Why Allele Harmonization Comes First
An ancestry-specific summary statistic is meaningful only when its effect estimate is attached to the correct variant and allele. One cohort may report the change in outcome per copy of allele G, while another reports the change per copy of allele A at the same biallelic site. Those coefficients have opposite signs under otherwise identical coding. If the labels are compared without recoding, a bookkeeping difference can look like a biological difference.
Harmonization means putting summary statistics into a common representation: the same variant, the same allele orientation, and a clearly stated effect allele. This is a data-alignment step, not a procedure for making effects equal across ancestries. Genuine differences may remain after the records are correctly aligned.
What Changes When the Effect Allele Changes?
For a biallelic variant, let \(X\) be the number of copies of the chosen effect allele for an individual, so \(X\) is 0, 1, or 2. In a model with an additive per-allele effect, the coefficient \(\beta\) describes the change in the modeled outcome for one additional copy of that allele, conditional on the other modeled terms. For a quantitative outcome, \(\beta\) has outcome units per allele copy. For a logistic model, it is a change in log odds per copy.
If the effect allele is changed to the other allele, the dosage becomes \(X'=2-X\). Rewriting the same model in terms of \(X'\) reverses the coefficient's sign. The standard error is unchanged because multiplying a coefficient by \(-1\) does not change its estimated uncertainty. The effect-allele frequency changes from \(f\) to \(1-f\). The test statistic changes sign, but a two-sided p-value does not change.
For a logistic model, the odds ratio per copy is \(\exp(\beta)\). Reversing the effect allele gives \(\exp(-\beta)=1/\exp(\beta)\), the reciprocal odds ratio. Thus, a positive coefficient is not inherently a risk-increasing effect: its meaning depends on which allele is counted and on how the outcome is coded.
A Practical Harmonization Sequence
Before comparing coefficients, determine whether records refer to the same biological variant. Then orient alleles and recode estimates as needed. Keep a record of each transformation so that a final table remains auditable.
Check the genome assembly, chromosome, position, and allele representation. Do not assume that equal identifiers guarantee equal records or that different identifiers necessarily mean different variants.
Where the reported strand is known, complement each allele from the reverse strand to the forward strand: A pairs with T, and C pairs with G. Complementing labels alone does not switch which biological allele is counted.
Compare the effect and other alleles after strand alignment. If the effect allele differs from the chosen one, reverse the sign of the coefficient and replace its frequency with one minus that frequency.
Verify allele frequencies against suitable reference data and check that standard errors and two-sided p-values are consistent with the reported statistics. Exclude or flag records whose identity or orientation cannot be resolved reliably.
A strand complement and an effect-allele swap are distinct operations. For example, a reverse-strand report of T/C corresponds to forward-strand A/G. If T is the effect allele, its forward-strand counterpart is A; the effect estimate does not change merely because the labels were complemented. If the desired common effect allele is G instead, then a genuine allele swap is also needed and the coefficient sign must be reversed.
Worked Examples
Worked Example: Reversing an Effect Allele
Two hypothetical cohorts report results for the same biallelic variant, with alleles A and G already placed on the same strand. Cohort 1 reports G as the effect allele: \(\beta=0.12\), \(SE=0.04\), and G frequency \(0.30\). Cohort 2 reports A as the effect allele: \(\beta=-0.10\), \(SE=0.05\), and A frequency \(0.62\). Express both results per copy of G.
Cohort 1 is already in the requested orientation, so its coefficient, standard error, and G frequency remain \(0.12\), \(0.04\), and \(0.30\). In cohort 2, A must be swapped for G. Therefore,
As a check, the original cohort 2 statistic is \(z=-0.10/0.05=-2.00\); after reorientation it is \(z=0.10/0.05=2.00\). The two-sided p-value is unchanged (approximately \(0.0455\)). Both harmonized estimates are now positive per G copy, but that does not establish that the effects are equal. It only makes their directions comparable.
Worked Example: Resolving a Nonpalindromic Strand Difference
A hypothetical association file reports a variant as T/C on the reverse strand, with T as the effect allele, \(\beta=0.08\), \(SE=0.03\), and T frequency \(0.27\). A second file reports the same variant on the forward strand as A/G. The chosen common effect allele is forward-strand A. Harmonize the first file.
Complementing the reverse-strand alleles gives T/C \(\rightarrow\) A/G. The reported effect allele T complements to A, so the effect allele already matches the chosen allele. No effect-allele swap is required. The harmonized result is \(\beta_{\text{A}}=0.08\), \(SE=0.03\), and \(f_{\text{A}}=0.27\). The p-value is also unchanged.
If the intended common effect allele had instead been G, an additional allele swap would be required: \(\beta_{\text{G}}=-0.08\), \(SE_{\text{G}}=0.03\), and \(f_{\text{G}}=1-0.27=0.73\). This distinction prevents a strand-label conversion from being mistaken for a sign-changing recoding.
Palindromic Variants and Frequency Checks
A palindromic single-nucleotide variant has alleles A/T or C/G. Complementing either allele pair gives the same unordered pair: A/T remains A/T, and C/G remains C/G. Consequently, allele labels alone cannot reveal whether a file used the forward or reverse strand. A palindromic record may still be harmonizable if reliable metadata establishes its strand or if frequency information resolves the orientation.
Frequency checks must use an appropriate comparison. Compare a reported allele frequency with a reference panel that is relevant to the cohort's ancestry and variant representation. Do not demand equal frequencies across ancestry groups: genuine population frequency differences can occur. If the frequency is near 0.5, or if the available reference is not suitable, frequency may not distinguish the two orientations. In that case, flag or exclude the record rather than choosing an orientation by guesswork.
Worked Example: A Palindromic Variant With an Ambiguous Orientation
A hypothetical cohort reports an A/T variant with T as the effect allele, frequency \(0.48\), and \(\beta=0.06\). The strand is undocumented. A suitable reference for this cohort has A frequency \(0.51\) and T frequency \(0.49\). Can the reported effect be safely placed on the forward-strand A allele?
The allele labels cannot resolve strand because A/T is palindromic. The reported T frequency, \(0.48\), is close to the reference T frequency, \(0.49\), and also close to the reference A frequency, \(0.51\). The small differences do not reliably distinguish the two possible orientations, especially given sampling and frequency-estimation uncertainty.
It would therefore be unjustified to declare that the reported T is forward-strand T and flip the sign, or to declare that it corresponds to A and keep the sign, on these data alone. The record should be marked unresolved and excluded from analyses requiring a known common effect allele unless additional strand metadata or more decisive frequency evidence is available.
Variant Identifiers and Representation
Allele alignment cannot repair a mistaken variant match. Variant identifiers can be missing, duplicated, outdated, or associated with different alleles across data sources. Coordinates are also tied to a genome assembly: the same biological variant can have different positions in different builds. A reliable comparison therefore uses the assembly and allele pair as well as an identifier, rather than matching on a label alone.
Indels and multiallelic sites require particular care. Equivalent indel descriptions can differ in their placement in repetitive sequence; normalization commonly represents an indel consistently relative to a specified reference assembly, including left alignment where applicable. A multiallelic site must be matched to the same alternate allele, not merely the same coordinate. If a record's allele representation cannot be reconciled, it should not be treated as the same biallelic variant.
Worked Example: Checking a Build and an Indel Record
Suppose two hypothetical files label an indel with the same rs identifier. File 1 uses assembly X and reports chromosome 4, position 100, reference A, alternate AT. File 2 uses assembly Y and reports chromosome 4, position 108, reference C, alternate CA. File 1 reports \(\beta=0.20\) per copy of AT, with \(SE=0.08\). Should the records be joined and the effect compared immediately?
No. The shared identifier is not enough to establish that the records are represented at the same position and with the same reference and alternate alleles. The assemblies differ, and the indel descriptions differ. First use a documented assembly conversion and consistent variant normalization to determine whether both records represent the same biological allele change. If that process establishes a match and shows that the effect allele in file 2 is the same allele as AT, retain \(\beta=0.20\) and \(SE=0.08\) per copy of that allele. If it establishes the opposite allele as the effect allele, report \(\beta=-0.20\) and \(SE=0.08\) per copy of the chosen allele. If identity remains uncertain, do not join the records.
Why Coding Errors Matter
An uncorrected allele swap reverses an effect estimate while leaving its standard error and two-sided p-value unchanged. Thus a table can show equally strong evidence against a zero effect but apparently opposite directions simply because different alleles were counted. If an effect estimate is transformed but its allele-frequency label is not, interpretation can be wrong even when the coefficient sign is correct.
Harmonization is especially important before comparing ancestry-specific estimates. A difference in signs is not evidence of effect heterogeneity until the same variant and effect allele are confirmed in each cohort. Conversely, successful harmonization does not imply that estimates must agree: this step removes representation errors, not real differences or other sources of uncertainty.
Check Your Understanding
For each question, distinguish allele-label changes from effect-allele swaps.
- A cohort reports \(\beta=0.15\) per copy of C and another reports \(\beta=-0.12\) per copy of T at the same C/T variant. What is the second coefficient when expressed per C?
- Why does complementing reverse-strand T/C labels to forward-strand A/G not, by itself, require changing the coefficient sign?
- What happens to the effect-allele frequency and standard error when changing the counted allele at a biallelic variant?
- Why are allele labels alone insufficient to resolve the strand of an A/T variant?
- A variant identifier matches across two files, but their genome assemblies and allele representations differ. What must be checked before comparing the estimates?