Why Conflicting Flags Need a Decision, Not a Vote
Genotype quality control often produces several flags for the same sample or variant. A sample might have slightly low call rate and unusually high heterozygosity, while its sex-check metrics agree with its recorded information. A variant might fail a Hardy–Weinberg screen but have high call rate and no other obvious warning. These combinations do not have a universal “two flags means exclude” interpretation. The flags can have different causes, operate at different levels, and affect different parts of the data.
The task is to make a defensible decision using the pattern of evidence. First identify whether the concern is about a sample, a variant, or both. Then ask whether the flags plausibly share a technical cause, whether they have a credible biological or design explanation, and what data would be lost under each action. A flag is a reason to investigate under a specified rule; it is not, by itself, proof that a sample or variant is erroneous.
Establish What Each Flag Measures
A sample call rate is the proportion of eligible variants for which that sample has a called genotype. A variant call rate is the proportion of eligible samples with a called genotype at that variant. These are different summaries: low sample call rate can implicate one sample across many loci, whereas low variant call rate can implicate one locus across many samples. The denominator should be stated, since excluded loci or samples change the rate.
A Hardy–Weinberg (HWE) result concerns genotype-count departures from the frequencies expected under the model used for the test. A small p-value is not the probability that the variant is erroneous or that the null is true. HWE departures can arise from technical problems, but also from population structure, relatedness, selection, or study design. The earlier material on Hardy–Weinberg deviations and failures provides the test context; here the key decision is whether the pattern is specific to a variant or part of a broader data problem.
Heterozygosity is a sample-level summary of the fraction of called genotypes that are heterozygous across selected loci. An extreme value can be consistent with contamination, inbreeding, ancestry differences, or other causes; it is not a diagnosis on its own. Sex checks compare genetic evidence, such as X- and Y-chromosome patterns, with recorded information. A discrepancy calls for careful investigation and should not be treated as evidence that a person’s recorded sex or identity is “wrong.” Relatedness estimates identify pairs with more genetic similarity than expected for unrelated samples. Such pairs may be duplicates, twins, relatives intentionally included in a family study, or apparent relationships influenced by ancestry or technical artifacts.
These metrics are not independent votes. For example, a sample with widespread missing calls may also have an unstable heterozygosity estimate because that estimate uses only its called genotypes. A variant with poor calls can distort HWE counts. Two flags that share one underlying cause should not be counted as two independent confirmations.
A Structured Decision Process
Determine whether the flag applies to a sample, a variant, or a sample pair. Check the denominator, analysis subset, and threshold that produced it.
Ask whether related metrics point to one technical issue. For instance, poor sample call rate across many variants is a different pattern from one variant with poor call rate across samples.
Consider ancestry, family structure, intentional study design, and recorded metadata. These explanations do not automatically clear a flag, but they affect its interpretation.
Retain when the evidence is compatible with the study and no material technical concern remains; investigate when evidence is ambiguous; filter when a prespecified rule and converging evidence support exclusion. Record the affected data and rationale.
Thresholds should be selected for the data and intended analysis, not invented after seeing an inconvenient result. A project may use a call-rate threshold, an HWE screening rule, or a heterozygosity outlier rule, but the threshold alone does not establish why a record failed. Compare results across relevant subsets where appropriate, and preserve enough information to reproduce the decision.
Worked Example: A Variant Has an HWE Flag but Good Call Rate
Suppose a hypothetical study has 1,000 eligible samples at one autosomal variant. Genotypes are called for 970 samples, so the variant call rate is
Assume the project’s prespecified variant call-rate threshold is 95%. The variant passes that rule. An HWE screen in the control subset nevertheless gives a p-value of 0.0008, below the project’s screening threshold of 0.001. These results are not contradictory: the call-rate rule measures missingness, while the HWE screen concerns genotype proportions among called samples.
Before filtering, check whether the HWE result is concentrated in controls as intended, whether the variant has an unusual missingness pattern across genotype groups, and whether population structure or relatedness could explain the departure. Suppose the hypothetical review finds the same departure within ancestry-defined control subsets and no evidence that missingness is concentrated in a genotype class. That review does not prove the variant is valid, but it weakens two simple technical explanations. A defensible decision is to retain it provisionally or investigate further under the study protocol, rather than automatically exclude it solely because its p-value is small. The HWE flag and the follow-up evidence should be recorded separately.
Resolve Sample-Level Conflicts in Context
For sample-level decisions, consider whether multiple measurements are calculated from overlapping calls. A low call rate can make summaries based on observed genotypes less stable. A heterozygosity outlier is more concerning when it is accompanied by other independent evidence of a sample problem, but a sex-check discrepancy or relatedness flag may instead reflect a metadata issue or intended family sampling. Follow-up should target the suspected cause, not merely repeat the same calculation.
Worked Example: Low Call Rate, High Heterozygosity, and an Agreed Sex Check
In a hypothetical dataset, one sample has calls at 8,800 of 10,000 eligible autosomal variants. Its call rate is
The project’s sample call-rate threshold is 95%, so this sample fails it. Among the 8,800 called genotypes, 3,520 are heterozygous, giving observed heterozygosity
Suppose the comparison group has mean heterozygosity 0.300 and standard deviation 0.020. The standardized deviation is
Thus the sample is five comparison-group standard deviations above the mean, an extreme value under that comparison. Its X- and Y-chromosome checks agree with its recorded information. That agreement does not negate the two autosomal flags, but it gives no additional support to a sex-chromosome discordance explanation.
The appropriate next step is to inspect whether the missingness and heterozygosity pattern is widespread across loci and whether the comparison group is suitable for this sample’s ancestry and study context. Because the low call rate breaches the stated threshold and heterozygosity is extreme, the sample warrants investigation; if the project’s prespecified rule requires filtering samples below 95%, it is filtered under that rule. The sex-check result is recorded as concordant, not used as a reason to override the call-rate decision. The heterozygosity result remains a diagnostic signal, not proof of a particular cause.
A common mistake is to treat concordant sex checks as a general “pass” that cancels unrelated sample concerns. Each metric addresses a different question. Likewise, an extreme heterozygosity value should not be interpreted against a pooled reference without considering whether ancestry composition or other population differences make that comparison inappropriate.
Relatedness Flags and the Cost of Filtering
A relatedness flag is a statement about a pair, not automatically a defect in either member. Before removing a sample, determine whether the relationship is expected and whether the analysis requires unrelated observations. In a family-based design, removing relatives can destroy the intended structure. In an analysis that assumes independent samples, a pair may need resolution, but the choice of which sample to retain should use sample quality, phenotype completeness, and study design rather than arbitrary order.
Worked Example: A Related Pair with Unequal Sample Quality
Suppose two hypothetical samples exceed a project’s relatedness threshold for flagging pairs. Sample A has 9,850 calls among 10,000 eligible variants, while Sample B has 9,200. Their call rates are
The project’s sample call-rate threshold is 95%, so A passes and B fails. The records also show that the pair are known first-degree relatives in a study designed to include family members. The relatedness flag is therefore expected and does not, by itself, justify removing either sample. If a particular downstream analysis requires unrelated samples, Sample B is the more defensible candidate for exclusion under the stated call-rate rule, subject to the analysis plan and availability of its phenotype data. If the analysis uses family structure, both may be retained. The decision follows the analysis goal and sample-level evidence, not a rule that every related pair must be broken.
Diagnostics, Alternatives, and Limitations
When flags conflict, useful diagnostics include verifying that the intended sample and variant subsets were used, checking missingness by relevant groups, and looking for clusters of affected samples or variants. A single outlier among otherwise well-behaved data suggests a different follow-up from a platform-wide pattern. If several metrics depend on the same calls, identify that dependence before describing them as converging evidence.
| Decision approach | Useful when | Main limitation |
|---|---|---|
| Apply a prespecified single-metric threshold | A project has a clear rule, such as a minimum call rate, and needs a reproducible screen | A threshold does not explain the cause of failure or resolve conflicts with other evidence |
| Require converging evidence | A costly exclusion needs support from distinct diagnostics | Metrics may be correlated, and a real problem may not trigger several flags |
| Investigate before deciding | Flags have plausible biological, design, or technical explanations | Follow-up takes time and may remain inconclusive; decisions still need a documented rule |
No combination of these metrics guarantees error-free data. Call rate alone cannot detect every incorrect genotype. HWE results depend on the tested group and model. Heterozygosity outlier status depends on a suitable comparison distribution. Sex checks are limited genetic checks, not comprehensive statements about identity or biology. Relatedness estimates can be affected by ancestry and technical quality. These limitations make a transparent, analysis-specific decision more reliable than a universal flag-count rule.
Common Mistakes
- Counting flags as independent votes. Two summaries based on the same missing calls may reflect one issue, not two independent confirmations.
- Using a variant flag to remove a sample, or vice versa. Check the unit and scope before acting; a locus-level issue does not establish that an individual is poor quality.
- Calling an HWE p-value the probability that a variant is bad. It measures evidence against the test’s HWE model under its assumptions, not the probability of error.
- Removing every related pair. Determine whether the relationship is expected and whether the analysis requires unrelated samples.
- Letting one reassuring metric erase an unrelated warning. A concordant sex check does not repair low autosomal call rate.
Check Your Understanding
For each situation, state the most defensible interpretation or next action. Solutions follow the questions.
- A variant has a 99% call rate and a small HWE p-value in controls. Does its call rate cancel the HWE flag?
- A sample has 9,400 calls among 10,000 eligible variants. Calculate its call rate and decide whether it passes a 95% threshold.
- A relatedness flag occurs for two known relatives in a family study. Is either sample automatically removed?
- A sample has extreme heterozygosity relative to a pooled group but ordinary heterozygosity within its ancestry group. What should be reconsidered?
- Why should a concordant sex check not be used to dismiss a low sample call rate?
Solutions. (1) No. The metrics measure different properties; investigate the HWE departure in context while reporting that the variant passes the call-rate rule. (2) \(9400/10000=0.940=94.0\%\), so it fails the stated 95% threshold. (3) No. The relationship may be intentional; decide based on the analysis design and any applicable sample-quality rules. (4) Reconsider whether the pooled comparison is appropriate; ancestry composition may explain the apparent outlier, though other evidence should still be reviewed. (5) The sex check and call rate address distinct questions, so agreement on one does not resolve missingness in autosomal calls.