Imputation Quality Across Allele Frequencies

Why Examine Quality by Allele Frequency?

Imputation quality is not necessarily uniform across variants. A single summary across all imputed variants can conceal a pattern in which common variants have high quality but rare variants have low quality. Conversely, low quality in every frequency range suggests a broader concern than a decline confined to the rarest variants. Examining quality within allele-frequency strata makes these patterns visible.

This is a descriptive diagnostic: it helps locate where quality is weak, but it does not by itself identify the cause, validate genotypes against truth, or determine which variants should be retained. The introductory material on INFO scores establishes that an INFO score summarizes imputation uncertainty, with details depending on the software and metric. Here, the focus is how to compare those scores across allele frequencies.

Definition: If an allele has frequency \(p\), its minor-allele frequency (MAF) is \(f=\min(p,1-p)\), so \(0\leq f\leq 0.5\). A frequency-stratified quality summary groups variants into stated MAF ranges and summarizes an imputation-quality metric separately within each range. For each stratum, report at least the number of variants and a measure of typical quality, such as the median INFO score.

Constructing a Frequency-Stratified Summary

Choose frequency ranges that suit the question and the available data. For example, one might use \(f<0.01\), \(0.01\leq f<0.05\), and \(f\geq0.05\). The cut points are descriptive choices, not universal biological boundaries. State them explicitly, use mutually exclusive ranges, and ensure each eligible variant is assigned to one range. Report how MAF was estimated: for example, from the study-sample dosages or from a specified reference panel. Different sources can yield different strata, particularly when allele frequencies vary across populations.

Within each stratum, useful summaries include the median INFO score, lower and upper quantiles, and the number of variants. The median describes a typical variant and is less sensitive than the mean to a small number of extreme values. Quantiles show whether low quality is widespread or concentrated in a tail. Counts matter because a summary based on few variants is less stable and represents a narrower set of variants than a summary based on many.

Compare strata using the same quality metric, computed on a consistent scale and in the same target sample. INFO implementations are not identical, so a difference in values is interpretable only with the metric and its definition in view. A stratum's median is not the quality of every variant in that stratum; distributions can overlap substantially.

Worked Example: A Decline Concentrated at Low Frequency

Suppose a hypothetical imputation run has the following variant-level INFO summaries. MAF is estimated from the study-sample dosages; all variants use the same INFO definition.

MAF stratum	Number of variants	Median INFO	10th percentile INFO
\(f<0.01\)	800	0.62	0.21
\(0.01\leq f<0.05\)	1,700	0.84	0.55
\(f\geq0.05\)	7,500	0.96	0.88

The median INFO rises from 0.62 in the lowest-frequency stratum to 0.84 and 0.96 in the next two strata. The 10th percentile is also lowest in the rarest stratum, so the pattern is not merely a low median paired with a uniformly high lower tail. The data indicate that quality is weaker among low-frequency variants in this hypothetical run, while common variants have substantially higher typical INFO. They do not establish why that difference occurs or show that every rare variant is poorly imputed.

The counts give important context: the first stratum contains 800 variants, not just a handful. Still, these are summaries across variants, not independent experimental replications. A narrow quantile or a large count does not alone establish that the observed difference will recur in another cohort or reference panel.

Why an Overall Summary Can Mislead

An overall mean is a weighted average of the stratum means. If stratum \(k\) contains \(n_k\) variants and has mean INFO \(\bar{I}_k\), then the mean across all \(N\) variants is

\bar{I}_{\mathrm{all}}=\sum_{k=1}^{K}\frac{n_k}{N}\bar{I}_k, \qquad N=\sum_{k=1}^{K}n_k.

Thus, a large number of common variants can dominate the overall mean and make a dataset appear to have high quality even when a smaller rare-variant stratum has low quality. The same weighting principle applies to an overall median only in the broad sense that the pooled distribution is dominated by the most numerous observations; an overall median cannot be recovered by averaging stratum medians.

Worked Example: High Overall Mean, Weak Rare-Stratum Mean

In a hypothetical set of 10,000 variants, 800 rare variants have mean INFO 0.60 and 9,200 common variants have mean INFO 0.96. Calculate the overall mean INFO and compare it with the stratum means.

The rare variants contribute \(800(0.60)=480\) INFO units, and the common variants contribute \(9{,}200(0.96)=8{,}832\). Therefore,

\bar{I}_{\mathrm{all}} =\frac{800(0.60)+9{,}200(0.96)}{10{,}000} =\frac{9{,}312}{10{,}000} =0.9312.

The overall mean is 0.9312, which is close to the common-variant mean of 0.96, even though the rare-variant mean is only 0.60. The reason is arithmetic: rare variants make up 8% of the variants, so their low scores have limited influence on this pooled mean. Reporting the overall mean alone would obscure the marked difference between strata. These are means for illustration; in practice, medians and quantiles can add robustness and distributional context.

Distinguishing Frequency-Specific and Broad Patterns

A useful first comparison is the shape of the quality profile across strata. If quality is low only in the lowest-frequency range and improves steadily with MAF, the concern is frequency-associated. If summaries are low across low-, intermediate-, and high-frequency strata, the evidence instead points to a broad loss of imputation quality. These are diagnostic descriptions, not causal conclusions: neither pattern identifies a particular pipeline component as responsible.

Worked Example: Comparing Two Hypothetical Runs

Two hypothetical runs are evaluated with the same INFO metric, target samples, and MAF strata. Each stratum contains enough variants for the displayed medians to be meaningful descriptively.

MAF stratum	Run A median INFO	Run B median INFO
\(f<0.01\)	0.65	0.69
\(0.01\leq f<0.05\)	0.88	0.70
\(f\geq0.05\)	0.97	0.72

For Run A, the rarest stratum has a lower median than the other two, while the common-variant median is 0.97. This is a profile with a pronounced low-frequency weakness, not uniformly low scores. For Run B, the medians are 0.69, 0.70, and 0.72. Because all three are low relative to Run A's common-variant median, the profile suggests a broader quality concern. The small increase with frequency in Run B does not erase the fact that its common-variant median is also low.

The comparison is informative only because the metric, target sample, and strata are held constant. If those differ, changes in medians could reflect the comparison setup rather than a difference in imputation performance.

Assumptions, Checks, and Failure Modes

A frequency-stratified comparison is most interpretable when the following conditions are checked:

Comparable quality values: INFO scores come from the same definition and software settings, or any differences are explicitly accounted for.
Consistent MAF source: MAF is estimated in a stated and consistent way. Using different frequency sources for different variants or runs can move variants between strata.
Clear variant eligibility: The set of variants included in each summary is defined. Report how many lack a usable MAF or quality score instead of silently omitting them.
Adequate stratum size: Very small strata can produce unstable summaries. Show counts and, when useful, quantiles or the distribution rather than relying on a single number.
Relevant population context: Allele frequencies and reference-panel representation can differ across populations. A profile averaged over a mixed sample may conceal differences between subgroups.

A profile that changes sharply at one cut point should prompt a check of the frequency estimates, bin boundaries, and variant counts before interpreting the change as a biological or technical pattern. Repeating the summary with reasonable alternative cut points can show whether the conclusion is robust or depends on a particular binning choice. Avoid interpreting a smoother-looking profile as stronger evidence if it was created by merging distinct frequency ranges.

Stratified summaries are not a substitute for examining variant-level results. A median can conceal a long low-quality tail, while a lower quantile can be sensitive to the number of variants. Nor does a low INFO score prove a specific cause, such as a reference-panel mismatch. The summaries locate patterns for further investigation; they do not validate the imputed genotypes against truth.

A Short Justification for Reporting Counts and Distributions

The weighted-mean formula follows directly by grouping the sum of all variant scores. Let \(I_{jk}\) be the INFO score for variant \(j\) in stratum \(k\), and let \(n_k\) be that stratum's count. The overall sum is the sum of the within-stratum sums:

\sum_{k=1}^{K}\sum_{j=1}^{n_k} I_{jk} =\sum_{k=1}^{K} n_k\bar{I}_k.

Dividing both sides by \(N=\sum_k n_k\) gives \(\bar{I}_{\mathrm{all}}=\sum_k(n_k/N)\bar{I}_k\). This makes explicit why a heavily populated stratum has greater influence on the pooled mean. Counts reveal those weights; quantiles add information that the mean, or the median alone, cannot show about within-stratum spread.

Common Mistakes

Using only a genome-wide average. A high overall value can coexist with low quality among rare variants. Show stratum-specific summaries and counts.
Treating a stratum median as a verdict on each variant. It describes the middle of that stratum's score distribution, not every variant. Inspect quantiles or variant-level results when the tail matters.
Calling a frequency association a cause. A decline with lower MAF describes a pattern; it does not establish why the pattern occurs.
Comparing non-equivalent summaries. Differences in INFO definitions, samples, or frequency sources can make a contrast misleading. Align the comparison or state the differences.
Ignoring sparse bins. A striking median based on very few variants may be unstable. Report the count and qualify the interpretation.

Key takeaway: Report imputation quality by explicit MAF strata, alongside variant counts and distributional summaries. A decline limited to low-frequency strata differs diagnostically from uniformly low quality, but neither pattern alone identifies a cause or validates individual variants.

Check Your Understanding

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

A dataset has 500 variants with mean INFO 0.55 in the rare stratum and 4,500 variants with mean INFO 0.95 in the common stratum. Calculate the overall mean.
Median INFO is 0.58, 0.86, and 0.95 across increasing MAF strata. Describe the pattern, and state one conclusion that cannot be drawn from it.
Why should a frequency-stratified summary report counts as well as medians?
Two runs have different median INFO values, but one uses study-sample MAF and the other uses reference-panel MAF. What limits a direct comparison?

Solutions. (1) The total is \(500(0.55)+4{,}500(0.95)=275+4{,}275=4{,}550\); dividing by 5,000 gives an overall mean of 0.910. (2) Quality is lower in the rarest stratum and increases with MAF, a frequency-associated pattern. It does not establish why the pattern occurs or show that every rare variant has low INFO. (3) Counts show how many variants underlie each summary and how strongly strata contribute to a pooled mean; a small stratum may give an unstable summary. (4) The MAF sources can assign variants differently to strata, so the summaries may describe different variant sets. A direct comparison is limited unless the frequency source and other comparison conditions are aligned or the difference is explicitly handled.