Switch Errors And Phasing Uncertainty

Why Phase Errors Matter

At a heterozygous variant, a diploid individual carries two different alleles, one on each chromosome copy. The genotype records both alleles but does not, by itself, say which allele lies on which copy. Phasing assigns the alleles to the two haplotypes. If that assignment is wrong at one part of a region, the resulting haplotype pattern can be wrong even though every observed genotype is correct. This matters for imputation because nearby typed variants help identify the local haplotype context used to infer an untyped allele.

A phase assignment has an arbitrary labeling: calling a chromosome copy “haplotype 1” rather than “haplotype 2” carries no biological meaning. Swapping those labels everywhere leaves the phase unchanged in substance. The relevant error is a switch in relative phase between successive heterozygous variants, not a global relabeling.

The denominator matters. Switch errors are usually evaluated between adjacent heterozygous sites, so a region with few such pairs has a small and potentially unstable denominator. A rate also depends on how the evaluated sites and intervals are defined. It is therefore useful to report the count, denominator, and genomic span or distance covered, rather than a percentage alone.

Representing Relative Phase

For a precise description, encode the phase at heterozygous site \(j\) by \(z_j\), where \(z_j=0\) or \(1\) indicates which allele is assigned to a chosen haplotype label. The particular choice of label is arbitrary. The relative phase across adjacent sites is represented by the exclusive-or operation, written \(\oplus\):

If both site labels are swapped, both \(z\) values change, but \(r_j\) does not. Thus, \(r_j\) describes the biologically relevant relationship and is unaffected by a global relabeling. Let \(\hat r_j\) denote the estimated relative phase and \(r_j\) the true relative phase. A switch error at this interval is the event \(\hat r_j\ne r_j\).

Switch Error and Phasing Uncertainty Are Different

A switch error compares an estimate with the true phase. In most applications, the true phase is not known, so the actual switch-error rate cannot be calculated for each sample directly. Instead, a phasing method may report uncertainty, such as a posterior probability that a pair of sites has a particular relative phase. That probability describes uncertainty given the data and model; it is not a direct observation of whether the phase is correct.

A site-level phase confidence and an interval-level confidence also answer different questions. Confidence about the allele assignment at one site does not alone determine confidence in the relationship between two sites. Conversely, a local phase relationship can be clear even though the choice of which haplotype receives label 1 is arbitrary.

How Uncertain Site Assignments Relate to Switch Risk

A short derivation clarifies why adjacent phase uncertainty is not simply the uncertainty at one site. Let \(e_j=1\) if the estimated orientation at site \(j\) is reversed relative to truth, and \(e_j=0\) otherwise. Reversing the orientation at exactly one endpoint changes the relative phase; reversing both endpoints, or neither, preserves it. Therefore, a switch error occurs exactly when \(e_{j-1}\oplus e_j=1\).

If the endpoint error events are independent, and their probabilities are \(p_{j-1}\) and \(p_j\), this becomes \(p_{j-1}(1-p_j)+(1-p_{j-1})p_j\). For example, if the endpoint error probabilities are 0.10 and 0.20, the interval switch probability under independence is \(0.10(0.80)+0.90(0.20)=0.26\). Without independence, the joint probabilities are needed. This calculation is a justification for the relationship between endpoint uncertainty and relative-phase uncertainty, not a substitute for the interval probabilities a phasing method may provide directly.

How a Switch Can Affect Nearby-Variant Imputation

Imputation uses observed information and a reference of haplotypes to infer alleles at untyped variants. The key point here is limited but important: phase determines which observed alleles are treated as belonging together on a chromosome copy. A switch can assign a segment’s alleles to the wrong local haplotype pattern. If the untyped allele is associated with one pattern more often than another, using the wrong pattern can change the inferred probability for that allele. The genotype calls at the typed sites need not be wrong for this to happen.

The effect is local in the sense that a switch changes the assignment of alleles across an interval, but its consequences can extend to multiple nearby untyped variants. The size and direction of the effect depend on the pattern of linkage in the reference data, the location of the switch, the available typed markers, and the uncertainty in phase. A switch error does not guarantee an incorrect imputed genotype: different phase configurations can imply similar target probabilities, and imputation may remain uncertain rather than making a confident wrong prediction.

Diagnostics and Comparisons

When true phase is available for a validation set, switch-error counts and rates directly assess relative-phase accuracy on the evaluated intervals. When it is unavailable, posterior phase uncertainty is a useful diagnostic, but it should not be presented as an observed error rate. In either setting, assess whether the summary matches the purpose of the analysis and the amount of information in the region.

• Check the denominator. Report how many adjacent heterozygous pairs were evaluated. A rate from a few pairs is imprecise and can change sharply with one additional error.
• Locate the errors. A single switch can reverse the apparent phase of several subsequent sites until another switch occurs. A regional average can conceal such a segment.
• Consider interval length. The same number of errors can have different implications across short and long genomic spans. State whether intervals are based on adjacent sites, physical distance, or another defined scale.
• Compare to available truth carefully. Pedigree, long-read, or other benchmark phase can support an empirical error calculation, but the benchmark itself may be incomplete or uncertain. Align sites and allele coding before comparing assignments.
• Inspect uncertainty calibration where possible. If intervals assigned similar confidence are repeatedly evaluated against truth, their empirical accuracy should broadly agree with their stated confidence. A mismatch may indicate overconfidence or a validation mismatch.

Switch-error rate is not the only useful summary. Per-site phase accuracy can count incorrect site orientations, but it can overcount a single persistent flip and is sensitive to arbitrary haplotype labels unless orientation is aligned. Genotype concordance assesses whether the two alleles at a site were called correctly, but does not establish which allele lies on which chromosome copy. Posterior phase probabilities describe model uncertainty, whereas switch-error rates describe agreement with known truth. These measures are complementary rather than interchangeable.

Common Pitfall and Limitations

A second pitfall is treating high reported confidence as proof of correct phase. Confidence is conditional on the method, data, and assumptions; systematic problems or limited information can produce confident errors. Sparse heterozygous sites, long intervals, rare variants, and differences between a study sample and its reference can all make phase less reliable. The uncertainty summary may also fail to reflect dependencies between neighboring intervals.

Switch rate is an average over evaluated pairs and does not by itself identify which imputed variants are affected. Nor does it establish that a phase error caused an observed imputation error. To make that claim requires appropriate validation and attention to other sources of error. Phase uncertainty is best treated as one component of imputation reliability, interpreted alongside the local marker information and the evidence available for validation.