What a Familial QTL Result Does—and Does Not—Say
A variance-component linkage analysis asks whether relatives’ trait covariances are better explained when they share more alleles identical by descent at a genomic position. Its result is evidence about a model of covariance and a location on the genome. It is not, by itself, identification of a causal variant, a molecular mechanism, or a precise interval containing a causal change. Interpretation therefore requires looking beyond a single likelihood-ratio statistic: examine the estimated component, its boundary behavior, the calibration of the test, the profile across positions, and the stability of the result under defensible modeling choices.
For a quantitative trait, a common model for the phenotype vector \(Y\) in a family is a multivariate normal model with mean \(X\beta\) and covariance \(V(\theta)=\sigma_A^2\,2\Phi+\sigma_Q^2\,\Pi(\theta)+\sigma_e^2 I\). Here \(X\beta\) represents fixed effects, \(2\Phi\) describes the modeled polygenic covariance, and \(\Pi(\theta)\) describes expected sharing identical by descent at position \(\theta\). The nonnegative variance components \(\sigma_A^2\), \(\sigma_Q^2\), and \(\sigma_e^2\) represent polygenic, position-specific QTL, and residual variation, respectively. For multiple independent families, their likelihood contributions are combined. If IBD is uncertain, the likelihood must account for that uncertainty in the manner specified by the analysis.
This interpretation assumes that the phenotype and covariance model are suitable, variance components are constrained to be nonnegative, and the fixed-effect structure is handled consistently under the null and alternative. If families were selected using their phenotypes, the likelihood must also handle ascertainment appropriately. A misspecified mean, unmodeled family structure, or incorrect treatment of IBD uncertainty can alter the evidence; a small p-value does not repair those problems.
Fixed-Position Evidence and Genome Scans
At one prespecified position \(\theta\), the null hypothesis is \(H_0:\sigma_Q^2=0\), and the alternative permits \(\sigma_Q^2\geq 0\), while holding the position fixed. Write \(\ell\) for the log likelihood. The fixed-position likelihood-ratio statistic is
The ellipses indicate the other parameters optimized under the corresponding model. The alternative likelihood in this statistic is maximized over the variance component and nuisance parameters, but not over genomic position. Because the null model is nested within the alternative and the alternative includes \(\sigma_Q^2=0\), the maximized alternative log likelihood cannot be lower than the null log likelihood. Thus \(T(\theta)\geq 0\).
A scan asks a different question. If \(\Theta\) is the set of positions examined, its statistic is the largest fixed-position statistic:
The position attaining this maximum is the estimated peak location, not necessarily a well-localized QTL position. Searching over positions changes the null distribution: under \(H_0\), the location is not identified, and the maximum can be large just by chance. A fixed-position p-value therefore does not generally calibrate a genome scan. A parametric bootstrap can calibrate the scan by simulating under the null and repeating the entire scan for each simulated data set. Any likelihood-ratio calibration must match the actual test and its assumptions.
Worked Example: A Fixed-Position Test at a Boundary
Suppose a hypothetical analysis tests one prespecified position \(\theta_0\). Under the null, the maximized log likelihood is \(-120.00\). At the same fixed position, the alternative maximum is \(-117.80\), with \(\widehat{\sigma}_Q^2(\theta_0)=0.60\) trait-units squared. Calculate and interpret the likelihood-ratio statistic.
The position remains \(\theta_0\) in both likelihoods, so \(T(\theta_0)=2\{(-117.80)-(-120.00)\}=2(2.20)=4.40\). The corresponding base-10 log likelihood ratio is \(4.40/(2\ln 10)\approx 0.955\), since the log likelihood difference is \(2.20\). Thus the alternative at this position fits better, with an improvement of 2.20 natural-log-likelihood units.
For illustration only, suppose the regular fixed-position boundary approximation is appropriate: under the null, the statistic has probability one-half at zero and probability one-half distributed as \(\chi_1^2\). The approximate tail probability is then \(0.5P(\chi_1^2\geq4.40)=0.5\{2[1-\Phi(\sqrt{4.40})]\}\approx0.0180\). This is evidence against the no-QTL-component null at this prespecified position under that approximation. It is not the p-value for a genome scan. In practice, calibration should reflect the study design and model, including ascertainment and any relevant departures from the approximation.
The estimate \(0.60\) is a covariance-component estimate in trait-units squared for the fitted model at \(\theta_0\). It does not show that a causal variant at \(\theta_0\) has been found. A QTL component can reflect linked genetic effects or model inadequacy, and the analysis does not identify a molecular cause.
Boundary Estimates and Their Interpretation
Variance components cannot be negative. Consequently, a fitted QTL component may land exactly at zero. This is a boundary estimate, not an ordinary estimate with a symmetric uncertainty range extending equally in both directions. If the best constrained estimate is zero, the alternative likelihood equals the null likelihood at that position and the likelihood-ratio statistic is zero. The data and model have not gained likelihood by adding that component there.
A zero estimate is not proof that no QTL exists. Limited information, uncertain IBD sharing, small samples, weak effects, or an unfavorable position can all leave the component at the boundary. Conversely, a positive estimate is not sufficient evidence of linkage: its size alone does not indicate whether the improvement over the null is unusual under the correct null distribution. Report the estimate with its units, the test statistic, the calibration used, and the position or scan context.
Worked Example: A Zero Estimate Is Not Proof of No QTL
At a second hypothetical prespecified position, the null log likelihood is \(-85.40\). An unconstrained numerical fit would put the QTL variance component at \(-0.12\) trait-units squared, but the model requires \(\sigma_Q^2\geq0\). What is the constrained estimate and likelihood-ratio statistic?
The constrained maximum cannot use the inadmissible value \(-0.12\). Suppose the likelihood is maximized over the permitted values at the boundary, \(\widehat{\sigma}_Q^2=0\), with log likelihood \(-85.40\). Then \(T=2\{(-85.40)-(-85.40)\}=0\). The fitted model supplies no likelihood improvement from adding a QTL component at this position.
The appropriate conclusion is that this analysis at this position did not provide positive likelihood-ratio evidence for a nonzero component. It is not that the probability of a QTL is zero, nor that a QTL has been ruled out. Those stronger claims do not follow from a constrained estimate at the boundary.
Diagnosing Scan Evidence and Localization
A scan peak and a localization result are different. A peak records where the likelihood ratio was largest among the positions examined. Localization asks how sharply the data distinguish that position from nearby alternatives. If many positions have similar likelihoods, the peak can be prominent while the location remains uncertain. The peak is also affected by marker information, recombination, IBD uncertainty, and the number and structure of informative families.
Do not treat the fixed-position significance calculation at the peak as if the peak had been specified before seeing the data. For a scan, calibration must reproduce the search: each null replicate should be analyzed across the same positions, and its maximum statistic recorded. The proportion of null maxima at least as large as the observed maximum estimates the scan-level tail probability. A profile across positions is useful descriptive evidence about localization, but a formal interval requires a method with appropriate coverage for the model and search; a generic likelihood drop should not automatically be called a confidence interval.
Worked Example: Why a Peak Needs Scan Calibration
Consider a hypothetical scan at three positions. The fixed-position statistics are \(T(\theta_1)=4.40\), \(T(\theta_2)=2.10\), and \(T(\theta_3)=3.00\). The largest is \(T_{\mathrm{scan}}=4.40\) at \(\theta_1\). To calibrate the scan, investigators simulate 1,000 null data sets and repeat the same three-position scan for each. In 120 simulations, the maximum statistic is at least 4.40.
Using the common plus-one estimate, the scan-level p-value is \((120+1)/(1000+1)=121/1001\approx0.121\). The peak statistic is 4.40, but the calibrated scan-level result is not the fixed-position approximation of about 0.018 from the preceding example. The latter addresses a prespecified position under its stated approximation; the bootstrap result accounts for selecting the largest of three statistics. The evidence is therefore not as strong after accounting for the search.
Suppose, in addition, the profile has statistics 4.40, 4.10, and 3.90 at three adjacent positions. Those values are close, so the data do not sharply distinguish the peak from its neighbors. The highest point remains \(\theta_1\), but this hypothetical profile supports weak localization, not a precise causal position. A valid interval, if required, needs a localization procedure suited to the design rather than an unsupported cutoff.
Model Sensitivity and Family Heterogeneity
A linkage result is conditional on its model. Check whether the signal persists under scientifically defensible alternatives for fixed effects, covariance structure, IBD uncertainty, ascertainment correction, and treatment of incomplete phenotypes. A model change can shift the estimated QTL component or the likelihood ratio because it changes which covariance patterns are attributed to the QTL. Such comparisons are diagnostic; results from models with different likelihood definitions or different data should not be compared as if they were automatically the same formal test.
A QTL variance estimate is also not the fraction of trait variance caused by a particular variant. For a simple fitted covariance model with components \(\sigma_Q^2,\sigma_A^2,\sigma_e^2\), one may calculate the model-based component fraction \(\sigma_Q^2/(\sigma_Q^2+\sigma_A^2+\sigma_e^2)\). This fraction describes the fitted decomposition under that model and at that position. It is not a causal attribution, and uncertainty in the estimates matters. Family-specific evidence may also vary: a pooled estimate can mask differences in informativeness, family structure, or genuine effect heterogeneity. Subgroup patterns are diagnostic prompts, not proof of biological heterogeneity; a formal claim requires an explicit model and suitable calibration.
Worked Example: Sensitivity of a Component Fraction
In a hypothetical analysis at a fixed position, the fitted components are \(\widehat{\sigma}_Q^2=0.80\), \(\widehat{\sigma}_A^2=1.20\), and \(\widehat{\sigma}_e^2=2.00\), all in trait-units squared. Calculate the model-based QTL component fraction.
The fitted total is \(0.80+1.20+2.00=4.00\) trait-units squared. Thus the component fraction is \(0.80/4.00=0.20\), or 20%. The interpretation is that, under this covariance model, the estimated position-specific component is one-fifth of the sum of the three fitted components. It does not mean that a specific causal variant explains 20% of the trait, or that the component is established as nonzero.
Now suppose a defensible alternative fixed-effect specification gives estimates \(0.25\), \(1.55\), and \(2.20\) trait-units squared. Its corresponding fraction is \(0.25/(0.25+1.55+2.20)=0.25/4.00=0.0625\), or 6.25%. This change is a warning that the estimated decomposition is sensitive to modeling choices. It does not, by itself, tell which model is correct or constitute a formal test between them; that requires an appropriate comparison and attention to their assumptions.
For family heterogeneity, inspect whether the pattern is driven by a small subset of families and whether families differ in informativeness or phenotype ascertainment. For example, if a pooled peak weakens greatly when one family is omitted, report that dependence and investigate data quality and model fit rather than presenting the pooled statistic alone. Do not infer different causal mechanisms merely because subgroup estimates differ: estimates can vary through sampling error, and a heterogeneity test must be specified and calibrated for the analysis.
Common Mistakes and a Practical Reporting Standard
- Equating the component with a causal variant. The QTL variance component describes a modeled covariance contribution linked to IBD sharing, not a discovered molecular cause.
- Using a scan maximum in a fixed-position formula. At a fixed position, hold \(\theta\) fixed in both likelihoods. Maximizing over \(\theta\) defines a scan statistic and requires scan-level calibration.
- Calling a boundary estimate proof of absence. A zero constrained estimate means no improvement at that fit; it does not establish that a QTL does not exist.
- Calling the peak a precise location. A maximum identifies the best-scoring tested position. A broad profile indicates that nearby positions remain difficult to distinguish.
- Reporting only a p-value. Include the estimated component and units, test type, position or scan, calibration method, and material sensitivity or heterogeneity findings.
Check Your Understanding
Try these questions using the definitions and distinctions above.
- In a fixed-position likelihood-ratio test, why must the alternative likelihood keep the genomic position fixed?
- What does a constrained estimate \(\widehat{\sigma}_Q^2=0\) imply about the likelihood-ratio statistic at that position, and what does it not prove?
- Why can a scan-level p-value differ from a fixed-position p-value computed at the scan peak?
- A fitted QTL component is 0.50 trait-units squared, with polygenic and residual components of 1.00 and 1.50. What is the model-based component fraction, and what causal interpretation should be avoided?
- What does a broad likelihood profile around the maximum suggest about localization?