Harmonizing Alleles and Effect Scales Before Meta-Analysis

Why Harmonization Comes Before Pooling

A meta-analysis can calculate a precise weighted average and still answer the wrong question if its input estimates are not comparable. In genetic association analysis, a coefficient is tied to a particular effect allele, phenotype definition, outcome scale, and genetic model. If one cohort reports the effect of allele G and another reports the effect of allele A at the same biallelic variant, the numerical signs can differ simply because the alleles were reversed. If one cohort measures an outcome in mmol/L and another in mg/dL, the coefficient magnitudes differ because the units differ.

Harmonization is the set of checks and transformations that put estimates on a common interpretation before meta-analysis. It does not make genuinely different phenotypes or estimands equivalent, and it cannot resolve every ambiguity from summary statistics alone. The goal is to establish that every included estimate refers to the same variant, effect allele, effect direction, outcome scale, and association model.

Definition: For a biallelic variant, the effect allele is the allele whose dosage is used to define the reported association coefficient; the other allele is the alternative allele. The effect-allele frequency is the proportion of counted alleles that are effect alleles. An estimate is harmonized when its allele orientation and effect scale have been made consistent with the target definition, or when it has been excluded because that consistency cannot be established.

What Must Match

Before combining estimates, record the variant identifier and alleles, the effect allele and its frequency, the coefficient and standard error, the phenotype definition and units, and the model that produced the estimate. Allele labels alone are not always enough: genomic strand conventions can make the same physical allele appear as its complement. Variant identity should therefore be checked using a stable variant identifier and, where possible, genomic position and reference build, not inferred from a gene name or nearby marker.

For a usual additive model, the coefficient describes the expected change in the modeled outcome for each additional copy of the effect allele, conditional on the covariates in the model. The cohorts need to target sufficiently comparable quantities: the same outcome or a scientifically justified equivalent, the same genetic coding, and compatible covariate adjustment. Different adjustment sets can change the estimand, particularly when covariates are associated with both genotype and outcome. Harmonizing a sign or unit does not remove that difference.

Allele frequencies are a diagnostic, not an exact matching rule. Frequencies can differ across ancestry groups, sampling frames, and finite samples, even with correct allele alignment. Large discrepancies can signal a swapped allele, strand issue, variant mismatch, or data error. They warrant investigation rather than an automatic correction based on frequency alone.

Changing the Effect Allele Reverses the Sign

Suppose a genotype dosage \(x\) counts copies of the effect allele, so \(x\) is 0, 1, or 2. If the alternative allele is chosen as the effect allele, its dosage is \(x'=2-x\). For a linear model with outcome \(Y\),

E(Y\mid x)=a+\beta x =a+2\beta-\beta x' =(a+2\beta)+(-\beta)x'.

Thus the coefficient for the reversed allele is \(-\beta\); the standard error is unchanged because the fitted contrast has only been reversed. The new effect-allele frequency is \(1-f\), where \(f\) is the original frequency. The same sign reversal applies to the slope in a logistic regression when the counted allele is reversed; the intercept changes as well.

Worked Example: Reversing an Effect Allele

Two hypothetical cohorts analyze the same variant and the same continuous outcome in standard-deviation units. Cohort 1 reports G as the effect allele, with \(\hat\beta_1=0.080\) outcome SD per G allele, \(SE_1=0.030\), and G frequency \(0.32\). Cohort 2 reports A as the effect allele, with \(\hat\beta_2=-0.060\) outcome SD per A allele, \(SE_2=0.025\), and A frequency \(0.69\). Assume the variant is biallelic G/A and both results use the same additive model and compatible phenotype definition.

To express both estimates per G allele, cohort 1 already has the target orientation. In cohort 2, G is the other allele, so reverse the coefficient and complement the frequency:

\hat\beta_{2,\mathrm{G}}=-(-0.060)=0.060,\qquad SE_{2,\mathrm{G}}=0.025,\qquad f_{2,\mathrm{G}}=1-0.69=0.31.

The G frequencies, \(0.32\) and \(0.31\), are close, which is consistent with the proposed alignment. The two positive coefficients now both describe an increase in the outcome SD per additional G allele. This frequency agreement supports the check but does not prove that every other aspect of the analyses is comparable.

Strand Ambiguity and Frequency Checks

For DNA alleles, strand complementation maps A to T and C to G. Variants with allele pairs A/T or C/G are called strand-ambiguous (or palindromic): the complement of the pair has the same labels with the alleles exchanged. Consequently, an A/T record without reliable strand information may not reveal whether its reported A corresponds to A on the reference strand or to T on that strand. Reversing an effect allele and complementing a strand are distinct operations, and the data record must establish which operation is needed.

For non-palindromic pairs, allele labels often make strand resolution possible when a reference convention is known. For palindromic pairs, an allele frequency far from 0.5 can sometimes help: a reported frequency near 0.20 is more readily distinguished from its complement near 0.80 than a frequency near 0.50. This is only a check. Frequencies from different ancestries may legitimately differ, and frequencies close to 0.5 cannot reliably distinguish orientations.

Worked Example: An A/T Variant Near Frequency One-Half

Two hypothetical cohorts report an A/T variant. Cohort 1 reports effect allele A with frequency \(0.48\); cohort 2 reports effect allele T with frequency \(0.51\). Their coefficients are \(0.04\) and \(-0.03\), respectively, in the same outcome units. The cohorts provide no strand-aware reference information.

A frequency comparison does not settle the orientation. If cohort 2's T frequency is complemented, the corresponding A frequency is \(1-0.51=0.49\), close to cohort 1's \(0.48\). But the uncomplemented frequency, \(0.51\), is also close to \(0.48\). Because these are palindromic alleles and the frequencies are near one-half, both orientations are plausible from the supplied information.

It would therefore be unjustified to flip the \(-0.03\) coefficient solely to make its sign agree with the \(0.04\) estimate, or to choose an orientation solely from this small frequency difference. The analysts should resolve strand and allele identity using reliable reference or genotype information. If that is not possible, the safe decision is to exclude this variant from the pooled analysis and report why.

Converting Outcome Units and Effect Scales

A linear rescaling of the outcome changes both the coefficient and its standard error. If \(Y^*=a+cY\), then an estimated slope becomes \(c\hat\beta\), and its standard error becomes \(|c|SE\). The constant \(a\) shifts the outcome but does not change the slope. If \(c\) is negative, the direction reverses as a consequence of the outcome transformation, not an allele change. Both the estimate and its standard error must be converted; changing only the estimate distorts its precision and therefore its meta-analysis weight.

Some effect scales require transformations that are not simple unit conversions. For a binary outcome, a logistic-regression coefficient is a log odds ratio per effect allele, and the corresponding odds ratio is \(\exp(\hat\beta)\). An odds ratio must be converted to its logarithm before it is combined with log-odds-ratio coefficients. A reported confidence interval can be used to recover an approximate standard error when it is a conventional symmetric interval on the log scale. An odds ratio is not interchangeable with a risk ratio, and converting between them generally requires additional information, such as baseline risk.

Worked Example: Converting Outcome Units

Suppose two hypothetical cohorts estimate the additive effect of the same allele on a blood measurement. Cohort 1 reports \(0.240\) mmol/L per allele with \(SE=0.080\) mmol/L. Cohort 2 reports \(9.27\) mg/dL per allele with \(SE=3.09\) mg/dL. Use the conversion \(1\) mmol/L \(=38.67\) mg/dL, and assume both cohorts use the same phenotype definition and model.

To express cohort 2 in mmol/L, divide both its coefficient and standard error by \(38.67\):

\hat\beta_{2,\mathrm{mmol/L}}=\frac{9.27}{38.67}=0.2397\ \mathrm{mmol/L}, \qquad SE_{2,\mathrm{mmol/L}}=\frac{3.09}{38.67}=0.07991\ \mathrm{mmol/L}.

These values are close to cohort 1's \(0.240\) and \(0.080\) mmol/L per allele, as expected if the measurements and estimates are comparable. The small differences reflect rounding in the reported values. Dividing the cohort 2 coefficient but leaving its standard error in mg/dL would produce a mismatched estimate and uncertainty, so the resulting weight would be invalid.

Worked Example: Converting an Odds Ratio to a Log Effect

A hypothetical cohort reports an odds ratio of \(1.20\) per copy of the effect allele for a binary outcome, with a 95% confidence interval from \(1.05\) to \(1.37\). Another cohort reports an additive logistic-regression coefficient on the log-odds scale. Convert the first result to that scale. Assume the interval is a conventional Wald interval and the allele orientation and outcome definition already match.

The log-odds coefficient is the natural logarithm of the odds ratio:

\hat\beta=\ln(1.20)=0.1823.

For a 95% Wald interval, the endpoints on the log scale are approximately \(\hat\beta\pm1.96SE\). Hence the standard error can be recovered from the log-transformed endpoints:

SE\approx\frac{\ln(1.37)-\ln(1.05)}{2(1.96)} =\frac{0.3147-0.04879}{3.92} =0.06784.

The result is therefore approximately \(0.1823\) log-odds units per effect allele with \(SE=0.06784\). This is on the scale needed to compare with the other cohort's log-odds coefficient. The conversion would not make the estimates comparable if the cohorts used different outcome definitions or genetic models.

Why Harmonization Matters for Weighted Estimates

Once estimates have a common meaning and scale, an inverse-variance weighted fixed-effect estimate can be written as a weighted average, with weight \(w_i=1/SE_i^2\). This formula presumes that the estimates are expressed on the same scale and refer to the same target effect. The need for harmonization follows directly from the weighted average: if some coefficients count the opposite allele, their signs cancel or distort the common direction; if they use different units, the numerical weighted average has no coherent unit.

For example, with aligned estimates \(\hat\beta_1\) and \(\hat\beta_2\), the pooled estimate is \((w_1\hat\beta_1+w_2\hat\beta_2)/(w_1+w_2)\). Replacing \(\hat\beta_2\) with the coefficient for the opposite allele changes its sign, while the weight is unchanged. That operation can change the pooled estimate substantially even though the underlying association evidence has not changed. Harmonization is therefore part of defining the analysis, not a cosmetic data-cleaning step.

Key takeaway: Pool only estimates that refer to a verified common variant, the same effect allele, a comparable phenotype and model, and a common effect scale. If allele orientation or scale cannot be resolved from reliable information, do not conceal the ambiguity by choosing the value that best fits the other cohorts.

Diagnostics, Alternatives, and Limits

A practical audit records the original and harmonized alleles, frequencies, coefficients, standard errors, transformations, and any exclusion reason. Check whether allele frequencies agree plausibly after the documented orientation change; inspect coefficient signs against the stated allele; verify units from the phenotype documentation; and confirm whether reported effects are raw, standardized, logarithmic, or exponentiated. Unusually large frequency discrepancies, sign patterns inconsistent with allele coding, or implausible effect magnitudes should trigger investigation before meta-analysis.

When ambiguity remains, reasonable alternatives include consulting strand-aware genotype data or a suitable reference, re-estimating effects from harmonized individual-level data, or excluding unresolved variants. Re-estimation can be preferable when cohorts used materially different phenotype definitions or genetic models, but it requires access to compatible data and a prespecified analysis. Simply converting units is not a remedy for differences in phenotype meaning, population composition, covariate adjustment, or association model.

Harmonization also does not establish that effects are homogeneous across cohorts, that the cohorts are independent, or that the variant is causal. Those are separate questions. In particular, overlapping participants can induce dependence between estimates; harmonizing their alleles does not remove that covariance. Nor should a mismatch be “fixed” by reversing an estimate because its direction disagrees with the majority: a genuine difference in association is possible, and arbitrary sign changes can manufacture agreement.

Common Mistake

A common mistake is to assume that matching variant IDs makes the coefficients directly comparable, then average them even when the effect alleles or units differ. The reasoning is wrong because the ID identifies a location or variant, not the direction or scale of the reported coefficient. First establish allele orientation, phenotype and model compatibility, and units; then convert both the estimate and its uncertainty as needed. If those details cannot be established, document the unresolved issue and do not pool that estimate.

Check Your Understanding

Answer each question, then compare your reasoning with the solutions below.

A cohort reports an effect of \(-0.12\) per A allele, with standard error \(0.04\) and A frequency \(0.27\). Express the coefficient and frequency per alternative allele.
Why is an A/T variant with effect-allele frequency \(0.50\) particularly difficult to orient from frequency alone?
A coefficient and standard error are reported in mg/dL, but the target scale is mmol/L with \(1\) mmol/L \(=38.67\) mg/dL. What must be converted?
A logistic model reports an odds ratio of \(0.90\) per effect allele. What coefficient scale is appropriate for pooling with log-odds estimates?
Two cohorts have well-aligned alleles but different phenotype definitions. Does allele harmonization make their estimates comparable? Explain.

Solutions. (1) The alternative-allele coefficient is \(-(-0.12)=0.12\) per allele, its standard error remains \(0.04\), and its frequency is \(1-0.27=0.73\). (2) For a palindromic A/T pair, strand complementation exchanges the labels, and a frequency of \(0.50\) is unchanged by taking its complement; it gives no useful orientation evidence. (3) Divide both the coefficient and its standard error by \(38.67\), so both are expressed in mmol/L. (4) Use the log odds ratio, \(\ln(0.90)\), with a standard error on the log-odds scale; the odds ratio itself is not the coefficient scale. (5) No. Allele alignment resolves orientation only; different phenotype definitions may represent different estimands and require scientific justification or separate analysis before pooling.