Why Select Tag Variants?
A study may measure many variants in a genomic region, even though several carry highly correlated information. A tag variant is a measured variant used as a proxy for one or more other variants. Selecting tags aims to reduce the number of markers while retaining a specified level of prediction for the variants not directly measured. The choice depends on the pairwise \(r^2\) values in the population being studied and on how much prediction quality is required.
This is a selection problem, not a claim that one variant causes another or that a tag identifies a causal variant. A high \(r^2\) indicates strong correlation between the allelic measurements at a pair of loci; it does not establish a biological mechanism. The earlier material on \(r^2\) as allelic correlation provides the pairwise measure used here. The focus now is how to use a matrix of those values to choose a compact set.
For a selected set \(S\), define the best pairwise prediction value for variant \(j\) as \(q_j(S)=\max_{k\in S}r^2_{jk}\). The threshold coverage proportion is the fraction of all variants covered:
where \(I\{\cdot\}\) is 1 when the condition in braces holds and 0 otherwise. This definition includes selected variants themselves. A panel with complete coverage has \(C(S,\tau)=1\). The mean best pairwise \(r^2\), \(m^{-1}\sum_{j\in V}q_j(S)\), is another useful summary, but it is not the same as threshold coverage: a panel may cover every variant at a modest threshold while having quite different average prediction quality.
A Jointly Consistent Example Matrix
Pairwise values cannot be chosen independently without checking whether they could arise from a common set of variables. For example, the signed correlations corresponding to the squared values must be capable of forming a positive semidefinite correlation matrix. The following hypothetical values are jointly consistent: they are constructed from standardized variables at angles \(0^\circ,15^\circ,35^\circ,55^\circ,\) and \(75^\circ\). If \(U\) and \(W\) are independent, mean-zero, variance-one variables, define \(X_j=\cos(\theta_j)U+\sin(\theta_j)W\). Each \(X_j\) has variance one, and the correlation between \(X_j\) and \(X_k\) is \(\cos(\theta_j-\theta_k)\). Squaring these correlations gives the values in the table, rounded to three decimals.
| Variant pair | Pairwise \(r^2\) | Variant pair | Pairwise \(r^2\) |
|---|---|---|---|
| A–B | 0.933 | B–C | 0.883 |
| A–C | 0.671 | B–D | 0.587 |
| A–D | 0.329 | B–E | 0.250 |
| A–E | 0.067 | C–D | 0.883 |
| C–E | 0.587 | ||
| D–E | 0.883 |
The construction establishes that these are mutually compatible squared correlations, rather than just a collection of plausible-looking pairwise numbers. It is a mathematical example for learning the selection procedure; an applied analysis would calculate \(r^2\) from the study population’s variant data. A diagonal entry of 1 is understood for each variant. The table is symmetric, so, for instance, \(r^2_{AB}=r^2_{BA}\).
Worked Example: Calculating Coverage at a Chosen Threshold
Suppose the target threshold is \(\tau=0.65\), and select variant C as the only tag. For each target variant, compare its \(r^2\) with C against the threshold. The values are \(r^2_{AC}=0.671\), \(r^2_{BC}=0.883\), \(r^2_{CC}=1\), \(r^2_{DC}=0.883\), and \(r^2_{EC}=0.587\).
Variants A, B, C, and D meet or exceed 0.65; E does not. Thus four of the five variants are covered:
The selected tag set therefore covers 80.0% of this five-variant panel at the stated threshold. This does not mean that every covered variant has \(r^2=0.65\): the threshold is a minimum, and the four values range from 0.671 to 1.
Choosing a Small Set of Tags
For fixed \(\tau\), each candidate tag \(k\) has a coverage set \(G_k=\{j\in V:r^2_{jk}\geq\tau\}\). Selecting tags means choosing some of these sets so their union covers as many target variants as possible. If the requirement is complete coverage, the objective is to find the smallest \(S\) for which \(\bigcup_{k\in S}G_k=V\). This is a set-cover problem.
For a small panel, one can inspect candidate sets or enumerate combinations. With many variants, a practical strategy is the greedy rule: repeatedly add the tag that covers the largest number of currently uncovered variants. The rule is easy to apply, but it is a heuristic; it need not produce the smallest possible tag set. If tags have different costs or eligibility restrictions, those must also be specified rather than silently treated as equal.
Worked Example: Finding a Minimum Tag Set at \(r^2=0.65\)
Use the table above and set \(\tau=0.65\). Include each candidate tag itself in its coverage set. Reading across the table gives \(G_A=\{A,B,C\}\), \(G_B=\{A,B,C\}\), \(G_C=\{A,B,C,D\}\), \(G_D=\{C,D,E\}\), and \(G_E=\{D,E\}\). For example, A covers C because \(r^2_{AC}=0.671\), but not D because \(r^2_{AD}=0.329\).
No single tag covers all five variants: the largest coverage set has four members, namely \(G_C\). Selecting C leaves only E uncovered, and E is covered by tag E because its self-value is 1. Thus \(\{C,E\}\) covers the whole panel using two tags. In fact, \(\{A,D\}\) also covers all five: \(G_A\cup G_D=\{A,B,C,D,E\}\).
One tag cannot suffice, since none of the five coverage sets equals the full set \(V\). Therefore two is the minimum number of tags for complete coverage at this threshold. A greedy procedure can find a two-tag solution here: C initially covers four variants, and adding E covers the remaining one. The existence of another minimum solution, \(\{A,D\}\), also shows that a minimum tag set need not be unique.
Thresholds, Prediction Quality, and Marker Count
The threshold \(\tau\) controls what counts as adequate pairwise prediction. Increasing it makes the requirement stricter: for any fixed tag set, coverage cannot increase as \(\tau\) rises. Consequently, the minimum number of tags needed for complete coverage cannot decrease when the threshold rises. There is no universally correct threshold for all studies; it should reflect the analysis goal and be reported alongside the resulting coverage.
Worked Example: Comparing a Stricter Threshold and Mean Prediction
First, raise the threshold for the same five variants from 0.65 to 0.90. At 0.90, only the A–B pair exceeds the threshold, with \(r^2_{AB}=0.933\). The other off-diagonal values are below 0.90, including the largest values of 0.883. Thus A covers A and B, while C, D, and E each require their own tag to be covered. A minimum complete-coverage set is \(\{A,C,D,E\}\), containing four tags. Three tags cannot cover all five because the three variants C, D, and E cannot cover one another across this threshold, and at least one of A or B is also needed to cover both A and B.
At the less strict threshold of 0.65, the two-tag set \(\{C,E\}\) already has complete coverage. Its best \(r^2\) for each target, in the order A through E, is \(0.671, 0.883, 1, 0.883,\) and \(1\). The mean is
A different two-tag solution, \(\{A,D\}\), also has complete coverage at 0.65. Its best \(r^2\) values are \(1, 0.933, 0.883, 1,\) and \(0.883\), giving
Both sets use two tags and cover 100% of the variants at 0.65, but \(\{A,D\}\) has a larger mean best \(r^2\) in this example. The result illustrates why reporting only the tag count or only threshold coverage can hide differences among solutions. It does not establish that \(\{A,D\}\) is preferable for every study; other considerations, such as marker availability and the intended use of predictions, may matter.
Practical Interpretation and Common Mistakes
A tag-selection result is conditional on the variant set, the population or sample used to estimate pairwise \(r^2\), the threshold, and the selection criterion. If the population changes, allele frequencies and correlations can change, so the same tags need not provide the same coverage. In reporting an analysis, state which variants were eligible as tags, the threshold, the number selected, and the coverage achieved. If useful, also report the distribution or mean of \(q_j(S)\), rather than implying that all covered variants are predicted equally well.
Pairwise \(r^2\) is a measure of correlation-based predictability, not a promise of error-free genotype imputation. It summarizes a pair of variants and does not by itself describe uncertainty in an estimated \(r^2\), rare-allele behavior, missing data, or performance in a different population. Those limitations matter when moving from a selection exercise to an applied study.
- Mistake: “The largest pairwise values look plausible, so they can be entered independently.” Pairwise correlations must be jointly compatible with a correlation matrix. Use values calculated from data or verify that a hypothetical matrix is jointly realizable; a matrix with impossible combinations cannot represent one common set of variables.
- Mistake: “A variant is covered if it has a high value with any variant in the panel.” It must have \(r^2\geq\tau\) with at least one selected tag. A strong pair that contains no selected tag does not contribute to coverage.
- Mistake: “Complete coverage means the selected tags predict every variant equally well.” It only means every variant meets the stated minimum threshold. Report the threshold and consider the actual \(q_j(S)\) values when comparing panels.
- Mistake: “A greedy selection is necessarily the smallest selection.” Greedy selection chooses the largest immediate gain. It can be useful, but for a small panel verify the minimum by checking feasible smaller sets; for a larger panel, distinguish a heuristic solution from a proven optimum.
Check Your Understanding
Use the definitions and five-variant matrix above where needed.
- At threshold 0.65, which variants are covered by tag B, and what is its coverage proportion?
- Explain why a variant not selected as a tag can still count as covered.
- At threshold 0.90, why can none of C, D, and E be represented by a different one of those three variants?
- For a fixed tag set, what happens to threshold coverage when \(\tau\) increases? Give the reason.
- Why is complete coverage at a threshold not the same as saying that every variant has the same prediction \(r^2\)?