Imputation Quality in Cross-Cohort Meta-Analysis

Why Imputation Quality Matters in Meta-Analysis

A cross-cohort meta-analysis can combine association estimates only when each estimate carries meaningful information about the variant being studied. In one cohort, a variant may be directly genotyped; in another, its genotype may be inferred from nearby markers using a reference panel. Even when both cohorts report an estimate for the same variant, their estimates may have very different precision because the imputation is more informative in one cohort than the other.

Imputation quality is therefore relevant to both the interpretation of cohort-specific results and their contribution to a meta-analysis. It is not, however, a universal correction factor for effect estimates. The quality statistic, the association model, and the standard error all need to be understood together.

Definition: For an individual, let $G$ be the unobserved genotype count for a specified effect allele, taking values 0, 1, or 2. An imputation procedure uses observed genetic data and a reference panel to estimate genotype probabilities. The dosage is the expected allele count given those data: $$ D = P(G=1\mid \text{observed data}) + 2P(G=2\mid \text{observed data}). $$ A dosage can be fractional. A cohort-specific imputation quality metric summarizes how informative the imputed dosages are about the underlying genotypes. Its precise definition depends on the software and method used.

A reference panel supplies haplotypes or genotypes that help infer unobserved variants from observed patterns of linkage disequilibrium. Imputation tends to be more informative when the panel represents the relevant genetic variation and linkage patterns in the cohort. A panel that is a poor match can yield less accurate or less well-calibrated dosages. Panel size alone does not guarantee good imputation: representation of relevant ancestry and haplotypes also matters.

What an Imputation Quality Statistic Does—and Does Not—Say

A common kind of quality statistic, often labelled INFO or $R^2$, estimates how strongly the imputed dosages track the underlying genotype variation. Some implementations use a ratio involving dosage variance and the expected genotype variance; others summarize posterior genotype uncertainty. The labels are not fully interchangeable across software, so analysts should check the metric’s definition and the software’s guidance before comparing values.

For a common variant with effect-allele frequency $f$, genotype variance under Hardy–Weinberg equilibrium is $2f(1-f)$. A variance-ratio-style quality measure can be illustrated as the variance of the dosages divided by this expected genotype variance. The following calculation uses that particular definition; it should not be assumed to reproduce every program’s INFO statistic.

Worked Example: Interpreting a Variance-Ratio Quality Measure

Suppose a hypothetical cohort has an effect-allele frequency of $f=0.20$. Its dosage variance is $0.288$, and the quality measure being used is dosage variance divided by expected genotype variance under Hardy–Weinberg equilibrium.

First calculate the expected genotype variance: $$ 2f(1-f)=2(0.20)(0.80)=0.32. $$

Then calculate the variance ratio: $$ \text{quality ratio}=\frac{0.288}{0.32}=0.900. $$

Under this definition, the imputed dosages retain about 90% of the expected genotype variation. This describes information in the dosages; it does not mean that the estimated genetic effect should be multiplied by 0.900, nor does it establish that the effect estimate is unbiased. The interpretation is conditional on the specified quality metric and its assumptions.

A quality statistic is not a guarantee that every genotype is correct, and it is not a direct measure of association significance. Low quality usually means less information and therefore less power. If the association test is appropriately calibrated, low quality alone does not imply an inflated false-positive rate. Conversely, a high quality score does not rule out other problems, such as a poorly specified association model or an allele-coding error.

Why Lower Quality Usually Means a Larger Standard Error

Consider a quantitative phenotype and a cohort-level additive association model, $Y_i=\mu+\beta G_i+\varepsilon_i$, where $G_i$ is the true effect-allele count and the residuals are independent with mean zero and common variance $\sigma^2$. With genotypes observed, the information about $\beta$ is approximately proportional to the sum of the genotype variances, divided by $\sigma^2$. With imputed dosages, the information is instead approximately proportional to the sum of the dosage variances, divided by $\sigma^2$.

If the quality measure $I$ is a reasonable approximation to the ratio of dosage information to genotype information, then the standard error is approximately inversely proportional to the square root of that ratio. This is an information approximation, not an exact identity for every imputation method or association model.

\frac{SE_{\text{imputed}}}{SE_{\text{genotyped}}} \approx \frac{1}{\sqrt{I}}.

For example, if $I=0.36$, the approximate standard-error ratio is $1/\sqrt{0.36}=1.67$. Under the conditions behind the approximation, this resembles having about 36% as much information as with directly observed genotypes. The effect estimate itself should not automatically be rescaled by this factor or by $I$. With dosage-based analysis, the association estimate can remain useful while its uncertainty increases; effect bias depends on the data and model, not on the quality score alone.

This distinction matters in practice. A standard error calculated from an association analysis using dosages will often already reflect the available dosage variation. If so, inverse-variance meta-analysis naturally assigns that cohort less weight. Applying an additional penalty based on INFO would count the same loss of information twice. If the reported standard error does not account for dosage uncertainty—for example, because a procedure uses hard calls or an unsuitable variance calculation—the analyst should resolve that issue rather than assume the reported precision is valid.

Worked Meta-Analysis With Different Quality Levels

For independent cohorts estimating a common effect on the same allele and phenotype scale, fixed-effect inverse-variance meta-analysis weights each estimate by the inverse of its squared standard error. This requires comparable estimands and correctly calculated standard errors. The earlier material on allele harmonization and phenotype scales is important here: quality differences do not excuse combining estimates that refer to different alleles or outcomes.

Worked Example: Letting the Standard Errors Reflect Quality

Suppose three hypothetical, independent cohorts report association estimates in phenotype standard deviations per effect allele. Their estimates and standard errors come from dosage-based analyses. Cohort C has notably lower imputation quality than the others.

Cohort A: $\hat{\beta}_A=0.10$, $SE_A=0.04$, quality $0.90$.
Cohort B: $\hat{\beta}_B=0.12$, $SE_B=0.05$, quality $0.80$.
Cohort C: $\hat{\beta}_C=0.30$, $SE_C=0.10$, quality $0.25$.

The inverse-variance weights are: $$ w_A=\frac{1}{0.04^2}=625,\qquad w_B=\frac{1}{0.05^2}=400,\qquad w_C=\frac{1}{0.10^2}=100. $$

The fixed-effect estimate is: $$ \hat{\beta}_{\text{FE}} = \frac{625(0.10)+400(0.12)+100(0.30)}{625+400+100} = \frac{140.5}{1125} = 0.1249. $$

Its standard error is: $$ SE_{\text{FE}}=\sqrt{\frac{1}{1125}}=0.02981. $$

The combined estimate is 0.1249 phenotype standard deviations per effect allele, with a standard error of 0.02981 in the same units. Cohort C contributes $100/1125=0.08889$, or about 8.9%, of the total weight. Its larger standard error already limits its influence, despite its larger point estimate. If these standard errors properly reflect dosage information, multiplying Cohort C’s weight by 0.25 again would be an unjustified second penalty.

A low-quality cohort can still matter if its estimate is valid and precise enough to contribute information. Inverse-variance weighting is not a substitute for quality control, however. It can reduce the influence of a noisy estimate, but it cannot repair a biased estimate, unaccounted-for uncertainty, or a systematic imputation error.

Reference-Panel Fit and Cohort Comparability

Imputation quality can differ across cohorts because their reference panels, observed marker sets, allele frequencies, and linkage patterns differ. A variant well tagged by observed markers in one population may be poorly tagged in another. Thus, a single overall quality value for a variant can hide substantial cohort-level differences. Review quality separately for each cohort and, where relevant, for the ancestry groups represented within it.

Quality metrics can also be difficult to compare across cohorts when their allele frequencies differ. Expected genotype variance depends on allele frequency, and some quality summaries use frequency-dependent quantities. A low value can reflect limited information, but analysts should verify that it is not being interpreted without regard to the metric’s definition, frequency, or sample size.

Reference-panel mismatch may create more than a larger standard error. If the imputation procedure produces systematically inaccurate dosages, the association estimate itself can be biased or poorly calibrated. A larger standard error does not necessarily correct such bias. This is why a quality review should consider the imputation pipeline and panel fit as well as the numerical score.

Worked Example: A Prespecified Quality Exclusion

Suppose a hypothetical analysis prespecifies that variants with an INFO value below 0.30 in any contributing cohort will be excluded from the primary cross-cohort meta-analysis. This is a study-specific quality rule, not a universal threshold. In one cohort of $n=4{,}000$, a variant has INFO $I=0.18$, and the cohort reports $\hat{\beta}=0.20$ with $SE=0.12$. The reference panel is also a poor match to the cohort’s ancestry.

As a rough information comparison, the effective sample size is: $$ n_{\text{effective}}\approx nI=4{,}000(0.18)=720. $$

This approximation suggests substantially less information than the nominal sample size of 4,000. The cohort estimate is also imprecise: its approximate 95% confidence interval is $$ 0.20\pm1.96(0.12) = 0.20\pm0.2352 = (-0.0352,\ 0.4352). $$

Because INFO is below the prespecified threshold, and the panel fit raises an additional concern, the appropriate decision under this analysis plan is to exclude the result from the primary meta-analysis and report it separately as uncertain. The reason is the quality rule and imputation concern, not that the interval includes zero. If the quality rule had not been prespecified, the analyst should investigate calibration and sensitivity to inclusion rather than invent a threshold after seeing the association result.

Choosing Whether to Retain, Downweight, or Exclude

A defensible decision distinguishes uncertainty from evidence of unreliability. When cohort estimates and standard errors are well calibrated, retaining a lower-quality result may be reasonable: its larger standard error will generally reduce its inverse-variance weight. A separate quality-based weight is appropriate only when the analysis method justifies it and the original standard errors have not already incorporated the same information loss.

Exclusion can be appropriate when a prespecified quality rule is failed, the imputation pipeline is not considered reliable for the variant, or the uncertainty cannot be represented adequately by the reported standard error. If exclusion is chosen, report the rule and the affected cohort or variant. Where uncertainty is material but evidence is insufficient for a definitive exclusion, report the cohort result separately and show how conclusions change when it is omitted. These choices should be based on quality evidence, not on whether the estimate is statistically significant or agrees with the other cohorts.

Check what the quality statistic means.
Identify the software-specific definition, the relevant allele and cohort, and whether values are comparable across the analyses.

Check the association uncertainty.
Confirm that the cohort’s standard error was calculated using an appropriate dosage-based procedure and reflects the information available.

Assess panel and cohort fit.
Consider whether the reference panel, marker data, and genetic variation support reliable imputation in that cohort.

Apply a transparent analysis decision.
Retain a valid estimate with its appropriate standard error, exclude it under a justified quality rule, or report it separately and assess sensitivity.

Common Mistakes and Key Takeaways

Multiplying the effect estimate by INFO. INFO is a measure of imputation information, not a universal effect-size correction. Do not rescale an estimate unless a specific, justified model calls for it.
Penalizing a standard error twice. If the dosage-based association analysis already reflects reduced information in its standard error, applying an extra INFO weight can over-penalize the cohort.
Assuming a large standard error fixes bias. Downweighting reduces the contribution of an imprecise estimate. It cannot guarantee correction of systematic error caused by poor imputation.
Using one quality value for all cohorts. Imputation quality is cohort-specific and may differ with reference-panel fit, allele frequency, and observed markers.
Excluding results because they are not significant. Apply quality criteria independently of the association result and report exclusions transparently.

Key takeaway: Evaluate imputation quality for each cohort, interpret it using its software-specific definition, and verify that association standard errors reflect dosage uncertainty. Use those standard errors appropriately; do not apply an additional quality penalty automatically. Exclude or separately report results when quality concerns make their estimates unreliable, using transparent criteria rather than significance-based decisions.

Check Your Understanding

Try these questions before reviewing the worked examples.

What does an imputed genotype dosage represent, and why can it be fractional?
Under the information approximation, what happens to the standard error when INFO is 0.25 rather than 1.00?
Why can multiplying a cohort’s inverse-variance weight by INFO be inappropriate when its reported standard error already reflects dosage uncertainty?
Give one reason that reference-panel mismatch might affect more than the precision of an association estimate.
Why should a quality-based exclusion rule not be chosen solely because a cohort’s association result is not statistically significant?