Why Allele Labels Matter
A SNP association coefficient has a direction only after the counted allele is specified. A positive coefficient for one allele can describe exactly the same fitted association as a negative coefficient for the other allele. This is a change in coding, not evidence that the underlying biological relationship changed. The practical challenge is to distinguish that legitimate sign reversal from a reporting or harmonization error.
This distinction matters when checking a result against genotype data, comparing reports, or interpreting an effect estimate and its confidence interval. A strand change can alter the letters used for alleles without changing which biological allele is being counted. An effect-allele swap changes the counted allele and, for a diploid additive dosage, reverses the coefficient sign. Allele labels and frequencies must therefore be checked together.
Dosage Recoding and the Sign Change
For a quantitative trait, consider the additive association model \(Y_i=\beta_0+\beta G_{B,i}+\boldsymbol{\gamma}^{T}\mathbf{Z}_i+\varepsilon_i\), where \(Y_i\) is the trait, \(\mathbf{Z}_i\) is the vector of covariates, and \(\beta\) is the per-copy association with allele \(B\). For a binary trait, a corresponding logistic model uses \(\log\{P(Y_i=1\mid G_{B,i},\mathbf{Z}_i)/[1-P(Y_i=1\mid G_{B,i},\mathbf{Z}_i)]\}=\beta_0+\beta G_{B,i}+\boldsymbol{\gamma}^{T}\mathbf{Z}_i\). In either model, the interpretation presumes that the stated dosage is actually the allele counted.
These models do not make a coefficient’s sign an intrinsic property of the SNP. If the outcome and covariate definitions are held fixed but the counted allele changes, the predictor changes. The following recoding shows exactly what happens.
Substituting \(G_B=2-G_A\) into the quantitative-trait model gives:
Thus, with the other model terms unchanged, recoding from \(B\) dosage to \(A\) dosage changes the slope from \(\beta\) to \(-\beta\) and changes the intercept to \(\beta_0+2\beta\). The same algebra applies to the linear predictor in logistic regression. Its slope changes to \(-\beta\); consequently, the per-copy odds ratio changes from \(\exp(\beta)\) to \(\exp(-\beta)=1/\exp(\beta)\). The fitted values do not change.
The standard error of the slope is unchanged by this exact recoding, since the dosage is multiplied by \(-1\) and shifted by a constant. The test statistic changes sign, but a two-sided test has the same p-value. A confidence interval \([L,U]\) for \(\beta\) becomes \([-U,-L]\) for the opposite dosage. This is a mathematical property of changing the predictor’s coding, not a reason to choose whichever direction is more convenient.
Worked Example: Reversing the Counted Allele
Worked Example: Reversing the Counted Allele
Suppose a hypothetical quantitative-trait analysis counts allele \(G\). Its fitted coefficient is \(0.30\) trait units per \(G\) allele, with a 95% confidence interval from \(0.05\) to \(0.55\) trait units per allele. The fitted model includes the same covariates in both codings. What are the corresponding results when counting \(A\), the other allele?
For every diploid genotype, \(G_A=2-G_G\). Therefore:
Reverse the interval endpoints and change their signs: \([-U,-L]=[-0.55,-0.05]\). The coefficient is now negative because an additional \(A\) copy corresponds to one fewer \(G\) copy. If the two-sided test was for a zero slope, its p-value is unchanged by this recoding. Neither result says that one allele causes the trait to increase or decrease.
Strand Changes Are Not Automatically Effect-Allele Swaps
A strand-complement operation replaces each nucleotide label with its complement: \(A\) with \(T\), and \(C\) with \(G\). By itself, that relabeling need not change the counted biological allele. For example, \(C\) on one strand corresponds to \(G\) on the other. If a report changes the label \(C\) to \(G\) but still counts that same allele, the coefficient should retain its sign. If instead the report counts the other allele, the coefficient should reverse sign.
A useful comparison therefore names both the alleles and which one is counted. The label “A1” alone is not enough: different files or software may use it for different alleles. Likewise, an allele pair written in a different order does not, by itself, tell you whether a coefficient should change sign.
Worked Example: Complementing Strand Labels
In a hypothetical association report, a SNP is labelled \(A/C\), with \(C\) as the effect allele, a coefficient of \(0.12\) trait units per copy, and an effect-allele frequency of \(0.27\). A second file labels the same SNP \(T/G\), with \(G\) as the effect allele. Assume the files refer to the same variant and that the second file’s frequency is \(0.27\).
The complement of \(C\) is \(G\), so the effect alleles correspond. Both results count the same allele, despite the different strand labels. The second coefficient should therefore be \(0.12\) trait units per copy, not \(-0.12\). The matching frequencies are consistent with this alignment, though frequency agreement alone does not prove that the files represent the same variant or analysis.
If the second file instead counted \(T\), that allele would correspond to \(A\) in the first label scheme—the allele opposite to \(C\). Provided the SNP and model are otherwise the same, the corresponding coefficient would be \(-0.12\) trait units per copy. Its effect-allele frequency would be \(1-0.27=0.73\) in the same sample.
Checking Genotype Counts and Allele Frequencies
For called diploid genotypes at a biallelic SNP, the frequency of allele \(B\) can be calculated directly from genotype counts. If \(n_{AA}\), \(n_{AB}\), and \(n_{BB}\) are the counts among \(n=n_{AA}+n_{AB}+n_{BB}\) individuals, then the total number of \(B\) copies is \(n_{AB}+2n_{BB}\). Dividing by the total number of allele copies, \(2n\), gives the sample frequency. The other allele’s frequency is one minus this value. This calculation assumes the genotypes are correctly called and the counts refer to the sample and individuals relevant to the reported frequency.
Allele frequency is a check on the direction of a dosage, not a substitute for allele metadata. If one report says the effect allele has frequency \(0.24\), a harmonized report for the same sample would ordinarily have frequency \(0.24\) for the same counted allele and \(0.76\) for its alternative. Differences in samples, ancestry, missingness, or frequency conventions can prevent exact agreement. In case-control data, for example, a frequency computed across all participants need not match one computed in controls only.
Worked Example: Verifying a Dosage From Genotype Counts
In a hypothetical sample of \(100\) individuals, the genotype counts are \(AA=60\), \(AC=32\), and \(CC=8\). Calculate the frequency of \(C\) and check which frequency should accompany a dosage counting \(A\).
There are \(32+2(8)=48\) copies of \(C\) among \(2(100)=200\) allele copies. Thus:
A dosage counting \(A\) should have an effect-allele frequency of \(0.76\) in this sample, while a dosage counting \(C\) should have a frequency of \(0.24\). If a report says it counts \(A\) but gives \(0.24\), the label, frequency definition, or sample needs checking before interpreting the coefficient. The counts alone do not identify which allele a software field calls “effect”; that information must be checked in the data or report metadata.
A Practical Verification Sequence
When a reported coefficient looks reversed, check the encoding and alignment systematically before deciding that the association differs. The checks below verify a single variant’s labels and dosage; they do not establish that two analyses used the same sample, phenotype scale, covariates, or model.
Confirm the variant identifier and the two alleles. Check genome build and variant position where available; mismatched variant records cannot be fixed by changing a sign.
Find the field that explicitly defines the effect allele or dosage allele. Do not infer it from allele order, a field name such as A1, or the coefficient sign.
For an ordinary diploid additive dosage, values should run from 0 to 2 copies of the named allele. Compute the allele frequency from called genotype counts when available and compare it with the reported frequency for the same sample.
Apply complements only when the strand convention requires them. Determine whether the resulting counted allele is the same allele or its alternative; use the frequency as a consistency check, not as the sole alignment rule.
If the counted allele is opposite, reverse the coefficient and confidence-interval endpoints. If the counted allele is the same, do not reverse the sign just because the strand letters differ.
Palindromic SNPs and Diagnostic Limits
Some allele pairs are palindromic: \(A/T\) and \(C/G\). Their complementary allele pairs have the same labels in reverse order, so strand orientation cannot be determined from the pair alone. A frequency far from \(0.5\) can sometimes help distinguish which allele is being counted, but only if the compared frequencies come from suitably comparable samples and use the same convention. When the frequency is near \(0.5\), or sample composition differs, frequency provides little direction.
Other warning signs include a dosage frequency near the complement of the reported frequency, an effect-allele label inconsistent with genotype counts, or coefficients with opposite signs despite a claimed identical counted allele and model. Investigate before interpreting. A sign difference can be explained by a real dosage swap, but may also indicate a variant mismatch, an incorrect strand transformation, a reversed frequency convention, or a data-processing error.
| Comparison | Expected coefficient relation | Frequency check |
|---|---|---|
| Same biological effect allele, different strand labels | Same sign and scale | Same counted-allele frequency in the same sample |
| Opposite allele counted at the same biallelic SNP | Opposite sign for an additive dosage | Frequencies sum to 1 in the same sample |
| Palindromic pair with uncertain strand or allele order | Cannot determine from allele letters alone | May help if comparable and not near 0.5, but is not definitive |
Common Mistake and Limitations
A common mistake is to see a negative coefficient in one report and a positive coefficient in another and conclude that the SNP has opposite effects. That conclusion is not justified until the counted alleles are aligned. If one coefficient is per copy of \(A\) and the other is per copy of \(B\), then for a biallelic diploid additive model they should have opposite signs even when the underlying fitted association is identical.
Conversely, it is also a mistake to flip a coefficient merely because allele letters differ after a strand change. A strand complement may preserve the biological allele being counted. Confirm the allele correspondence first. These sign transformations presume the same variant, trait scale, sample definition, covariates, and additive dosage model. They do not reconcile differences in those features, and a matching frequency does not prove that the association analyses are otherwise equivalent.
Check Your Understanding
Answer each question, then compare with the short solutions.
- A dosage counts \(T\), with coefficient \(0.08\) trait units per copy. What is the coefficient for dosage counting the other allele at the same diploid SNP?
- A report changes the label \(C\) to \(G\) because of a strand complement and still counts the corresponding allele. Should the coefficient sign change?
- In \(50\) individuals, genotype counts are \(AA=20\), \(AB=20\), and \(BB=10\). What is the frequency of \(B\)?
- One report gives an effect-allele frequency of \(0.18\), and another gives \(0.82\). What might explain this, and what else should be checked?
- Why can allele letters alone fail to resolve strand orientation for an \(A/T\) SNP?
Solutions. (1) The coefficient is \(-0.08\) trait units per copy of the other allele, by \(G_{\text{other}}=2-G_T\). (2) No. A complement changes the label, but the same biological allele is still counted. (3) There are \(20+2(10)=40\) copies of \(B\) among \(100\) total allele copies, so \(\hat{p}_B=40/100=0.40\). (4) The reports may count opposite alleles, whose frequencies sum to 1 in the same sample; check variant identity, counted-allele definitions, sample, and frequency convention. (5) Complementing \(A/T\) produces the same pair in reverse order, so the labels do not identify which strand was used.