Reference Panel Mismatch and Ancestry Bias

Why Reference-Panel Match Matters

An ancestry estimate is conditional on the reference populations used to construct it. The panel is not a neutral backdrop: its allele frequencies help define what the model means by each ancestry component. If a reference group is a poor proxy for a source population, the model may assign too much or too little ancestry to that component. If a source is omitted entirely, the model may explain its genetic variation using the available components, even when none is a good fit.

This is a problem of model specification, not merely sampling variability. A larger panel can reduce uncertainty about the allele frequencies of the sampled reference groups, but it cannot make those groups representative of a different source population. The practical question is therefore not only whether an estimate is numerically precise, but also whether it is stable under reasonable changes to the panel.

For a simple illustration, suppose a target population has ancestry proportion \(q\) from source A and \(1-q\) from source B. At marker \(i\), let \(f_{iA}\) and \(f_{iB}\) be the allele frequencies in the two true sources. Under a two-source mixture model, the expected target allele frequency is

This relation assumes that the target is represented by the stated two sources, that the allele frequencies refer to the same marker and allele, and that the mixture proportion applies across the markers being considered. In real estimation, many markers contribute, genotypes are sampled, and the model may incorporate additional structure. The equation isolates the basic issue: the model substitutes panel frequencies for the source frequencies, and a mismatch changes the fitted proportions.

How Mismatch Creates Bias

At one marker, suppose the observed target frequency is \(p_i\), while the panel provides frequencies \(f^{*}_{iA}\) and \(f^{*}_{iB}\). If the fitted two-source model matches that frequency exactly, its estimate for the A proportion is

The denominator condition matters: if the panel frequencies are equal at this marker, the marker contains no information for distinguishing the two sources. Even when they differ, a small difference makes the estimate highly sensitive to small frequency errors. Estimates based on many markers combine information across loci, but the same principle remains: the fitted ancestry proportions depend on the reference frequencies and on how informative the markers are.

There is no universal direction to this bias. It depends on which allele frequencies are shifted, how those shifts relate to the target, the markers included, and the fitting procedure. Some loci can push an estimate upward and others downward. A model that reports one genome-wide proportion can therefore conceal substantial disagreement among marker-level signals.

Omitted Sources and Poor Fit

If the target contains ancestry from a source absent from the panel, a model with only the available sources must allocate that variation somewhere. The resulting proportions may be numerically well-defined while still failing to describe the target adequately. In a simple frequency model, an omitted source can make the target frequency fall outside the range attainable by mixtures of the represented sources. With multiple markers, the target may be compatible with the panel at some loci but not others, so no single mixture proportion fits them all.

A poor fit does not by itself identify which population is missing or prove that a particular alternative panel is correct. It is evidence that the stated panel and model may not adequately represent the target. Other explanations, such as marker quality issues or frequency-estimation error, should also be considered. The useful response is to examine whether the mismatch signal persists across markers and under reasonable panel alternatives.

Local Ancestry Calls Can Also Shift

Global ancestry proportions summarize contributions across the genome. Local ancestry methods assign or estimate ancestry along genomic segments, using marker information and, in many methods, information about neighboring positions. They rely on reference frequencies or haplotypes to distinguish sources. If those reference patterns are mismatched, the evidence at a segment can be misweighted: a segment may be assigned to the wrong represented source, or the evidence may become less decisive.

The following single-marker calculation illustrates the emission evidence, not a complete local ancestry algorithm. Assume one observed allele is modeled as Bernoulli within each source, and give the two sources equal prior probability at the position. Under these assumptions, the posterior probability for A is proportional to the A allele likelihood. Actual local ancestry methods may combine many markers and account for transitions between neighboring positions.

Assessing Panel Sensitivity

A sensitivity analysis asks whether the substantive conclusion persists when plausible choices about the reference panel change. The alternatives should be scientifically defensible, not selected only because they produce a preferred estimate. Depending on the data and question, useful comparisons include a more geographically or genetically relevant panel, a panel with a relevant group added, a panel with a questionable proxy removed, and panels formed by reasonable subsets of reference samples. A comparison should keep other analysis choices as consistent as possible so that differences can be attributed to panel choice.

For global ancestry, report the estimates from each panel and the change in each component in its natural units, such as percentage points. For local ancestry, compare segment or position-level posterior probabilities, call changes, and regions where the evidence changes. A change in a hard call may be more informative when accompanied by the corresponding probabilities: a shift from 0.51 to 0.49 changes a thresholded label but indicates weak evidence in both analyses.

A stable estimate across tested panels is reassuring about sensitivity to those particular choices, but it does not prove that the panel represents every relevant source. Conversely, a changing estimate is not automatically evidence that one panel is wrong; it shows that the result depends on panel specification and deserves qualification. Panel sensitivity is distinct from uncertainty caused by small reference samples: the latter concerns imprecision in estimated frequencies, while mismatch concerns whether the represented frequencies are appropriate in the first place. Both can occur together.

Common Mistake and Practical Interpretation

Wrong reasoning: “The ancestry estimate has a small standard error, so the reference panel must be appropriate.” A small standard error can describe precision conditional on the fitted model and its inputs. It does not establish that the model includes the relevant sources or that its reference frequencies match them.

Correction: Treat precision and representativeness as separate questions. Examine alternative panels and fit diagnostics, and state clearly that ancestry proportions are estimates relative to the specified reference framework. They are not direct measurements of a person's complete ancestry, and a change in a component is not evidence of a biological cause or of a change in identity.

These checks matter because ancestry estimates may be used to describe samples, compare groups, or inform downstream genetic analyses. When panel mismatch is ignored, apparent differences between people or groups can partly reflect the chosen reference rather than the target samples alone. Transparent panel comparisons make that dependence visible and help prevent overconfident interpretation.