Why Imputation Uses a Reference Panel
A genotyping assay records alleles at a selected set of sites, while many other variants remain untyped. As discussed in the introductory material on untyped variants and missing genotypes, these are not necessarily missing calls at sites that were measured: they may be sites the assay did not measure at all. Imputation uses information from typed sites to estimate genotypes at untyped sites.
A reference panel contributes observed genetic patterns against which the typed data can be compared. In particular, a panel can provide phased haplotypes: sequences of alleles along a single chromosome copy, with the alleles assigned to that copy. The panel therefore records which alleles occur together across its included sites, not just the separate frequency of each allele. That joint information can reveal that a particular pattern at typed sites tends to occur with one allele at an untyped site.
The panel does not directly reveal the genotype of a person who is not in the panel. Its contribution is evidence: if a target chromosome has a pattern at measured sites that is also seen among panel haplotypes, the alleles accompanying that pattern in the panel can inform an estimate at an untyped site. The usefulness of that evidence depends on which variants the panel includes, how many relevant chromosomes it represents, and how well those chromosomes reflect the target sample.
What the Haplotype Information Adds
Suppose a panel contains a typed marker and an untyped target variant. If the panel’s phased chromosomes with marker allele A mostly carry target allele T, then observing marker allele A on a target chromosome is evidence in favor of T on that chromosome. This evidence is stronger when the pattern is observed repeatedly in a panel that adequately represents the target sample. It is weaker when the pattern is rare, the panel is small, or the target chromosomes differ from the panel chromosomes in how alleles are associated.
For a simple illustration, let \(h\) denote a pattern at typed sites and let \(T\) denote a specified allele at the untyped site. If \(M_h\) panel haplotypes carry pattern \(h\), and \(K_{h,T}\) of those also carry \(T\), the panel’s observed conditional frequency is
This fraction summarizes how often the allele accompanies that typed pattern in the panel. It is not a guarantee that a target chromosome with pattern \(h\) carries \(T\), nor is it automatically the correct probability for every target sample. Treating it as relevant requires the panel to be sufficiently representative of the target chromosomes and the observed pattern to be measured consistently.
Worked Example: Reading a Conditional Haplotype Frequency
Consider hypothetical phased panel data. Among 12 panel haplotypes with typed-marker pattern \(h\), 9 carry allele T at an untyped target site and 3 carry allele C. Calculate the observed conditional frequency of T and interpret it.
Here \(M_h=12\) and \(K_{h,T}=9\), so
In this panel, 75% of the haplotypes with pattern \(h\) carry T at the target site. If a target chromosome has the same typed pattern, this panel offers evidence for T, but the evidence is not certain: 3 of the 12 matching panel haplotypes carry C. The value 0.75 is a panel-based conditional frequency, not a statement that 75% of all target individuals have a particular genotype.
This interpretation assumes that the typed pattern and target allele are recorded correctly and that panel haplotypes with pattern \(h\) are informative for the target chromosomes. A panel drawn from a substantially different genetic background might have a different conditional frequency.
Coverage, Size, and Representation Do Different Jobs
These three properties are related but not interchangeable. Coverage determines whether the relevant target variant and useful typed markers are present in the panel. Size affects how many examples of a haplotype pattern may be available. Representation affects whether those examples reflect the target sample. More individuals cannot supply direct allele observations at a site the panel does not include, and a large panel can still be poorly matched to the target chromosomes.
| Panel feature | What it contributes | What it does not guarantee |
|---|---|---|
| Variant coverage | Observed alleles at included sites, including the target site and potentially informative typed sites | That every relevant variant or useful marker is included |
| Sample size | More observed chromosome copies and, often, more examples of haplotype patterns | That the target’s particular pattern is represented or that the panel matches the target sample |
| Population representation | Greater relevance of the panel’s haplotype patterns and allele associations to the target chromosomes | Perfect correspondence for every individual or genomic region |
For a site to be directly represented, the panel must have data at that site. If it does not, additional panel individuals do not reveal the site’s alleles. A target allele may still be predictable indirectly from other measured sites if the panel contains informative patterns, but that is a different kind of evidence and may be less reliable. Conversely, including a target site is not enough on its own: useful estimation also requires relevant panel chromosomes and information that connects observed typed patterns to the target allele.
Worked Example: More Haplotypes Improve Frequency Precision
Suppose two hypothetical panels have the same conditional frequency of T for pattern \(h\), but different numbers of matching haplotypes. Panel A has 30 T alleles among 50 matching haplotypes; Panel B has 1,200 T alleles among 2,000 matching haplotypes. Compare the observed frequencies and their approximate standard errors under a simple independent Bernoulli sampling model.
Both observed frequencies equal 0.60:
Under the stated model, the standard error of a sample proportion is approximately \(\sqrt{\hat{q}(1-\hat{q})/M_h}\). Thus,
The larger panel gives a more precise estimate of this conditional frequency under the model, while the observed estimate is the same in both panels. This calculation assumes independent, representative haplotypes with a common conditional frequency. Real panel chromosomes may not be independent, for example because some donors are related or because haplotypes share long stretches of sequence. The calculated standard errors can therefore overstate the effective information in an actual panel.
Representation matters because allele patterns and their frequencies can differ among populations and within populations. Broad population labels are imperfect proxies for an individual’s genetic background, and representation can vary across genomic regions. The practical question is not simply whether a panel has a label that matches the target cohort, but whether it contains haplotypes relevant to the target chromosomes and sites being imputed.
Worked Example: A Larger but Less Representative Panel
Suppose a hypothetical validation set contains 30 target chromosomes with typed pattern \(h\); 20 carry T at the target site. Two panels provide the following conditional frequencies for the same pattern: a large Panel A has \(\hat{q}_{T\mid h}=0.25\), while a smaller Panel B has \(\hat{q}_{T\mid h}=0.70\). Compare their estimates with the validation frequency.
The validation frequency is
Panel A differs from the validation frequency by \(|0.25-0.667|=0.417\). Panel B differs by \(|0.70-0.667|=0.033\). For this particular pattern and validation set, Panel B’s conditional frequency is closer, despite its smaller size. This comparison illustrates why more panel chromosomes cannot compensate automatically for poor representation.
The validation set is hypothetical and small, so its frequency is itself uncertain; closeness in this example does not establish that Panel B will perform better for other patterns, sites, or cohorts. A sound comparison would use validation data relevant to the intended target sample and would examine more than one pattern or site.
Assumptions, Diagnostics, and Failure Modes
Interpreting panel evidence requires several conditions. The panel and target data must use compatible variant identities and allele coding; otherwise, apparent matches or differences may be artefacts of inconsistent records. The target’s typed markers must be measured reliably enough to identify relevant patterns. The panel must include the target variant for direct information about its allele, or it must contain other informative sites for indirect evidence. Finally, panel haplotypes used to inform the target should be relevant to its genetic background. These are assumptions about data quality and applicability, not automatic properties of a panel.
Warning signs include a target pattern absent or rare in the panel, a target variant absent from panel data, substantial differences in allele or haplotype frequencies between panel and target data, and results that vary markedly across plausible panels or target subgroups. Such signs do not prove that every imputed estimate is wrong, but they indicate that the panel may provide weak or poorly calibrated evidence. Relevant checks include confirming variant and allele alignment, counting the panel haplotypes matching common target patterns, and evaluating predictions in an appropriate held-out or validation sample when available.
A common mistake is to reason, “This panel is much larger, so its imputations must be more accurate.” Size can improve the precision with which the panel’s frequencies are estimated and make rarer patterns more likely to be observed. It cannot make an absent variant directly observed, ensure that the target pattern occurs, or make an unrepresentative panel representative. A second mistake is to equate a variant’s presence in the panel with accurate imputation: coverage supplies observations, but accuracy also depends on whether the typed patterns distinguish the target alleles in the target sample.
How the Panel Role Compares With Alternatives
A reference panel is one source of information, not the only possible source. The comparison below concerns where the evidence comes from; it does not prescribe a particular imputation algorithm.
| Source of information | Potential contribution | Main limitation |
|---|---|---|
| External reference panel | Phased haplotypes and observations at sites not typed in the target data | Applicability depends on coverage and representation |
| Target cohort alone | Patterns and allele frequencies observed among the study participants | May have few chromosomes, limited variant coverage, or few examples of uncommon patterns |
| Direct measurement of a variant | Observed genotype at the measured site rather than an estimate based on other sites | Only available for variants actually assayed with adequate data quality |
These sources answer different needs. Direct measurement avoids relying on panel-based inference for that site, while a panel can provide information beyond the target cohort’s own measured sites. A target cohort may be more relevant to its own participants but too small to represent uncommon patterns well. An external panel may be large yet less applicable. These trade-offs should be assessed for the variants and target sample of interest rather than reduced to a single panel-size number.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions below.
- A panel has 16 haplotypes with typed pattern \(g\), of which 12 carry allele A at an untyped site. What is the observed conditional frequency of A?
- Two panels have the same conditional frequency, but one has four times as many matching haplotypes. What can be said about the expected sampling precision, and what assumption is needed?
- A target variant is absent from the panel, but several typed markers are present. Does a larger number of panel individuals provide direct observations of the target variant? Explain.
- A large panel has conditional frequency 0.30 for a pattern, while a relevant validation set has frequency \(9/12\). Calculate the absolute difference and state what it suggests, without overgeneralizing.
Solutions. 1. The conditional frequency is \(12/16=0.75\), or 75% of the panel haplotypes with pattern \(g\). 2. Under independent, representative sampling with the same underlying conditional frequency, four times as many observations gives approximately half the standard error, because standard error scales as \(1/\sqrt{M_h}\). Dependence or poor representation can undermine that comparison. 3. No. A larger panel still lacks direct allele observations at an excluded site; typed markers might provide indirect evidence, but not direct coverage. 4. The validation frequency is \(9/12=0.75\), so the absolute difference is \(|0.30-0.75|=0.45\). This substantial discrepancy suggests that the panel’s conditional frequency may not represent this validation set for this pattern; it does not establish the panel’s performance at other sites or patterns.