University at Buffalo - The State University of New York
Skip to Content
Genotype Imputation | Annual Review of Genomics and Human Genetics
Access Provided by University at Buffalo Libraries
Access Provided by University at Buffalo Libraries

Genotype Imputation

Annual Review of Genomics and Human Genetics

Vol. 10:387-406 (Volume publication date September 2009)


Genotype imputation is now an essential tool in the analysis of genome-wide association scans. This technique allows geneticists to accurately evaluate the evidence for association at genetic markers that are not directly genotyped. Genotype imputation is particularly useful for combining results across studies that rely on different genotyping platforms but also increases the power of individual scans. Here, we review the history and theoretical underpinnings of the technique. To illustrate performance of the approach, we summarize results from several gene mapping studies. Finally, we preview the role of genotype imputation in an era when whole genome resequencing is becoming increasingly common.


Identifying and characterizing the genetic variants that affect human traits, ranging from disease susceptibility to variability in personality measures, is one of the central objectives of human genetics. Ultimately, this aim will be achieved by examining the relationship between interesting traits and the whole genome sequences of many individuals. Although whole genome resequencing of thousands of individuals is not yet feasible, geneticists have long recognized that much progress can be made by measuring only a relatively modest number of genetic variants in each individual. This type of “incomplete” information is useful because data about any set of genetic variants in a group of individuals provide useful information about many other unobserved genetic variants in the same individuals.

The idea that data on a modest set of genetic variants measured in a number of related individuals can provide useful information about other genetic variants in those individuals forms the theoretical underpinning of both genetic linkage mapping in pedigrees and haplotype mapping in founder populations (23, 24, 50). These studies typically use <10,000 genetic markers to survey the entire human genome by identifying stretches of chromosome inherited from a common ancestor. The shared stretches usually span several megabases and include thousands of genetic variants. Both approaches have been highly successful in identifying genes responsible for single-gene Mendelian disorders (9). In contrast, they have had only limited success in mapping genes that influence complex traits, although success stories do exist (40, 42, 75, 83).

More recently, technological advances have made genome-wide association studies (GWAS) possible (39, 67, 109). Rather than <10,000 variants, these studies typically genotype 100,000–1,000,000 variants in each of the individuals being studied. Since >10 million common genetic variants are likely to exist (104), even these detailed studies examine only a fraction of all genetic variants. Whereas in traditional genetic linkage and founder haplotype mapping studies geneticists expect to identify long stretches of shared chromosome inherited from a relatively recent common ancestor, in GWAS focusing on apparently unrelated individuals geneticists expect to identify only relatively short stretches of shared chromosomes. Remarkably, genotype imputation can use these short stretches of shared haplotype to estimate with great precision the effects of many variants that are not directly genotyped.

In this review, we first attempt to provide the reader with an intuition for how genotype imputation approaches work and for their theoretical underpinnings. We start with the relatively intuitive setting of imputing missing genotypes for a set of individuals using information on their close relatives. We then examine how genotype imputation works when applied to more distantly related individuals. Next, we survey results of studies that have used genotype imputation to study complex disease susceptibility. We attempt to provide the reader with critical information to assess the merits of genotype imputation-based analyses and to provide guidance to analysts attempting to implement these approaches. Finally, we survey potential uses of imputation-based analyses in the context of whole genome resequencing studies that we believe will soon become commonplace.


Family samples constitute the most intuitive setting for genotype imputation. Genotypes for a relatively modest number of genetic markers can be used to identify long stretches of haplotype shared between individuals of known relationship. These stretches of shared haplotype (or regions “identical-by-descent”, IBD) are typically used to evaluate the evidence for linkage. Specifically, genetic linkage implies that family members who share a region IBD will be more similar to each other than will family members with the same degree of relatedness who do not share the region IBD. In the context of genotype imputation, we characterize each of these stretches in detail by genotyping additional markers in one or more individuals in the family. Genotypes for these markers can then be propagated to other family members who are only typed at a minimal set of markers.

The approach is illustrated in Figure 1. In the figure, all individuals have been genotyped for a set of genetic markers indicated in red; a subset of individuals in the top two generations has been genotyped at additional markers indicated in black (panel a). Genotypes for the red markers, available in all individuals, can be used to infer the segregation of haplotypes through the family (panel b). Finally, most of the missing genotypes for individuals in the bottom generation can be inferred by comparing the haplotypes they inherited with copies of the same haplotypes that are IBD and present in other individuals in the family (panel c).

The idea that family members share long stretches of haplotype that are IBD underpins nearly all methods of linkage analysis. Furthermore, many early approaches for association analysis in pedigree data implicitly impute missing genotypes by considering the distribution of potential genotypes of each individual jointly with that of other individuals in the same pedigree (35, 45). The extension of this idea to the imputation of missing genotypes (as outlined above) was first described by Burdick and colleagues (12), who coined the term in silico genotyping to describe the idea that computational analyses could be used to replace laboratory-based procedures in the determination of individual genotypes. To illustrate the potential of the approach, they reanalyzed the data of Cheung and colleagues (17). Cheung and colleagues sought to identify genetic variants associated with regulation of gene expression by examining RNA transcript levels and genotype data for individuals in the top two generations of the Centre d'Etude du Polymorphism Humain (CEPH) pedigrees (21). The CEPH pedigrees are three-generation pedigrees with a structure similar to that of the cartoon pedigree in Figure 1.The top two generations of several of these pedigrees were genotyped at more than 830,000 genetic markers in the first phase of the International HapMap Project (103). Using genotypes for approximately 6500 genetic markers genotyped by the SNP consortium in all three generations of the pedigrees (85), Burdick and colleagues imputed genotypes for most of the HapMap Project markers in the third generation of these pedigrees (12). They showed that this imputation-based analysis was more powerful than the original analysis, which examined only directly genotyped markers for each individual.

Several formal statistical descriptions of genotype imputation procedures for association analyses in families have now been published (15, 108), and the procedures to support genotype imputation are implemented in packages such as MERLIN (2, 3) and MENDEL (52, 53). In principle, these procedures can be implemented using the infrastructure of the Lander-Green (48) or Elston-Stewart (29) algorithms, or one of the many other pedigree analysis algorithms, including those that are based on Monte Carlo sampling (38, 96). An important observation from these more formal treatments of the problem is that even when genotypes cannot be imputed with high confidence, partial information about the identity of each of the true underlying genotypes can be productively incorporated in association analysis (15, 108). For example, when genotypes are measured directly, observed allele counts are often used in regression analyses to estimate an additive effect for each marker (1, 8, 34). These observed allele counts indicate the number of copies of the allele of interest (0, 1, or 2) carried by each individual. When genotypes are not measured directly, these discrete counts can be replaced with an expected allele count for each marker (a real number between 0 and 2) (15).

The approach has been successfully used to study several quantitative traits in a sample of closely related individuals from four villages in Sardinia (77). Among study participants, 1412 individuals were genotyped with the Affymetrix 500K mapping array set (which assays ∼500,000 SNPs) and a further 3329 individuals were genotyped with Affymetrix 10K SNP mapping arrays (which assay ∼10,000 SNPs) (94). The data were then used to study the genetic architecture of a variety of quantitative traits, ranging from body mass index (94) to fetal hemoglobin levels (106) to personality traits (101). Clearly, family-based genotype imputation will be maximally useful in samples that include very large numbers of related individuals. In these settings, genotypes for a relatively modest number of individuals can be propagated to many other additional individuals, thereby increasing power. However imputing genotypes for known relatives of the individuals included in a GWAS of mostly unrelated individuals will always increase power (15) and should be considered whenever phenotyped relatives for the individuals to be genotyped in a scan are available. This will often be the case when individuals to be scanned are selected from a larger sample of related individuals that was previously phenotyped for linkage analyses or family-based association testing.


Analyses of related individuals provide the intuition behind genotype imputation: Whenever a particular stretch of chromosome is examined in detail in at least one individual, we learn about the genotypes of many other individuals who inherit that same stretch IBD. When studying samples of apparently unrelated individuals, the same approach can be utilized. The major difference is that, when studying apparently unrelated individuals, shared haplotype stretches will be much shorter (because common ancestors are more distant) and thus may be harder to identify with confidence. The intuition that short stretches of haplotype provide useful information about untyped genetic markers provides the justification for the power gains suggested for many haplotype analysis strategies (22, 60, 91, 115).

The mechanics of genotype imputation in unrelated individuals are illustrated in Figure 2. Here, study samples genotyped for a relatively large number of genetic markers (perhaps 100,000–1,000,000) are compared to a reference panel of haplotypes that includes detailed information on a much larger number of markers (Figure 2a). To date, the HapMap Consortium database has typically served as the reference panel (104), but we expect that in the future larger sets of individuals characterized at larger numbers of markers will be available. Stretches of shared haplotype are then identified (Figure 2b) and missing genotypes for each study sample can be filled in by copying alleles observed in matching reference haplotypes (Figure 2c). In analyses of samples of European ancestry, comparisons with genotypes for the HapMap CEU panel typically yield shared haplotypes that range from about 100–200 kb in length. Thus, in a GWAS that examines 300,000 SNP markers, these shared stretches generally include 10–20 genotyped markers. When there is ambiguity about which haplotype stretch should be “copied” to fill in missing genotypes for a particular individual, most imputation programs provide an answer that summarizes this ambiguity probabilistically (for example, in 60% of reconstructions genotype A/A was observed at a specific site, whereas in the remaining 40% a different genotype A/C was observed).

In principle, any of the methods typically used to estimate missing haplotypes—whether based on a simple heuristic (18), on an E-M algorithm (30), or on more sophisticated coalescent models (99) could be used to impute missing genotypes. In fact, most haplotyping programs will automatically “impute” missing genotypes during the haplotype estimation process. In practice, most researchers now use one of the tools that have been specifically enhanced to facilitate genotype imputation-based analyses. These tools generally provide convenient summaries of the uncertainty surrounding each genotype estimate or, perhaps, convenient built-in association testing. Genotype imputation tools fall into two main categories: (a) computationally intensive tools such as IMPUTE (64), MACH (59), and fastPHASE/BIMBAM (92, 95) that take into account all observed genotypes when imputing each missing genotype, and (b) computationally more efficient tools such as PLINK (80), TUNA (71), WHAP (114), and BEAGLE (11) that usually focus on genotypes for a small number of nearby markers when imputing each missing genotype. Tools in the first category can be further subdivided into those that compare the potential haplotypes for each individual with all other observed haplotypes (e.g., IMPUTE and MACH) and those that compare potential haplotypes for each individual to a representative set of haplotypes (e.g., fastPHASE). Tools that consider all available markers and all available haplotypes can require substantially more intensive computation but generally do better at estimating missing genotypes, particularly for rare polymorphisms. Table 1 provides a partial list of recent genome-wide association scans that used genotype imputation, together with the method(s) used for imputing missing genotypes in each scan.

Table 1

Examples of GWAS that have used genotype imputation


Our first experience with genotype imputation in the context of a genetic association study occurred when fine-mapping the Complement Factor H susceptibility locus for age-related macular degeneration (58). This locus shows evidence for multiple disease-associated alleles and haplotypes (58, 63). Since multimarker association analyses are much more convenient in the absence of missing genotype data (5), we used the program PHASE (97, 98) and an early version of our MACH software (59) to fill in missing genotypes in our sample. In the absence of missing data, it is much easier to compare the evidence for association at different markers and to interpret the results of conditional association analyses that seek to identify independently associated markers. To validate our imputation approach, we masked 5% of the genotypes at the locus and showed that these could be imputed correctly >99% of the time by comparing each individual with a missing genotype to other individuals who shared a common haplotype or haplotypes.

The first few applications of genotype imputation on a genome-wide scale also involved considerable effort to validate the accuracy of imputed genotypes. For example, in the first published account of the performance of genotype imputation in the context of a genome-wide scan, Scott et al. (93) genotyped a set of type 2 diabetes cases and controls at approximately 300,000 SNPs. They then imputed genotypes at an additional >2 million SNPs to facilitate comparisons with the results of two other genome-wide association scans for type 2 diabetes that relied on a different genotyping platforms (90, 117). To evaluate the accuracy of imputed genotypes, they contrasted imputed genotypes generated in silico with experimental genotypes generated in the lab for >500 SNPs, including 16 SNPs with imputation-based p-values of <10−5 (see online supplementary material in Reference 93). Their results showed excellent concordance between genotype calls, estimated allele frequencies, and test statistics for both types of data, with an overall allelic discrepancy rate of <1.50% between genotyped and imputed SNPs.

Similar comparisons with newer genotyping platforms, which can provide better coverage of the genome because they include larger numbers of tag SNPs, show that imputed genotypes can achieve even greater accuracy. For example, in the GAIN psoriasis study (69), imputed and experimentally derived genotypes were compared at >660,000 SNPs in 90 individuals with an overall allelic discrepancy rate of <0.90% and an r2 correlation between observed and imputed allele counts that averaged 0.93. The r2 correlation coefficient is a particular useful summary of the effect of genotype imputation on power: In the context of the GAIN psoriasis study we expect that, on average, imputing genotypes for one of the 660,000 evaluated markers in 1000 individuals would provide similar information to that obtained by genotyping the same marker in 930 individuals (69).


One obvious use of genotype imputation-based analysis is to accelerate fine-mapping studies. Once an association signal has been identified and confirmed, genotype imputation can be used to evaluate the evidence for association at each of several nearby SNPs and help focus the search for potential causal variants. An example of this approach occurs in the fine-mapping study of Orho-Melander et al. (76). To fine-map an association signal linking SNPs in the glucokinase regulatory protein (GCKR) gene and triglyceride levels in blood, Orho-Melander et al. examined evidence for association with genotyped and imputed SNPs in the region and showed that an imputed common missense variant in the GCKR gene was more strongly associated with triglyceride levels than any other nearby SNP, a result that was subsequently confirmed by direct genotyping (76).

Although we agree that examining evidence for association at imputed markers can be extremely useful in the context of fine-mapping association signals, genotype imputation is also expected to increase the power of GWAS. For example, Willer et al. (111) and Kathiresan et al. (43) showed that rs6511720, a common variant in the low-density lipoprotein receptor gene (LDLR), was strongly associated with blood low density lipoprotein (LDL) cholesterol levels (Figure 3). The association signal was missed in an initial analysis that considered only genotyped SNPs because rs6511720 is not included in the Affymetrix arrays used to scan the genome in the majority of their samples and is only poorly tagged by individual SNPs on the chip (the best single marker tag is rs12052058 with pairwise r2 of only 0.21). Another example we have encountered concerns the genome-wide association analysis of G6PD activity levels in a sample of Sardinian individuals (77, 94). There, analysis of directly genotyped SNPs revealed two sets of SNPs strongly associated (p < 5 × 10−8) with G6PD activity levels, one near the G6PD gene locus on chromosome X and another near the HBB locus on chromosome 11. Genotype imputation revealed a strong additional signal (also with p < 5×10−8) upstream of the 6PGD locus on chromosome 1 (M. Uda, S. Sanna & D. Schlessinger, personal communication) (Figure 4). The three signals (near G6PD, HBB, and 6PGD) all fit our understanding of the biological basis of measurements of G6PD activity: (a) The role of variants near G6PD in the regulation of G6PD activity in Sardinia and elsewhere is well established (25), (b) variants in the HBB locus can influence the life span and rate of turnover of red blood cells and because G6PD activity is higher in younger red blood cells the same polymorphisms indirectly influence G6PD activity levels (70) and (c) 6PGD activity levels are commonly used as an internal standard in assays for G6PD activity so that changes in 6PGD activity will influence estimates of G6PD activity (13, 31, 68a).

Overall, the LDLR and 6PGD loci, together with many other anecdotal examples, suggest that genotype imputation can improve the power of genome-wide association analyses. Nevertheless, accurately estimating the impact of genotype imputation on the power of a GWAS is more challenging. We have tried to accurately quantify this potential power gain in two ways: first, by generating and analyzing simulated datasets; and second, by analyzing datasets that combine genome-wide genotype data and large-scale surveys of gene expression. The second approach is especially attractive because true positive associations between genetic variants and transcript levels are easy to identify (they often map to the locus encoding the transcript). Both approaches suggest that genotype imputation can increase the power of gene-mapping studies, particularly when the associated variants have frequencies of <10%–20%. When we imputed genotypes and then reanalyzed the gene expression data of Dixon et al. (28), we mapped, on average, 10% more genome-wide association peaks to the locus surrounding each transcript than before imputation (L. Liang, W.O.C. Cookson & G.R. Abecasis, unpublished data).


Perhaps the most dramatic illustration of the utility of genotype imputation has been the ability of researchers to conduct meta-analyses of genome-wide association scans even in samples that were originally genotyped using several different platforms. Genotype imputation was first used to combine GWAS for blood lipid levels (43, 111) and height (89) and soon thereafter to combine data across genome-wide scans for type 2 diabetes (116), body-mass index (62), and Crohn's disease (6). The success of these meta-analyses can be dramatic: in the case of blood lipid levels (43, 111) a meta-analysis of three studies with relatively modest findings (each identifying one to three strongly associated loci) resulted in a total of 19 strongly associated loci including 7 loci not previously implicated in regulating cholesterol and lipoprotein levels in humans. Because it greatly simplifies issues related to examining data collected on multiple different platforms, genotype imputation also makes it simple for researchers to compare results of GWAS that target related traits. In this way, it has been possible to contrast results from genetic studies of blood lipid levels (111) with those of previous studies of coronary artery disease (105), to compare results of studies of blood glucose levels in nondiabetic individuals (79) with those of previous case-control studies of type 2 diabetes (116), and to compare results of studies of height (89) to those of previous studies of osteoarthritis (68). We expect that these sorts of contrasts between the results of genome-wide studies for different traits will become ever more commonplace and that they will ultimately provide useful insights about the genetic basis of many complex human traits.


While most GWAS completed to date have focused on populations of European ancestry (see Table 1 for examples), we expect that many GWAS will be conducted in much more diverse groups of samples. The success of genotype imputation depends critically on the choice of reference population from which densely characterized haplotypes are drawn. For studies of European ancestry samples, it is now clear that the HapMap CEU samples (102104) usually constitute an appropriate reference panel. Similarly, we expect the HapMap CHB + JPT (102104) samples will constitute a good reference for imputing genotypes in samples of East Asian ancestry and that the HapMap YRI (102104) samples will constitute a good reference for imputing genotypes in populations of West African ancestry.

Studies of populations that are genetically more distinct from those examined by the HapMap consortium will require more careful consideration in the design of strategies for genotype imputation. For example, we expect that when imputing missing genotypes in Middle Eastern samples, Native American samples, or even samples from the Indian subcontinent, it will be advantageous to use a reference panel that includes all HapMap haplotypes, rather than just the CEU, just the YRI, or just the CHB + JPT haplotypes (59). Fortunately, whenever the choice of reference panel is unclear, it is possible to mask a subset of the available genotype data, run genotype imputation using each of the different reference panels being considered, and then contrast imputed and masked genotypes to identify the strategy that provides the most accurate genotypes. Table 2 summarizes the results of a recent analysis (59) that sought to identify the most appropriate reference panel for a series of samples in the Human Genome Diversity Panel (19).

Table 2

Recommended choices of HapMap reference panel haplotypes for imputing genotypes in Human Genome Diversity Panel different samples

An alternative to using the HapMap samples as a reference is to genotype a subset of study samples for additional markers of interest and then use these as templates for genotype imputation in the remaining samples. This approach was used by Chambers et al. (14) to combine data across three different platforms in a recent study of the genetics of obesity focused on individuals of South Asian ancestry. Compared to approaches that use the HapMap as a reference, this strategy can greatly reduce imputation error (14).


In this review, we have tried to provide readers with an intuition about why genotyping imputation methods work, described their history in the context of GWAS, and summarized some examples of current uses of genotype imputation. For readers who are encouraged to attempt genotype imputation in their own samples, we now summarize important practical issues to consider when carrying out genotype imputation-based analyses. In particular, we focus on issues encountered when developing, implementing, and supporting our Markov Chain Haplotyping (MACH) software package for haplotype estimation and genotype imputation. As with other analyses of genetic association data, we recommend that a standard set of quality filters be used to exclude markers with poor-quality genotypes prior to imputation. These quality filters typically flag markers that have low call rates, significant evidence for deviations from Hardy-Weinberg equilibrium, a large rate of discrepancies between duplicate genotypes, or evidence for non-Mendelian inheritance (67).

When using an external reference panel as a template for imputation, the most important challenge for successfully imputing genotypes in genome scan samples is to ensure that alleles are labeled consistently (that is, on the same strand) in the reference panel and in the samples being imputed. MACH checks that allele frequencies are similar in the reference panel and in the samples, but it cannot catch all errors. In practice, we have found it extremely useful to genotype a small number of HapMap samples as part of each genome-wide scan—this helps both to evaluate genotyping error rates and to ensure that consistency of allele labels can be easily checked.

Once this first hurdle has been surpassed, the next step is to impute missing genotypes for each sample. As noted above and in Table 2, a key step is to select an appropriate set of reference haplotypes. Different choices of reference panel can be assessed by masking a subset of the available genotypes and checking whether these can be recovered accurately. After a reference panel has been selected and imputation is complete, a key issue is to decide which markers to take forward for analysis. Typically, not all markers can be accurately imputed, and several different measures have been proposed to help identify well-imputed markers. The simplest of these measures focuses on the average probability that an imputed genotype call is correct—in this context for example, one might look for markers where genotypes are imputed with >90% accuracy. We do not recommend this type of measure because it is not very meaningful when comparing markers with different allele frequencies (for example, if a marker has an allele frequency of <5%, it should be possible to achieve 90% accuracy by simply assigning the most common genotype to every individual). Instead, we typically recommend measures that try to capture the correlation between imputed genotype calls and the true underlying genotypes—usually expressed as an r2 coefficient. Generally these measures are calculated by comparing the variance in a set of imputed allele counts to theoretical expectations based on Hardy-Weinberg equilibrium (because imputed allele counts for poorly imputed markers show less variability than expected based on allele frequency).

The final step is to analyze the resulting imputed “genotypes.” MACH and other genotype imputation programs represent imputation results in a variety of ways. Usually, imputed genotypes are probabilistic rather than discrete. For example, a particular individual might have a 90% probability of carrying genotype A/A and a 10% probability of carrying genotype A/C at a specific marker—corresponding to 1.9 expected copies of allele A. We do not recommend transforming these “probabilistic” genotype calls into discrete genotypes as that can result in a substantial loss of information, especially for less common alleles. Generally, imputed allele counts for each allele (e.g., 1.9 expected copies of allele A) can conveniently be tested for association with quantitative or discrete traits using an appropriate regression model. Of course, as in other genetic association analyses, adequate adjustment for potential population stratification is essential (27, 36, 78). If ancestry-informative principal components are to be estimated from genetic data (78), this should be done before imputation.

If results from multiple studies are to be combined, each study should be analyzed individually and results should then be meta-analyzed across studies using standard approaches (see 89, 111, 112, 116 for examples). We do not recommend pooling data across studies as an alternative to meta-analysis, especially when these have been genotyped using different platforms.


The technologies used in human genetic studies are rapidly improving. We expect several enhancements to genetic imputation technologies. First, as better characterized reference panels are developed, it should become possible to use genotype imputation methods to study not only single-nucleotide polymorphisms but also other types of genetic variants, such as copy-number variants (33, 66) or classical HLA types (55). Second, improved algorithms for genotype imputation will likely be developed to tackle ever more complex problems. Similar pressures previously motivated continued development of methods for pedigree analysis, both for large pedigrees (29, 51, 54, 73) and for smaller ones (2, 37, 4648, 65). Still, the most useful advance, in the context of genotype imputation-based analyses, would be the development of larger reference panels. As illustrated in Figure 5, the accuracy of genotype imputation-based analyses should increase substantially as the size of reference panels increases. This increase in accuracy occurs because haplotype stretches shared between study samples and samples in the reference panel increase in length and are easier to identify unambiguously with a larger reference panel.


So far, we have focused our discussion on the analysis of genotype data. However, genome sequencing technologies are also improving extremely rapidly. Whereas the first two human whole genome assemblies took years to complete (49, 107), several additional genomes have been assembled just in the past 18 months (7, 57, 110). These advances in whole genome sequencing have resulted from the deployment of massive throughput sequencing technologies, which differ from standard Sanger-based sequencing (88) in many ways. For example, the data produced by these new technologies typically have somewhat higher error rates (on the order of 1% per base). Since these technologies produce very large amounts of data, one typically accommodates these error rates by resequencing every site of interest many times to achieve a high-quality consensus.

We expect that the continued deployment of these technologies will significantly change how genotype imputation is used. An example is given by the 1000 Genomes Project (see, which aims to deliver whole genome sequences for >1000 individuals from several different populations in the next 12 months. To do this in a cost-effective manner, the project is using a strategy that combines massively parallel shotgun sequencing technology with the same statistical machinery used to drive genotype imputation-based analyses. Specifically, a relatively modest amount of shotgun sequence data is being collected for each individual: Each of the target bases will be resequenced only 2–4x on average (statistical fluctuations around this average mean that many bases will not be covered even once), rather than the 20–40x used in conventional applications of these technologies to whole genome resequencing. To accurately call polymorphisms in each genome, the Project will then use imputation-based techniques to combine information across individuals who share a particular haplotype stretch. Using simulations, we have predicted that when 400 diploid individuals are sequenced at only 2x depth (1x per haploid genome) and the data are analyzed using approaches that combine data across individuals sharing similar haplotype stretches, polymorphic sites with a frequency of >2% can be genotyped with >99.5% accuracy (Y. Li & G. Abecasis, unpublished data). Note that the same 2x average depth would not be useful for genotype calling when examining a single individual—since, by chance, ∼37% of alleles would not be sampled. For another example of how genotype imputation can be combined with sequence data, see Reference 72.

The ability to combine relatively modest amounts of sequence data across many individuals to generate high-quality sequence data for all may become one of the most common uses of imputation technologies in the near future. For a given sequencing effort, genotype imputation-based analyses may allow an increase in the number of individuals to be sequenced by five- to tenfold with minimal loss of accuracy in individual genotypes. Such an increase in sample size is critical when attempting to map the genes for complex diseases. Of course, even before massively parallel sequencing technologies are deployed more widely, one immediate change will occur with the completion of the 1000 Genome Project (see Specifically, we expect these data will provide accurate genotype information on >10 million common variants and quickly replace the HapMap Consortium genotypes as the reference panel of choice for imputation studies. Thus imputation-based analyses will be able to examine even more genetic markers, and each of these markers will, on average, be imputed much more accurately.


In the past two years, genotype imputation-based analyses have become a key tool for the analysis of human genetic data. They have been used to aid fine-mapping studies, to increase the power of genome wide association studies, to extract maximum value from existing family samples, and to facilitate meta-analysis of genome-wide association data. In the next few years, we expect these imputation-based analyses will become a key tool in the analysis of massively parallel shotgun sequence data, enabling geneticists to rapidly deploy these technologies to analyze large samples and dissect the genetic basis of complex disease.

disclosure statement

The authors are not aware of any affiliations, memberships, funding, or financial holdings that might be perceived as affecting the objectivity of this review.


We thank S. Kathiresan, K. Mohlke, D. Schlessinger, and M. Uda for the example relating common variants near LDLR and LDL-cholesterol levels. We thank D. Schlessinger and M. Uda for the example relating variants near 6PGD to G6PD activity levels. Finally, we thank M. Boehnke, K. Mohlke, and FUSION colleagues for the data used to generate Figure 5. Yun Li was supported in part by a Rackham Merit fellowship. Cristen Willer was supported in part by an American Diabetes Association Fellowship. Gonçalo Abecasis is a Pew Scholar for the Biomedical Sciences. This research was supported in part by research grants HG-2651, HL-84729, and MH-84698.

literature cited