Equal-Variance Versus Equal-Influence Block Weights

Why Two Equal-Weighting Goals Differ

When omics blocks are concatenated, a block with more features can contribute more to a distance or variance-based analysis, even if its features have the same average variance as those in a narrower block. This follows because total variation adds across features. A weighting rule therefore needs a clearly stated target: should blocks have the same average variance per feature, or should each block contribute the same total variance?

These goals are not interchangeable. If two blocks have the same average feature variance but one has four times as many features, its total variance is four times as large. Equalizing average variance preserves that difference. Equalizing total variance removes it, which means the wider block must have a lower average variance after weighting. The choice changes what “balanced” means in a combined analysis.

Assume the blocks contain the same \(n\) matched samples, so each row refers to the same sample across blocks. Let \(X_b\) be the centered data matrix for block \(b\), with \(n\) rows and \(p_b\) features. Centering means that each feature has sample mean zero. Write \(s_{bj}^2\) for the sample variance of feature \(j\) in block \(b\). In this tutorial, a block is multiplied by one nonnegative scalar \(a_b\); this scalar changes every feature in the block by the same factor.

The squared factor arises because variance is measured in squared units: multiplying a feature by \(a_b\) multiplies its variance by \(a_b^2\). These summaries describe marginal feature variances. They do not assert that features within a block are independent, nor do they measure biological importance by themselves.

Deriving the Two Weighting Rules

For equal average variance, choose a common target \(c>0\) and require every weighted block to have average feature variance \(c\). Solving \(a_b^2V_b=c\) gives \(a_b=\sqrt{c/V_b}\). A convenient reference is \(c=1\), yielding \(a_b=1/\sqrt{V_b}\), provided \(V_b>0\). The target value is a scale convention; the essential feature is that each block’s average variance is matched.

For equal total influence, where influence is defined here by total variance, choose a common target \(C>0\) and require \(a_b^2T_b=C\). Solving gives \(a_b=\sqrt{C/T_b}\). With \(C=1\), the weight is \(a_b=1/\sqrt{T_b}\). Since \(T_b=p_bV_b\), this rule depends on both the block’s average variance and its feature count.

The formulas assume centered data, finite and positive variance summaries, and the same sample set when comparing blocks. A zero-variance block has no variation to rescale: dividing by its zero summary is undefined, so it must be investigated or excluded under a stated preprocessing rule. If blocks have different missingness patterns or sample sets, their summaries may not be comparable without addressing that difference first.

What Equal Total Influence Changes

Equal total influence uses the same starting blocks but sets both weighted totals to a common value. In the preceding example, choosing \(C=100\) gives block A weight \(\sqrt{100/400}=0.5\), and block B weight \(\sqrt{100/100}=1\). Their weighted totals are both \(100\). Their weighted average variances are \(100/100=1\) for A and \(100/25=4\) for B. Equal total influence therefore does not equalize average feature variance when feature counts differ.

The target \(C\) can be chosen for convenience. Using \(C=1\) would multiply both weights in this example by an additional common factor of \(1/10\); it would preserve their relative weighting and still give equal totals. A common change of scale may affect the numerical units of a combined representation, but it does not change the ratio of the block weights.

Why Total Variance Predicts a Distance Contribution

Total variance has a direct interpretation for analyses based on squared Euclidean distances. Let \(x_{ib}\) be the centered feature vector for sample \(i\) in block \(b\), and suppose two distinct samples are independent draws from the same distribution with covariance matrix \(\Sigma_b\). Their expected squared distance in that block is twice the sum of the feature variances. The result does not require features within a sample to be independent.

This is a population argument under the stated sampling assumptions; it explains why total variance is a useful proxy for expected squared-distance contribution. It does not guarantee identical contributions for every pair of observed samples. Equalizing this proxy can be appropriate when the goal is to give blocks comparable total scale in a distance-based analysis, but that goal should not be confused with equalizing each feature’s typical variance.

Choosing a Target, Checking It, and Recognizing Failure

Equal average variance is useful when the intended unit of balance is a typical feature: after weighting, the average marginal variance per feature is the same across blocks. It still allows a wider block to contribute more total variation. Equal total influence is useful when the intended unit is the block as a whole: each block has the same summed marginal variance, so a wider block’s average feature variance is lower. Neither choice is universally preferable.

Other approaches answer different questions. Leaving blocks unweighted preserves the variation on the scales produced by earlier preprocessing, but can let a high-variance or wide block dominate. Scaling every feature individually to unit variance makes each feature’s variance equal, but a block with \(p_b\) such features then has total variance \(p_b\); this is not equal total influence. A block-level scalar preserves the relative variances among features within that block, unlike individual feature scaling. The earlier material on feature scaling and concatenated analyses motivates these distinctions; the decision here is specifically which block-level summary to equalize.

After applying weights, calculate both \(a_b^2V_b\) and \(a_b^2T_b\). The first checks the average-variance target, and the second checks the total-influence target. Also inspect each block’s contribution to representative squared distances, and check whether a small number of unusual features dominate a block summary. A block can meet its target while still having substantial within-block heterogeneity.

The calculation can fail to represent the intended comparison if blocks use different samples, if preprocessing has left major technical artifacts, or if very small variance estimates produce extreme weights. Weighting cannot repair poor measurements or make blocks equally informative. In addition, total variance sums marginal variances; it is not a measure of signal relevance, reliability, or cross-block agreement. Those properties require separate scientific or statistical considerations.

• Common mistake: “Both blocks have average variance one, so both have equal influence.” This ignores feature count: total variance is \(p_b\) times average variance. The correction is to compute the totals explicitly and use equal-total weighting if matching those totals is the actual objective.
• Common mistake: “Equal total influence means every feature has the same variance.” It means the sums of feature variances match; the averages will generally differ when feature counts differ. Inspect the per-feature variances if equal feature-level scale is the goal.