Coverage-Driven Missingness in Rare-Variant Data

Why Missing Genotypes Need Their Own Diagnostic

A genotype that is not called is not the same as a reference genotype. In rare-variant analysis, confusing the two can change allele counts, carrier counts, and comparisons between groups. Sequencing coverage is one reason a genotype may be missing: when too few usable reads overlap a site, the evidence may not support a reliable call. The tutorial on depth of coverage introduced the idea of callable bases; here the question is what to do when callability differs across samples or groups.

The central diagnostic is to count missing calls against a clearly defined set of genotype opportunities. That denominator must be comparable across the samples or groups being compared. A missingness difference is a signal to investigate, not by itself proof that coverage caused the difference or that an association is false. Its importance is that results based only on called genotypes can represent a selected subset of the intended data.

The set of sites must be stated before comparing rates. For example, a panel of 100 sites supplies 100 opportunities per sample if every sample is evaluated at every site, not 100 opportunities per sample for each of several samples combined. If some sites are excluded from a sample's denominator, report why and ensure that the comparison remains meaningful. A rate over a common set of sites is usually easier to interpret than rates based on different, sample-specific sets.

Summarizing Missingness Across Samples

A sample-level missingness table can reveal unusually low call rates that an overall average would conceal. Consider a hypothetical panel with 100 prespecified site opportunities for each of three samples. The counts and rates are as follows.

Sample B has 18 missing genotypes among its own 100 opportunities, so its missingness rate is 0.180, or 18.0%. It does not have 18 missing calls out of the three samples' combined opportunities. The denominator is part of the result: a missing count without its opportunity count cannot be interpreted as a rate.

A high overall call rate does not guarantee that every site or every group is equally well observed. Missingness can cluster at particular sites, among particular samples, or in a subgroup. Useful summaries therefore include both sample-level rates and site-level rates. When many opportunities come from the same sample or site, they are not automatically independent observations; a simple test that treats every entry as independent may understate uncertainty.

Comparing Missingness Between Groups

A group comparison asks whether the proportion of missing calls differs between groups, such as cases and controls. For a simple two-group comparison at one locus, each person contributes one genotype opportunity. A two-by-two table records called and missing genotypes in each group. The Pearson chi-square test compares observed cell counts with counts expected under equal missingness proportions. Its standard approximation assumes independent individuals and sufficiently large expected cell counts. If expected counts are small, an exact test is generally preferable.

The test concerns missingness, not the genetic association itself. A statistically detectable difference indicates that callability varies with group membership in the analyzed data. It does not establish whether group status caused the difference, whether coverage is the mechanism, or what the true genotypes were among missing calls.

The same calculation should not be applied mechanically to a large set of site-by-sample entries as if every entry were independent. Entries from one person can share factors affecting coverage, and sites can differ in their tendency to be callable. For multi-site summaries, inspect sample and site patterns and use a comparison method that accounts for the sampling structure when formal inference is needed. The simple example above avoids that complication by using one opportunity per independent person.

How Differential Callability Can Distort Variant Summaries

A called-genotype summary describes the people whose genotypes passed the calling process. It describes the full target population only when the selection into the called set does not systematically favor particular genotypes, or when a suitable adjustment addresses that selection. Coverage-driven missingness can be associated with other characteristics too: for example, if group membership is associated with sequencing quality, the groups may have different sets of callable samples.

The following derivation makes the selection issue precise at one diploid locus. Let \(G\) be the true alternate-allele count for a randomly selected person, with possible values 0, 1, or 2. Let \(C=1\) if the genotype is called and \(C=0\) otherwise. Assume called genotypes are correct and the target population is represented by the random person in this model. Among called people, the expected alternate-allele frequency is the expected allele count among called people divided by two:

Since \(E(GC)=E(G)E(C)+\operatorname{Cov}(G,C)\) and \(E(C)=P(C=1)\), it follows that:

The population allele frequency is \(p=E(G)/2\). Thus, if callability is unrelated to genotype, \(\operatorname{Cov}(G,C)=0\) and the called-set frequency equals the population frequency in expectation under this model. If higher alternate-allele counts are more likely to be called, the covariance is positive and the called-set frequency is higher; if they are less likely to be called, it is lower. This is a statement about selection into the observed set, not a claim that coverage always depends directly on genotype.

Differential missingness can distort comparisons even when the underlying genotype distributions are the same in two groups. If cases and controls have different call rates, the observed case-control contrast may partly reflect which genotypes were available for comparison. Conversely, similar overall call rates do not guarantee unbiased summaries: the missing calls could still concentrate in particular sites or genotype classes. Rates are therefore a first diagnostic, not a proof that missingness is harmless.

A Practical Diagnostic Sequence

A useful audit separates the description of missingness from explanations for it. The following sequence keeps denominators and interpretations explicit.

Common Mistakes and Interpretation

• Using the wrong denominator. Dividing a sample's missing count by a total pooled across samples gives the wrong sample-level rate. Use that sample's genotype opportunities as the denominator.
• Assuming missing means reference. A missing genotype supplies no observed allele count. Recode it as reference only if an explicitly justified data rule establishes that interpretation; ordinarily it must remain missing.
• Calling a difference causal. A higher missingness rate in cases is evidence of a group difference in callability, not proof that case status caused low coverage. Other features may distinguish the groups.
• Equating a nonsignificant test with equal callability. A large p-value means the data do not provide strong evidence against the tested null under the model; it does not demonstrate equivalence or rule out site-specific missingness.
• Ignoring the level of dependence. Repeated genotype opportunities within a person are not necessarily independent. A test that treats them as independent may give misleading uncertainty.