Molecular Ecology (1994) 3, 91-99

Analysis of population genetic structure with RAPD

markers
M. LYNCH and B. G. MILLIGAN"
Department of Biology, University of Oregon, Eugene, OR 97403 and *Department of Botany, University of Texas, Austin, TX
78773, U S A

Abstract
Recent advances in the application of the polymerase chain reaction make it possible to
score individuals at a large number of loci. The RAJ?D (random amplified polymorphic
DNA) method is one such technique that has attracted widespread interest. The analysis
of population structure with RAPD data is hampered by the lack of complete genotypic
information resulting from dominance, since this enhances the sampling variance
associated with single loci as well as induces bias in parameter estimation. We present
estimators for several population-genetic parameters (gene and genotype frequencies,
within- and between-population heterozygosities, degree of inbreeding and population
subdivision, and degree of individual relatedness) along with expressions for their
sampling variances. Although completely unbiased estimators do not appear to be
possible with RAPDs, several steps are suggested that will insure that the bias in
parameter estimates is negligible. To achieve the same degree of statistical power, on the
order of 2 to 10 times more individuals need to be sampled per locus when dominant
markers are relied upon, as compared to codominant (RFLP,isozyme) markers. Moreover,
to avoid bias in parameter estimation, the marker alleles for most of these loci should be
in relatively low frequency. Due to the need for pruning loci with low-frequency null
alleles, more loci also need to be sampled with RAPDs than with more conventional
markers, and some problems of bias cannot be completely eliminated.
Keywords: gene diversity, inbreeding, molecular markers, structure, RAPDs, relatedness
Received 29 March 1993; revision accepted 15 July 1993

Introduction
Estimates of genetic variation increasingly are being
based upon information at the DNA level. Whereas
some DNA markers, such as restriction fragment length
polymorphisms, can be analysed with traditional meth-
ods developed for isozymes, others, such as DNA-
fingerprints, need to be analysed in new ways (Lynch
1988, 1990; Stephens et al. 1992). A recent method that is .
being assimilated rapidly by empiricists is the RAPD
(random amplified polymorphic DNA) technique (Welsh
& McClelland 1990; Williams et al. 1990; Caetano-AnollCs
et a/. 1991; Deragon 1992; Hadrys et al. 1992).
RAPDs are generated by applying the polymerase
chain reaction (PCR) to genomic DNA samples, using
randomly constructed oligonucleotides as primers. Since
the technique is relatively easy to apply to a wide array of
Correspondence: Michael Lynch +1-503-346-2364
plant and animal taxa, and the number of loci that can be
examined is essentially unlimited, RAPDs are viewed as
having several advantages over RFLPs and DNA
fingerprints. When the primers are of intermediate size
(on the order of 10 base pairs), multiple amplifiable
fragments (from different loci) are usually present for
each set of primers in each genome. The fragments can be
separated by size on a standard agarose gel and
visualized by ethidium bromide staining, eliminating
the need for radiolabeled probes. Since the primers consist
of"random sequences, and do not discriminate between
coding and noncoding regions, it is reasonable to expect
the technique to sample the genome more randomly than
conventional methods.
Despite these advantages, RAPD analysis presents
some practical problems. As in the case of DNA
fingerprinting where multiple markers appear on the
same gel, there can be uncertainty in assigning markers to
92 M. LYNCH and B. G. MILLIGAN
speclfic loci in the absence of a preliminary pedigree
analysis. In addition, the possibility that products of
different loci will have similar molecular weights, and
therefore be indistinguishable on a gel (because of
comigration), is of concern. In some cases, these
problems can be overcome by practical means (Hadrys
et al. 1992). However, a third issue seems to be
unavoidable - the dominance of RAPD markers. If one
of the “alleles” at a RAPD site is unamplifiable, then
marker/marker homozygotes cannot be distinguished
from marker/null heterozygotes. Provided there is only a
single amplifiable allele per locus, this does not prevent
the estimation of allele frequencies necessary for popula-
tion-genetic analysis, but it does reduce the accuracy of
such estimation relative to analysis with codominant
markers.
Our purpose is to develop the general theory needed
for the analysis of population structure with dominant
markers such as W D s , and to lay out formally some of
the difficulties that such markers present. Formulae are
presented for the estimation of conventional measures of
population structure. All of the estimators take into
account bias due to small sample size and are accom-
panied by approximate expressions for their sampling
variance.
Assumptions
In order to focus on the salient statistical issues, we start
with several assumptions. First, we make the generous
assumption that the interpretation of banding patterns on
gels can be accomplished in a completely unambiguous
manner. That is, we assume that marker alleles from
different loci do not comigrate to the same position on a
gel, and that the investigator is fully capable of matching
bands from different lanes within and among gels.
Secondly, we assume that each locus can be treated as a
two-allele system, with only one of the alleles per locus
being amplifiable by the PCR. The ‘’null‘’ allele may fail to
amplify either because of loss of a primer site or because
an insertion has caused the distance between primer sites
to exceed the capacity of the PCR; the specific mechanism
is not relevant to the theory that follows. Without
extensive segregation analysis, it is difficult to validate
the tyo-allele assumption, although a possible way to
resolve the issue is presented in the discussion. Most
current users of RAPDs seem to be of the opinion that the
existence of multiple amplifiable alleles at a locus is
relatively rare. It is important to realize, however, that
this opinion is usually based on rather circumstantial
evidence, and even if true, it does not rule out the
possibility that the pool of unamplifiable alleles is
heterogeneous. We will return to the potential impor-
tance of this issue later.
We will restrict our attention to sexual diploid
populations. Letting M and m denote the marker and
null alleles at a locus, and p and q = 1-p their frequencies,
the expected genotype frequencies in a segregating pop-
+
ulation are P M M = p2(1 - F I S ) pFls, &,,, = 2pq(l - F I S ) ,
+
and P,, = q2(1 - F l s ) qFIs, where Fls measures the
departure from Hardy-Weinberg equilibrium. Under
random mating, FIs=O, and the expected genotype
frequencies follow the Hardy-Weinberg proportions.
Under complete inbreeding, Fls = 1, there are no hetero-
zygotes, and the frequencies of the MM and mm
homozygotes are p and q, respectively. Fls >0 is usually
interpreted as a measure of the degree of inbreeding
under the assumption that the markers are selectively
neutral and in gametic-phase equilibrium with any
selected loci. However, Fls can take on negative values
when the frequency of heterozygotes exceeds the Hardy-
Weinberg expectation.
Because the MM and Mm genotypes cannot be
discriminated from each other by RAPD analysis, the
marker allele is ”dominant” to the null allele on a gel. (In
principle, such discrimination might be possible through
the detection of staining intensity differences, but this
degree of resolution would be very difficult to attain
without carefully controlled assays.) Thus, the only
observable quantity for a locus is the fraction
of individuals in the population with ( I - x ) and without
(x) the marker. Our goal is to demonstrate how observed
marker frequencies can be used to estimate conventional
measures of population-genetic structure.
Throughout the paper, we will discriminate estimates
from true population parameters by placing hats ( ” )
on the former. All of the estimators, as well as approxi-
mate expressions for their sampling variances, were
obtained by use of second-order Taylor expansions.
Although tedious at times, this approach is fairly
straightforward, and we omit the algebraic details.
Similar procedures were used by Nei & Roychoudhury
(1974) and Nei (1978) to obtain some estimators with
codominant markers.
In several places, we will require the sampling variance
of a mean of parameter estimates. To save on space, we
will simply refer to the following equation, denoting the
appropriate terms to substitute for the sample size s, the
individual estimates z(i), and the mean of those estimates
-
Z,
. 4
Estimation of gene and genotype frequencies
If the population is in Hardy-Weinberg equilibrium,
x = q2, and x”’
is the null allele frequency. However, i1/2,
 ANALYSIS OF RAPD DATA 93

where 32 is the proportion of the N sampled individuals PCR-based, the offspring generation can be assayed at
that do not exhibit the marker, provides a downwardly a very early stage (e.g. seeds), prior to the operation
biased estimate of q. A less biased (and asymptotically of selection.
unbiased) estimator is Letting io be the observed value of x in an offspring
population resulting from random mating, and assuming
the genotypic frequencies are not influenced by selection,
q can be estimated by substituting f, into eqn 2a. The
deviation from Hardy-Weinberg equilibrium can then be
where Var(32) = 2(1 - f ) / N is the sampling variance of
estimated with
the frequency of null homozygotes. The sampling
variance of the allele frequency is approximately
1-32
Var(q) = -
4N ' (34
The square root of this quantity provides an estimate of
the standard error of 4 and may be used as an
approximate indicator of the accuracy of estimation.
However, it should be noted that the construction of (0) Esbmalns 01 q

confidence limits for q based on this variance estimator

and on the assumption of normality may be inappropriate
when the sample size is small. Resampling procedures
(Crowley 1992) provide a more general approach. This
caveat applies to confidencelimit construction for all of
the other estimators discussed below.
The term in parentheses in eqn 2a accounts for
much, but not all, of the bias in the estimation of q due
to small sample size and can be substantial when the
nuU allele is rare (x<O.l). If the expected number of
null homozygotes in the sample (Nq') is less than two,
1::
0.1
1 p = 0.3
q = 0.7
&If: p = 0.3
q = 0.7 -
-
_____
even eqn 2a yields an estimate of 9 that can be
downwardly biased by 5% or more (Fig. 1).Addition of
higher-order terms to the Taylor expansion does not
improve the situation. However, eqn 2a always yields a
better estimate of q than ?/', and if Nq2>3, it yields an
estimate that is essentially unbiased (Fig. 1). The clear
implication of this result is that unbiased estimates of
population-genetic parameters can be achieved with
RAPDs provided the analysis is restricted to markers
that are not too common. Specifically, we recommend the
restriction of analyses to bands whose observed frequency
is less than 1 -(3/N), i.e. markers whose incidence is less
than 0.70 for N = 10, 0.94 for N=SO, etc.
The situation is more complicated if the population is
locally inbred, for then x = q2(1-Frs)+@rs. Contrary to the.
usual situation when all three genotypes are observable,
0 1 2 3
[2&( I

4 5
H/[2q(l - q)] f = 0.3
-a]/[Zq(

6
-
I q)J. q ==o0.3

7
7 ..

q = 0.7

8
-
-
9
i
I
10

there is no way to evaluate whether Frs = 0 with RAPD
analysis of a single generation. If, however, the study
population, or a sample of it, can be mated randomly,
then Hardy-Weinberg equilibrium will be attained in
the offspring generation, and disequilibrium in the
parent population will be revealed by a change in x
across generations. This approach to testing for Hardy-
Weinberg equilibrium may be practically feasible in
many empirical settings. Since the RAPD technique is
Number 01 Homozygetas
Erp~ted
Fig. 1. (a) Expected estimates of the gene frequency, 4 obtained
by eqn 2a relative to the parametric value 9, as function of
the expected number of null homozygotes in the sample, Nq2
(solid hes). Results obtained by numerical solution on a
computer are given for q=O.3 and 0.7; results for 9=0.1 and
0.2 are essentially identical to those for q=O.3. Also plotted
(dotted lines) are the expected values of 2'/2/q. (b)Slmilar results
for the expected estimates of the with-population gene
diversity, H obtained by eqn 4a relative to the parametric
value, Q(1-q).
94 M. LYNCH and B. G. MILLIGAN
where x,, is the estimated value of x in the parental
generation, and the term in large parentheses again
accounts for small sample size bias. The sampling
variance of P I S at a particular locus is approximately
A precise statistical test of the deviation of the parental
generation from the Hardy-Weinberg expectation can be
achieved via a x2 or log-likelihood ratio test by evaluating
the observed parental frequencies, f, and (1 -?,), against
their expectations, io and (1-201, with one degree of
freedom. If there is no compelling evidence that FIs # 0,
then it is reasonable to pool the estimates of q from both
generations. In the latter case, the sampling variance of
+
the pooled estimate of q is [Var(Lj,) Var(Q,,)]/4, provided
the sampled members of the two generations are not
related.
A mean estimate of FIs over all sampled loci, &, can be
obtained by averaging the f~s(i)acquired for each of the
i = 1,. . .,L markers. The sampling variance of this mean
estimate, Var(Fls), is obtained by letting s = L,
z ( i ) = fIs(i), and 2 = F I ~in eqn 1, and this accounts for
sampling of both individuals and loci. Provided the
number of loci sampled is fairly large, the mean estimate
Frs should be roughly normally distributed around its
expectation with a standard error equal to [var(&)]”*.
Similar logic can be used to obtain a mean (and sampling
variance) of the FIS estimates acquired for multiple
populations.
In the following, we assume that the population has in
some way been analyzed at a time when the genotype
frequencies are in Hardy-Weinberg equilibrium, since no
further progress with RAPD analysis is possible if that is
not the case. Our assumption that each locus is biallelic
implies that every observable band in the population
sample represents a unique locus. In the following, q(i)
represents the frequency of the null allele at the locus
involving the ith marker.
Gene diversity within populations
With the estimates of allele frequencies in hand, it is
possible to estimate the gene diversity within a popula-
tion. (We use the term gene diversity in an informal sense
here since it is likely that RAPDs will often contain only
noncoding DNA.) We employ the conventional measure
of gene diversity, H,(i) = 2qj(i)[1- q/(i)], which is the
probability that two genes, randomly drawn from
population j , differ at the ith locus. This measure is
equivalent to the expected heterozygosity under Hardy-
Weinberg equilibrium, and under the assumptions of our
model, can be viewed as the probability that a random
pair of alleles will contain one marker and one null. An
asymptotically unbiased estimator of this quantity is
Hl(i) = q l ( i ) [- +
l ci,(i)] ~ a r [ q , ( i ) ] , (44
and its sampling variance is approximately
Var[fij(i)]= 4[1 - Zjj(i)]’Var[Lj;(i)]. (4b)
The results shown in Fig. 1 indicate that, as in the case of
our estimator for gene frequency, eqn 4a yields an
essentially unbiased estimate of gene diversity provided
Nq2>3. The figure also shows that eqn 4a provides a
much better estimate of Hi(i) than the first-order estimate
2qj(i)[l- Ljj(i)].
Averaging over all L observed loci, the mean observed
gene diversity in the jth population is
(5)
The sampling variance of H, is a function of the variance
due to the sampling of a finite number of individuals a t
each locus, described by eqn 4b, and of the variance in
gene diversity among loci. Both sources of variance
contribute to the variance among the L observed k;(i)
within the population. That is, the total sampling variance
of the gene diversity within a population, Var(H,),can be
obtained by applying eqn 1 with s = L, z ( i ) = H , ( i ) and
ii = fi]. The square-root of this quantity provides an
estimate of the standard error of Hi.
Partitioning the total sampling variance of H, into its
two sources can provide insight into preferred experi-
mental designs for estimating the gene diversity (e.g.
Lynch & Crease 1990). The expected contribution to the
sampling variance of the average gene diversity owing to
sampling a finite number of individuals per locus is
where the Var[Hj(i)] have been defined in eqn 4b. The
sampling variance of H, due to variation in gene diversity
among loci is then
varL (H/)= Var (HI - Var1(HI) . (6b)
The within-locus sampling variance can be reduced to
zero by increasing the number of individuals sampled,
but if there are true interlocus differences in gene
diversity, as will almost always be the case, any further
reduction in the variance of would require an increase
in the number of loci scored.
Finally, if n populations have been sampled, the mean
within-population gene diversity can be estimated by

(7)
Its sampling variance, Var(kw), is a function of the
sampling of finite numbers of populations, loci, and
individuals, all three of which are accounted for when eqn
1 is used with s = n, z(i) = H,, and Z = Aw. The square
root of the resdtant quantity provides an estimate of the
standard error of f i w . Following the logic used above, the
total sampling variance of f i w can be partitioned into
contributions from sampling individuals (I), loci (L),and
populations (P),
Population subdivision
When data are available from more than one population,
it is usually of interest to evaluate the degree to which the
total gene diversity partitions into its within- and
between-population components. The heterozygosity
between populations j and k at the ith locus,
H;k ( i ) = 9/ ( i )[l - q k (i)]+ q k (i)[ 1 - q, (i)], is the probability
that a gene randomly drawn from population j differs
from one randomly drawn from population k. An
asymptotically unbiased estimator of this quantity is
and its variance due to sampling a finite number of
individuals is approximately
In the absence of population subdivision, the gene fre-
quencies in all populations are the same, so
Hik(1) = H, ( 1 ) = Hk (i). Thus, a more meaningful defini-
tion of the between-population gene diversity is the het-
eroiygosity in excess of that observed within populations,
the sampling variance of which is
ANALYSIS OF W D DATA 95
All of the terms of t h s expression have been defined
above except the last two, and these are of the form
Such covariance exists because the same individuals from
population j are used to estimate fjifi) and ff&(i).
Averaging over all loci, the estimated mean gene
diversity between populations j and k is
and its variance, Var(fi,k), is obtained by letting s = L,
z(i) = fijk(i), and Z = f i , k in eqn 1. The variance of can
be obtained in a parallel manner.
The mean between-population gene diversity is ob-
tained by averaging over all distinct pairs of populations,
13a)
The sampling variance of is given by
13b)
This expression accounts for individual, locus, and
population sampling. It is somewhat more complicated
than the variance formulations given above because not
all of the painvise population comparisons contributing
to H B are independent. Gene diversities between over-
lapping pairs of populations (eg. HI2 and H I 3 ) are likely
to be correlated positively. The term V B represents the
variance among diversity measures involving nonover-
lapping population pairs (e.g. 1212, H,, &6, fij.8, etc.) The
term CB represents the covariance among diversity
measures involving overlapping population pairs (e.g.
pairs [ f i l z , fin],[HIz ,H i , . . . , fin, fix I . . . [ f i ~
I .
&sIt
etc.) Both VB and CR are obtained by applying the stan-
dard definitions of a variance and covariance to the fi&.
Wright's (1951) measure of population subdivision,
FST= HB/HT, where HT = H B + H ~takes , on extreme values
of 0 when all populations have identical gene frequencies
and 1 when all populations are completely homozygous
for alternate alleles. An asymptotically unbiased estimate
of FST can be obtained by using
(14a)
96 M. LYNCH and B. G. MILLIGAN
All of the elements of this equation have been defined
above except the sampling covariance of the within- and
between-population esimates of gene diversity,
Such covariance exists because the same populations, loci,
and individuals are used to estimate HB and Hw. The
sampling variance of & is approximately
Genetic distance
Nei (1972) introduced a measure of genetic distance
between populations, Dik, that provides an estimate of the
mean number of mutations separating the genes from two
populations. Nei's genetic distance accounts for multiple
mutations per locus, and in the ideal case of isolated
populations and a constant rate of substitution, is
proportional to the mean coalescence time for genes in
different populations. Letting ],=l-H, be the gene
identity within population 1, and rjk = 1 -H;k be the
gene identity between populations j and k, both averaged
over all loci, then Djk = - l n ( J j k / a ) is Nei's genetic
distance. An unbiased estimator of Djk is given by
(154
and its sampling variance is approximately
All of the elements of this formula have been defined
above except
I I.
An alternative procedure for estimating nucleotide
divergence with RAPDs has been introduced by Clark
& Lanigan (1993). They assume that the time since
separation is sufficiently small that there is a negligible
probability of multiple mutational changes per locus.
Relatedness
Since related individuals are expected to have more
similar genotypes than nonrelatives, it stands to reason
that the fraction of loci for which two individuals have
identical phenotypes must increase with the degree of
relatedness. Assuming noninbred individuals, and letting
41 and 42 be the probabilities of sharing one or two alleles
identical by descent at a locus, then the relatedness of
individuals a and b is defined to be r a b =+lo h/2 ++ & b r
which takes on values of 0.5 for parents and offspring and
for full sibs, 0.25 for half sibs, etc.
Taking into account the joint probabilities of all
genotypes, and assuming a random-mating population,
the probability that two individuals, a and b, share the
same phenotype (both have the marker, or both do not) at
the ith locus is
where O ( i ) = 1 - 2$(i)[1 - q2(i)] is the probability that
two nonrelatives have matching phenotypes at the locus
and H ( i ) = 2q(i)[l - q(i)] is the locus-specific heterozyg-
osity. This expression has an undesirable property - it is a
function of the relatedness Tab, which we want to estimate,
and the identity coefficient q&, which is unknown. Since
blabis a measure of identity by descent, not identity in
state, it cannot be estimated from observational data.
Unfortunately, we have been unable to obtain any
expression for that does not have a similar problem.
Note, however, that the final term in eqn 16a can be no
greater than 1/16 (this only being the bias in the case of
parent-offspring relationships when H(11=0.5). Thus, if
one is willing to accept a bias of this magnitude,
provides an approximate relationship between band-
sharing and Tab. Taking into account uncertainty in the
estimation of 8, a relatedness estimator using data from
the ith locus is
where Snb(i)=lor 0 denotes whether a and b have the
 ANALYSIS OF RAPD DATA 97

same phenotype, and
Estimates of relatedness based on single loci are
extremely noisy. A more accurate estimate is obtained
by averaging over all L loci,
Such an analysis is only applied to polymorphic loci.
When a large number of markers is used, the sampling
distribution of should be roughly normal. Its variance
can be estimated by substituting s=L, z(i) = fob(i),and
t = i,b in eqn 1.
The utiIity of RAPD markers for discriminating pairs of
individuals related to different degrees depends on the
overlap between the distributions of similarity for
different degrees of relatedness. If there are L informative
markers for a pair of individuals, the mean (over loci)
similarity (over loci) for pairs of individuals with
relatedness r should be approximately binomially dis-
tributed with expectation
and sampling variance
These expressions can be used to explore the utility of
band-sharing in the estimation of relatedness.
The degree of overlap in the distributions of S
associated with different r depends strongly on the
frequency of markers in the population (Fig. 2). If the
markers are present in high frequency (all x =O), they will
provide little discrimination among relatives, since nearly
all members of the population will exhibit the marker a t
each locus. The use of very low frequency markers does
not improve the situation. Since the distribution of
similarity depends only on cl(i), which is symmetrical in
X ( I ] and l-x(i), it takes on the same properties for
q2(i)= x(i)= 0.1 as for q2(i)= x(i) = 0.9. The best resolution
occurs when markers are at intermediate frequency, i.e.
x(D = 0.5. However, even in this case, the degree of over-
lap in distributions of S is very broad, and based on the
results in the figure, it can be seen that the problem is not
eliminated until very large numbers of markers (well in
excess of 50) are scored. We conclude that the utility of
RAPDs for relatedness estimation is rather limited.
Discussion
Because of the dominance property of RAPDs, gene
frequency estimates for such loci are necessarily less
accurate than those obtained with codominant markers
such as isozymes and RFL.Ps. In the latter case, the
sampling variance of the gene frequency is q(l - q ) / 2 N .
Comparing this with eqn 2b, the ratio of the sampling
variance of gene frequency for RAPDs relative to a
comparable locus with codominance is (l+q)/(2q), where
9 is the frequency of the null allele. Comparing eqn 4b to
the result of Nei & Roychoudhury (1974), this can also be
shown to be the ratio for the sampling variance of the

04; X=O1.09 0.3 i X = 0.5

L =20 ' L=20
I
03-
0.2 4 ~..~
. .__

0.0
0.0 0.2 0.4. 0.6 0.8 1.0

Fig. 2. Expected distributions of similarity in

band sharing for random pairs of individuals of
different degrees of relatedness T . It is assumed
that individuals of a given relationship have
been scored at L = 20 or L = 100 markers, each
associated with a locus with the same value of x.
The mean of each binomial distribution is given
by eqn 19a, and the variance by eqn 19h.
-1

00 0.2 0.4 0.6 0.8 1.0 1 .o Variation among loci in the frequency of the
marker allele will increase the variance beyond
Similarity that illustrated.
98 M. LYNCH and B. G. MILLIGAN
within-population gene diversity using both types of
markers. This ratio can be viewed as the proportional
increase in sampling effort that needs to be applied to a
RAPD locus in order to achieve a genetic-parameter
estimate as accurate as would be acquired with a locus
with codominant markers with the same frequency.
Recall from above, that to avoid biased parameter
estimates, RAPDs for which x < 3/N should be avoided in
population-genetic analysis. Assuming this is done, then
the critical value of the above ratio is
(1 + m)/(2m), which takes on values of 1.4 for
a sample size of N=10, 3.4 for N=100, and 9.1 for
N = 1000. Thus, for acceptable loci alone, on the order of 2
to 10 times more individuals should be sampled per RAPD
locus than per RFLP or isozyme loci.
The pruning of loci for which the null phenotype
frequency is below 3 / N also implies that more loci need to
be sampled in RAPD analysis than in a survey involving
more conventional codominant markers. The number of
loci that are rejected using the 3/N criterion will neces-
sarily depend on the allele-frequency distribution, but it is
clear that this number must increase with decreasing N.
The pruning of loci raises an additional issue. Although
our procedure insures that the estimates of population
parameters will be essentially unbiased for the pool of
acceptable loci, it also tends to reject loci that are highly
homozygous. Moreover, the loci that are rejected may
differ from population to population. Consequently,
estimates of genome-wide parameters, such as average
within- and between-population gene diversities, derived
via RAPD analysis are likely to be biased in complicated
ways whether or not the pruning procedure is applied. In
either case, the obvious way to minimize the problem of
bias is to sample large numbers of individuals per
population (ideally at least loo), since this reduces the
fraction of loci that will yield biased parameter estimates.
In addition, it seems quite advisable in comparative
studies to equalize the sample sizes in different popula-
tions to minimize the possibility that observed differences
between populations are simply artifacts of bias resulting
from sample size differences.
Certainly, if even with large N, 10% or more of the
sampled loci still fail to meet the 3/N criterion, then any
population-parameter estimates derived from the data ’
should be interpreted with an appropriate degree of
caution. If it is deemed desirable not to prune loci from
the final analysis, then the best that can be said is that the
bias of a parameter estimate averaged over all loci cannot
exceed the worst locus-specific bias.
A critical assumption of the preceding theory is that
there are a t most two alleles per locus (one marker and
one null). If that is not the case, the approach we have
employed will lead to an overestimate of the frequency of
alleles, and an underestimate of the within-population
gene diversity. The bias resulting from violations of this
assumption depends on the frequency distribution of
alleles, but a qualitative assessment can be made for the
simple situation in which n markers at a locus all have
equal freqdencies, (1-q)/n. If the investigator focuses on
a single marker and views the heterozygosity as
(1-x”‘), where x is the frequency of individuals not
exhibiting that marker, the quantity 2(1 -q)(n - l+q)/n2
would actually be estimated. The ratio of this quantity to
the actual heterozygosity, 1-q2-n[(l -q)/n12, is
which shows that in this extreme case the estimated
heterozygosity would be on the order of (2/n)th or less
than the true value.
A simple way to avoid the problem of multiple allelic
markers is to focus on markers that are common. If the
sum of the frequencies of individuals in whch two
markers are uniquely found exceeds one, then the
markers cannot be allelic. So if this criterion is applied
to all pairs of markers, the multiple-marker problem can
be eliminated. This, of course, does not remove the issue
of multiple unamplifiable (null) alleles, a problem that
seems much more difficult to resolve. Moreover, by
eliminating loci with high allele frequencies, this proce-
dure raises the additional issues of bias associated with
pruning, already discussed above.
Finally, we note that our emphasis has been on the
development of unbiased parameter estimators. Such
estimators do not always minimize the mean squared
deviations between parameters and their estimates.
Alternative procedures, such as maximum likelihood,
may provide minimum-variance estimators, although
such estimators are not necessarily unbiased.
Acknowledgements
This work has been supported by NSF grants BSR 89-
11038 and BSR 90-24977, and PHS grant GM36827-01 to
ML, and by NSF grant BSR 89-06129 to BGM. We are
especially grateful to K. Ritland for pointing out some
fundamental problems in an earlier draft of this paper,
and to E. Martins for helpful comments.
References
Caetanc-Anoll6s G, Bassam BJ, Gresshoff PM (1991) DNA
amplification lingerprinting usmg very short arbitrary
oligonucleotide primers. Bio/Technology, 9, 553-557.
Clark AG, Lanigan CMS (1993) Prospects for estimating
nucleotide divergence with RAPDs. Molecular Biology and
Evolution, 10, 1096-1111.
Crowley PH (1992) Resampling methods for computation-

intensive data analysis in ecology and evoluhon. Annual
Rmiew of Ecology and Systematics, 23, 405-447.
Deragon JM, Landry BS (1992) RAPD and other PCR-based
analyses of plant genomes using DNA extracted from small
leaf disks. PCR Methods and Applications, 1, 175-180.
Hadrys H, Dalick M, Scluerwater B (1992) Applications of
random amplified polymorphic DNA (RAPD) in molecular
ecology. Molecular Ecology, 1, 55-64.
Lynch M (1988) Estimation of relatedness by DNA iingerprinting.
Molecular Biology and Evolution, 5, 584-599.
Lynch M (1990) The similarity index and DNA fingerprinting.
Molecular Biology and Evolution, 7, 478-484.
Nei M (1972) Genetic distance between populations. American
Naturalist, 106, 283-292.
Nei M (1978) Estimation of average heterozygosity and genetic
distance from a small number of individuals. Genetics, 89,
583-590.
Nei M, Roychoudhurj AK (1974) Sampling variances of
heterozygosity and genetic distance. Genetics, 76, 379-390.
ANALYSIS OF RAPD DATA 99
Stephens JC, Gilbert DA, Yuhki N, OBrien SJ (1992) Estimation of
heterozygosity for single-probe multilocus DNA fingerprints.
Molecular Biology and Evolution, 9, 729-743.
Welsh J, McClelland M (1990) Fingerprinting genomes using PCR
with arbitrary primers. Nuclezc Acids Research, 18, 7213-7218.
Williams JGK, Kublelik AR, Livak KJ, Rafalski JA, Tingey SV
(1990) DNA polymorphisms amplified by arbitrary primers
are useful as genetic markers. Nucleic Acids Research, IS,6531-
6535.
Wright S (1951) The genetical structure of populations. Annals of
Eugenics, 15, 323-354.
Much of the author's research 1s concerned with the ecologcal
and evolutionary determinants of population-genetic structure,
and with the use of molecular markers for quantifying such
structure.