Common Imputation Pipeline Failures

When Imputation Output Looks Implausible

Imputation predicts unobserved genotypes from observed markers and patterns in a reference panel. A pipeline can complete without an obvious software error and still produce output that is biologically implausible: allele frequencies may look reversed, high-quality variants may occur at unexpected positions, or prediction quality may collapse in a region or sample group. The useful question is not simply whether a metric is low, but which failure could produce the pattern observed?

This tutorial focuses on four common causes: allele-orientation errors, genome-build mismatches, sparse marker coverage, and a mismatch between the study samples and reference panel. These problems can overlap. For example, a build mismatch can reduce the number of markers that match the panel, making a region appear sparsely covered. Diagnostics therefore narrow the possibilities; they do not automatically prove a cause.

Quality metrics such as INFO or dosage \(R^2\) summarize aspects of imputation uncertainty or performance; their exact calculation and interpretation depend on the metric and software. Earlier material on imputation quality metrics, INFO scores, and validation by masking provides their definitions and uses. Here, treat a low value as a signal to investigate, not as a diagnosis. A high value is not proof that alleles, positions, or sample assignments are correct.

Four Failure Patterns and Their Diagnostics

Allele flips and strand-orientation errors

For a biallelic variant, switching which allele is counted changes each dosage from \(d_{ij}\) to \(2-d_{ij}\). If \(p\) is the frequency of the counted allele, its frequency under the opposite coding is \(1-p\). Thus, a striking discrepancy between study and reference frequencies may be a coding reversal rather than a biological difference. Allele labels and strand alignment matter: A/G and T/C can describe complementary strands, while A/T and C/G variants are palindromic and cannot be oriented from allele labels alone.

A flip may produce frequency discrepancies close to complements, reversed effect directions when downstream analyses use inconsistent coding, or inconsistent genotype counts after matching. It does not necessarily cause poor imputation: a pipeline that correctly harmonizes alleles can use a different dosage orientation and still impute correctly. Check allele labels, strand conventions, frequency comparisons, and dosage coding together. Do not flip a variant solely because its frequency differs from the panel; real population differences can also produce such a discrepancy.

Genome-build mismatches

A chromosome and position identify a locus only relative to a particular genome build. If study coordinates use one build and panel or imputation software expects another, matching by position can select the wrong variant or fail to match a real one. Warning signs include unusually many unmatched variants, variants assigned to unexpected nearby positions, allele combinations that do not agree with the intended locus, and abrupt regional drops in usable markers or quality.

Check the build recorded for each input, compare coordinates against a trusted build-specific reference, and inspect several neighboring variants rather than relying on a single identifier. Variant identifiers can be reused or stale, and an identifier match alone does not establish that the position and alleles match. A build mismatch can affect a whole file or chromosome, while isolated discrepancies may instead reflect variant-specific errors.

Sparse marker coverage

Imputation relies on observed markers that carry information about nearby unobserved variation. When few usable markers are present, the local haplotypes are less well anchored, so imputation quality can fall across a stretch of the region. A pattern of low quality across consecutive variants, accompanied by a gap in observed markers, is more consistent with sparse coverage than a single isolated low-quality variant. Check marker density and the number of markers that actually matched and passed quality control; the raw count before matching can be misleading.

Sparse coverage is not established by low INFO alone. A poor panel match, incorrect coordinates, or low-quality input genotypes can also reduce imputation quality. Conversely, dense markers do not guarantee good results if the markers are misaligned or the relevant haplotypes are poorly represented.

Reference-panel mismatch

A reference panel can be technically well formed yet represent study samples poorly. Differences in ancestry, population history, or relatedness can mean that study haplotypes are not well represented. The resulting pattern may include lower masked-genotype concordance in some sample groups, larger frequency differences at many variants, or particularly weak performance for less common variants. These are clues, not proof: frequency differences may be real, and rare-variant performance is sensitive to sample size and validation design.

Compare validation results across relevant sample groups and genomic regions, and examine frequency differences across many well-matched variants rather than diagnosing from one locus. Review the panel’s documented composition and the study sample composition. A panel mismatch may coexist with adequate overall average performance while leaving a subgroup or variant class poorly imputed.

Worked Diagnostic Examples

A Practical Diagnostic Logic

A useful investigation compares several kinds of evidence, starting with checks that can affect the whole dataset and then examining local or sample-specific patterns. Keep the original data and settings available so a proposed correction can be tested rather than assumed.

Why the Patterns Matter—and Where They Can Mislead

A diagnostic pattern is most useful when it predicts more than one observation. A flip predicts complementary frequency coding and reversed dosage orientation at the affected variant. A build mismatch predicts coordinate-related matching problems, often across a set of variants. Sparse coverage predicts a local relationship between marker gaps and weaker performance. Panel mismatch predicts performance differences associated with sample groups or haplotype representation that may persist despite adequate marker matching.

These signatures overlap, so avoid treating one statistic as a verdict. A low INFO score does not say whether the cause is sparse markers, a poor panel match, or incorrect variant alignment. A frequency difference is not automatically a flip. A high aggregate score can conceal a badly performing region or subgroup. Examine results at the scale where the suspected problem occurs: variant, region, sample group, or whole dataset.

The main alternatives are not mutually exclusive. More markers may help when coverage is sparse, but cannot repair incorrect allele coding. Correcting coordinates can restore matches, but cannot make an unrepresentative panel contain missing haplotypes. Choosing a better-matched panel may improve representation, but does not resolve a build error. The limitation of output-based diagnosis is that multiple causes can produce similar symptoms; targeted checks and independent validation are needed to distinguish them.

A common mistake is to see a low quality value and immediately discard all variants in the region. The reasoning is that low quality means the imputation has failed for one known reason. The correction is that quality indicates uncertainty or performance, not its cause. First check identity, build, coverage, and panel representation; then use the quality metric and validation evidence to decide what should be retained or investigated further.