Tutorials › Statistical Genetics › Common Interpretation Errors In Structural-Variant Association

Home  /  Tutorials  /  Structural Variants and Copy Number

Structural Variants and Copy Number · Tutorial 496 of 989

Common Interpretation Errors In Structural-Variant Association

Interpret structural-variant association results carefully by checking what was counted, how copy-number direction and call uncertainty were handled, and what a regional burden signal can establish.

Advanced 11 min read

What You'll Learn

  • Distinguish an event call from a person carrying a qualifying structural variant
  • Interpret copy-number coefficients with a clearly defined baseline and direction
  • Recognize when collapsing deletions and duplications can hide opposing effects
  • State the conditional-independence assumption needed for a common genotype-uncertainty calculation
  • Separate evidence for regional association from evidence that a specific event is causal
  • Check conclusions against call quality, study design, and the scope of the tested burden

Why Interpretation Requires More Than a Small P-Value

Structural-variant (SV) association results can be difficult to interpret because the data have several layers: a sequence of observed call records, an underlying genotype or copy-number state, and an association statistic for an individual event or a group of events. Confusing any two layers can change the apparent number of carriers, the direction of an estimated effect, or the strength of the conclusion.

An association result describes evidence for a relationship under a specified analysis and its assumptions. It does not by itself establish that a particular SV is causal, that every recorded call is a true genotype, or that a detected relationship will replicate in another cohort. Careful interpretation starts by identifying what was counted, how uncertain calls were handled, and exactly which variants contributed to the test.

Definition: An event call is a record produced by an SV-calling or harmonization procedure, such as a deletion call with specified breakpoints in one sample. A carrier is an individual whose genotype meets a stated rule for carrying a qualifying event or event class. One individual can have multiple event calls, and the number of event calls need not equal the number of carriers. A regional burden is a defined summary or test that combines qualifying events in a region; its interpretation depends on the event-inclusion and weighting rules.

Event Counts Are Not Carrier Counts

Association tests often compare individuals, not call records. If the outcome is whether a person carries at least one qualifying event, the relevant variable is an individual-level indicator such as \(C_i=1\) when person \(i\) meets the carrier definition and \(C_i=0\) otherwise. Counting call records instead can count a person several times, for example when that person has multiple SVs or when a complex event has been represented by more than one call.

The carrier rule must be specified before interpreting an estimate. It may require a call to meet quality criteria, fall within a particular region or mask, and satisfy a stated event-type rule. In a region-based test, a person with several qualifying events might still be counted once as a carrier, or might contribute a summed or weighted burden. These definitions answer different questions and should not be interchanged.

Worked Example: Converting Event Records to Carrier Counts

Hypothetical data. A study compares 80 cases with 120 controls for a specified SV class. After applying the same quality and region rules to both groups, the caller has produced 17 qualifying event records in cases and 10 in controls. Review shows that the 17 case records belong to 12 distinct people, while the 10 control records belong to 8 distinct people. The association is defined using the indicator “at least one qualifying event.”

The case carrier proportion is \(12/80=0.150\), and the control carrier proportion is \(8/120=0.0667\). The corresponding sample odds ratio is calculated from the numbers of people with and without a qualifying event:

$$ \widehat{\mathrm{OR}} = \frac{12/68}{8/112} = \frac{12\times112}{68\times8} = \frac{1344}{544} \approx 2.471. $$

In this hypothetical sample, the odds of being a case among people classified as carriers are estimated to be 2.471 times the odds among noncarriers, for the stated carrier definition. This is an association estimate, not a causal effect. Substituting 17 and 10 for the carrier counts would incorrectly treat event records as distinct people and would not calculate the specified person-level odds ratio.

Copy-Number Direction Must Stay Visible

For a diploid autosomal region, copy number \(CN_i=2\) is the usual baseline. One simple signed dosage is \(D_i=CN_i-2\): a one-copy deletion has \(D_i=-1\), and a three-copy duplication has \(D_i=+1\). A model using this dosage retains direction. By contrast, an indicator for “any non-diploid event” assigns both states the value 1 and discards the distinction between loss and gain.

For example, in a linear regression of a quantitative phenotype \(Y_i\) on signed dosage and covariates \(X_i\), a coefficient \(\beta_D\) describes the modelled change in mean phenotype per one-copy increase in \(CN_i\), holding the included covariates fixed. This interpretation relies on the model’s linear dosage specification and applies to the analyzed population and region. It does not imply that all copy-number states have equal effects or that the association is causal. In regions where the expected baseline differs from two, dosage should be defined relative to the appropriate baseline rather than by blindly using \(CN_i-2\).

Worked Example: Opposing Effects Can Disappear in a Carrier Comparison

Hypothetical data. Consider an autosomal region and a quantitative phenotype measured in arbitrary units. There are 10 people with one copy, 10 with two copies, and 10 with three copies. Their group mean phenotypes are 14, 10, and 6 units, respectively. For this illustration, suppose every person in a group has that group’s stated phenotype value.

Using signed dosage \(D=CN-2\), the three dosage values are \(-1,0,+1\), and the sample mean dosage is \(\bar D=0\). The mean phenotype is \(\bar Y=(14+10+6)/3=10\) units because the groups have equal sizes. The ordinary least-squares slope is

$$ \widehat{\beta}_D = \frac{\sum_i(D_i-\bar D)(Y_i-\bar Y)} {\sum_i(D_i-\bar D)^2} = \frac{10(-1)(14-10)+10(0)(10-10)+10(1)(6-10)} {10(-1)^2+10(0)^2+10(1)^2} = \frac{-80}{20} = -4. $$

The fitted linear dosage association is a decrease of 4 phenotype units per additional copy in this constructed data set. Now collapse the one-copy and three-copy groups into a single “any non-diploid carrier” group. Its mean is \((14+6)/2=10\) units, equal to the 10-unit mean in the two-copy group. That collapsed comparison shows no difference in group means, despite the strong opposing pattern across signed dosage.

This example does not establish that real SV effects are linear. It shows why a result based on carrier status alone cannot be described as a direction-specific deletion or duplication effect. The analysis must preserve the direction and coding needed for the claim being made.

Call Uncertainty Requires an Explicit Model

An SV call is evidence about a genotype, not always a perfectly observed genotype. Let \(G_i\) denote the underlying genotype or copy-number state, \(R_i\) the observed sequencing evidence or call information, \(Y_i\) the phenotype, and \(X_i\) relevant covariates and study-design information. A calibrated genotype probability \(P(G_i=g\mid R_i,X_i)\) describes uncertainty about \(G_i\) given the observed evidence and covariates. Calibration alone does not determine the relationship between phenotype and call evidence once genotype is known.

The general law of total probability is:

$$ P(Y_i\mid R_i,X_i) = \sum_g P(Y_i\mid G_i=g,R_i,X_i) P(G_i=g\mid R_i,X_i). $$

The frequently used simplification that replaces \(P(Y_i\mid G_i=g,R_i,X_i)\) with \(P(Y_i\mid G_i=g,X_i)\) requires the conditional-independence assumption \(Y_i\perp R_i\mid G_i,X_i\). In words, after accounting for the true genotype and relevant covariates or design factors, the remaining call evidence must not carry additional information about the phenotype. That assumption can fail, for example, if call quality or sequencing batch is associated with case status. An analysis should state and justify the assumptions it uses rather than treating calibrated genotype probabilities as sufficient by themselves.

Worked Example: Why Calibrated Genotype Probabilities Do Not Guarantee Independence

Hypothetical illustration. Let the true genotype \(G\) be binary with \(P(G=1)=0.5\), and suppose \(G\) is independent of both phenotype \(Y\) and observed call information \(R\). Let \(Y\) be binary with \(P(Y=1)=0.5\), and, in this deliberately simplified example, let the observed call information equal the phenotype: \(R=Y\). The genotype probabilities conditional on the call remain calibrated in the sense that \(P(G=1\mid R=1)=0.5\).

The actual conditional probability is \(P(Y=1\mid R=1)=1\), because \(R=Y\). But the simplified expression that drops \(R\) from the outcome model gives

$$ \sum_{g\in\{0,1\}}P(Y=1\mid G=g)P(G=g\mid R=1) = (0.5)(0.5)+(0.5)(0.5) = 0.5. $$

The two quantities differ because \(Y\) and \(R\) are not conditionally independent given \(G\). In an actual association analysis, this toy setup is a warning about possible phenotype-linked batch or call processes, not a model to use. If the analysis assumes \(Y_i\perp R_i\mid G_i,X_i\), the relevant design factors must be represented in \(X_i\), and the assumption must be scientifically defensible. Otherwise, the general expression retains \(R_i\) in the outcome term; uncertainty handling does not automatically remove bias from an informative calling process.

A Regional Burden Signal Does Not Identify a Causal Event

A regional burden test asks whether a defined aggregate of qualifying events is associated with an outcome. It may combine different breakpoints, event classes, or copy-number states according to a stated rule. A positive result is evidence about that tested aggregate under the model and analysis used. It does not, without further evidence, identify which constituent event drives the signal, show that every included event has the same effect, or prove that any one event causes the outcome.

This distinction matters when events are rare, correlated, or unevenly distributed. A signal might reflect one event, several events with similar effects, a mixture of effects, or a source of bias such as population structure or differential call quality. Even if the burden test is correctly calibrated, its regional conclusion is not automatically an event-specific conclusion. Event-level follow-up requires an analysis designed to evaluate the individual event, with adequate data and appropriate control of uncertainty and confounding.

Worked Example: Interpreting a Significant Regional Burden

Hypothetical data. A region-based carrier test defines a carrier as a person with at least one qualifying deletion or duplication in a prespecified region. Among 200 cases, 30 are carriers; among 300 controls, 18 are carriers. The remaining 170 cases and 282 controls are noncarriers. The sample odds ratio is

$$ \widehat{\mathrm{OR}} = \frac{30/170}{18/282} = \frac{30\times282}{170\times18} = \frac{8460}{3060} \approx 2.765. $$

The estimated odds of case status are about 2.765 times as high for carriers as for noncarriers under this regional definition. Suppose the prespecified burden test also returns a small p-value. That result supports an association between the tested regional carrier definition and case status, subject to the test’s assumptions and the study’s design.

It does not show that one particular deletion or duplication is causal. The 30 case carriers could carry different events, and the combined result alone does not reveal which events contribute to it. Nor does the odds ratio show that every qualifying event increases risk by the same amount. The appropriate conclusion remains at the level actually tested: the defined regional burden is associated with case status in this hypothetical analysis.

A Practical Interpretation Check

Before writing a conclusion, trace the analysis from observed calls to the reported statistic. Check the unit counted (records or people), the carrier rule, the copy-number coding, the treatment of uncertain calls, and the exact burden definition. Then match the wording of the conclusion to the level of the test: an event-level result concerns that event as analyzed, while a regional burden result concerns the aggregate.

1
Identify the analysis unit.
Confirm whether the data are counted per call record, per person, per genotype, or per region-level burden.
2
Inspect the coding.
Check whether the predictor preserves deletion-versus-duplication direction, represents carrier status only, or applies weights and masks.
3
Review uncertainty and design.
Determine what call evidence and genotype probabilities represent, and whether the assumptions connect those probabilities to the outcome model appropriately.
4
Limit the conclusion to the tested claim.
Describe the association and its population or region as analyzed; do not turn a regional burden result into a causal claim about one event.

Common Interpretation Errors

  • Wrong: “There were 17 event calls, so there were 17 carriers.” Correction: Count distinct people who meet the stated carrier rule; one person can contribute multiple calls.
  • Wrong: “The carrier coefficient describes the effect of a deletion.” Correction: A binary carrier coding may combine deletions and duplications. Use a direction-preserving coding and model if the claim concerns copy-number direction.
  • Wrong: “The genotype probabilities are calibrated, so the phenotype likelihood can always average \(P(Y_i\mid G_i)\) over them.” Correction: The general total-probability expression uses \(P(Y_i\mid G_i,R_i,X_i)\). Dropping \(R_i\) requires a suitable conditional-independence assumption, with study design and ascertainment handled appropriately.
  • Wrong: “The region passed its burden test, so the lead event is causal.” Correction: A regional test supports an association for its defined aggregate; event-specific causality is a stronger claim not established by that result alone.
  • Wrong: “A small p-value rules out artifacts.” Correction: A p-value summarizes incompatibility with a specified null model; it does not certify call validity, eliminate confounding, or prove a causal mechanism.
Key takeaway: Interpret the result at the level the analysis actually measures. Separate call records from carriers, preserve copy-number direction when it matters, justify assumptions used to handle uncertain genotypes, and treat regional burden evidence as evidence about a region-level aggregate—not automatic proof that a specific event is causal.

Check Your Understanding

Use the definitions and distinctions in this tutorial to assess each situation.

  1. A cohort has 24 qualifying event calls belonging to 17 distinct individuals. For a test of “at least one qualifying event,” which count defines the number of carriers, and why?
  2. In a diploid autosomal region, how do the signed dosages \(D=CN-2\) for a one-copy deletion and a three-copy duplication differ? What information is lost by coding both as “carrier = 1”?
  3. State the general law-of-total-probability expression for \(P(Y_i\mid R_i,X_i)\) after summing over \(G_i\). What conditional-independence assumption permits replacing the outcome term with \(P(Y_i\mid G_i,X_i)\)?
  4. A burden test combines several deletions and duplications in one region and reports an association. Name one conclusion it supports and one event-specific conclusion it does not establish.
  5. Why does a small p-value not by itself show that an SV call is genuine or that an association is causal?

Previous

Next

Back to Tutorials
Clinical Biostats