Heredity (2015) 115, 63–72

& 2015 Macmillan Publishers Limited All rights reserved 0018-067X/15

www.nature.com/hdy

ORIGINAL ARTICLE
Measuring individual inbreeding in the age of genomics:
marker-based measures are better than pedigrees
M Kardos1,2, G Luikart3 and FW Allendorf1

Inbreeding (mating between relatives) can dramatically reduce the fitness of offspring by causing parts of the genome to be
identical by descent. Thus, measuring individual inbreeding is crucial for ecology, evolution and conservation biology. We used
computer simulations to test whether the realized proportion of the genome that is identical by descent (IBDG) is predicted
better by the pedigree inbreeding coefficient (FP) or by genomic (marker-based) measures of inbreeding. Genomic estimators of
IBDG included the increase in individual homozygosity relative to mean Hardy–Weinberg expected homozygosity (FH), and two
measures (FROH and FE) that use mapped genetic markers to estimate IBDG. IBDG was more strongly correlated with FH, FE and
FROH than with FP across a broad range of simulated scenarios when thousands of SNPs were used. For example, IBDG was more
strongly correlated with FROH, FH and FE (estimated with ⩾ 10 000 SNPs) than with FP (estimated with 20 generations of
complete pedigree) in populations with a recent reduction in the effective populations size (from Ne = 500 to Ne = 75).
FROH, FH and FE generally explained 490% of the variance in IBDG (among individuals) when 35 K or more SNPs were used.
FP explained o80% of the variation in IBDG on average in all simulated scenarios, even when pedigrees included 20
generations. Our results demonstrate that IBDG can be more precisely estimated with large numbers of genetic markers than with
pedigrees. We encourage researchers to adopt genomic marker-based measures of IBDG as thousands of loci can now be
genotyped in any species.
Heredity (2015) 115, 63–72; doi:10.1038/hdy.2015.17; published online 18 March 2015

INTRODUCTION with few chromosomes and short genetic maps (Franklin, 1977; Stam,
Biologists have long recognized that inbred individuals (those whose 1980; Hill and Weir, 2011; Figure 1 and Supplementary Figure S1).
parents are closely related) often have lower fitness than the offspring of Furthermore, FP cannot account for inbreeding caused by distant
unrelated parents (Darwin, 1868; Ives and Whitlock, 2002). The ancestors not included in a pedigree.
cumulative effects of inbreeding on individual fitness can reduce the The increasing availability of large-scale molecular genetic data
population growth rate and the probability of persistence (O'Grady et al., might make it possible to more precisely estimate IBDG with genetic
2006; Kenney et al., 2014). Effects of inbreeding on individual fitness and markers than with pedigrees (Keller et al., 2011; Hoffman et al., 2014).
population dynamics make measuring individual inbreeding a crucial For example, IBDG can be estimated as the increase in individual
part of many studies in ecology, evolution and conservation biology. homozygosity relative to Hardy–Weinberg expected homozygosity
Inbred individuals have lower genome-wide heterozygosity because (FH) (Purcell et al., 2007). Other measures of IBDG use mapped
a fraction of loci are ‘identical by descent’ (IBD). A locus is identical genetic markers to identify IBD chromosome segments (referred to
by descent if it carries two gene copies that both originated from a here as IBD tracts), which are characterized by long contiguous runs of
single copy in a common ancestor of the parents. All measures of homozygosity (ROH) at mapped genetic markers (Chapman and
individual inbreeding seek to predict the proportion of the genome Thompson, 2003; McQuillan et al., 2008). The FROH statistic estimates
that is identical by descent (IBDG). The classical measure of individual IBDG as the proportion of the genome within inferred ROH that
inbreeding is the pedigree inbreeding coefficient FP (Wright, 1922; satisfy user-defined criteria (for example, number of SNPs, physical
Keller and Waller, 2002). FP predicts the IBDG due to the known length and SNP density) (Purcell et al., 2007; McQuillan et al., 2012).
common ancestors of parents and assumes that the pedigree founders Leutenegger et al. (2003) introduced a maximum likelihood estimator
are unrelated and non-inbred. FP has often been considered as the best (FE) that uses mapped genetic markers and a hidden Markov model to
measure of individual inbreeding (Pemberton, 2004,2008). However, identify putative IBD tracts and to estimate IBDG.
FP may be an imprecise measure of IBDG because IBDG can vary A thorough evaluation of the relative performance of pedigree-
substantially among individuals with the same pedigree (for example, versus marker-based measures of IBDG in small populations would
siblings, Figure 1 and Supplementary Figure S1) (Franklin, 1977; Hill advance our understanding of how individual inbreeding should be
and Weir, 2011; Forstmeier et al., 2012). The standard deviation of measured in natural populations of conservation concern as we enter
IBDG among individuals with the same pedigree is highest in species an era when thousands of loci can be genotyped for any organism.

1
Division of Biological Sciences, University of Montana, Missoula, MT, USA; 2Department of Evolutionary Biology, Evolutionary Biology Centre (EBC), Uppsala University, Uppsala,
Sweden and 3Flathead Lake Biological Station, Division of Biological Sciences, University of Montana, Polson, MT, USA
Correspondence: Dr M Kardos, Division of Biological Sciences, University of Montana, 822 Defoe Street, Missoula, MT, 59802, USA.
E-mail: [email protected]
Received 22 July 2014; revised 3 February 2015; accepted 13 February 2015; published online 18 March 2015
 Measuring individual inbreeding
M Kardos et al
64

deviation of IBDG among individuals with the same pedigree

σ(IBDG)) when genetic map lengths are in the range we simulated
(Supplementary Figure S1). We used two different recombination
rates to evaluate the effect of genetic map length on the precision of
FP, FROH, FE and FH. We simulated genomes with 1.2 cM/Mb (that is,
an average of approximately 0.012 recombinations per Mb per
meiosis) which is similar to humans (Dumont and Payseur, 2008)
and resulted in a genetic map length of 3600 cM. We also simulated
genomes with 0.27 cM/Mb which is typical of the lower end of the
distribution of recombination rates among mammals (Dumont and
Payseur, 2008) and resulted in a genetic map length of 800 cM.
Our model of recombination follows the simulations of Chapman
and Thompson (2003) and is based on Fisher’s theory of junctions
(Fisher, 1965). We assume no interference, that the number of
crossovers along a chromosome is Poisson distributed, and that the
recombination probability is constant across the genome and among
individuals.
Population founders were unrelated and non-inbred. Thus, IBDG
for an individual is relative to the population founders. Founders for
each simulated population were drawn from a hypothetical source
population in Hardy–Weinberg equilibrium. Allele frequencies at
400 K (K = 1000) SNPs in the source populations were assigned so
that the mean expected heterozygosity (He) was approximately 0.3 at
the end of the simulations before SNP filtering steps (see below). To
Figure 1 The distribution of IBDG among 5000 simulated offspring of full
account for the effects of SNP heterozygosity (Miller et al., 2014) on
siblings with 20 chromosomes of equal length, and 800 cM (upper panel) or
3600 cM (lower panel) genomes. our results, we repeated simulations of populations with 3600 cM
genomes with founder allele frequencies assigned so that He was only
Using simulations, Keller et al. (2011) found that the number of approximately 0.2, and analyzed these data without filtering based on
homozygous rare alleles within an individual was more strongly minor allele frequency (MAF) (beyond removing fixed loci). SNP
correlated with marker-based measures of inbreeding than with FP. positions were determined using a random number generator so that
each nucleotide position in the genome was equally likely to be
However, they simulated large populations with 220 cM genomes and
polymorphic in the source population. Individual founder genotypes
only 2 chromosomes. Thus, their results likely underestimate the
were assigned by randomly choosing two alleles at each locus based on
precision of FP as a measure of IBDG in humans and other species
their frequencies in the source population (see Kardos et al., 2014 for
with much longer genetic maps and more chromosomes (Figure 1 and
further details on the assignment of founder genotypes). Simulation
Supplementary Figure S1; Franklin, 1977).
output included IBDG and genotypes at 300–350 K diallelic SNPs,
Our objective was to test whether IBDG is predicted better by
which were left after removing loci with MAFo0.05 for each
pedigrees or by genetic markers in small populations. Specifically, we
individual in the final generation. SNPs with MAFo0.05 were
evaluated the precision and bias of FP, FROH, FE and FH in small
removed to maximize the information content of the genotypes.
populations that are of the greatest concern for conservation with
Statistical analyses were performed only on individuals from the last
genomic characteristics typical of mammals. Our simulations generation to hold the number of pedigree generations used to
accounted for the depth of the pedigree used to estimate FP, the estimate FP constant among individuals within each analysis, and
number of markers used to estimate FROH, FE and FH, the genetic map to hold the sample sizes for statistical tests constant across all
length of the genome and demographic history. We emphasize simulations. An R script for the simulations is available in Dryad
comparisons of the performance of FP versus the marker-based (see Data Accessibility).
measures of IBDG.
Defining IBDG
MATERIALS AND METHODS We defined IBDG as the realized proportion of the physical genome in
Computer simulation model IBD tracts. IBDG was measured without error because the exact
We used a stochastic, individual-based simulation program (Kardos boundaries of IBD tracts were known to the base pair. Thus, IBDG is
et al., 2014) for R version 3.0.1 (R Core Team, 2013). We simulated a a parameter representing the true individual genomic inbreeding. The
sexually reproducing, random mating, non-selfing species with non- IBDG was calculated from the simulated genomic data once for each
overlapping generations. We simulated hermaphrodites to avoid individual. The mean and variance of IBDG for each simulated
temporal fluctuations in the effective populations size (Ne) due to scenario are shown in Supplementary Table S2.
stochastic variation in the sex ratio. We simulated recombination and
kept track of the ancestral origin of each chromosome segment, which Simulated demographic scenarios
was known to the base pair. We simulated genomic characteristics We simulated two demographic scenarios to account for differences in
typical of mammals, including 3 Gb genomes with 20 chromosomes of the variance of IBDG, and to mimic populations where researchers are
equal size, and mutation with a base change probability of 1.2 × 10 − 8 often interested in inbreeding and its effects on fitness. First, we
per base pair per generation. We simulated only 20 chromosomes simulated partially isolated populations with a constant local Ne of 75
because chromosome number does not strongly affect the standard for 150 generations, receiving occasional immigrants (m = 0.013, one

Heredity
 Measuring individual inbreeding
M Kardos et al
65

immigrant per generation on average). This scenario mimics small used, regardless of the parameter settings used in PLINK. For example,
populations on habitat islands with occasional immigration from a large fewer than two SNPs are expected to occur in an 8-Mb IBD tract on
source population. Immigrants were unrelated to each other and to average when only 500 randomly distributed SNPs are used in a 3-Gb
residents. Immigrants were drawn from the same hypothetical source genome. We adjusted the required number of SNPs and SNP density
population as the founders and thus had genotypes drawn from the (maximum Kb per SNP on average within an ROH) per inferred ROH
same allele frequency distribution as the founders. Second, we simulated based on the number of loci used (Supplementary Table S1). We
closed populations that were large (constant Ne = 500) for 90 genera- allowed a single heterozygous position in any inferred ROH to account
tions, and experienced a reduction in size to Ne = 75 for 10 additional for the occasional mutation occurring within otherwise IBD chromosome
generations. We ran 20 replicate simulations for each scenario. segments. Initial testing showed that adjusting PLINK parameters
We wanted to test whether our results were substantially different according to the number of SNPs and LD pruning substantially
when all inbreeding was due to recent ancestors. Therefore, we reran improved the performance of FROH when ⩽ 25 K SNPs were used.
the simulations as described above, except now running the simula- Consistent with our definition of IBDG above, we estimated FROH as
tions for only 20 generations (not 100–150 generations as above). Here, the sum of the lengths of all detected ROH detected by PLINK divided
all inbreeding is due to recent (within 20 generations) ancestors, and FP by the physical genome size (3 Gb).
estimated with a 20 generation pedigree accounts for all common The second measure of IBDG we used (FH) measures the excess in
ancestors of parents. For these simulations, partially isolated popula- the observed number of homozygous genotypes within an individual
tions had a constant local Ne of 75 for 20 generations, and received one relative to the mean number of homozygous genotypes expected
immigrant per generation on average (m = 0.013). Populations with under random mating:
recently reduced Ne were again closed (m = 0) and had Ne = 500 for the
first 10 generations, then Ne = 75 for the last 10 generations. OðHomi Þ  EðHomÞ
F Hi ¼ ;
m  EðHomÞ
Linkage disequilibrium pruning where O(Homi) is the observed number of homozygous loci for the ith
Before estimating FROH, FE and FH, we used the—indep-pairwise individual, andEðHomÞis the Hardy–Weinberg expected mean num-
function in PLINK version 1.07 (Purcell et al., 2007) to reduce linkage ber of homozygous genotypes across m loci (Keller et al., 2011). The
disequilibrium among the 300–350 K SNPs remaining after MAF- average OðHomÞ is expected to equal EðHomÞ on average in random
based filtering. We conducted ‘LD pruning’ because FE assumes all mating populations. FH can thus take negative values (like some other
markers are in linkage equilibrium (Polašek et al., 2010). Additionally, marker-based measures of inbreeding or relatedness; Wang, 2014),
FROH is typically estimated after LD pruning to avoid detecting fixed and will be centered near zero in random mating populations when
or nearly fixed ROH arising due to ancient population processes (for allele frequencies are taken from the current sample of individuals.
example, selective sweeps) (for example, McQuillan et al., 2012). For FH should thus be considered as a proxy rather than
LD pruning, we used a sliding window of 25 SNPs, a step size of 10 a direct estimator of IBDG, which is a proportion with range 0–1.
SNPs and we removed one SNP from each pair in a window above the We estimated FH using the—het function in PLINK.
LD threshold of r2 = 0.5. Our results were not substantively affected by The third marker-based estimator (FE) is implemented in the
increasing the size of the sliding window for LD pruning to 50 SNPs, program FEstim and uses maximum likelihood and a hidden Markov
or by using a more stringent LD pruning threshold (r2 = 0.25) model to determine the values of IBDG that maximize the likelihood
(data not shown). of the observed genotypes at mapped loci (Leutenegger et al., 2003). FE
We conducted statistical analyses on 500 SNPs to 100 K SNPs has a theoretical range of 0–1 and is thus directly proportional to
randomly sampled from those remaining after MAF- and LD-based IBDG. Starting values for both FE and the IBD rate of change
filtering. We used only up to 100 K SNPs because preliminary parameter were 0.1. Default FEstim settings were used for other
simulations showed that using more loci did not substantively parameters. FEstim did not consistently run to completion on data
affect our results (data not shown). Mean He after filtering based on from partially isolated populations when more than 10 K SNPs were
MAF and LD was 0.31 in partially isolated populations with 3600 cM used, even when using more stringent LD pruning settings in PLINK
genomes, and 0.33 on average in partially isolated populations with (LD threshold of r2 = 0.25, and a sliding window size of 50 SNPs). This
800 cM genomes. He was 0.35 on average in populations with recently is likely due to higher LD remaining among closely linked SNPs in the
reduced Ne. These levels of SNP heterozygosity are similar to partially isolated populations than in populations with recently
large SNP chip data sets from mammals and birds (for example, reduced Ne after LD pruning (Supplementary Figure S2). Using more
Deatwyler et al., 2010; Kawakami et al., 2014). In our simulations stringent LD pruning thresholds (for example, r2 = 0.1) left too few
of populations with low heterozygosity SNPs (He≈0.2), the mean SNPs to usefully compare FE and the other IBDG estimators. Thus, we
realized He after removing fixed loci and LD pruning was 0.21 in present results from FE in partially isolated populations for 10 K and
populations with recently reduced Ne and 0.19 in partially isolated fewer SNPs only. Allele frequencies used to estimate FH and FE were
populations. from the 75 individuals in the last generation of each simulation. We
used the kinship2 package in R to estimate FP for each individual using
Estimators of IBDG 3–20 generations of pedigree.
We used three marker-based measures of IBDG. We used
the—homozyg function in PLINK to estimate FROH using 5–100 K Comparing the performance of FP, FH, FE and FROH
SNPs. PLINK uses a sliding window approach and user defined criteria We used the proportion of variance in IBDG explained by FROH, FE,
to identify ROH, which are inferred to be putative IBD tracts. FROH is FH and FP (r2) from linear regression models to evaluate the precision
estimated as the proportion of the genome in ROH, and ranges from of FROH, FE, FH and FP (N = all 75 individuals from the last simulated
zero to one. Thus, FROH is directly proportional to IBDG. We used generation). We conducted separate regressions of FROH, FH, FE and
only ⩾ 5 K SNPs to estimate FROH because we believed that SNP FP versus IBDG for individuals in the final generation of each
density would be too low to reliably detect ROH when fewer loci are simulated population. We used two-tailed t-tests to determine whether

Heredity
 Measuring individual inbreeding
M Kardos et al
66

the mean r2 from regressions with IBDG (among 20 replicate RESULTS

simulations) was statistically significantly different for FROH, FH, FE We first present results on the precision and bias of FP, FH, FE and
and FP. We compared all possible combinations of pedigree depth (for FP) FROH in simulations of 3600 cM genomes where inbreeding was due to
and number of SNP used (for FH, FE, and FROH) when testing if the both recent and distant ancestors (that is, simulations ran for 100–150
mean r2 with IBDG was statistically significantly different for FP, FH, FE generations). We then summarize how the results change when
and FROH. We also used t-tests to determine whether the mean r2 from inbreeding is due only to recent ancestors (that is, simulations ran
regressions with IBDG was statistically significantly different for FROH, for only 20 generations), when genetic map length is shortened to
FE, FH and FP when the genetic map length was 800 cM instead of 800 cM, and finally when SNP He≈0.2 (instead of He≈0.3).
3600 cM, and when simulations were run for 20 generations instead of
100–150 generations. Precision of IBDG estimators
We measured bias of the IBDG estimators as the mean amount by FP had low precision in both demographic scenarios (Figure 2). The
mean r2 from regressions of FP versus IBDG across 20 simulation
which FP, FH, FE and FROH over- or under-estimated IBDG (that is, the
repetitions (r 2F P IBDG ) was 0.72 in partially isolated populations and
mean of the values of FP FH, FE or FROH minus IBDG). Note that FH is
0.71 in populations with recently reduced Ne when pedigrees included
expected to underestimate IBDG when allele frequencies are taken
20 generations. Using more than five pedigree generations had little
from the current population and mating occurs at random. Thus, we
effect on r 2F P IBDG in populations with recently reduced Ne. However,
expect FH to underestimate IBDG in our simulated random mating
r 2F P IBDG increased when deeper pedigrees were used in partially
populations. We also wanted to measure bias of the estimated isolated populations.
differences in IBDG among individuals. For this we estimated the FROH and FH had high precision in both demographic scenarios
ratio of the standard deviation of each of the estimators (FP, FH, FE when large number of SNPs were used. The mean r2 from regressions
and FROH) to the standard deviation of IBDG to measure bias of the with IBDG was statistically significantly higher for FH (r 2F H IBDG ) than
estimated differences in IBDG among individuals. A ratio of the for FE (r 2F E IBDG ) or FROH (r 2F ROH IBDG ) in partially isolated populations
standard deviation of an estimator to the standard deviation of IBDG for all numbers of loci used (Po0.05, Figure 2). r 2F H IBDG was 40.9
near 1 indicates that the estimated differences in IBDG among when 10 K or more SNPs were used in both demographic scenarios.
individuals are unbiased. Ratios of below or above 1 indicate that r 2F ROH IBDG was 40.9 when 100 K SNPs were used in partially isolated
estimated differences in IBDG among individuals are substantially populations, and when ⩾ 15 K SNPs were used in populations with
under- or overestimated. recently reduced Ne. r 2F E IBDG was always statistically significantly

Figure 2 Barplots of the mean r2 (±1 s.d. among 20 simulated populations) from regressions of FP, FE, FH and FROH versus IBDG. The data shown are from
20 partially isolated populations (top row) and also 20 populations with recently reduced Ne (bottom row) and 3600 cM genomes. Dashed lines at r2 = 0.9 are
to aid comparison of r2 for FP, FH, FE and FROH.

Heredity
 Measuring individual inbreeding
M Kardos et al
67

higher than r 2F H IBDG and r 2F ROH IBDG in populations with recently recently reduced Ne. FROH underestimated IBDG by 0.14 when 5 K
reduced Ne (Po0.001). SNPs were used, and by 0.05 on average when 100 K SNPs were used
FROH, FE and FH were always more precise than FP when large in populations with recently reduced Ne. FH was the most downwardly
numbers of SNPs were used. For example, r 2F H IBDG measured with biased of all (Po0.0001) and the number of SNPs had no effect on
⩾ 5 K SNPs and r 2F ROH IBDG measured with ⩾ 35 K SNPs were both bias. FH underestimated IBDG by 0.21 on average in partially isolated
statistically significantly higher than r 2F P IBDG based on 20 generation populations and by 0.16 in populations with recently reduced Ne.
pedigrees in partially isolated populations (Po0.001; Figure 2). The Both FE and FROH were less downwardly biased than FP in both
mean r2 from regressions with IBDG was statistically significantly demographic scenarios when large numbers of SNPs were
higher for FROH, FE and FH estimated with ⩾ 10 K SNPs than for FP used (Figure 3). FROH estimated with 10 K and 50 K SNPs was less
estimated with 20 generations in populations with recently reduced Ne downwardly biased than FP estimated with 3 and 20 generations,
(Po0.001). respectively, in populations with recently reduced Ne (Po0.0001).
FE estimated with 500 and 10 K SNPs was less downwardly biased than
Bias of IBDG estimators FP estimated with 3 and 20 generations, respectively, in populations
FP consistently underestimated IBDG in both simulated demographic with recently reduced Ne (Po0.0001). FH had larger downward bias
scenarios (Figure 3). For example, FP underestimated IBDG by 0.19 than FP in all cases (Figure 3). The precision and bias of FROH, FE, FH
with 3 generation pedigrees, and by 0.1 with 20 generation pedigrees and FP relative to IBDG in a representative population is shown in
on average in partially isolated populations. FP was less downwardly Figure 4.
biased in populations with recently reduced Ne where FP under-
estimated IBDG by 0.13 and 0.07 on average when pedigrees included Bias in estimated differences in IBDG among individuals
3 and 20 generations, respectively. The mean ratio of the standard deviation of FP to the standard
Marker-based estimators were all substantially downwardly biased deviation of IBDG (sF P =sIBDG ) across 20 simulation repetitions was
when few SNPs were used, and this downward bias was largest in o1 in all scenarios (Figure 5). sF P =sIBDG was 0.61 in partially isolated
partially isolated populations (Figure 3). FE and FROH became less populations and 0.86 in populations with recently reduced Ne when
downwardly biased when larger numbers of SNPs were used. For pedigrees included 20 generations. Using pedigrees of more than five
example, FE underestimated IBDG by 0.11 when 500 SNPs were used, generations did not substantively affect sF P =sIBDG in populations with
and by only 0.02 when 100 K SNPs were used in populations with recently reduced Ne.

Figure 3 Barplots of the mean bias of FP, FH, FE and FROH (±1 s.d. among 20 simulated populations). Bias was measured as the mean error (FP, FH, FE or
FROH minus IBDG) across all individuals in each simulated population. The data shown are from 20 partially isolated populations (top row) and 20
populations with recently reduced Ne (bottom row) and 3600 cM genomes.

Heredity
 Measuring individual inbreeding
M Kardos et al
68

Figure 4 FP (a), FH (b), FROH (c) and FE (d) versus IBDG in a population with recently reduced Ne and 3600 cM genomes. FP was estimated with five
generations. FH, FE and FROH were estimated with 25 K SNPs. The dashed lines have intercept of zero and slope of one. Points below the lines represent
underestimates of IBDG.

Figure 5 Ratios of the standard deviation of FP, FE, FH and FROH to the standard deviation of IBDG. The data shown are from partially isolated populations
(top row) and populations with recently reduced Ne (bottom row) and 3600 cM genomes. The dashed lines represent a ratio of 1. Bars extending above and
below the line represent over- and underestimates, respectively, of the difference in IBDG among individuals.

Heredity
 Measuring individual inbreeding
M Kardos et al
69

FH always overestimated the differences in IBDG among individuals, 800 cM genomes than in partially populations with 3600 cM genomes
particularly when relatively few SNPs were used (Figure 5). For (Po0.0001, Figure 5 and Supplementary Figure S5). sF E =sIBDG was
example, the mean ratio of the standard deviation of FH to the 0.16 units higher (closer to 1) on average with 800 cM genomes than
standard deviation of IBDG (sF H =sIBDG ) across 20 simulation with 3600 cM genomes in partially isolated populations (Po0.002).
repetitions was approximately 1.7 when 5 K SNPs were used, and
1.25 when 15 K SNPs were used in partially isolated populations. Performance of IBDG estimators when inbreeding is due only to
The mean ratio of the standard deviation of FROH to the standard recent ancestors
deviation of IBDG (sF ROH =sIBDG ) was consistently below 1 in partially FP was imprecise when all inbreeding was due to recent ancestors (that
isolated populations (Figure 5). sF ROH =sIBDG was 0.70 when 100 K is, in simulations run for only 20 generations). For example, r 2F P IBDG
SNPs were used in partially isolated populations. However, in estimated with 20 generation pedigrees was 0.78 in partially isolated
populations with recently reduced Ne, sF ROH =sIBDG was approximately populations with 3600 cM genomes (Supplementary Figure S6) and
1.05 on average when ⩾ 25 K SNPs were used (Figure 5). The mean 0.53 in partially isolated populations with 800 cM genomes
ratio of the standard deviation of FE to the standard deviation of IBDG (Supplementary Figure S9).
(sF E =sIBDG ) was 1.1 when 500 SNPs were used, and 1.0 when 100 K The precision of the marker-based measures of IBDG in 20
SNPs were used in populations with recently reduced Ne (Figure 5). generation simulations was generally very similar to the results
However, sF E =sIBDG was approximately 0.72 with 10 K SNPs in described above for simulations that ran for 100–150 generations
partially isolated populations. (Supplementary Figures S6 and S9). However, r 2F E IBDG was considerably
sF E =sIBDG and sF ROH =sIBDG were always 4 sF P =sIBDG when large higher in partially isolated populations when the simulations ran for
numbers of SNPs were used (Figure 5). For example, sF E =sIBDG was only 20 generations instead of 150 generations (Figure 2 and
statistically significantly higher than sF P =sIBDG estimated with 20 Supplementary Figure S6).
generations when 10 K SNPs were used in partially isolated popula- FROH, FE, FH and FP had lower bias in simulations that ran for only
tions (P = 0.009) and when ⩾ 500 SNPs were used in populations with 20 generations (Supplementary Figures S7 and S10) than in simula-
recently reduced Ne (Po0.0001). sF ROH =sIBDG was statistically tions that ran for 100–150 generations (Figure 3 and Supplementary
significantly higher than sF P =sIBDG estimated with 20 generations Figure S4). In particular, FP estimated with pedigrees including all 20
when 100 K SNPs were used in partially isolated populations generations was unbiased, as expected when pedigree founders are
(P = 0.003), and when ⩾ 15 K SNPs were used in populations with unrelated and non-inbred.
recently reduced Ne (P ⩽ 0.025). The standard deviations of FROH, FE and FP more closely matched
In the following sections, we summarize how altering simulation the standard deviation of IBDG in simulations of partially isolated
parameters (genetic map length, number of simulated generations and populations that ran for only 20 generations instead of 150 generations
SNP heterozygosity) caused results to differ substantially from those (Figure 5 and Supplementary Figure S8). For example, sF P =sIBDG was
presented above. Detailed results from these scenarios are provided in always o0.65 in partially isolated populations with 3600 cM genomes
Supplementary Figures S3–S14. simulated for 150 generations. However, sF P =sIBDG was 40.8 when
FP was estimated with ⩾ 5 generations of pedigree in partially isolated
Effects of the genetic map length on precision and bias populations simulated for only 20 generations. sF E =sIBDG and
r 2F P IBDG was always statistically significantly lower in populations with sFROH =sIBDG were always o0.75 in partially isolated populations
800 cM genomes than in populations with 3600 cM genomes run for 150 generations. However, in partially isolated populations
(P ⩽ 0.005), except when pedigrees included only 3–5 generations in simulated for only 20 generations, sF E =sIBDG was 40.9 for all
partially isolated populations (Figure 2 and Supplementary Figure S3). numbers of loci used, and sFROH =sIBDG was 40.9 when FROH was
For example, r 2F P IBDG based on three pedigree generations was 0.54 estimated with 100 K SNPs.
when the genetic map was 3600 cM, and 0.36 when the genetic map
was 800 cM in populations with recently reduced Ne. Effects of SNP heterozygosity
FE and FROH always had statistically significantly higher precision in FH and FROH were less precise in simulations with lower hetero-
partially isolated populations with 800 cM genomes than with 3600 cM zygosity (He≈0.2) when relatively few loci were used. For example,
genomes (Po0.002), with the exception of when FROH was estimated r 2F ROH IBDG was 0.35 with He≈0.2 and 0.74 with He≈0.3 when 5 K
with 100 K SNPs (Figure 2 and Supplementary Figure S3). For SNPs were used in populations with 3600 cM genomes and recently
example r 2F E IBDG and r 2F ROH IBDG were 0.83 and 0.88, respectively, reduced Ne (Figure 2 and Supplementary Figure S12). However, the
when 10 K SNPs were used in partially isolated populations with number of SNPs necessary for high precision (r2 ⩾ 0.9) was unaffected
800 cM genomes. However, r 2F E IBDG and r 2F ROH IBDG were 0.65 and 0.5, by SNP heterozygosity for any of the marker-based measures of IBDG.
respectively, when 10 K SNPs in partially isolated populations with
3600 cM genomes. FROH and FE were often statistically significantly less DISCUSSION
downwardly biased when genomes were 800 cM instead of 3600 cM, Pedigrees have long been the preferred tool to measure individual
but these differences in bias were always small (⩽0.06) (Figure 3 and inbreeding (Pemberton, 2004,2008). However, we found that the
Supplementary Figure S4). realized IBD proportion of the genome (IBDG) is better predicted by
sF P =sIBDG based on pedigrees including five or more generations large numbers of genetic markers than by pedigrees. Our results are
was always statistically significantly lower when genomes were 800 cM consistent with two recent studies that used different approaches to
than when genomes were 3600 cM in populations with recently evaluate the precision of marker- versus pedigree-based measures of
reduced Ne (P ⩽ 0.001; Figure 5 and Supplementary Figure S5). For individual inbreeding when using thousands of SNPs. Hoffman et al.
example, sF P =sIBDG based on pedigrees including five generations was (2014) found a moderately high correlation between heterozygosity
0.90 with 3600 cM genomes, and 0.75 with 800 cM genomes in estimated at 413 000 SNPs and FP (r2 = 0.74) in oldfield mice
populations with recently reduced Ne. sF ROH =sIBDG was 0.2 units (Peromyscus polionotus). However, the predicted r2 between hetero-
higher (closer to 1) on average in partially isolated populations with zygosity and IBDG (based on estimated identity disequilibrium in their

Heredity
 Measuring individual inbreeding
M Kardos et al
70

oldfield mice) was ≈1; thus, IBDG was likely more precisely predicted FH was always very strongly correlated with IBDG (r2 near or above
by heterozygosity than by FP in their study population. Keller et al. 0.9) when 10 K or more SNPs were used (Figure 2 and Supplementary
(2011) found that the number of homozygous rare alleles within an Figure S3). However, as expected, FH underestimated IBDG (Figure 3
individual was more strongly correlated with marker-based estimates and Supplementary Figure S4) and the magnitude of this bias was
of inbreeding than with FP in large simulated populations with short unaffected by the number of SNPs. This bias can be explained by a
genetic maps (220 cM) and two chromosomes. discrepancy between the assumed and actual base populations used to
We expanded on previous work (Keller et al., 2011; Hoffman et al., estimate FH (Wang, 2014). As mentioned above, FH can take negative
2014) by evaluating the performance of marker- and pedigree-based values, and is expected to be centered near zero when allele
measures of IBDG using simulations where IBDG was known for each frequencies are estimated from the current population and mating is
individual without error. Additionally, we accounted for variation in random (for example, Figure 4). For FH to provide unbiased estimates
the genetic map length, population demography, heterozygosity of the of IBDG, allele frequencies must be derived from a base population
genetic markers and pedigree depth. Finally, we specifically simulated where all individuals would be unrelated if mating occurred randomly
relatively small populations that are of the greatest concern for (Supplementary Figure S16). Allele frequencies usually can only be
conservation. Our results suggest that marker-based measures of IBDG estimated using the individuals for which inbreeding measurements
computed from thousands of loci should be preferred over measures are desired. Mean pairwise relatedness can be high in populations with
based on FP in future studies of inbreeding depression in natural small Ne (Wang, 2014), which are often the focus of studies of
populations. inbreeding and inbreeding depression. Thus, researchers should expect
FH to substantially underestimate IBDG in most small natural
Precision and bias populations.
FP had low precision and underestimated the differences in IBDG FP consistently underestimated the differences in IBDG among
among individuals in all simulated scenarios. Additionally, FP was individuals (Figure 5 and Supplementary Figure S5). This is likely due
always downwardly biased except when pedigrees included 20 to pedigrees failing to account for the variation in IBDG due to distant
generations in simulations that ran for only 20 generations and ancestors, and linkage and recombination. We found that FROH and FE
pedigree founders were thus unrelated and non-inbred had lower bias than FP in estimated differences in IBDG among
(Supplementary Figures S7 and S10). Errors in FP have two major individuals when large numbers of SNPs were used (for example,
sources. As discussed above, σ(IBDG) can be high among individuals Figure 5). Downward bias in estimates of the differences in IBDG
with the same pedigree (Figure 1; Franklin, 1977). High σ(IBDG) (among individuals) might lead to overestimation of the strength of
partially explains why multiple locus heterozygosity is usually only inbreeding depression (Keller, 1998) because of a greater decrease in
weakly related to FP, and may also explain why heterozygosity-fitness fitness per unit increase in FP than per unit increase in IBDG. This
correlations are sometimes observed when FP-fitness correlations are suggests that the magnitude of inbreeding depression (that is, the
absent (Forstmeier et al., 2012). Linkage and a finite genetic map number of lethal equivalents, Keller and Waller, 2002) might be
length are expected to weaken the correlation between FP and IBDG, overestimated in pedigree-based tests for inbreeding depression.
but not to cause FP to be downwardly biased. However, estimates of the number of lethal equivalents might have
The downward bias in FP can easily be explained by relatedness and lower bias when using FROH and FE with large numbers of SNPs. The
inbreeding of pedigree founders. When founders are related, inbreed- statistical performance of marker- versus pedigree-based estimates of
ing due to common ancestors deeper in an individual’s ancestry is not the strength of inbreeding depression should be formally evaluated in
accounted for, thus causing FP to underestimate IBDG. FP will also future research.
underestimate IBDG for individuals whose parents have a common
ancestor that is an inbred founder. Moving the base population further How many SNPs are needed?
back in time (that is, including more generations in a pedigree) The number of markers necessary for precise and low-bias estimates of
accounts for a larger proportion of the common ancestors of parents IBDG in a particular study system will depend on the genome size,
and the inbreeding of those common ancestors, thus decreasing the whether inbreeding due to distant ancestors is of interest, marker
downward bias of FP (Figure 3 and Supplementary Figure S15). High heterozygosity and proportion of missing genotypes, and the variance
variance in founder relatedness may reduce the precision of FP by of IBDG. The ability to reliably detect IBD tracts depends on marker
causing variation among individuals in the magnitude of the under- density (SNPs/cM); thus, the necessary number of markers increases
estimation of IBDG. A focus of future research should be to determine with genome map length. Additionally, higher density markers are
the relative influence of related or inbred pedigree founders versus necessary to detect short IBD tracts caused by distant ancestors or a
linkage and recombination on the imprecision of FP. very high recombination rate (Supplementary Figure S17). The
All of the marker-based measures of IBDG were highly precise when expected correlation between heterozygosity-based measures of
large number of SNPs were used. Additionally, genetic markers individual inbreeding and IBDG increases with heterozygosity and
showed lower bias than FP in estimated differences in IBDG among the variance of IBDG (Miller et al., 2014). This was evident in our
individuals. Thousands of SNPs can currently be typed for any results where FROH and FH often had higher precision when He was
organism, for example, with SNP chips (for example, Kawakami high. Thus, more loci may be necessary for marker-based measures of
et al., 2014), restriction site-associated DNA sequencing (RAD-seq) IBDG to have higher precision than FP in studies where He is low.
(Davey et al., 2011), or by whole genome sequencing (Ellegren, 2014). The discussion above suggests that there is not a single number of
These technologies along with the increasing number of whole loci that will always achieve high precision and low bias estimates of
genome reference assemblies make it possible to precisely measure IBDG. Simulations with genomic and demographic parameters set to
IBDG with molecular markers in any species. Genotyping large match a particular study system can be used to evaluate the number of
numbers of SNPs is still expensive, but this cost will almost always markers necessary to precisely estimate IBDG. However, we believe the
pale in comparison with the expense and difficulty of obtaining deep number of SNPs typically available with SNP chips or RAD-seq (⩾104
pedigrees in free-ranging natural populations. SNPs, Allendorf et al., 2010) should in most cases provide highly

Heredity
 Measuring individual inbreeding
M Kardos et al
71

precise estimates of IBDG when used with one or more of the marker- FE had higher precision and often lower bias than FH and FROH
based measures of IBDG included in the present study. when relatively few markers were used (for example, ⩽ 10 K SNPs)
(Figures 2,3 and 5). Therefore, FE may often be the most powerful
Effects of genetic map length on the performance of IBDG measure of IBDG in studies employing relatively few mapped SNPs.
estimators However, the current implementation of FE (FEstim) almost always
We found that the advantages of marker-based measures of IBDG over failed to run to completion when 410 K SNPs were used in partially
FP are greatest in organisms with short genetic map lengths. The isolated populations, presumably due to high LD remaining among
reduced precision of FP among organisms with shorter map lengths is linked SNPs (Supplementary Figure S2). Lowering the Ne of the
easily attributed to the higher σ(IBDG) because of fewer crossovers per simulated populations from Ne = 75 to Ne = 20, which increases the
meiosis on average (Franklin, 1977). Unlike FP, FE and FROH were magnitude of LD (Hill, 1981), resulted in FEstim failing to run to
often more strongly correlated with IBDG in populations with shorter
completion in the great majority of simulation repetitions in both
genetic maps (Figure 2 and Supplementary Figure S3). The increased
demographic scenarios (data not shown). This suggests that studies of
precision of FE and FROH in populations with shorter genetic maps
populations with very small Ne (where strong LD extends across the
likely occurs because the longer IBD tracts resulting from a lower
entire genome) will often be limited to using small numbers of SNPs
recombination rate (Supplementary Figure S17) are more easily
detected by the hidden Markov model and sliding window algorithms with FEstim.
used to estimate FROH and FE. These results demonstrate that the
marker-based measures of IBDG are sensitive to the higher variance in Assumptions and limitations
IBDG in organisms with shorter genetic maps. Therefore, marker- Our simulations assumed no crossover interference. However, inter-
based measures of IBDG should be even more strongly preferred over ference can be strong in some vertebrates (Broman and Weber, 2000).
FP in organisms with short genetic maps. We are not aware that the effects of interference on the spatial and
We simulated genomes typical of mammals in terms of the physical length distributions of IBD tracts have been investigated. Therefore, it
size (3 Gb), number of chromosomes and genetic map length is unclear how interference affects the performance of pedigree- and
(800–3600 cM; Dumont and Payseur, 2008). FP is will be less precise marker-based measures of IBDG.
in species with shorter genetic maps or fewer chromosomes due to We also assumed pedigrees and genotypes were error free. However,
higher variance in IBDG among individuals with the same pedigree genotyping (Pompanon et al., 2005) and pedigree errors (Pemberton,
(Franklin, 1977; Supplementary Figure S1). Thus, the advantage of 2008) are common. Incorrect heterozygous versus homozygous calls
marker-based measures of IBDG over FP is expected to be greater in will increase errors in marker-based statistics. Likewise, incorrect or
species with fewer chromosomes or shorter genetic maps. However, missing parentage assignments will increase errors in FP (Pemberton,
the precision of FP will increase in species with longer genetic maps or 2008). The effects of genotyping and pedigree errors were beyond the
more chromosomes, and more loci will be required to have equivalent scope of this study. However, plausible rates of genotyping or pedigree
or higher precision than FP in such cases. errors should be considered in empirical inbreeding studies. Further,
we did not simulate inbreeding depression. It is not clear how
Relative value of FROH, FE and FH inbreeding depression, where the most highly inbred individuals are
The major advantages of FH are that it does not require mapped less likely to survive and be sampled in natural populations, might
genetic markers and it was highly precise in all simulated scenarios affect the performance of marker- and pedigree-based measures of
when using as few as 10 K SNPs (Figure 2 and Supplementary Figure IBDG. Finally, we simulated genetic markers with He in the approximate
S3). Physical genome maps are currently available for few taxa
range of 0.2–0.3. Using loci with lower heterozygosity is expected to
(Ellegren, 2014). Therefore, our results are encouraging for studies
decrease the precision of marker-based estimates of IBDG.
on organisms lacking genome maps and have no choice but to
measure individual inbreeding with unmapped loci (for example,
CONCLUSIONS
Hoffman et al., 2014).
Our results show that marker-based measures of IBDG are substan-
Though IBDG can be measured precisely with unmapped genetic
tially more precise and often less biased than FP. This and other recent
markers, there are advantages to using mapped markers when they are
findings (for example, Hoffman et al., 2014; Keller et al., 2011) are in
available. In addition to showing high precision and low bias when
large numbers of markers are used, FROH can distinguish inbreeding stark contrast to previous studies that used few genetic markers and
due to recent versus distant ancestors by including only specific ROH did not consider the effects of linkage and recombination on the
length categories in statistical analyses (for example, Kirin et al., 2010). relative performance of pedigree- versus marker-based measures of
Inbreeding depression could be caused mainly by inbreeding due to inbreeding (for example, Balloux et al., 2004; Slate et al., 2004; Santure
recent ancestors when a large fraction of deleterious recessive alleles et al., 2010). The current explosion of DNA sequence data from non-
are purged over many generations. For example, Szpeich et al. (2013) model organisms is providing the tools necessary to precisely measure
found that very long ROH (caused mainly by recent ancestors) was IBDG with genetic markers in any organism. We encourage researchers
more highly enriched for deleterious recessive alleles than short ROH to adopt marker-based measures of IBDG as large numbers of markers
(caused mainly by distant ancestors) in humans. Therefore, tests for and genome maps quickly become available for non-model organisms.
inbreeding depression may often be more powerful when FROH is
estimated only with very long ROH. However, it could be important DATA ARCHIVING
to account for inbreeding due to distant ancestors (for example, by Our simulation program is available in Dryad: http://dx.doi.org/
including short ROH in FROH, or by using FE or FH) when purging is 10.5061/dryad.f8p63.
inefficient at removing the genetic load (for example, when inbreeding
depression is caused by many deleterious recessive alleles with CONFLICT OF INTEREST
small effects). The authors declare no conflict of interest.

Heredity
 Measuring individual inbreeding
M Kardos et al
72

ACKNOWLEDGEMENTS Keller LF, Waller DM (2002). Inbreeding effects in wild populations. Trends Ecol Evol 17:
MK was supported by the National Science Foundation (NSF) through the 230–241.
Keller MC, Visscher PM, Goddard ME (2011). Quantification of inbreeding due to distant
Montana Institute on Ecosystems (EPS-1101342) and the Montana Ecology of ancestors and its detection using dense single nucleotide polymorphism data. Genetics
Infectious Diseases IGERT program (DGE-0504628). GL and FWA were 189: 237–249.
partially supported by NSF grants DEB-0742181 and 1258203. GL was also Kenney J, Allendorf FW, McDougal C, Smith JLD (2014). How much gene flow is
partially supported by DEB-0723928. We greatly appreciate helpful comments needed to avoid inbreeding depression in wild tiger populations? Proc Biol Sci 281:
20133337.
from the reviewers. Kirin M, McQuillan R, Franklin CS, Campbell H, McKeigue PM, Wilson JF (2010). Genomic
runs of homozygosity record population history and consanguinity. PLOS ONE 5:
e13996.
Leutenegger A-L, Prum B, Génin E, Verny C, Lemainque A, Clerget-Darpoux F et al. (2003).
Allendorf FW, Hohenlohe PA, Luikart G (2010). Genomics and the future of conservation Estimation of the inbreeding coefficient through use of genomic data. Am J Hum Genet
genetics. Nat Rev Genet 11: 697–709. 73: 516–523.
Balloux F, Amos W, Coulson T (2004). Does heterozygosity estimate inbreeding in real McQuillan R, Eklund N, Pirastu N, Kuningas M, McEvoy BP, Esko T et al. (2012). Evidence
populations? Mol Ecol 13: 3021–3031. of inbreeding depression on human height. PLoS Genet 8: e1002655.
Broman KW, Weber JL (2000). Characterization of human crossover interference. Am J McQuillan R, Leutenegger AL, Abdel-Rahman R, Franklin CS, Pericic M, Barac-Lauc L
Hum Genet 66: 1911–1926. et al. (2008). Runs of homozygosity in European populations. Am J Hum Genet 83:
Chapman N, Thompson E (2003). A model for the length of tracts of identity by descent in 359–372.
finite random mating populations. Theor Popul Biol 64: 141–150. Miller J, Malenfant R, David P, Davis C, Poissant J, Hogg J et al. (2014). Estimating
Daetwyler H, Hickey J, Henshall J, Dominik S, Gredler B et al. (2010). Accuracy of genome-wide heterozygosity: effects of demographic history and marker type. Heredity
estimated genomic breeding values for wool and meat traits in a multi-breed sheep 112: 240–247.
population. Anim Prod Sci 50: 1004–1010. O'Grady JJ, Brook BW, Reed DH, Ballou JD, Tonkyn DW, Frankham R (2006). Realistic
Darwin C (1868). The Variation of Animals and Plants Under Domestication, Vol 2. Murray: levels of inbreeding depression strongly affect extinction risk in wild populations. Biol
London. Cons 133: 42–51.
Davey JW, Hohenlohe PA, Etter PD, Boone JQ, Catchen JM, Blaxter ML (2011). Genome- Pemberton J (2004). Measuring inbreeding depression in the wild: the old ways are
wide genetic marker discovery and genotyping using next-generation sequencing. Nat the best. Trends Ecol Evol 19: 613–615.
Rev Genet 12: 499–510. Pemberton JM (2008). Wild pedigrees: the way forward. Proc Biol Sci 275: 613–621.
Dumont B, Payseur B (2008). Evolution of the genomic rate of recombination in mammals. Polašek O, Hayward C, Bellenguez C, Vitart V, Kolčić I, McQuillan R et al. (2010).
Evolution 62: 276–294. Comparative assessment of methods for estimating individual genome-wide
Ellegren H (2014). Genome sequencing and population genomics in non-model organisms. homozygosity-by-descent from human genomic data. BMC Genom 11: 139.
Trends Ecol Evol 29: 51–63. Pompanon F, Bonin A, Bellemain E, Taberlet P (2005). Genotyping errors: Causes,
Fisher RA (1965). The Theory of Inbreeding, 2nd ed. Oliver & Boyd: Edinburgh. consequences and solutions. Nat Rev Genet 6: 847–859.
Forstmeier W, Schielzeth H, Mueller JC, Ellegren H, Kempenaers B (2012). Heterozygos- Purcell S, Neale B, Todd-Brown K, Thomas L, Ferreira MA, Bender D et al. (2007). PLINK:
ity–fitness correlations in zebra finches: microsatellite markers can be better than their a tool set for whole-genome association and population-based linkage analyses. Am J
reputation. Mol Ecol 21: 3237–3249. Hum Genet 81: 559–575.
Franklin I (1977). The distribution of the proportion of the genome which is homozygous by Core Team R (2013). R: A Language and Environment for Statistical Computing.
descent in inbred individuals. Theor Popul Biol 11: 60–80. R Foundation for Statistical Computing: Vienna, Austria. URL http://www.R-project.
Hill WG (1981). Estimation of effective population size from data on linkage disequili- org/.
brium. Genet Res 38: 209–216. Santure AW, Stapley J, Ball AD, Birkhead TR, Burke T, Slate J (2010). On the use of
Hill WG, Weir BS (2011). Variation in actual relationship as a consequence of Mendelian large marker panels to estimate inbreeding and relatedness: empirical and simulation
sampling and linkage. Genet Res 93: 47–64. studies of a pedigreed zebra finch population typed at 771 SNPs. Mol Ecol 19:
Hoffman JI, Simpson F, David P, Rijks JM, Kuiken T, Thorne MA et al. (2014). High- 1439–1451.
throughput sequencing reveals inbreeding depression in a natural population. Proc Natl Slate J, David P, Dodds KG, Veenvliet BA, Glass BC, Broad TE et al. (2004).
Acad Sci USA 111: 3775–3780. Understanding the relationship between the inbreeding coefficient and
Ives AR, Whitlock MC (2002). Inbreeding and metapopulations. Science 295: 454–455. multilocus heterozygosity: theoretical expectations and empirical data. Heredity 93:
Kardos M, Allendorf FW, Luikart G (2014). Evaluating the role of inbreeding depression in 255–265.
heterozygosity-fitness correlations: how useful are tests for identity disequilibrium? Stam P (1980). The distribution of the fraction of the genome identical by descent in finite
Mol Ecol Res 14: 519–530. random mating populations. Genet Res 35: 131–155.
Kawakami T, Backström N, Burri R, Husby A, Olason P, Rice AM et al. (2014). Estimation Szpiech ZA, Xu J, Pemberton TJ, Peng W, Zöllner S, Rosenberg NA et al. (2013). Long runs
of linkage disequilibrium and interspecific gene flow in Ficedula flycatchers by a newly of homozygosity are enriched for deleterious variation. Am J Hum Genet 93: 90–102.
developed 50k single-nucleotide polymorphism array. Mol Ecol Res 14: 1248–1260. Wang J (2014). Marker‐based estimates of relatedness and inbreeding coefficients: an
Keller LF (1998). Inbreeding and its fitness effects in an insular population of song assessment of current methods. J Evol Biol 27: 518–520.
sparrows (Melospiza melodia). Evolution 52: 240–250. Wright S (1922). Coefficients of inbreeding and relationship. Am Nat 56: 330–338.

Supplementary Information accompanies this paper on Heredity website (http://www.nature.com/hdy)

Heredity