Panel Churn and Composition Drift

A panel dataset does not usually cover the same fixed set of companies forever. Vendors add new merchants to a credit-card panel, drop stores that stop reporting, and revise their sampling methodology every so often without announcing it loudly. Panel churn is the rate at which entities enter and leave the dataset, and composition drift is the effect of that churn on what the panel, as a whole, represents at any given moment — even when no single company's own data has changed at all.

Separating churn from change

For a panel metric averaged across all covered names at time $t$:

$$\bar{X}_t = \frac{1}{N_t}\sum_{i \in \text{panel}_t} X_{i,t}$$

In words: the panel average at any date depends on both the values $X_{i,t}$ and on which set of $N_t$ names happens to be in the panel that date — and that set, $\text{panel}_t$, is itself changing underneath the average. A jump in $\bar{X}_t$ could mean the covered companies genuinely changed, or it could mean the panel added or dropped a batch of names with systematically different values.

Worked example

A card-spending panel's average year-over-year sales-growth reading jumps from 4% to 11% in a single month. Checking the raw panel size for the same month shows the merchant count rose from 18,000 to 26,000 — the vendor onboarded a new category of fast-growing e-commerce merchants. Restricting the average to only the 18,000 merchants present in both months shows growth actually held flat at 4.2%; the entire apparent acceleration was new, faster-growing names entering the panel, not existing merchants accelerating.

What this means in practice

Track panel size and composition alongside any aggregate metric derived from it, and re-run key results on a "balanced panel" — the fixed subset of names present throughout the whole period — as a check against the full, churning panel. If a signal's performance looks materially different on the balanced panel, the full-panel result was partly an artifact of composition.