Why Panel Size and Coverage Both Matter
Imputation uses observed genotypes and reference haplotypes to infer genotypes at untyped variants. A larger reference panel can offer more examples of the haplotypes that occur in the target data. But the number of reference samples is only one part of the information available: the panel must also represent relevant variants and the haplotype patterns around them. These features vary across the genome, so the effect of adding reference samples can differ markedly from one locus to another.
Here, panel size means the number of sampled chromosomes or individuals available in the reference data. Because each diploid individual typically contributes two chromosomes at an autosomal locus, it is useful to distinguish the number of individuals from the number of sampled haplotypes. They are not necessarily equivalent to the same number of independent observations: relatedness and shared ancestry can make chromosomes dependent. Variant coverage means which variants are present and usable in the panel, together with how well the panel's typed markers represent the local haplotype structure. A large panel can have limited coverage at a particular site, while a smaller panel may contain informative variants there.
The following probability calculation gives one precise, deliberately simplified way to understand the benefit of size. It is not a complete model of imputation accuracy. In practice, imputation also depends on how target haplotypes compare with reference haplotypes, the local pattern of linkage disequilibrium, phasing, and the informativeness of observed markers.
A Sampling Model for Haplotype Representation
The assumptions are important: sampled haplotypes are treated as independent draws from one population, and the class frequency \(q\) is taken as fixed. Real reference panels can include related individuals, unequal sampling across subpopulations, or haplotypes that do not match the target population. In those settings, the formula is a useful benchmark rather than a guarantee.
The probability of no representative is \((1-q)^N\): each of the \(N\) independent draws misses the class with probability \(1-q\), so all \(N\) miss it with probability equal to the product of those probabilities. Subtracting from one gives the probability of at least one representative. The expected number of representatives is \(Nq\), since each draw has probability \(q\) of being in the class. Neither quantity says whether that haplotype will be distinguishable from other haplotypes using the target's observed markers.
Worked Example: A Less Common Haplotype Class
Suppose a relevant haplotype class has frequency \(q=0.02\), and a panel contains \(N=100\) independently sampled haplotypes. Under the sampling assumptions, the probability of seeing at least one copy is
So the model gives about an 86.7% probability that the class is represented. If the panel instead has \(N=200\) independent haplotypes, the calculation is
The larger panel makes representation much more likely in this example. It does not imply that imputation at every locus is 98.2% accurate: a represented haplotype may not be distinguishable from alternatives with the available markers, and the sampling assumptions may not hold.
Why Added Samples Help Some Loci More Than Others
The representation calculation depends on \(q\). If an added panel samples haplotypes that are genuinely independent draws from the same population, a class with a moderate frequency may already be represented in the existing panel. Additional samples can then add relatively few new examples of that class. A class that was previously absent has a larger opportunity to become represented. The potential gain therefore depends on the frequency and distribution of relevant haplotypes at that particular locus.
For example, a locus may have common haplotypes that are already well represented, so added samples mainly repeat patterns already present. At another locus, a less frequent local haplotype may be absent from the smaller panel and appear in the expanded panel. A third locus may involve target haplotypes that are poorly represented because the panel's sampled population differs from the target population. Increasing the count of samples does not, by itself, ensure that the missing target-relevant patterns enter the panel.
Worked Example: Contrasting Representation Gains at Two Loci
Consider two hypothetical haplotype classes, each sampled under the independent-draw model. At locus A, the relevant class has frequency \(q=0.20\); at locus B, it has frequency \(q=0.01\). Compare the probability of representation in \(N=20\) and \(N=100\) haplotypes.
For locus A, the probabilities are
The class is already very likely to occur in the smaller panel, so the increase in representation probability is about \(1.0000-0.9885=0.0115\), or 1.15 percentage points.
For locus B, the corresponding probabilities are
The increase is \(0.6340-0.1821=0.4519\), or about 45.2 percentage points. Thus the same increase in panel size has very different effects on the probability that these particular classes are represented. This comparison concerns representation only; it does not rank the imputation accuracy of the two loci.
The calculation also illustrates diminishing gains in the probability of representation: once that probability is already near one, it cannot increase much further. This is not a universal claim that every measure of imputation performance must improve with diminishing returns. It is a consequence of this particular probability, whose maximum is one. Other aspects of performance depend on the panel's haplotypes and marker information.
Coverage: Which Variants and Patterns Are Available?
Panel coverage has at least two related aspects. First, the reference data may or may not include the variant being imputed. If the variant is directly represented in the reference, its observed alleles can contribute to the haplotype patterns used for imputation. If it is absent, imputation must rely on relationships between that variant and variants that are represented, as reflected in the reference haplotypes. Second, the panel's markers must provide enough information to distinguish relevant haplotypes in the target data. Simply counting reference individuals does not resolve either limitation.
Marker coverage is local. A region with many informative markers can distinguish haplotypes more finely than a long gap where different chromosomes share the same observed marker pattern. Even in a densely covered region, markers are useful only insofar as their allele combinations track the variant of interest. Local recombination and variation in linkage disequilibrium mean that marker density alone is not a complete measure of how informative a region is.
The distinction between panel size and coverage is practical. Increasing the number of individuals while keeping the same set of measured variants can add new examples of the existing marker patterns. It cannot create direct observations of a variant omitted from the panel. Conversely, adding more measured variants can improve the representation of local structure, but a newly included variant may still be poorly represented if the relevant alleles or haplotypes are scarce in the samples.
Worked Example: A Variant Inside a Sparse Marker Gap
Suppose a target array observes markers at positions 1,000,000 and 1,100,000 base pairs, but not at an intermediate variant at 1,050,000 base pairs. In a hypothetical reference panel, the intermediate variant is included, and several reference haplotypes carry different combinations of the two flanking marker alleles and the intermediate allele.
The target's flanking-marker genotypes can then be compared with reference haplotypes to find plausible matches. If a target haplotype's marker pattern is shared by reference haplotypes carrying different alleles at the intermediate variant, those markers do not uniquely identify the allele. If the marker pattern instead tracks the intermediate allele consistently among relevant reference haplotypes, it provides more information for inference.
Now suppose the panel is expanded by adding many individuals, but the added individuals repeat the same marker and intermediate-variant patterns already present. The panel contains more copies of known patterns, but no new distinction between the competing haplotypes. If the added individuals contribute previously unseen, relevant patterns, they may provide new information. The result depends on which haplotypes are added, not only on the revised sample count. No numeric accuracy value follows from the positions alone: that would require information about the haplotype patterns and an assessment against observed genotypes.
Assumptions, Diagnostics, and Failure Modes
The sampling model is most informative when haplotypes are independent and representative of the target population. Relatedness reduces the amount of independent information relative to the raw chromosome count. Population structure can also make one overall frequency \(q\) misleading: a haplotype common in one group may be rare or absent in another. In these cases, a panel's nominal size can overstate how well it represents the target haplotypes relevant to a locus.
A useful conceptual diagnostic is to ask, separately for the region of interest: Is the variant directly present in the panel? Are nearby markers present and informative? Does the panel include haplotypes resembling those in the target data? Are added samples contributing new patterns, or mostly repeating patterns already observed? These questions identify possible bottlenecks; a large total number of panel samples does not answer them.
Signs that a simple size-based expectation may be misleading include uneven coverage across genomic regions, sparse marker gaps, target haplotypes not seen in the reference, and strong dependence among reference samples. Further, the presence of a haplotype class does not guarantee that it can be matched unambiguously from the target's observed data. A reference variant may be present but carried on several marker backgrounds, or two different haplotypes may look identical at the available target markers.
A full empirical assessment needs observed genotypes at variants whose imputation is being evaluated. Imputation quality metrics and validation procedures provide ways to quantify performance, but they address a different question from the representation probability derived here. The key distinction is between opportunity—whether the panel contains potentially useful variants and haplotypes—and performance—how well the imputation procedure recovers genotypes in the target data.
Comparing the Main Ways to Improve Representation
| Change | Potential benefit | What it cannot guarantee |
|---|---|---|
| Add reference samples | May add examples of haplotypes and make less frequent patterns more likely to be represented. | Cannot guarantee new informative patterns, target-population representation, or improved inference at every locus. |
| Add variants to the panel | Can directly include previously absent sites and reveal finer local haplotype patterns. | Cannot guarantee that relevant alleles are frequent enough or distinguishable in the target data. |
| Increase marker density in the target data | Can provide more observed information for matching target haplotypes to reference patterns. | Cannot compensate for a reference panel that lacks relevant haplotypes or the variant itself. |
These approaches address different limitations and are not interchangeable. Panel size concerns how many sampled chromosomes provide examples; reference coverage concerns which variants and patterns those examples contain; target marker density concerns what can be observed in the samples being imputed. A gain in one dimension may leave a bottleneck in another unchanged.
Common Mistake
Check Your Understanding
Answer each question, then compare your reasoning with the solutions below.
- Under independent sampling, a haplotype class has frequency \(q=0.05\). What is the probability it is represented in \(N=40\) sampled haplotypes?
- Why can adding many reference individuals have little effect at a locus where the relevant haplotype class is already almost certain to be represented?
- A variant is absent from the reference panel, but nearby markers are present. What information can still support its imputation, and what limitation remains?
- Why might a panel with many related individuals provide less new haplotype information than its raw sample count suggests?
Solutions. 1. The probability is \(1-(1-0.05)^{40}=1-0.95^{40}\approx1-0.1285=0.8715\), or about 87.2%. 2. The probability of representation cannot exceed one, so there is little room for it to increase once it is already near one; added samples may also repeat patterns already present. 3. Nearby marker patterns and their relationships to the variant in reference haplotypes may support inference, but the variant is not directly observed in the panel, and shared or non-informative marker patterns can leave its allele ambiguous. 4. Related chromosomes are not independent draws; they can repeat existing haplotypes, so the effective number of distinct patterns may be smaller than the raw count suggests.