assertions that large
geographical range is the best predictor of early species
description, because it increases the encounter rate with
collectors [37]. However, the high detectability of initially
widespread species before decline, and their low
detectability after decline, makes sense if they contracted
to a very small range that was not at the site where
they were last seen but one anywhere within the
former wide distribution, or at a remote edge of it
[35,38].
Our finding that, throughout historical time, species
with small ranges are unlikely to be rediscovered is not
an effect of island endemics being extirpated by
introduced predators. Unlike birds, which are dispropor-
tionately exterminated by predators introduced to islands
[39,40], invasive predators have had continental-scale
impacts on mammal extinction rates [2,41,42]. Being
restricted to islands was not correlated with the prob-
ability of rediscovery in our analyses (60% of
rediscoveries were on continents). We found that, overall,
mammals were unlikely to be rediscovered if the cause of
extinction was an introduced predator or disease, but they
were likely to be rediscovered if the cause was habitat loss.
This conclusion parallels recent findings in birds.
Although more birds are classified as threatened to
some degree by habitat loss than by biological
invasion, bird families threatened mainly by invasive
species are more extinction-prone, and families contain-
ing species primarily threatened by habitat loss are less
extinction-prone [39].
Moderate search effort was associated with increased
rediscoveries, in comparison with low search effort. We
could not separate this from the effect of century, because
all species missing in the 19th century and before were
subject to low search effort (two or fewer expeditions),
except for the Talaud flying fox (Acerodon humilis),
which was missing in 1897 and found alive in 1999
after three searches. Most missing mammals have not
been adequately searched for, but a few flagship species
(charismatic large mammals) received disproportionately
high numbers of searches. The highest search effort in
our dataset was confined to a handful of species that
remain missing, namely the thylacine (Thylacinus
cynocephalus), wild horse (Equus ferus, extinct in the
wild), kouprey (Bos sauveli) and Baiji (Lipotes vexillifer).
We suggest that this is because it is possible to keep
searching indefinitely without success if the species is
actually extinct. These large-bodied mammals all
declined mainly from overkill. However, body size did
not independently predict rediscovery rate in any of our
models, although persecuted and harvested species are
predominantly large and conspicuous [11,13,43]. A
species must be identifiable and detectable to be perse-
cuted, exploited or harvested, so publicity about its
supposed extinction is also more likely, which might
result in more search effort. Our data suggest that mam-
mals purportedly exterminated by overkill receive more
attention, because they were targeted by more than
twice as many reported searches on average as the more
enigmatic species that declined from habitat loss or intro-
duced predator impacts. Mammals that declined from
human persecution were more likely to be rediscovered
than those presumed to have been driven extinct by
introduced species, particularly in the 20th century.
Increased public attention and searching probably explain
why species that declined in the 20th century tended to be
rediscovered more frequently, especially if they declined
from overkill.
Our major findings are robust to varying definitions
and time scales, because the same conclusions were
important whether we used the overall dataset with a
broad definition of missing species, or subsections (19th
and 20th centuries only, species with one reported
threat only, or the restricted dataset of species with a nar-
rower definition of ‘missing’).
5. CONCLUSIONS
Rediscovery in purportedly extinct and missing mammals
is not a random process, but the chance of success
depends on search effort, search area, time missing and
traits known to be associated with extinction risk such
as population density and range size, which interacts
with the cause of extinction as predicted by theory. Past
effort has focused on a handful of species. Rather than
allocating even more effort to these charismatic mammals
affected by overkill that are certainly extinct, such as the
thylacine, we recommend particularly targeting neglected
Rediscovery bias in mammals D. O. Fisher & S. P. Blomberg 1095
Proc. R. Soc. B (2011)
on February 24, 2011rspb.royalsocietypublishing.orgDownloaded from

species missing later than the 18th century, with relatively
large ranges, threatened by habitat loss. It is most likely
that some of these species survive, and locating them
will enable us to protect their final habitats and avert
extinction.
We thank Kate Jones, Jaime Jiminez, David & Meredith
Happold, Andrew Cockburn, Hideki Endo, Rainer
Hutterer, Carla Kishinami, Stefan Klose, Friederike
Spitzenberger, Craig Hilton-Taylor and Richard Fuller for
providing or helping us to locate data, and Kerrie Wilson,
Peter Baxter, Hugh Possingham, James Brown, Anne
Goldizen and anonymous reviewers for discussions and/or
comments. This work was supported by funding from the
Australian Research Council (ARF DP0773920).
REFERENCES
1 Keith, D. A. & Burgman, M. A. 2004 The Lazarus effect:
can the dynamics of extinct species lists tell us anything
about the status of biodiversity? Biol. Conserv. 117,
41–48. (doi:10.1016/S0006-3207(03)00261-1)
2 MacPhee, R. D. E. & Flemming, C. 1999 Requiem aeter-
nam: the last five hundred years of mammalian species
extinctions (database: http://www.creo.amnh.org). In
Extinctions in near-time (ed. R. D. E. MacPhee),
pp. 333– 371. New York, NY: Kluwer Academic/Plenum.
3 Diamond, J. M. 1987 Extant unless proven extinct? Or,
extinct unless proven extant? Conserv. Biol. 1, 77– 79.
(doi:10.1111/j.1523-1739.1987.tb00012.x)
4 Collar, N. J. 1998 Extinction by assumption; or, the
Romeo Error on Cebu. Oryx 32, 239 – 244. (doi:10.
1046/j.1365-3008.1998.d01-51.x)
5 Mace, G. M., Collar, N. J., Gaston, K. J., Hilton-Taylor,
C., Akcakaya, H. R., Leader-Williams, N., Milner-
Gulland, E. J. & Stuart, S. N. 2008 Quantification of
extinction risk: IUCN’s system for classifying threatened
species. Conserv. Biol. 22, 1424– 1442. (doi:10.1111/j.
1523-1739.2008.01044.x)
6 Solow, A. R. 2005 Inferring extinction from a sighting
record. Math. Biosci. 195, 47–55. (doi:10.1016/j.mbs.
2005.02.001)
7 Roberts, D. L., Elphick, C. S. & Reed, J. M. 2010
Identifying anomalous reports of putatively extinct
species and why it matters. Conserv. Biol. 24, 189 – 196.
(doi:10.1111/j.1523-1739.2009.01292.x)
8 Scott, J. M., Ramsey, F. L., Lammertink, M., Rosenberg,
K. V., Rohrbaugh, R., Wiens, J. A. & Reed, J. M. 2008
When is an ‘extinct’ species really extinct? Gauging the
search efforts for Hawaiian forest birds and the Ivory-
billed Woodpecker. Avian Conserv. Ecol. 3,3.
9 Cardillo, M., Mace, G. M., Gittleman, J. L., Jones, K. E.,
Bielby, J. & Purvis, A. 2008 The predictability of
extinction: biological and external correlates of decline
in mammals. Proc. R. Soc. B 275, 1441 – 1448. (doi:10.
1098/rspb.2008.0179)
10 Davidson, A. D., Hamilton, M. J., Boyer, A. G., Brown,
J. H. & Ceballos, G. 2009 Multiple ecological pathways
to extinction in mammals. Proc. Natl Acad. Sci. USA
106, 10 702 –10 705. (doi:10.1073/pnas.0901956106)
11 Fisher, D. O. & Owens, I. P. F. 2004 The comparative
method in conservation biology. Trends Ecol. Evol. 19,
391–398. (doi:10.1016/j.tree.2004.05.004)
12 Hockey, P. A. R. & Curtis, O. E. 2009 Use of basic
biological information for rapid prediction of the
response of species to habitat loss. Conserv. Biol. 23,
64–71. (doi:10.1111/j.1523-1739.2008.01028.x)
13 Owens, I. P. F. & Bennett, P. M. 2000 Ecological basis of
extinction risk in birds: habitat loss versus human
persecution and introduced predators. Proc. Natl Acad.
Sci. USA 97, 12 144 –12 148. (doi:10.1073/pnas.
200223397)
14 Purvis, A., Gittleman, J. L., Cowlishaw, G. & Mace,
G. M. 2000 Predicting extinction risk in declining
species. Proc. R. Soc. Lond. B 267, 1947 – 1952. (doi:10.
1098/rspb.2000.1234)
15 Finlayson, H. H. 1932 Rediscovery of Caloprymnus
campestris (Marsupialia). Nature 129, 871 – 871. (doi:10.
1038/129871b0)
16 Carr, S. G. & Robinson, A. C. 1997 The present status
and distribution of the desert rat-kangaroo Caloprymnus
campestris (Marsupialia: Potoroidae). S. A. Naturalist
72, 4–27.
17 IUCN. 2009 IUCN red list of threatened species. Version
2009.2. See http://www.iucnredlist.org.
18 Jones, K. E. et al. 2009 PanTHERIA: a species-level
database of life-history, ecology and geography of extant
and recently extinct mammals. Ecology 90, 2648 – 2648.
(doi:10.1890/08-1494.1)
19 Schultz, M. 2002 National recovery plan for the
Christmas Island shrew, Crocidura attenuata trichura.
Canberra, Australia: Department of the Environment
and Heritage.
20 Gordon, G. 2008 Desert bandicoot. In The Australian
Museum complete book of Australian mammals (eds S. V.
Dyck & R. Strahan), pp. 185– 186. Sydney, Australia:
Reed Books.
21 Allen, G. M. 1942 Extinct and vanishing mammals of the
Western Hemisphere with the marine species of all the
oceans. New York, NY: American Committee for
International Wild Life Protection.
22 Clavero, M. & Garcia-Berthou, E. 2005 Invasive species
are a leading cause of animal extinctions. Trends Ecol.
Evol. 20, 110– 110. (doi:10.1016/j.tree.2005.01.003)
23 Hosmer, D. W. J. & Lemeshow, S. 2000 Applied survival
analysis: regression modeling of time to event data. Malden,
MA: Wiley-Interscience.
24 Nakagawa, S. & Freckleton, R. P. 2008 Missing in action:
the dangers of ignoring missing data. Trends Ecol. Evol.
23, 592–596. (doi:10.1016/j.tree.2008.06.014)
25 Little, R. J. A. & Rubin, D. B. 2002 Statistical analysis
with missing data. Hoboken, NJ: John Wiley and Sons Inc.
26 Fisher, D. O., Blomberg, S. P. & Owens, I. P. F. 2003
Extrinsic versus intrinsic factors in the decline and extinc-
tion of Australian marsupials. Proc. R. Soc. Lond. B 270,
1801– 1808. (doi:10.1098/rspb.2003.2447)
27 R Development Core Team. 2009 R: A language and
environment for statistical computing. Vienna, Austria: R
Foundation for Statistical Computing. See http://www.
R-project.org.
28 van Buuren, S. & Oudshoorn, C. G. M. 2009 MICE:
multivariate imputation by chained equations. See http://
www.stefvanbuuren.nl/publications/MICEinR-Draft.pdf.
29 Rubin, D. B. 1987 Multiple imputation for nonresponse in
surveys. New York, NY: John Wiley and Sons.
30 Lumley, T. 2009 mitools: tools for multiple imputation of
missing data. Vienna, Austria: R Foundation for Statistical
Computing.
31 Grambsch, P. & Therneau, T. 1994 Proportional hazards
tests and diagnostics based on weighted residuals.
Biometrika 81, 515 – 526. (doi:10.1093/biomet/81.3.515)
32 Fox, J. et al. 2009 Car: companion to applied regression.
R package version 1.2-16. Vienna, Austria: R project.
See http://www.CRAN.R-project.org/package=car.
33 Menard, S. 1995 Applied logistic regression analysis.
Thousand Oaks, CA: Sage Publications.
34 Kinzig, A. P. & Harte, J. 2000 Implications of endemics-
area relationships for estimates of species extinctions.
Ecology 81, 3305–3311.
1096 D. O. Fisher & S. P. Blomberg Rediscovery bias in mammals
Proc. R. Soc. B (2011)
on February 24, 2011rspb.royalsocietypublishing.orgDownloaded from

35 Fisher, D. O. In press. Trajectories from extinction:
where are missing mammals rediscovered? Glob. Ecol.
Biogeogr.
36 Cardillo, M., Mace, G. M., Jones, K. E., Bielby, J.,
Bininda-Emonds, O. R. P., Sechrest, W., Orme,
C. D. L. & Purvis, A. 2005 Multiple causes of high
extinction risk in large mammal species. Science 309,
1239–1241. (doi:10.1126/science.1116030)
37 Collen, B., Purvis, A. & Gittleman, J. L. 2004 Biological
correlates of description date in carnivores and primates.
Glob. Ecol. Biogeogr. 13, 459– 467. (doi:10.1111/j.1466-
822X.2004.00121.x)
38 Channell, R. & Lomolino, M. V. 2000 Dynamic
biogeography and conservation of endangered species.
Nature 403, 84–86. (doi:10.1038/47487)
39 Clavero, M., Brotons, L., Pons, P. & Sol, D. 2009
Prominent role of invasive species in avian biodiversity
loss. Biol. Conserv. 142, 2043– 2049. (doi:10.1016/j.
biocon.2009.03.034)
40 Duncan, R. P. & Blackburn, T. M. 2007 Causes of
extinction in island birds. Anim. Conserv. 10, 149 – 150.
(doi:10.1111/j.1469-1795.2007.00110.x)
41 Johnson, C. N. 2006 Australia’s mammal extinctions:
a 50 000 year history. Cambridge, UK: Cambridge
University Press.
42 Johnson, C. N., Isaac, J. L. & Fisher, D. O. 2007 Rarity
of a top predator triggers continent-wide collapse of
mammal prey: dingoes and marsupials in Australia.
Proc. R. Soc. B 274, 341 – 346. (doi:10.1098/rspb.2006.
3711)
43 Isaac, N. J. B. & Cowlishaw, G. 2004 How
species respond to multiple extinction threats.
Proc. R. Soc. Lond. B 271, 1135 – 1141. (doi:10.1098/
rspb.2004.2724)
Rediscovery bias in mammals D. O. Fisher & S. P. Blomberg 1097
Proc. R. Soc. B (2011)
on February 24, 2011rspb.royalsocietypublishing.orgDownloaded from