Home | About | Journals | Submit | Contact Us | Français |

**|**HHS Author Manuscripts**|**PMC2732762

Formats

Article sections

Authors

Related links

Genet Epidemiol. Author manuscript; available in PMC 2010 September 1.

Published in final edited form as:

Genet Epidemiol. 2009 September; 33(6): 508–517.

doi: 10.1002/gepi.20403PMCID: PMC2732762

NIHMSID: NIHMS93961

Department of Biostatistics and Center for Statistical Genetics, School of Public Health, University of Michigan, Ann Arbor, Michigan 48109-2029

Corresponding author: Gonçalo R. Abecasis, Ph.D., Department of Biostatistics, School of Public Health, University of Michigan, 1420 Washington Heights, Ann Arbor Michigan 48109-2029, (734)763-4901, Email: ude.hcimu@olacnog

See other articles in PMC that cite the published article.

Genome-wide association studies are helping to dissect the etiology of complex diseases. Although case-control association tests are generally more powerful than family-based association tests, population stratification can lead to spurious disease-marker association or mask a true association. Several methods have been proposed to match cases and controls prior to genotyping, using family information or epidemiological data, or using genotype data for a modest number of genetic markers. Here, we describe a genetic similarity score matching (GSM) method for efficient matched analysis of cases and controls in a genome-wide or large-scale candidate gene association study. GSM is comprised of three steps: 1) calculating similarity scores for pairs of individuals using the genotype data; 2) matching sets of cases and controls based on the similarity scores so that matched cases and controls have similar genetic background; and 3) using conditional logistic regression to perform association tests. Through computer simulation we show that GSM correctly controls false positive rates and improves power to detect true disease predisposing variants. We compare GSM to genomic control using computer simulations, and find improved power using GSM. We suggest that initial matching of cases and controls prior to genotyping combined with careful re-matching after genotyping is a method of choice for genome-wide association studies.

With the success of the International HapMap Project [The International HapMap Consortium, 2007], a dense set of single nucleotide polymorphisms (SNPs) throughout the human genome is now available for genetic studies of complex diseases, and many genome-wide association studies are being undertaken and published [Klein et al., 2005; Maraganore et al., 2005; Cheung et al., 2005; Sladek et al., 2007; Scott et al., 2007; Saxena et al., 2007; Zeggini et al., 2007].

Although case-control association tests are in principle more powerful for detecting disease variants than family-based association tests, population stratification can lead to spurious disease-marker association or mask true association [Li, 1972]. In genome-wide association studies, thousands of samples are typically used to ensure adequate power to identify disease predisposing variants, making it difficult to guarantee genetic homogeneity of the sample [Freedman et al., 2004]. Ancestry information on the sampled individuals may be unavailable to the researchers, and even when available, may not fully specify the underlying population genetic structure, due to vague definitions of ancestry groups and imperfect accuracy of self-report information.

Several methods have been proposed to adjust for the possible confounding effects of population substructure. Family-based association tests, such as the transmission/disequilibrium test (TDT) [Spielman et al., 1993], assess the transmission of alleles from parents to affected offspring. Comparisons are made within parent-offspring trios, and the resulting association test is immune to potential genetic heterogeneity between families. However, collecting trios can be difficult and expensive, and may simply be impractical for late-onset diseases. For unrelated case-control samples, approaches have been proposed to adjust the standard chi-square contingency test statistics according to a non-central χ^{2} distribution [Devlin et al., 1999; Gorroochurn et al., 2006], to infer population structure [Pritchard et al., 2000], or to cluster the similarity estimates into several components [Zhang et al., 2002]. A few more recent approaches [Price et al., 2006; Epstein et al., 2007; Kimmel et al., 2007; Luca et al., 2008] focus specifically on genome-wide association studies.

In this paper, we propose a different approach, genetic similarity score matching (GSM), to correct population stratification using individual-based matching rather than clustering. The huge amounts of data in genome-wide association studies have the potential to provide extremely accurate matching of individuals who share similar ancestries. We match cases with controls based on genetic (dis)similarity scores calculated from the genotype data available in a genome-wide association study or a large-scale candidate gene study and test the resulting matched sets for disease-marker association by conditional logistic regression. This matching-association framework builds on our previous work [Guan et al. 2005] and is similar to that of Luca et al. [2008]. Luca et al. [2008] derive the dissimilarity (distance) scores based on principal components of the variance matrix of genotypes, while our approach obtains the dissimilarity scores based on identity-by-state (IBS) measures. Simulations show that GSM results in false positive rates at the desired nominal level while retaining high power to detect disease associated markers. We find that with large-scale association data, the calculated genetic similarity scores differentiate subpopulations well, and that matching can be done with high accuracy even for samples that are mixtures of genetically similar populations. We further demonstrate that when population stratification is present, association tests based on GSM-matched case-control data can have a higher power than those that rely on either the standard trend test or the genomic-control method.

GSM includes three basic components:

*Genetic similarity score:*We calculate genetic similarity scores between pairs of cases and controls across all loci. Large scores should reflect pairs with similar genetic backgrounds.*Matching:*Based on the matrix of similarity scores calculated in 1), we conduct optimal full matching [Rosenbaum, 2002] which groups one case with one or more controls, or one control with one or more cases to maximize the overall similarity of matched cases and controls.*Association tests:*We use conditional logistic regression to assess the association between candidate markers and disease status. For ease of exposition, we consider here only single marker association tests, but other genetic or environmental factors can be easily incorporated into the regression.

We define a genetic similarity score for a pair of individuals which measures the degree of similarity of their genotype data. Individuals with similar genetic backgrounds will generally have higher scores. For simplicity, we consider *M* biallelic genetic markers each with alleles “A” and “a”; the scores can easily be generalized to multiallelic markers. We consider three similarity scores.

The first score calculates the proportion of marker alleles shared identical by state (IBS). If *IBS _{k}* is the number of alleles shared at marker

$${S}_{\mathit{\text{IBS}}}=\frac{1}{2{M}^{\ast}}\sum _{k=1}^{{M}^{\ast}}IB{S}_{k}$$

(1)

where 1 ≤ *M** ≤ *M* is the number of markers that are successfully genotyped in both individuals.

While *S _{IBS}* has the virtue of simplicity, we may want to allow different markers to make different contributions to measure similarity. For example, we may wish to weight sharing a rare allele more strongly than sharing a common allele. We define our second score as:

$${S}_{\mathit{\text{freq}}}=-\frac{1}{2{M}^{\ast}}\sum _{k=1}^{{M}^{\ast}}\sum _{i\in \{A,a\}}IB{S}_{k,i}\cdot log({q}_{k,i})$$

(2)

where *q _{k,i}* is the frequency of allele

In a random mating population, markers are expected to follow Hardy-Weinberg Equilibrium (HWE). When population subdivision is present, tests of HWE tend to be significant owing to excess homozygosity. Our third score takes advantage of this by weighting markers based on their one-sided (excess homozygosity) HWE test p-value *p _{k}* [Wigginton et al., 2005]:

$${S}_{\mathit{\text{HWE}}}=-\frac{\text{1}}{\text{2}{M}^{\ast}}\sum _{k=1}^{{M}^{\ast}}IB{S}_{k}\cdot log({p}_{k})\phantom{\rule{2em}{0ex}}$$

(3)

To avoid the impact of genotyping error that may lead to strong deviation from HWE, we exclude the markers that fail quality control; practically speaking, this might mean using markers with HWE p-value satisfying *p*>10^{-6}.

As an example, suppose 3 cases and 3 controls are genotyped at 3 loci, as listed in Table 2. Then the similarity scores *S _{IBS}* are as listed in Table 3.

For matching, we may use all genotyped markers, or a selected subset. For example, we might pick the markers with the smallest p-values in a HWE test for excess homozygosity, excluding those that fail quality control, in the hope that the selected markers provide maximal information about population stratification in the sample. Further, to avoid selecting markers which are highly correlated, we might choose at most one marker in every *n*-marker window or per linkage disequilibrium group.

In our analyses, matching relies on a transformed dissimilarity score, defined as:

$${D}_{ij}=f({S}_{ij})={\left(\frac{max-{S}_{ij}}{max-min}\right)}^{2}$$

(4)

where max = max* _{i,j} S_{ij}* and min = min

We use the chosen (dis)similarity score to identify optimal matches between cases and controls. The simplest matching scheme is a 1:1 match in which each case is matched to a unique control. This approach is widely used but has obvious drawbacks. For example, when the numbers of cases and controls are not equal, some subjects must be discarded, resulting in a loss of information. Further, samples from various subpopulations often are not equally represented among the cases and controls, leading to forced mismatches if only 1:1 matching is allowed.

Instead, we consider an optimal matching approach that minimizes the total dissimilarity score:

$$T=\sum _{s=1}^{S}\sum _{i\in {A}_{s},j\in {B}_{s}}{D}_{ij}$$

Here, *A _{s}* and

The problem of minimizing the total dissimilarity score *T* is analogous to the classic minimum cost flow (MCF) problem in computer science [Rosenbaum, 1991; Hansen, 2004; Hansen et al., 2006] (Appendix), and can be solved using the RELAX-IV algorithm [Bertsekas et al., 1994; Frangioni et al., 2006]. Given pre-calculated dissimilarity scores and an upper bound on *m*, determining the optimal matched set takes on the order of *n*^{3} log *n* operations, where *n* is the total number of subjects. The choice of parameter *m* constrains the size of matched sets and is somewhat arbitrary; we typically require *m≤5* when numbers of cases and controls are comparable (see Discussion). Prior to matching, we may exclude a few individuals with maximum similarity scores that are extremely small (this is the *caliper* parameter recommended by Hansen et al., 2006). In datasets including ~2,000 individuals, the matching typically takes <1 minute on a modern PC workstation.

To continue with the previous example, we calculate the dissimilarity scores in Table 3, and perform both 1:1 matching and optimal matching. In 1:1 matching, the best match yields three pairs: (1, 4), (2, 5), and (3, 6). The total dissimilarity score is 1/36 + 16/36 + 0 = 17/36. In contrast, the optimal full match has two matched sets: (1, 2, 4) and (3, 5, 6). The matched sets include 4 case-control pairs: (1, 4), (2, 4), (3, 5), and (3, 6). The total dissimilarity score is 1/36 + 0 + 1/36 + 0 = 2/36. In this example, the individuals within group (1, 2, 4) and (3, 5, 6) are similar to each other, and less similar to the individuals in the other group. Full matching offers an obvious matching advantage over 1:1 matching here. In the general case, full matching is guaranteed to produce a total dissimilarity score that is no greater than that obtained using 1:1 matching.

Once matching is done, a natural choice for matched-set analysis is to use conditional logistic regression to test for disease-marker association. We employ an additive model for association by assigning values of 0, 1, and 2 to genotypes AA, Aa, and aa, respectively. Other genotyping coding schemes could be considered, corresponding for example to dominant, recessive, or general models. The regression can easily incorporate genotype, covariate, and interaction effects.

In a genome-wide association scan, we apply conditional logistic regression analysis to each marker separately. The multiple testing problem can be addressed using Bonferroni correction, permutation, or false-discovery rates.

We simulated case-control data influenced by genotypes at a disease locus with alleles D and d, under six additive disease models (Table 4). We assumed sampling from a population that consisted of two subpopulations. We randomly sampled 500 cases and 500 controls from this mixed population. For each model, the relative risk (RR) of the predisposing variant allele is set to be the same in different populations. For models 1 and 2, the disease prevalences K_{1}=K_{2} and predisposing variant allele frequencies *q*_{1}=*q*_{2}; these models represent the scenario of no population stratification. For models 3 and 4, K_{1}<K_{2}, creating population stratification in the simulated data. For models 5 and 6, K_{1}<K_{2} and *q*_{1}≠*q*_{2}. For model 5, the first population has lower prevalence but higher predisposing variant allele frequency (K_{1}=.07, *q*_{1}=.55), than the second population (K_{2}=.13, *q*_{2}=.45). For model 6, the population with higher prevalence also has higher predisposing variant allele frequency (K_{2}=.13, *q*_{2}=.55) than the other population (K_{1}=.07, *q*_{1}=.45). For each model, we simulated 500 datasets.

We simulated autosomal SNPs using GENOME, a coalescent-based simulator [Hudson, 1983; Hudson, 1990; Donnelly et al., 1995; Liang et al., 2007]. Assuming discrete generations, GENOME simulates the genealogy of a sample of sequences. As the algorithm proceeds backwards in time, coalescence, recombination, and migration events are simulated. Multiple events can occur in the same generation. We set the effective population size as 10,000, the recombination rate as 10^{-8} per base pair, and the mutation rate as 10^{-9} per base pair, assuming the infinite-site mutation model [Kimura, 1969]. We set the rate of migration between subpopulations to .0025 per individual per generation, which resulted in a distribution of allele frequency differences similar to that observed when comparing HapMap Han Chinese (HCB) and Japanese (JPT) samples (www.hapmap.org). In particular, the mean allele frequency difference between the two simulated populations is .0470, compared to .0477 between the HCB and JPT samples. The simulated genome scans surveyed autosomal genomes of ~2866 Mb comprised of 22 chromosomes, whose lengths approximate the actual lengths of the human autosomes (NCBI build 33, www.ncbi.nlm.nih.gov/genome/seq/). We randomly selected 300,000 SNPs with minor allele frequencies > .05, and choose a disease liability locus with the desired allele frequencies.

To calculate the similarity scores, we used 10,000 markers with the smallest one-sided HWE p-values, choosing no more than one marker from each 10-marker window. We set the maximum size of matched groups (*m*) to 6. We compared the type I error and power of GSM, the trend test, genomic control, and EIGENSTRAT for each simulated setting. Given that the simulated samples were drawn from two subpopulations, we used the first principal component to adjust for stratification in EIGENSTRAT; using additional principal components gave similar results. The estimated type I error rates are the proportion of simulated SNPs in which the association test p-value is less than the nominal value 10^{-6}, a significance threshold similar to that typically used in genome-wide scans. In this evaluation of type I error rates, we only considered SNPs that were effectively unlinked to the disease locus. We calculated power as the proportion of simulated replicates where the empirical p-value is < 10^{-6} at the disease locus using a threshold obtained by inspection of test statistics at the null loci.

We applied GSM to genome-wide association data from the Pritzker Consortium bipolar study (unpublished data). We selected 717 independent bipolar I European American cases and 779 independent European American controls from NIMH Human Genetics Initiative (www.nimhgenetics.org); controls were carefully matched to cases by self-reported ethnicity prior to genotyping. In addition, we downloaded genotype data on 3,182 independent European American controls from Illumina iControlIDB database (www.illumina.com/pages.ilmn?ID=231). All individuals were genotyped using the Illumina HumanHap550 BeadChip. 505,796 autosomal SNPs passed quality-control criteria in the Prtizker bipolar study: 1) HWE p-values > 10^{-5}; 2) genotype call rate > 95%; and 3) no more than 1 non-mendelian inheritance or inconsistency among 15 father-mother-offspring trios and 30 duplicate samples. Of these, we excluded 1,632 SNPs due to allele frequency differences > .05 between the Illumina and Pritzker control samples. We applied GSM and trend tests for association on the Pritzker samples alone and then on the combined Pritzker and Illumina samples. In GSM, we used the 100,000 markers that passed quality control and have the smallest p-values from the one-sided HWE test to calculate the similarity scores. Given the relatively large control:case ratio of 3,961/717≈5.5, we set the upper limit of the group sizes (*m*) to 30.

We first examined the performance of our similarity scores in the HapMap dataset. We calculated our three similarity scores for all pairs of the 89 independent Han Chinese (CHB) and Japanese (JPT) individuals in the HapMap sample, using 100,000 HapMap phase I autosomal SNPs with MAF>.05, selected based on one-sided Hardy-Weinberg equilibrium test p-values of 4.3×10^{-6} to .11. In Figure 1, we showed plots from using multidimensional scaling (MDS) on the similarity score matrices. All three scores showed good separation between the two populations, except for one JPT individual residing in between the two clusters in the plots. The same individual is at a similar position in principal component analysis when plotting the first two principal components. While *S _{IBS}* and

Our experiences in simulations and real data (unpublished results) suggest that *S _{freq}* may perform slightly better than

For the six simulation models, mismatch rates are calculated as the proportion of individuals from population 1 matched to individuals from population 2. The minimal degree of mismatch in the simulations (Table 5) suggests accurate matching given the similarity measures and numbers of markers used.

Average false positive rate and power of GSM, trend test (Chisq), and genomic control (GC) given 500 cases and 500 controls, 300,000 SNPs with MAF > .05, significance level = 10^{-6}.

In the absence of population stratification (models 1 and 2), all three methods give false positive rates close to the nominal value of 10^{-6}. The power of our GSM method is typically ~2% lower than the trend test and genomic control, assumedly due to the unnecessary grouping of samples. When population stratification is present (models 3-6), the type I error rate of the trend test is ~30 times greater than the nominal value, while GSM and genomic control maintain the type I errors at or lower than the nominal value. Using empirical type I error rates, the power of the trend test is equal to that of genomic control, but significantly lower than that of GSM for models 3-4. For models 5 and 6, where population stratification is present, the variation of disease variant frequency may mask the association (model 5) or increase the power to detect association (model 6). For model 5, power of the trend test and genomic control drop ~30% compared to model 3, while GSM maintains the same level of power. For model 6, although the type I error is inflated, the trend test has adjusted power comparable to that of GSM. EIGENSTRAT has power similar to GSM in all simulation settings examined.

We also compared the frequency with which the disease variant is the most strongly associated marker, or among the most strongly associated 10, 100, and 1000 markers, in the trend test or GSM (Figure 2). The results are consistent with the observations above. In the absence of population stratification (models 1 and 2), the trend test identifies the disease variant slightly more frequently than GSM. When population stratification is present, GSM picks the correct disease variant more frequently for models 3-5. For model 6, GSM picks the correct disease variant almost as frequently as the trend test.

We first applied standard trend tests to the Pritzker bipolar case and control samples. The estimated genomic control variance inflation factor λ of the test statistics was 1.03, close to the expected value of 1 when there is no population stratification [Devlin et al., 1999], arguing that the matching based on self-reported ethnicity resulted in a sample with only limited population stratification. Applying GSM reduced the estimated λ slightly to 1.02. However, when we added the Illumina control samples to the analysis, the estimated λ from standard trend tests became 1.51, indicative of strong population stratification between the cases and controls. We then applied our GSM method on the combined samples, excluding one Illumina control sample that had a noticeably high similarity score with one Pritzker case sample (*S _{IBS}*=0.85), consistent with a first degree relationship. Using GSM, the estimated λ dropped to 1.072 when we used

As a comparison, we also applied EIGENSTRAT and another principal component-based method (Luca et al. [2008], GEM) to the bipolar data, using 10 principal components. Without removing any potential outliers, EIGENSTRAT gave an estimated λ of 1.074, comparable to our results. GEM removed 132 samples as outliers and gave a slightly better estimated λ of 1.063. When we applied our method to the same set of samples used in GEM, we obtained an estimate λ of 1.065. Although the removal of these samples decreased the inflation of type I error rates, its impact on power requires further investigation.

Population stratification, which can result in high false-positive rates and mask true associations, poses a potential problem for case-control association studies. In this paper, we propose GSM, a practical approach to correct for population stratification for large-scale association studies that uses information at thousands of genotyped genetic markers to group case and control subjects according to their similarity. Simulation studies show that GSM can control the false positive rates in the presence of population substructure, while maintaining power to detect disease loci.

GSM is computationally efficient. The computational time for similarity score calculation is linear in the number of markers used and in the number of all case-control pairs, and the time for matching is approximately cubic in the number of individuals.

We have compared the performance of GSM to the commonly used genomic control method [Devlin et al., 1999]. Genomic control assumes that a scaled test statistic (dividing the standard test statistic by a global correction factor λ) has an approximate central χ^{2} distribution. When stratification is modest, the genomic control procedure is able to control the false-positive rate at the nominal level through λ, but does not change the relative order of the test statistics along the genome. As shown in our simulations (model 3-5), when stratification masks the association, genomic control can be quite conservative. Another popular approach to correct for population stratification is structured association [Pritchard et al., 2000] which infers population structure using a set of independent makers. We did not evaluate this method in our simulations due to its computational intensity. Structured association also requires an assumption about the number of underlying subpopulations in the sample. EIGENSTRAT [Price et al., 2006] is an approach for genome-wide association studies based on principal components analysis (PCA). It has been shown that the K-1 principal components can be related to the solution to the K-way clustering solution [Ding et al., 2004]. EIGENSTRAT is less sensitive to the number of components than structured association (if the number is sufficiently large) because of orthogonality of the axes of variation, but the interpretation of the axes is less intuitive.

Our new GSM method tackles the stratification problem by matching at the individual level, without assuming an explicit population structure. Effectively, it treats every sample as a single population and compares it to the most similar counterparts. For samples from clearly distinguished subpopulations, such as the HapMap HCB and JPT populations or the two subpopulations in our simulations, GSM performs almost as well as cluster-based matching or EIGENSTRAT, with little loss of power. In real GWA studies, where sampled individuals may often derive from continuous mixtures of ancestral populations, the individual-based matching in GSM should be more flexible than cluster-based matching. Luca et al. [2008] (GEM) also applied full matching to correct for population stratification, but used a different score calculated from the top eigenvectors from PCA. They showed that outliers may greatly inflate type I errors of association tests using EIGENSTRAT and need to be carefully removed beforehand. The similarity scores in GSM can be used like the GEM scores to identify outliers, but are more intuitive in measuring genetic similarity, compared to the abstract measures from eigenvectors used in GEM. In addition, PCA analysis is very sensitive to the independence of samples, while GSM can actually help to identify related samples through IBS scores. In our Pritzker study example, we found one pair of individuals with large similarity score of 0.85 (*S _{IBS}*), which strongly suggested a potential first-degree relative. Although the two samples showed strong correlation in their PC scores, they were not identified as outliers by EIGENSTRAT or GEM because their scores did not show strong deviation from the center of the score distributions in the top 10 PCs.

The success of our GSM procedure depends on the accuracy of matching. Incorrectly grouping individuals from different populations could inflate the type I error rate, decrease the power to detect the susceptibility genes, or both. To ensure correct matching, a well-defined similarity measure and a substantial number of markers in which to compute this measure are both important. Our simulations analysis and practical experience, show that similarity measures derived from the distribution of IBS between pairs of individuals, which are simple to calculate and do not require much computing power, provide an effective means of matching individuals. Furthermore, we found that weighting IBS estimates by a function of the marker allele frequencies (*S _{freq}*) improved the accuracy of matching. Other score metrics also exist and can be easily incorporated into our approach to substitute the IBS-based scores presented. As an experiment, we considered similarity scores based on pairwise IBD estimates calculated using an E-M algorithm, the average mismatch rates using IBD-based scores were slightly higher than those for IBS-based scores. A weakness of IBD based scores is that they are truncated at zero: when many pairs of individuals are assigned IBD ~0, it becomes difficult to select optimal pairings. Figure 3 demonstrates the relationship between the IBD scores and IBS scores (

The number of markers used in score calculation is another factor that affects the matching. We prefer to calculate the scores based on a large set of markers (typically including 10,000 – 100,000 SNPs). However, using too many markers increases the computational load while not necessarily improving the accuracy of matching. In our simulations, 10,000 markers with the smallest p-values from one-sided HWE test can correctly match the individuals from closely related populations such as Han Chinese and Japanese, with zero or almost zero mismatch (Table 5). In this example, using 30,000 markers worked as well as using 10,000 markers, while using only 1,000 markers led to incorrect grouping of individuals from different populations with up to ~10% mispaired individuals. For samples with subtle differences in genetic ancestry, such as the European American samples in the bipolar data, more markers (50,000 to 100,000, passing quality control) may help to obtain better matching. Inspecting the genomic control parameter λ on its closeness to the expected value of 1 from different analysis strategies can help to determine the appropriate number of markers for controlling stratification. To select the subset of markers, we usually prefer those with smaller p-values from one-sided HWE tests, because they tend to be more informative about population structure. However, we need to be cautious regarding data quality, since markers with high error rates may show strong deviation from HWE and then give incorrect information about the genetic background of sampled individuals. A reasonable compromise is to exclude SNPs with extreme deviations from HWE (say, *p*<10^{-6}) but focus on those with mild deviations (say, 10^{-2}<*p*<10^{-6}) to evaluate stratification. GSM does not require that *all* markers should be independent of disease status, since in a typical genome-wide setting the vast majority of markers will meet this criterion and the impact of disease-associated markers on the similarity scores is negligible and can be ignored. Furthermore, since our similarity scores are a function of the mean (weighted) IBS values across a large number of markers, it is also not critical that the assessed SNPs should be independent of each other.

We chose not to include X-linked markers in our matching scheme to avoid any possible biases due to differences by gender. Given genome-wide association data, the autosomal markers provide ample information for accurate matching.

When there is no population stratification, our simulations showed a small loss of power in GSM due to unnecessary matching. Studies have shown that when the population is indeed homogeneous, random matching by pairs (1:1) can do almost as well as the unmatched test [Chase, 1968]. Additional power may be lost when the matching is not balanced, so that multiple controls are compared to a single case subject or multiple cases are compared to a single control (i.e., 1:*m* or *m*:1 when *m* > 1). However, when stratification is present, larger values of *m* are preferred to decrease the chance of matching errors. It is then a trade-off of efficiency and bias that we need to consider in practice. In our GSM method, the objective function (*T*) we choose for optimal matching favors smaller groups, minimizing loss of efficiency. Although the original optimal matching [Rosenbaum, 1991] is unconstrained (*m* = ∞) so that all controls are allowed to be matched to a single case or all cases to a single control, Hansen [2004] showed that the matching with restriction on *m* can reduce the variance of estimated parameters with little increase in bias, and suggested a linear search for good values of *m* that are as close to 1 as possible. In our simulations, a large proportion of the matched sets are 1:1 matches even when the proportions of the two populations in cases and controls are not equal, and the average size of matched sets does not vary much for different values of the upper bound of *m*. For example, for simulated setting 3, the average matched set size is 2.44 and 2.47 when the upper limits of *m* are set as 2 and 5, respectively.

Although the full matching scheme is flexible, cases (or controls) from a population without a corresponding partner among the controls (or cases) will decrease power and may lead to spurious association if matching is forced. Further, 1:1 matching is more efficient than *m*:1 for *m*>1. Therefore, we still strongly encourage careful sample selection during the study design. Skol et al. [2005] showed that the self-reported ethnicity can be a good predictor for population structure, consistent with our results based on the NIMH case and control samples alone.

In summary, we propose a new framework to match case and control samples by their genetic similarity and adjust for the underlying population substructure. Our GSM method is specifically designed to use the full information provided by the large number of genotypes in genome-wide association studies or large-scale candidate gene studies. Our method can correctly control the false positives, while maintaining considerable power to detect the disease-marker association. Our individual-based matching scheme can reflect the continuous mixing of ancestral populations. By comparing each case to one or more controls sharing the most genetic backgrounds, we hope our method may increase the chance to identify the genetic variants that influence disease risk. Our GSM software is available freely with C++ source code at http://www.sph.umich.edu/csg/liang/gsm/. The package allows the users to automatically calculate matching score matrices, conduct full matching with a range of parameter choices, and carry out association analyses. We expect our method will aid analyses of large-scale genome-wide association studies.

We thank the Pritzker Neuropsychiatric Disorders Research Consortium, which is supported by the Pritzker Neuropsychiatric Disorders Research Fund L.L.C. We thank NIMH Human Genetics Initiative for providing the case and control samples, and the Illumina database and its contributors for providing the additional control samples. This research was supported by grants HG000376 (MB) and HG02651 and HL084729 (GRA) from the National Institutes of Health.

In a minimum cost flow (MCF) problem, we define a directed graph consisting of nodes, *i* N, and arcs connecting the nodes, (*i*, *j*) A. For each arc (*i*, *j*), an integer *a _{ij}* denotes the cost and a positive integer

$$\sum _{(i,j)\in \text{A}}{a}_{ij}{x}_{ij}$$

subject to the constraints on capacity:

$$\begin{array}{l}\text{\u2211}_{\left\{j|(i,j)\in \text{A}\right\}}{x}_{ij}-\text{\u2211}_{\left\{j|(j,i)\in \text{A}\right\}}{x}_{ji}={s}_{i},\phantom{\rule{2em}{0ex}}\text{forall}i\in \text{N}\\ 0\le {x}_{ij}\le {c}_{ij},\phantom{\rule{2em}{0ex}}\text{forall}(i,j)\in \text{A}\end{array}$$

It is easy to see the equivalence between the MCF and the optimal matching (Figure 4). The nodes in a directed graph correspond to the cases and controls, *a _{ij}* is the dissimilarity measure between

In full matching, the numbers of case-control pairs vary across matched sets, so the supply of nodes (*s _{i}*) cannot be predetermined. To deal with this complication, we include an “overflow” node to the graph to balance the flows from or to the case or control nodes. Parameters U and U

The translation is demonstrated in Figure 4. The MCF problem is then solved by iteratively updating a dual cost vector and the flow vector ** x** (Bertsekas et al., 1994; Frangioni et al., 2006).

- Bertsekas DP, Tseng P. RELAX-IV: A faster version of the RELAX code for solving minimum cost flow problems Technical report LIDS-P-2276. Department of Electrical Engineering and Computer Science, Massachusetts Institute of Technology; Cambridge, MA: 1994.
- Cheung VG, Spielman RS, Ewens KG, Weber TM, Morley M, Burdick JT. Mapping determinants of human gene expression by regional and genome-wide association. Nature. 2005;437:1365–1369. [PMC free article] [PubMed]
- Ding C, He X. K-means clustering via principal component analysis. Proc Intl Conf Machine Learning (ICML 2004); 2004. pp. 225–232.
- Donnelly P, Tavaré S. Coalescents and genealogical structure under neutrality. Annu Rev Genet. 1995;29:401–421. [PubMed]
- Devlin B, Roeder K. Genomic control for association studies. Biometrics. 1999;55:997–1004. [PubMed]
- Epstein MP, Allen AS, Satten GA. A simple and improved correction for population stratification in case-control studies. Am J Hum Genet. 2007;80:921–930. [PubMed]
- Frangioni A, Manca A. A computational study of cost reoptimization for min cost flow problem. INFORMS J Comput. 2006;18:61–70.
- Freedman ML, Reich D, Penney KL, McDonald GJ, Mignault AA, Patterson N, Gabriel SB, Topol EJ, Smoller JW, Pato CN, Pato MT, Petryshen TL, Kolonel LN, Lander ES, Sklar P, Henderson B, Hirschhorn JN, Altshuler D. Assessing the impact of population stratification on genetic association studies. Nat Genet. 2004;36:388–393. [PubMed]
- Gorroochurn P, Heiman GA, Hodge SE, Greenberg DA. Centralizing the non-central chi-square: a new method to correct for population stratification in genetic case-control association studies. Genet Epidemiol. 2006;30:277–289. [PubMed]
- Gu XS, Rosenbaum PR. Comparison of multivariate matching methods: structures, distances, and algorithms. J Comput Graph Stat. 1993;2:405–420.
- Guan W, Liang L, Boehnke M, Abecasis GR. Matching cases and controls using genotype data from a whole genome association study. ASHG 2005 Annual Meeting #2395 (poster).2003.
- Hansen BB. Full matching in an observational study of coaching for the SAT. J Am Stat Assoc. 2004;99:609–618.
- Hansen BB, Klopfer SO. Optimal full matching and related designs via network flows. J Comput Graph Stat. 2006;15:609–627.
- Hudson RR. Properties of a neutral allele model with intragenic recombination. Theor Popul Biol. 1983;23:183–201. [PubMed]
- Hudson RR. Gene genealogies and the coalescent process. Oxford Surv Evol Biol. 1990;7:1–44.
- Khlat M, Cazes MH, Genin E, Guiguet M. Robustness of case-control studies of genetic factors to population stratification: magnitude of bias and type I error. Cancer Epidemiol Biomarkers Prev. 2004;13:1660–1664. [PubMed]
- Kimmel G, Jordan MI, Halperin E, Shamir R, Karp RM. A Randomization Test for Controlling Population Stratification in Whole-Genome Association Studies. Am J Hum Genet. 2007;81:895–905. [PubMed]
- Kimura M. The number of heterozygous nucleotide sites maintained in a finite population due to steady flux of mutations. Genetics. 1969;61:893–903. [PubMed]
- Klein RJ, Zeiss C, Chew EY, Tsai JY, Sackler RS, Haynes C, Henning AK, Sangiovanni JP, Mane SM, Mayne ST, Bracken MB, Ferris FL, Ott J, Barnstable C, Hoh J. Complement factor H polymorphism in age-related macular degeneration. Science. 2005;308:385–389. [PMC free article] [PubMed]
- Li CC. Population subdivision with respect to multiple alleles. Ann Hum Genet. 1972;33:23–29. [PubMed]
- Liang L, Zöllner S, Abecasis GR. GENOME: a rapid coalescent-based whole genome simulator. Bioinfomatics. 2007;23:1565–1567. [PubMed]
- Luca D, Ringquist S, Klei L, Lee AB, Gieger C, Wichmann HE, Schreiber S, Krawczak M, Lu Y, Styche A, Devlin B, Roeder K, Trucco M. On the use of general control samples for genome-wide association studies: genetic matching highlights causal variants. Am J Hum Genet. 2008;82:453–463. [PubMed]
- Maraganore DM, de Andrade M, Lesnick TG, Strain KJ, Farrer MJ, Rocca WA, Pant PVK, Frazer KA, Cox DR, Ballinger DG. High-resolution whole-genome association study of Parkinson disease. Am J Hum Genet. 2005;77:685–693. [PubMed]
- Price AL, Patterson NJ, Plenge RM, Weinblatt ME, Shadick NA, Reich D. Principal components analysis corrects for stratification in genome-wide association studies. Nat Genet. 2006;38:904–909. [PubMed]
- Pritchard JK, Stephens M, Rosenberg NA, Donnelly P. Association mapping in structured populations. Am J Hum Genet. 2000;67:170–181. [PubMed]
- Rosenbaum PR. A characterization of optimal designs for observational studies. J Roy Statist Soc Ser B. 1991;53:597–610.
- Rosenbaum PR. Observational Studies. 2nd. New York: Springer; 2002.
- Saxena R, Voight BF, Lyssenko V, Burtt NP, de Bakker PI, Chen H, Roix JJ, Kathiresan S, Hirschhorn JN, Daly MJ, Hughes TE, Groop L, Altshuler D, Almgren P, Florez JC, Meyer J, Ardlie K, Bengtsson K, Isomaa B, Lettre G, Lindblad U, Lyon HN, Melander O, Newton-Cheh C, Nilsson P, Orho-Melander M, Rastam L, Speliotes EK, Taskinen MR, Tuomi T, Guiducci C, Berglund A, Carlson J, Gianniny L, Hackett R, Hall L, Holmkvist J, Laurila E, Sjogren M, Sterner M, Surti A, Svensson M, Svensson M, Tewhey R, Blumenstiel B, Parkin M, Defelice M, Barry R, Brodeur W, Camarata J, Chia N, Fava M, Gibbons J, Handsaker B, Healy C, Nguyen K, Gates C, Sougnez C, Gage D, Nizzari M, Gabriel SB, Chirn GW, Ma Q, Parikh H, Richardson D, Ricke D, Purcell S. Genome-wide association analysis identifies loci for type 2 diabetes and triglyceride levels. Science. 2007;316:1331–1336. [PubMed]
- Scott LJ, Mohlke KL, Bonnycastle LL, Willer CJ, Li Y, Duren WL, Erdos MR, Stringham HM, Chines PS, Jackson AU, Prokunina-Olsson L, Ding CJ, Swift AJ, Narisu N, Hu T, Pruim R, Xiao R, Li XY, Conneely KN, Riebow NL, Sprau AG, Tong M, White PP, Hetrick KN, Barnhart MW, Bark CW, Goldstein JL, Watkins L, Xiang F, Saramies J, Buchanan TA, Watanabe RM, Valle TT, Kinnunen L, Abecasis GR, Pugh EW, Doheny KF, Bergman RN, Tuomilehto J, Collins FS, Boehnke M. A genome-wide association study of type 2 diabetes in Finns detects multiple susceptibility variants. Science. 2007;316:1341–1345. [PMC free article] [PubMed]
- Skol AD, Xiao R, Boehnke M, Veterans Affairs Cooperative Study 366 Investigators An algorithm to construct genetically similar subsets of families with the use of self-reported ethnicity information. Am J Hum Genet. 2005;77:346–354. [PubMed]
- Sladek R, Rocheleau G, Rung J, Dina C, Shen L, Serre D, Boutin P, Vincent D, Belisle A, Hadjadj S, Balkau B, Heude B, Charpentier G, Hudson TJ, Montpetit A, Pshezhetsky AV, Prentki M, Posner BI, Balding DJ, Meyre D, Polychronakos C, Froguel P. A genome-wide association study identifies novel risk loci for type 2 diabetes. Nature. 2007;445:881–885. [PubMed]
- Spielman RS, McGinnis RE, Ewens WJ. Transmission test for linkage disequilibrium: the insulin gene region and insulin-dependent diabetes mellitus (IDDM) Am J Hum Genet. 1993;52:506–513. [PubMed]
- The International HapMap Consortium. A second generation human haplotype map of over 3.1 million SNPs. Nature. 2007;449:851–861. [PMC free article] [PubMed]
- Wigginton JE, Cutler DJ, Abecasis GR. A note on exact tests of Hardy-Weinberg equilibrium. Am J Hum Genet. 2005;76:887–893. [PubMed]
- Zeggini E, Weedon MN, Lindgren CM, Frayling TM, Elliott KS, Lango H, Timpson NJ, Perry JR, Rayner NW, Freathy RM, Barrett JC, Shields B, Morris AP, Ellard S, Groves CJ, Harries LW, Marchini JL, Owen KR, Knight B, Cardon LR, Walker M, Hitman GA, Morris AD, Doney AS, McCarthy MI, Hattersley AT. Replication of genome-wide association signals in U.K. samples reveals risk loci for type 2 diabetes. Science. 2007;316:1336–1341. [PMC free article] [PubMed]
- Zhang S, Kidd KK, Zhao H. Detecting genetic association in case-control studies using similarity-based association tests. Statistica Sinica. 2002;12:337–359.

PubMed Central Canada is a service of the Canadian Institutes of Health Research (CIHR) working in partnership with the National Research Council's national science library in cooperation with the National Center for Biotechnology Information at the U.S. National Library of Medicine(NCBI/NLM). It includes content provided to the PubMed Central International archive by participating publishers. |