A Bayesian Characterization of Hardy-Weinberg Disequilibrium

Corresponding author: B. S. Weir, Program in Statistical Genetics, Department of Statistics, North Carolina State University, Raleigh, NC 27695-8203., weir{at}stat.ncsu.edu (E-mail).

Communicating editor: A. G. CLARK

	ABSTRACT

TOP ABSTRACT BAYESIAN APPROACH HARDY-WEINBERG EQUILIBRIUM METHODS RESULTS DISCUSSION LITERATURE CITED

A Bayesian method for determining if there are large departures from independence between pairs of alleles at a locus, Hardy-Weinberg equilibrium (HWE), is presented. We endorse the view that a population will never be exactly in HWE and that there will be occasions when there is a need for an alternative to the usual hypothesis-testing setting. Bayesian methods provide such an alternative, and our approach differs from previous Bayesian treatments in using the disequilibrium and inbreeding coefficient parameterizations. These are easily interpretable but may be less mathematically tractable than other parameterizations. We examined the posterior distributions of our parameters for evidence that departures from HWE were large. For either parameterization, when a conjugate prior was used, the prior probability for small departures was itself small, i.e., the prior was weighted against small departures from independence. We could avoid this uneven weighting by using a step prior which gave equal weighting to both small and large departures from HWE. In most cases, the Bayesian methodology makes it clear that there are not enough data to draw a conclusion.

ASSESSING independence of alleles within loci, Hardy-Weinberg equilibrium (HWE), has long been of interest to population geneticists in a variety of contexts. In evolutionary theory, nonindependence is a starting point for determining what forces may be acting to affect variation or changes in allele frequencies (ALLARD et al. 1972 ; CLEGG et al. 1972 ); in forensic science, independence is assumed by the product rule for estimating genotypic frequencies (NATIONAL RESEARCH COUNCIL 1996 ). Studies of departures from independence at single loci go back at least as far as HARDY 1908 and WEINBERG 1908 .

There are many tests for independence of alleles within loci using classic methods: for reviews see EMIGH 1980 , HERNANDEZ and WEIR 1989 , and WEIR 1996 . Bayesian approaches to HWE have been given by PEREIRA and ROGATKO 1984 , LINDLEY 1988 , CHOW and FONG 1992 , and AYRES and BALDING 1998 . We were motivated to consider a Bayesian approach for several reasons. First, this approach allows us to determine the degree of departure from HWE. The classic approach addresses this question indirectly by testing the null hypothesis of zero departure. Second, in describing departures from HWE, we are not directly interested in the allele frequencies; however, they are necessary to define some measure of disequilibrium, which is of primary interest. Bayesian methods provide a convenient way to deal with parameters like allele frequencies, called nuisance parameters, by accounting for their different possible values. Third, it is sometimes the case that we have a set of data in which we observe zero, or very few, individuals with certain genotypes. Classic methods may have trouble accommodating these situations and Bayesian methods offer an alternative to exact tests (e.g., ZAYKIN et al. 1995 ).

This article presents a Bayesian approach to HWE based on the disequilibrium parameter and on the within-population inbreeding coefficient. These parameterizations have the advantage of being easy to interpret biologically. The disadvantage is that for some questions they may be less tractable mathematically.

The traditional approach, to which we have contributed previously and that we still employ, is to test the hypothesis of no departure from HWE. We recognize that failure to reject the hypothesis is not equivalent to declaring HWE, and power calculations allow us to quantify likely bounds on the strength of any departure. An attractive alternative is based on the recognition that real populations are not in HWE and that data, when coupled with prior knowledge, lead to probability distributions for the levels of departure. Bayesian methods replace traditional point estimates and tests of hypotheses about HWE departures with these posterior distributions. We use the posterior distributions to find the probabilities that disequilibrium or inbreeding parameters lie inside or outside intervals of interest. Discussions of Bayesian statistics may be found in BERNARDO and SMITH 1994 , BERGER 1985 , LEE 1989 , O'HAGAN 1994 , and GELMAN et al. 1995 .

	BAYESIAN APPROACH

TOP ABSTRACT BAYESIAN APPROACH HARDY-WEINBERG EQUILIBRIUM METHODS RESULTS DISCUSSION LITERATURE CITED

In this section, we outline a Bayesian approach, and compare it to the classic hypothesis-testing framework in which the HWE problem is usually cast. In both Bayesian and classic methods, the parameters of interest, , are considered fixed and unknown. In addition, the conditional distribution of the data given particular values of the parameters, (X|), is the same. This distribution has the same form as the likelihood, (|X). In the HW testing framework, the data are counts, and are assumed to have a multinomial distribution. Bayesian statistics differ from classic statistics in the mode of inference. In the classic approach, we make statements about values of the parameters from the distribution of the data given the parameters. For example, a P value is a long run probability statement about the data under the null hypothesis. In a Bayesian setting, we make statements about probability distributions on the parameters. All the inference is made according to the posterior distribution on the parameters given the data, (|X). To make such statements, it is necessary to formally express knowledge (or uncertainty) about the parameters using a probability distribution. This distribution, (), is called the prior distribution. If the prior uncertainty is large, the prior distribution is diffuse. It is critical to understand that by using a prior (and posterior) distribution, we are not stating that we think the parameters themselves have a distribution; the parameters are considered unknown and fixed. Instead, the prior distribution on the parameter represents our uncertainty about the fixed value of the parameter. The posterior distribution of the parameters represents our uncertainty, updated by the data, about the parameter values. Using Bayes' rule, the posterior distribution of the parameters is proportional to the product of the prior distribution and the conditional distribution of the data:

(1)

To make the two sides equal, the right-hand side of the equation is divided by (X).

Prior distributions may be constructed based on prior information, mathematical convenience, or ignorance about the parameters (BERNARDO 1979; KASS and WASSERMAN 1993 ; GELMAN et al. 1995 ). Different prior distributions imply different levels of uncertainty about the parameters, not only when the parameters are considered together (joint prior distribution), but also when the parameters are considered separately (marginal prior distribution), or when one of the parameters is considered conditional on values of the other parameters (conditional prior distribution). It is useful to consider what each distribution implies about the parameters and to decide which one(s) to emphasize when constructing the joint prior distribution. This concept is illustrated below in our choice of prior distributions.

There are several ways to report features of the posterior distribution, and thus, there are several ways to compare results from Bayesian and classic approaches. The interpretation of a Bayesian result is completely different from the classic result to which it is being compared. With that caveat, one could report the posterior probabilities that the parameter lies inside an interval of interest. This probability statement is not interpreted as a long term frequency, but as an uncertainty. The most natural classic result to which to compare the posterior probability is a P value, the probability, under the null hypothesis, of obtaining data that are as extreme or more extreme than the observed result. Another way to summarize the posterior distribution is to specify an interval in which most (95%, for example) of the posterior distribution lies. One way to do this is to calculate a highest posterior density region. The most appropriate comparison to a 95% highest posterior density is a 95% confidence interval. In this article, we calculate the posterior probability that the parameter of interest lies within an interval suggested by the NATIONAL RESEARCH COUNCIL 1996 . Therefore, the classic result to which we compare our Bayesian result is a P value.

In a Bayesian framework there are no preset critical values. Of course, this statement is also true in a classic setting—the traditional critical values of 0.05 or 0.01 in a hypothesis-testing framework are not intrinsically part of the method. They are determined by convention or by the experience of the scientist. If specific cutoff values are desired, one could extend the analysis to a decision theory framework (BERGER 1985 ) by factoring in costs associated with making different decisions.

	HARDY-WEINBERG EQUILIBRIUM

TOP ABSTRACT BAYESIAN APPROACH HARDY-WEINBERG EQUILIBRIUM METHODS RESULTS DISCUSSION LITERATURE CITED

In a large, randomly mating population in which the evolutionary forces such as selection, migration, and mutation are not acting, allele and genotypic frequencies do not change. Furthermore, they are related in a simple way. At a locus with two alleles, A, B, with frequencies p_A, 1 - p_A, and genotypes AA, AB, BB with genotypic frequencies P_AA, P_AB, P_BB, the relationship is P_AA = p²_A, P_AB = 2p_A(1 - p_A), and P_BB = (1 - p_A)², and the locus is said to exhibit HWE. One generation of random mating is sufficient to produce HWE.

Classic tests for HWE include goodness-of-fit, exact, likelihood ratio, and z-tests (HALDANE 1954 ; EMIGH 1980 ; HERNANDEZ and WEIR 1989 ; MAISTE 1993 ). Simulations show that one of the most powerful tests for HWE is the exact test, particularly when the number of alleles is large (MAISTE 1993 ).

There are various ways of describing departures from HWE. One way is to use a disequilibrium parameter D_A (HERNANDEZ and WEIR 1989 ):

(2)

where the bounds on D_A are functions of the allele frequencies,

Note that a negative disequilibrium coefficient corresponds to a deficiency of homozygotes relative to HWE, while a positive D_A corresponds to an excess of homozygotes. With m alleles, there are the m(m - 1)/2 independent disequilibrium coefficients.

Another parameterization uses f_A, the inbreeding coefficient within populations (WEIR 1996 ),

(3)

and the bounds on f_A are

Only the lower bound of f_A depends on the allele frequencies. A negative value of f_A corresponds to a deficiency of homozygotes, while a positive value reflects an excess of homozygotes. Despite its name, the f_A parameterization may be used to describe departures from HWE due to any cause, not only departures due to inbreeding.

Two Bayesian studies (PEREIRA and ROGATKO 1984 ; LINDLE 1988) that address departures from HWE at loci with two alleles used a third parameterization

(4)

PEREIRA and ROGATKO 1984 framed their analysis in terms of Bayesian analogs to classic estimation, hypothesis testing, and confidence intervals. They used the mathematically convenient Dirichlet prior on genotype frequencies and estimated . A Bayes factor, which can be thought of as a likelihood ratio, and Bayesian confidence intervals provided the bases for hypothesis testing. LINDLEY 1988 considered various priors, both uniform and nonuniform, on two new parameters

His formulation has the advantage that and ß have bounds that do not depend on the other, making it straightforward to determine the marginal posterior distribution for each variable separately. The posterior distribution of and Bayes' factor = 0 were used for estimation and to draw conclusions regarding HWE.

AYRES and BALDING 1998 were interested in determining departures from HWE and in evaluating the inbreeding model. They used the f_A parameterization and relied upon Markov Chain Monte Carlo methods to sample from the posterior distribution of f_A. They treated loci with multiple alleles.

CHOW and FONG 1992 were interested in comparing estimators for genotype frequencies under HWE. They compared Bayes' estimators, using different priors, to two classic estimators, the uniformly minimum variance-unbiased estimator and the maximum likelihood estimator. The estimators were compared on the basis of risk; risk functions provide means for evaluating a decision, and require a loss function, which quantifies the consequences of making a decision given the true state of nature (O'HAGAN 1994 ). The estimator with the lowest risk was considered the best estimator. The authors showed that, under a standard loss function, Bayesian estimators corresponding to two of the priors were better than classic estimators if the allele frequency was not near zero or one.

	METHODS

TOP ABSTRACT BAYESIAN APPROACH HARDY-WEINBERG EQUILIBRIUM METHODS RESULTS DISCUSSION LITERATURE CITED

Let n₁, n₂, and n₃ be the observed numbers of AA, AB, and BB genotypes from a simple random sample. The counts have a multinomial distribution with independent parameters P_AA, P_AB and sample size n = n₁ + n₂ + n₃:

(5)

We assume that the population is never exactly in HWE and ask whether departures from HWE are large enough to be of importance in the relevant context. In this article, a relevant context is forensic science in which the genotype proportions are estimated as products of allele proportions. The United States National Research Council (NRC) report (NATIONAL RESEARCH COUNCIL 1996 ) suggested that f_A = 0.03 might be the upper limit of departure from HWE in human populations. Strictly, the report assumes that departures make f_A positive, but we will interpret the suggestion to mean |f_A| 0.03. The lower bound must be modified to allow for the dependence on allele frequencies:

(6)

Expression 6 represents a region of the parameter space, and means that the lower bound of f_A is -0.03 unless p_A lies outside the interval [0.03, 0.97]. Since D_A = p_A(1 - p_A)f_A, the bounds of the disequilibrium parameterization become

(7)

We can work directly with the posterior distribution of f_A to calculate the probability that f_A lies in the region of interest. In contrast, we must work with the joint posterior distribution of D_A, p_A to find the corresponding probability for D_A. In each case, however, we use posterior distributions to determine the posterior probabilities of regions 6 or 7. We take high values of these probabilities as evidence that departures from HWE are no larger than |f_A| = 0.03. If the probabilities are low, we view this as evidence that the departures are not within the NRC range. Probabilities that are neither high nor low will be regarded as indicating that the data provide an insufficient basis for drawing a conclusion.

We examine three prior distributions for each parameterization: one based on mathematical convenience, and two based on the question of interest; i.e., the posterior probabilities of regions 6 or 7.

We will examine first the D_A parameterization and then the f_A parameterization.

D_A parameterization:
The first of the three priors we examine is the mathematically convenient Dirichlet prior, which we use because it is conjugate to the multinomial. Use of a conjugate prior results in a posterior that has the same form as the prior. We apply the Dirichlet prior to the genotypic proportions, P_AA, and P_AB, and then transform to D_A, p_A. We write the Dirichlet parameters as _i(i = 1,2,3):

(8)

where · = E³_{i=1 i} . The corresponding joint prior distribution of D_A and p_A is

(9)

Combining the prior density with the likelihood, the posterior density of the genotype proportions given the data is Dirichlet (_i + n_i), i = 1,2,3. After transformation to p_A, D_A the joint posterior density becomes

(10)

In the absence of prior information and without regard to the question of interest, we choose to set the _i's equal to 1, i.e., a uniform prior for the genotype proportions. In this case, the joint prior of D_A and p_A is uniform over the parameter space (Figure 1, top). We refer to this prior as prior A. The second and third prior distributions were based on the question about departures from HWE. We wanted the prior probability that D_A lies inside the NRC range to be 0.5, reflecting our ignorance about whether or not departures from HWE are within the NRC range. For the second prior, we let the prior distribution for p_A, (p_A), be Beta(, ß), which has been shown by WRIGHT 1951 to be appropriate under certain conditions. We set = ß = 1, the uniform prior. The prior distribution for D_A given p_A, (D_A|p_A) was a step function uniform over each of 3 ranges for D_A|p_A, with the corresponding weights 0.25, 0.50, and 0.25 (Table 1). We refer to this prior as prior B. The joint prior distribution (D_A, p_A) is (p_A)(D_A|p_A) and is shown in Figure 2 (top).

View larger version (38K): In this window In a new window Download PPT slide   Figure 1. Joint prior (top) and posterior (bottom) distributions of DA, pA based on prior A. For the posterior distribution, the data were from the LDLR locus in the FBI Caucasian data base.

View larger version (43K): In this window In a new window Download PPT slide   Figure 2. Joint prior (top) and posterior (bottom) distributions of DA, pA based on prior B. For the posterior distribution, the data were from the LDLR locus in the FBI Caucasian data base.

The third prior distribution was constructed so that the joint prior distribution of D_A, p_A was uniform over each of the three ranges listed in Table 1, with the corresponding weights. We refer to this prior as prior C (Figure 3, top).

View larger version (34K): In this window In a new window Download PPT slide   Figure 3. Joint prior (top) and posterior (bottom) distributions of DA, pA based on prior C. For the posterior distribution, the data were from the LDLR locus in the FBI Caucasian data base.

f_A parameterization:
The development of the three priors for this parameterization paralleled that for the D_A parameterization. The Dirichlet prior distribution on the genotype frequencies, P_AA and P_AB, was transformed to f_A and p_A:

(11)

If the _i's are set equal to 1, the resulting prior distribution, in contrast to that for the D_A parameterization, is not uniform over the parameter space (Figure 4, top).

View larger version (44K): In this window In a new window Download PPT slide   Figure 4. Joint prior (top) and posterior (bottom) distributions of fA, pA based on prior A. For the posterior distribution, the data were from the LDLR locus in the FBI Caucasian data base.

Combining the joint prior distribution with the likelihood, the joint posterior distribution of the genotype proportions, given the data, is Dirichlet (_i + n_i), i = 1,2,3. After the transformation, the joint posterior density is

(12)

The second and third priors were constructed so that the prior probability of region 6 was 0.5. For the second prior, prior B, the prior for p_A was Beta(, ß), with = ß = 1, as for the D_A, p_A parameterization, and the prior for f_A given p_A was uniform over each of the three pieces corresponding to the ranges in Table 1. The joint prior distribution (f_A, p_A) is (p_A)(f_A|p_A), and is shown in Figure 5 (top).

View larger version (35K): In this window In a new window Download PPT slide   Figure 5. Joint prior (top) and posterior (bottom) distributions of fA, pA based on prior B. For the posterior distribution, the data were from the LDLR locus in the FBI Caucasian data base.

The third prior, prior C, was constructed so that the joint prior distribution of f_A, p_A was uniform over the ranges, with corresponding weights, listed in Table 1. The joint prior distribution of f_A, p_A is shown in Figure 6 (top).

View larger version (40K): In this window In a new window Download PPT slide   Figure 6. Joint prior (top) and posterior (bottom) distributions of fA, pA based on prior C. For the posterior distribution, the data were from the LDLR locus in the FBI Caucasian data base.

Under either prior A or prior B, the posterior probabilities of regions 6 and 7 were the same. In the case of prior A, it is because the Dirichlet prior is placed on the genotypes, and not directly on D_A, p_A or f_A, p_A. In the case of the B prior, it is because the relationship between D_A (or f_A) conditional on p_A is linear. Despite the identity of the posterior probabilities for specific regions, we think it useful to present both parameterizations: questions may be more conveniently phrased in terms of one of the two parameterizations, depending upon the context. In addition, it is possible to construct priors, such as the C prior, in which the posterior distributions will not be identical, because the priors were constructed by focusing not on the conditional distributions of D_A (or f_A) given p_A, but on the joint distributions of D_A (or f_A) and p_A.

For both parameterizations, we evaluated the posterior distributions by using a midpoint rule with 100,000 points (STOER and BULIRSCH 1980 ).

	RESULTS

TOP ABSTRACT BAYESIAN APPROACH HARDY-WEINBERG EQUILIBRIUM METHODS RESULTS DISCUSSION LITERATURE CITED

It is critical to understand that the prior and posterior distributions reflect uncertainty about the fixed but unknown parameter values. It is not the case that the parameters themselves have a distribution. Any statements regarding prior and posterior distributions or probabilities should be interpreted in this context.

The three different prior distributions describe uncertainty about the parameters D_A (or f_A) and p_A in different ways. Each prior distribution implies information about the uncertainty of D_A (or f_A) and p_A together (joint distribution) or D_A (or f_A) and p_A separately (marginal distribution), or D_A (or f_A) conditional on a fixed value of p_A (conditional distribution). The first prior was chosen for convenience. It places a Dirichlet(1,1,1), i.e., a uniform prior distribution, on the genotypes, which translates to a different prior on parameters of interest (Figure 1 and Figure 4, top). In the D_A parameterization, the joint distribution is also uniform over all possible pairs of D_A and p_A values. This means that if we consider uncertainty about D_A and p_A together, each possible pair of D_A, p_A values is equally uncertain. In the case of the f_A parameterization, a uniform prior on the genotypes does not translate into a uniform prior on f_A and p_A (Figure 4, top). Instead, pairs of f_A, p_A with central values have a higher probability than pairs with extreme values. The marginal prior distributions of D_A and p_A are symmetric with modes of 0 and 0.5, respectively; the marginal prior distribution of f_A implies that, under this prior, large negative values are less certain than small negative values, and that positive values of f_A are (approximately) as certain as the smallest negative value. Finally, the prior probability that D_A is inside the NRC range is only 0.04, compared to a prior probability of 0.50 when either prior B or C is used. The low prior probability of 0.04 reflects the fact that the first prior is more diffuse, and so the prior probability of any small interval, including the one in which we are interested, is small. This low prior probability makes it difficult to interpret the posterior probability (of regions 6 or 7) because it is weighted against being in the NRC range.

Prior B (Figure 2 and Figure 5, top) and prior C (Figure 3 and Figure 6, top) were constructed so that the probability that D_A (or f_A) is in the NRC range is equal to the probability that it is outside the NRC range. In addition, under this prior, the probability that pairs of values D_A, p_A (or f_A, p_A) are inside the NRC range is equal to the probability that they are outside the NRC range. However, the second prior emphasizes uncertainty about the marginal distribution of p_A and the conditional distribution, D_Ap_A (or f_A|p_A). The marginal distribution of p_A, which is uniform, implies that each value of p_A is equally uncertain. The conditional distribution of D_A|p_A (or f_A|p_A) implies that, for each fixed value of p_A, values of D_A (or f_A) in the NRC range are equally uncertain, values of D_A (or f_A) below the NRC range are equally uncertain, and values of D_A (or f_A) above the NRC range are equally uncertain. The third prior emphasizes the joint distribution; it implies that all pairs of D_A (or f_A) and p_A within the NRC range are equally uncertain, all pairs below the NRC range are equally uncertain, and all pairs above the NRC range are equally uncertain.

We applied the above methods to the published genotypic data shown in Table 2. The "FBI" data are given in BUDOWLE et al. 1995 and the "Cellmark" data are provided by WEIR et al. (B. S. WEIR, J. STOREY, R. COTTON, C. BASTEN, J. BUCKLETON and J. MORRIS, unpublished results). Three two-allele loci were used: D7S8 (GYLLENSTEN and ERLICH 1988), LDLR (YAMAMOTO et al. 1984 ), and GYPA (SIEBERT and FUKUDA 1987 ).

For locus LDLR in the FBI Caucasian sample, Figure 1 Figure 2 Figure 3 (bottom) show the posterior distribution of D_A, p_A given the data, using the three different priors. In Figure 1, since the joint prior on D_A, p_A was uniform, the posterior represents the transformed likelihood, standardized to integrate to 1. As expected, the joint posterior was in three sections when the prior B or C was used. In these cases (Figure 2 and Figure 3, bottom), the volume in the middle section represents the probability that D_A lies in the NRC range. Figure 4 Figure 5 Figure 6 (bottom) show the analogous posteriors for f_A and p_A.

Table 3 shows the posterior probabilities that D_A or f_A are inside the NRC range as well as the P values for an exact test of HWE. As expected, except for very small differences because of the numerical integration, the posterior probabilities were the same for either parameterization when the Dirichlet or prior B was used; thus, we reported only one number in these cases (columns P_A, P_B). There is a dramatic difference between the posterior probabilities when the mathematically convenient prior was used versus when priors B or C were used. This reflects the very different prior probabilities for the parameters lying within the NRC range—0.04 for the prior A and 0.50 for priors B and C. In no cases were the posterior probabilities for the parameters lying within the NRC range very high, although in a few cases, they approached high values. In one case (LDLR locus, Cellmark African American), the posterior probability approached a low value, which would lend support to the view that the parameters lie outside the NRC range.

	DISCUSSION

TOP ABSTRACT BAYESIAN APPROACH HARDY-WEINBERG EQUILIBRIUM METHODS RESULTS DISCUSSION LITERATURE CITED

Bayesian methods are not new to genetics. Animal breeders have used Bayesian methods in assessing response to selection, estimating breeding values and/or variance components (GIANOLA and FERNANDO 1986 ; SORENSEN et al. 1994; WANG et al. 1994 ) and locating markers linked to quantitative trait loci (HOESCHELE and VANRADEN 1993A , HOESCHELE and VANRADEN 1993B ; UIMARI et al. 1996 ). In population genetics, Bayesian approaches have been used to to estimate allele frequencies (GUNEL and WEARDEN 1995), classify genotypes (ALEXANDER et al. 1995 ), and distinguish disomic from tetrasomic inheritance (OLSON 1997 ). LANGE 1995 used empirical Bayes methods to estimate allele frequencies in different subpopulations. Bayesian methods have been used in linkage analysis in recombinant inbred lines (SILVER and BUCKLER 1986 ), pine trees (GEBUREK and VON WUEHLISCH 1989 ) and humans, the latter in the context of risk assessment (ROGATKO 1995A , ROGATKO 1995B ). In molecular evolution, Bayesian methods for inferring phylogenetic trees have been proposed (MAU and NEWTON 1997 ; YANG and RANNALA 1997 ). Our goal was to develop a Bayesian approach to HWE in a framework familiar to population geneticists. Our method differs from those of previous authors PEREIRA and ROGATKO 1984; LINDLEY 1988; AYRES and BALDING 1998) by using the disequilibrium and inbreeding parameterizations, and in defining levels of disequilibrium that are of interest in one context.

The posterior probabilities for prior A and priors B or C were quite different, indicating that the prior had a large effect with the sample sizes for the data we considered. Using prior A, the prior probabilities of the regions 6, 7 were only 0.04, while with priors B or C, the prior probability was 0.50. Therefore, the larger posterior probabilities for the latter priors are not surprising. Because the uniform prior was more diffuse than priors B or C, it weighted against departures from HWE that included only a small part of the parameter space, including the small departures from HWE in which we were interested. It is therefore difficult to draw conclusions from the analysis when the prior A is used. Because priors B and C were chosen to reflect our ignorance of whether or not the NRC recommendation is conservative, the resulting analyses are easier to interpret.

For most of the loci, there were not enough data in Table 1 to draw conclusions. For the LDLR locus in the Hispanic data base, the NRC recommendation may be conservative. In one case (LDLR locus in the Cellmark African American data base), the NRC recommendation may not be conservative. This was the same locus for which we would reject the HWE hypothesis using a classical test. These results do not preclude the use of these loci for forensic calculations because genotype proportions can be substituted for HWE proportions. Note that, under the Bayesian analysis, power calculations are not needed to distinguish the case in which there are not enough data from the case in which the departures are within the NRC range. In addition, we are asking directly the question of interest, i.e., what is the probability of the parameter being in the NRC range, instead of asking the question indirectly via a hypothesis test.

We have presented both the D_A, p_A and f_A, p_A parameterizations, because some researchers may find it more helpful to phrase their question in the context of one parameterization or the other. In our case, the NRC recommendation is phrased in terms of the f_A, p_A parameterization. Although the lower bound of f_A depends on the nuisance parameter p_A, because of the way the question is posed, we can work directly with the posterior distribution of f_A, although we find it more helpful to present graphs for the joint rather than the marginal posterior distribution. In contrast, for the D_A, p_A parameterization, we cannot work directly with the marginal distribution of D_A, and instead must work with the joint posterior distribution of D_A, p_A. Because we have only two parameters, the methods used to evaluate the joint posterior are not difficult to implement. However, with multiple alleles, the problem could be significant, in which case Markov Chain Monte Carlo methods would be useful (AYRES and BALDING 1998 ).

In summary, we have proposed a Bayesian method for examining departures from HWE based on the disequilibrium and inbreeding parameterizations. The priors based on the conjugate Dirichlet distribution were heavily weighted against small departures from HWE. It is difficult to interpret the posterior probabilities in these cases. We need to find other priors that will more accurately reflect prior information or the question of interest. The priors may be formulated as step priors and depend on the question of interest. In the present context, we based our question on recommendations by the NRC and asked whether there was evidence that the inbreeding (or disequilibrium) parameter lay inside the NRC range. In many of the samples we examined there were not enough data to draw a conclusion about that question.

	ACKNOWLEDGMENTS

We appreciate the advice of Drs. IAN EVETT and JOHN MONAHAN. The manuscript was improved by comments from Drs. JAMES CURRAN, JEFF LEIPS, CHRIS TRIGGS, and two anonymous reviewers. Thanks are due to DMITRI ZAYKIN for use of his program to calculate probabilities for the exact test. This work was supported in part by U.S. Public Health Service grants P01 GM43544 and T32 GM08443.

Manuscript received November 18, 1997; Accepted for publication April 27, 1998.

	LITERATURE CITED

TOP ABSTRACT BAYESIAN APPROACH HARDY-WEINBERG EQUILIBRIUM METHODS RESULTS DISCUSSION LITERATURE CITED

ALEXANDER, H., N. SLADE, and R. GOMULKIEWICZ, 1995  A Bayesian approach to the inference of diploid genotypes using haploid genotypes. Theor. Appl. Genet. 91:1284-1287.

ALLARD, R. W., A. L. KAHLER, and B. S. WEIR, 1972  The effects of selection on esterase allozymes in a barley population. Genetics 72:489-503[Abstract/Free Full Text].

AYRES, K. L. and D. J. BALDING, 1998  Measuring departures from Hardy-Weinberg: a Markov Chain Monte Carlo method for estimating the inbreeding coefficient. Heredity 80:769-777.

BERGER, J. O., 1985 Statistical Decision Theory and Bayesian Analysis. Ed. 2, Springer-Verlag, New York.

BERNADO, J. M., 1979  Reference posterior distributions for Bayesian inference (with discussion). J. Roy. Stat. Soc. B 41:113-147.

BERNARDO, J. M., and A. F. M. SMITH, 1994 Bayesian Theory. Wiley, Chichester.

BUDOWLE, B., J. LINDSEY, J. DECOU, B. KOONS, and A. GIUSTI et al., 1995  Validation and population studies of the loci LDLR, GYPA, HBGG, D7S8 and Gc (PM) loci, and HLA-DQ using a multiplex amplification and typing procedure. J. Forensic Sci. 40:45-54[Medline].

CHOW, M. and D. K. H. FONG, 1992  Simultaneous estimation of the Hardy-Weinberg proportions. Can. J. Stat. 3:291-296.

CLEGG, M. T., R. W. ALLARD, and A. L. KAHLER, 1972  Is the gene the unit of selection? Evidence from two experimental plant populations. Proc. Natl. Acad. Sci. USA 69:2474-2478[Abstract/Free Full Text].

EMIGH, T. H., 1980  A comparison of tests for Hardy-Weinberg equilibrium. Biometrics 36:627-642.

GEBUREK, T. and G. VON WUEHLISCH, 1989  Linkage analysis of isozyme gene loci in Picea abies (L.) Karst. Heredity 62:185-191.

GELMAN, A., J. CARLIN, H. STERN and D. RUBIN, 1995 Bayesian Data Analysis. Chapman and Hall, London.

GIANOLA, D. and R. L. FERNANDO, 1986  Bayesian methods in animal breeding theory. J. Anim. Sci. 63:217-244[Abstract/Free Full Text].

GUNEL, E. and S. WEARDON, 1995  Bayesian estimation and testing of gene frequencies. Theor. Appl. Genet. 91:534-543.

GYLLENSTERN, U. B. and H. A. ERLICH, 1988  Generation of single-stranded DNA by the polymerase chain reaction and its application to direct sequencing of the HLA-DQ alpha locus. Proc. Natl. Acad. Sci. USA 85:7652-7656[Abstract/Free Full Text].

HALDANE, J. B. S., 1954  An exact test for randomness of mating. J. Genet. 52:631-635.

HARDY, G. H., 1908  Mendelian proportions in a mixed population. Science 28:49-50[Free Full Text].

HERNÁNDEZ, J. L. and B. S. WEIR, 1989  A disequilibrium approach to Hardy-Weinberg testing. Biometrics 45:53-70[Medline].

HOESCHELE, I. and P. M. VANRADEN, 1993a  Bayesian analysis of linkage between genetic markers and quantitative trait loci. I. Prior knowledge. Theor. Appl. Genet. 85:946-952.

HOESCHELE, I. and P. M. VANRADEN, 1993b  Bayesian analysis of linkage between genetic markers and quantitative trait loci. II. Combining prior knowledge with experimental evidence. Theor. Appl. Genet. 85:953-960.

KASS, R. E., and L. WASSERMAN, 1993 Formal Rules for Selecting Prior Distributions: A Review and Annotated Bibliography. Tech. Report 583, Department of Statistics, Carnegie Mellon University.

LANGE, K., 1995  Applications of the Dirichlet distribution to forensic match probabilities. Genetica 96:107-111[Medline].

LEE, P. M., 1989 Bayesian Statistics: An Introduction. Oxford University Press, Oxford.

LINDLEY, D., 1988  Statistical inference concerning Hardy-Weinberg equilibrium. Bayesian Stat. 3:307-326.

MAISTE, P. J., 1993 Comparison of Statistical Tests for Independence at Genetic Loci with Many Alleles. Ph.D Thesis, North Carolina State University, Raleigh, N.C.

MAU, B. and M. A. NEWTON, 1997  Phylogenetic inference for binary data on dendograms using Markov chain Monte Carlo. J. Computat. Graph. Stat. 6:122-131.

NATIONAL RESEARCH COUNCIL, 1996 The Evaluation of Forensic DNA Evidence. National Academy Press, Washington, D. C.

O'HAGAN, A., 1994 Kendall's Advanced Theory of Statistics. Volume 2B. Bayesian Inference. Edward Arnold, London.

OLSON, M. S., 1997  Bayesian procedures for discriminating among hypotheses with discrete distributions: Inheritance in the tetraploid Asilbe biternata.. Genetics 147:1933-1942[Abstract].

PEREIRA, C. and A. ROGATKO, 1984  The Hardy-Weinberg equilibrium under a Bayesian perspective. Rev. Bras. Genet. 4:689-707.

ROGATKO, A., 1995a  Risk prediction with linked markers: Theory. Am. J. Med. Genet. 59:14-23[Medline].

ROGATKO, A., 1995b  Risk prediction with linked markers: Pedigree analysis. Am. J. Med. Genet. 59:24-32[Medline].

SIEBERT, P. D. and M. FUKUDA, 1987  Molecular cloning of human glycophorin B cDNA: Nucleotide sequence and genomic relationship to glycophorin A. Proc. Natl. Acad. Sci. USA 84:6735-6739[Abstract/Free Full Text].

SILVER, J. and C. E. BUCKLER, 1986  Statistical considerations for linkage analysis using recombinant inbred strains and backcrosses. Proc. Natl. Acad. Sci. USA 83:1423-1427[Abstract/Free Full Text].

SORENSON, D., C. WANG, J. JENSEN, and D. GIANOLA, 1994  Bayesian analysis of genetic change due to selection using Gibbs sampling. Genet. Sel. Evol. 26:333-360.

STOER, J., and R. BULIRSCH, 1980 Introduction to Numerical Analysis. Springer-Verlag, New York.

UIMARI, P., G. THALLER, and I. HOESCHELE, 1996  The use of multiple markers in a Bayesian method for mapping quantitative trait loci. Genetics 143:1831-1842[Abstract].

WANG, C., D. GIANOLA, D. SORENSON, J. JENSEN, and A. CHRISTIANSEN et al., 1994  Response to selection for litter size in Danish landrace pigs: a Bayesian analysis. Theor. Appl. Genet. 88:220-230.

WEINBERG, W., 1908  Über den Nachweis der Vererbung beim Menschen. Jh. Ver. vaterl. Naturk. Württemb. 64:369-382.

WEIR, B. S., 1996 Genetic Data Analysis II. Sinauer, Sunderland, MA.

WRIGHT, S., 1951  The genetical structure of populations. Ann. Eugen. 15:323-354.

YANG, Z. and B. RANNALA, 1997  Bayesian phylogenetic inference using DNA sequences: A Markov chain Monte Carlo method. Mol. Biol. Evol. 14:717-724[Abstract].

YAMAMOTO, T., C. G. DAVIS, M S, W. J. BROWN, M. L. SCHNEIDER, and M. L. SCHNEIDERCASEY et al., 1984  The human LDL receptor: a cysteine-rich protein with multiple Alu sequences in its mRNA. Cell 39:27-38[Medline].

ZAYKIN, D., L. ZHIVOTOVSKY, and B. S. WEIR, 1995  Exact tests for association between alleles at arbitrary numbers of loci. Genetica 96:169-178[Medline].