Covariance From Shared Data

Why Shared Data Determine Covariance

At successive interim looks, a cumulative score statistic at the later look reuses the observations already analyzed at the earlier look. The two statistics are therefore generally dependent, even if newly observed data bring independent information. Their covariance can be derived directly by separating the shared data from the later-only data.

The key result concerns raw cumulative statistics. It is not yet a formula for the correlation of standardized \(Z\) statistics: standardization changes the scale and is a separate calculation. The assumptions behind the covariance calculation also matter. In particular, adding individual variances is valid only when the corresponding shared contributions are independent.

For the basic shared-data result, assume the score contributions have finite second moments and that the shared block \(S_i\) is uncorrelated with the later-only block \(T_{ij}\). Independence between the two blocks is a sufficient, but stronger, assumption. No normality assumption is needed to establish the covariance identity. Independence or dependence within the shared block determines how its variance is calculated.

Deriving the Shared-Block Covariance

Covariance is bilinear, and the covariance of a random variable with itself is its variance. Applying these facts to the decomposition of \(U_j\) gives the result.

Thus, when the shared contributions are independent, the formula simplifies to \(\operatorname{Cov}(U_i,U_j)=\sum_{r\in A_i}\operatorname{Var}(X_r)\). If they are also identically distributed with variance \(v\), and there are \(n_i\) shared contributions, this becomes \(n_i v\). These are nested formulas: the first identity needs uncorrelated shared and later-only block totals; the simplification to a sum of individual variances additionally needs independence within the shared block.

If the block totals are not uncorrelated, the omitted term must be retained: \(\operatorname{Cov}(U_i,U_j)=\operatorname{Var}(S_i)+\operatorname{Cov}(S_i,T_{ij})\). For example, this can arise when an observation in the later-only block is associated with a shared observation. Merely labeling data “later-only” does not guarantee zero covariance.

Worked Covariance Calculations

When the Formula Applies—and When It Does Not

The shared-block calculation relies on the score contribution for a shared observation being the same in both statistics. It also relies on the shared and later-only block totals being uncorrelated. These conditions often hold when new independent participants enter at each look and the same prespecified score contribution is applied to each participant. If contributions are clustered, repeated, or otherwise dependent, the covariance structure must be represented at the correct unit of independence.

A useful alternative is to define one total score per independent cluster and use those cluster totals as the contributions. If clusters are independent, the variance of the shared total is then the sum of cluster-total variances. Within-cluster dependence is already incorporated in each cluster-total variance, as in the preceding example. This avoids falsely treating correlated individuals as independent units.

In likelihood-based settings, the variance of a score contribution may equal its expected information contribution under appropriate model conditions. When that identity holds, and the independent shared contributions have information contributions \(v_r\), their covariance is \(\sum_{r\in A_i}v_r\), the total information contributed by the shared data. The identity should not be assumed automatically: it depends on the statistical model and the conditions supporting the information identity.

If the score is recalculated at each look using different estimates, weights, or nuisance-parameter adjustments, the shared observations may not contribute the same \(X_r\) at both looks. In that case, the shared part of the covariance must use the look-specific contributions. For example, if the shared contribution at look \(i\) is \(X_r^{(i)}\) and at look \(j\) is \(X_r^{(j)}\), covariance terms have the form \(\operatorname{Cov}(X_r^{(i)},X_s^{(j)})\); they cannot be replaced automatically by \(\operatorname{Var}(X_r)\). This tutorial's simple formula is not a substitute for deriving those terms under the actual score definition.

Diagnostics, Alternatives, and Limitations

Before applying the shared-data formula, check the analysis sets and the dependence assumptions explicitly:

• Verify that the analysis sets are nested and identify exactly which observations or independent units are shared.
• Confirm that a shared observation has the same score contribution at both looks. If it does not, derive the cross-look covariance from the actual definitions.
• Determine whether shared contributions are independent. If not, include their pairwise covariance terms or work with independent cluster totals.
• Check whether the shared block and later-only block are uncorrelated. If that condition fails, include \(\operatorname{Cov}(S_i,T_{ij})\).
• Check units: a covariance of score statistics has squared score units, while a variance is nonnegative. A computed covariance that violates the applicable variance or covariance bounds signals an arithmetic or modeling error.

The direct shared-data derivation is the most transparent approach when the cumulative statistics have an explicit sum representation. A second approach is to calculate \(\operatorname{Cov}(U_i,U_j)\) from the full joint covariance matrix of all contributions; this is useful when there are complex dependence patterns, but it requires all relevant covariance entries. For independent increments, another approach is to write the later statistic as the earlier statistic plus an independent increment; that gives the same shared-block result under its assumptions. The more specialized independent-increments representation is not justified merely by nested sets if the block totals are dependent.

A common mistake is to write “the covariance is the number shared times the variance of one contribution” whenever data overlap. That reasoning silently assumes a common variance, independence among shared contributions, unchanged contributions across looks, and no covariance between shared and later-only totals. The correction is to start from \(\operatorname{Cov}(U_i,U_j)=\operatorname{Var}(S_i)+\operatorname{Cov}(S_i,T_{ij})\), then expand the shared-block variance with all necessary covariance terms. Covariance alone also does not determine the correlation of the standardized statistics; their marginal variances are needed for that calculation.