Why the Choice Matters
When a relatedness screen flags a pair of samples, the next decision is not simply whether one sample has a larger QC number. The choice determines which person or record contributes to the analysis. Keeping the sample with more complete genotypes may be sensible when the pair represents duplicate records, but it may not be the best choice if that sample lacks the phenotype needed for the primary analysis or if removing it damages an important feature of the study design.
This tutorial addresses a specific setting: an analysis requires a subset with no pair above the chosen relatedness threshold, and a flagged pair cannot both be retained. The goal is to choose one member of that pair for the analysis, using genotype call rate, phenotype completeness, and study-design priorities. The decision should be based on the analysis plan and comparable measurements, not on a universal rule that one metric always wins.
A choice between the pair is meaningful only after the QC quantities have been defined on a comparable basis. For call rate, use the same set of eligible variants and the same genotype filters for both samples. For phenotype completeness, specify which fields matter to the planned analysis before comparing records. A sample can be complete for age and sex but missing the primary outcome; a single unqualified percentage could conceal that distinction.
Conditions for a Defensible Pairwise Decision
The selection procedure assumes that the pair has been flagged using a relatedness estimate and threshold suitable for the intended analysis, as covered in the earlier material on pairwise relatedness and relatedness thresholds. It also assumes that the samples have passed any mandatory QC requirements or that those requirements are being applied consistently. If one sample fails a hard exclusion rule, it is not a candidate to retain merely because it has a more useful phenotype.
Before ranking the pair, confirm that the records represent the situation the analysis assumes. A pair of duplicate records or monozygotic twins may call for different interpretation than a pair of other close relatives. Verify sample identifiers, phenotype linkage, and any known duplicate or twin status. If the relationship estimate is uncertain, or the pair’s identity is unresolved, the metrics alone do not settle the question.
Study design also matters. For example, a case-control analysis may need to preserve case representation, while a study with several recruitment sites may need to preserve coverage across sites. These are not automatic priorities: the analysis plan determines which features are essential. A small call-rate advantage should not silently override a prespecified design need, and a design preference should not justify retaining a sample that fails a required QC threshold.
A Transparent Decision Rule
A useful approach is to make the decision in stages. First apply hard eligibility rules. Then compare call rate and completeness for the analysis variables. Finally consider prespecified design priorities, such as case status, site, or a required follow-up measurement. Record the reason for retaining one sample and removing the other.
Exclude any sample that fails a mandatory call-rate threshold, has an unresolved identity problem, or lacks a field essential to the primary analysis.
Use call rates computed over the same eligible variant set, and assess completeness for the same prespecified phenotype and covariate fields.
If both samples remain eligible, use the priority stated in the analysis plan. Do not invent a priority after seeing which sample it favors.
Record both sample identifiers, the reason the pair was flagged, the metrics compared, and the reason for the retain/remove choice.
The staged rule is a lexicographic decision: a higher-priority criterion is considered before a lower-priority one. This is often clearer than combining unlike quantities into a single score. A call rate difference is measured in percentage points, while a design priority may concern whether a rare case group remains represented; adding these as if they were interchangeable requires arbitrary weights.
Call rate does have a direct interpretation for genotype data. If \(m\) eligible variants are assessed and a sample’s call rate is \(c\), its expected number of missing calls over that set, given the observed rate, is \(m(1-c)\). This does not measure every aspect of genotype quality, but it makes the scale of a call-rate difference concrete.
For example, with \(m=100{,}000\) variants, a call rate of \(0.996\) corresponds to \(100{,}000(1-0.996)=400\) missing calls. This calculation is useful only when both samples use the same \(m\) and the same eligible variant definition. It does not imply that the sample with fewer missing calls must always be retained.
Worked Pairwise Comparisons
Worked Example: Call Rate Decides When Other Factors Are Tied
Two hypothetical duplicate sample records are flagged. Both have the same phenotype fields, the same case-control status, and no design distinction relevant to the analysis. Each is evaluated at \(100{,}000\) eligible variants. Sample A has call rate \(0.996\); Sample B has call rate \(0.982\). Both pass the mandatory minimum call-rate requirement of \(0.970\).
For Sample A, the expected number of missing calls over the evaluated variants is \(100{,}000(1-0.996)=100{,}000(0.004)=400\). For Sample B, it is \(100{,}000(1-0.982)=100{,}000(0.018)=1{,}800\). The difference is \(1{,}800-400=1{,}400\) missing calls.
Because the records are duplicates, both have equivalent analysis-relevant phenotype and design value, and both pass the hard threshold, retaining Sample A is justified by its substantially higher call rate. The choice is not based on the label “A”; it follows from the stated comparison and assumptions.
Worked Example: Phenotype Completeness Outweighs a Small Call-Rate Difference
Two hypothetical related participants have the same case status and recruitment site. Both pass all mandatory QC thresholds. Sample C has call rate \(0.988\) and all 6 of the 6 prespecified primary-analysis fields observed. Sample D has call rate \(0.991\), but only 4 of those 6 fields are observed. The primary analysis requires all six fields.
Sample C’s phenotype completeness is \(6/6=1.00\), or \(100\%\). Sample D’s is \(4/6=0.667\), or \(66.7\%\) to three significant figures. At \(100{,}000\) eligible variants, C has \(100{,}000(1-0.988)=1{,}200\) missing calls, while D has \(100{,}000(1-0.991)=900\). D’s advantage is \(300\) fewer missing calls.
The study’s prespecified rule requires complete primary-analysis fields where possible, and both samples meet the call-rate minimum. Retain C: its more complete phenotype record makes it usable for the planned analysis, while D’s modest call-rate advantage does not compensate for missing two required fields. If the missing fields could be recovered reliably before analysis, the comparison should be repeated with the corrected records rather than treating missingness as permanent.
Worked Example: Study Design Resolves Competing QC Metrics
A hypothetical case-control study flags a related pair from the same recruitment stratum. Sample E is a control with call rate \(0.995\) and complete covariates. Sample F is a case with call rate \(0.989\) and the same required fields complete. Both exceed the call-rate minimum of \(0.980\). The analysis plan, written before this comparison, prioritizes retaining the case because cases in this stratum are difficult to recruit and the control group has adequate representation.
At \(100{,}000\) eligible variants, E has \(100{,}000(1-0.995)=500\) missing calls; F has \(100{,}000(1-0.989)=1{,}100\). E therefore has \(600\) fewer missing calls. However, the stated design priority favors preserving F, and both samples satisfy the hard QC and phenotype requirements.
Under that prespecified rule, retain F and remove E from this analysis. This is not a claim that lower call rate is preferable, nor that the case should always be retained. If case-control balance were not a concern, or if F failed a mandatory threshold, the decision could differ. The reason for the choice and the resulting change in sample counts should be documented.
Justification, Diagnostics, and Alternatives
The staged procedure follows from the goal of making the retained data usable for the intended analysis. Hard requirements define eligibility; comparing measurements computed on the same basis avoids attributing a difference to the sample when it comes from different denominators. Among eligible candidates, a declared design priority protects a feature the analysis plan has identified as important. When no such priority distinguishes the samples, better call rate and more complete required phenotypes provide relevant, interpretable grounds for choosing.
A combined score can be used only if its components, scaling, weights, and tie-breaking rules are specified in advance and are appropriate to the scientific objective. For instance, an unweighted sum of a call-rate percentage and a phenotype-completeness percentage has no natural statistical interpretation: the numerical result depends on how each quantity was scaled. A lexicographic rule is more transparent when the study can clearly state which criterion takes precedence.
Before finalizing a decision, check whether the measurements could be misleading. A call-rate difference may reflect batch, plate, or processing differences rather than a stable difference in sample quality. Confirm that both records were evaluated with the same variant set and filters. Phenotype completeness should be calculated for the variables needed by the actual analysis, not for an indiscriminate list of every field collected. Check whether a missing phenotype can be corrected from source records and whether the chosen sample leaves the intended case, control, or site representation intact.
The main alternatives depend on the analysis objective. For a conventional analysis that requires an unrelated sample set, retaining one member is a pruning decision. An analysis designed to use related participants may instead retain both and account for their dependence, but that is a different analysis strategy, not a justification for ignoring the stated filtering requirement. Removing both samples is a conservative option when neither is trustworthy or the pair cannot be resolved, but it sacrifices more data and should have a clear reason.
A pairwise decision also has a limitation: choices made independently for many pairs may conflict when a sample appears in more than one flagged pair. This tutorial’s rule explains how to compare two candidates; it does not guarantee that a set of pairwise choices produces an optimal retained cohort. Record decisions and review the resulting sample set rather than assuming every local choice is globally harmless.
Common Mistakes
- Always retaining the sample with the higher call rate. That ignores required phenotype fields and prespecified design priorities. Compare all relevant criteria after applying hard thresholds.
- Comparing percentages with different denominators. A call rate over different variant sets is not a fair pairwise comparison. Recalculate both on the same eligible set.
- Treating every phenotype field as equally important. Completeness should be defined for the primary analysis requirements; missing an essential outcome is not equivalent to missing an unused field.
- Creating a priority after seeing the pair’s metrics. A post hoc rule can favor a desired sample. State priorities in the analysis plan and document exceptions.
- Assuming the retained record is the “better person.” The decision is specific to an analysis and its requirements. It does not establish that the removed participant’s data are invalid for every purpose.
Check Your Understanding
For each question, identify the decision rule and then compare your reasoning with the solutions.
- Two samples are assessed at 80,000 eligible variants. Their call rates are \(0.990\) and \(0.985\), with equal phenotype completeness and no design distinction. How many fewer missing calls does the higher-call-rate sample have, and which would you retain if both pass the threshold?
- Sample G has call rate \(0.997\) but is missing the primary outcome. Sample H has call rate \(0.986\), passes the required minimum of \(0.980\), and has the outcome observed. The primary analysis requires the outcome. Which sample is eligible for that analysis?
- A related pair contains one case and one control. The control has a slightly higher call rate, but the prespecified analysis plan prioritizes preserving cases in a sparse recruitment stratum. Both samples pass hard QC and have complete required fields. What should guide the choice?
- Why is it invalid to compare call rates if one was computed across 50,000 variants and the other across 60,000 variants selected by different filters?
Solutions. (1) The difference is \(80{,}000(0.990-0.985)=80{,}000(0.005)=400\) missing calls. With other criteria tied and both eligible, retain the sample with call rate \(0.990\). (2) H is eligible for the stated analysis; G lacks a required outcome, despite its higher call rate. (3) Apply the prespecified design priority and retain the case, assuming the case is the sample in the sparse stratum; document the choice and its effect on counts. (4) The rates have different denominators and variant eligibility rules, so their difference can reflect the evaluated sets rather than sample-level missingness. Recompute both using the same eligible variant set before comparing.