The Price of Granularity

This chapter treats reporting granularity as a decision in the cross-channel advertising measurement system, not as a simple upgrade from summary to detail. The focus is the trade-off among cell size, privacy noise, suppression, effect heterogeneity, and the strength of the claim a team wants to make. The practical result is a way to choose a partition for a stated business threshold while keeping IMA’s reconstruction evidence separate from privacy-robust causal certification.

A finer report can tell a marketer more about a campaign audience, geography, or customer-history group. It can also create smaller denominators, more suppressed cells, more simultaneous comparisons, and wider uncertainty. A coarse report can be more stable while hiding the differences that make a segment decision useful. Neither level dominates in every setting.

More detail is not automatically more usable

The first criterion is the decision itself. If the question is whether a channel deserves a budget change, a larger aggregate cell may be appropriate. If the question is whether a particular segment is safe to target, a channel-wide result is not enough. The reporting partition should be chosen after stating the claim, its outcome, and its business threshold—not before.

Coarse cells usually contain more randomized units. When reporting noise is added to arm totals and those totals are converted to means, a larger denominator reduces the effect of that noise. Larger cells are also less likely to fall below an aggregation threshold. These advantages can make an aggregate lower bound more stable and reduce the number of missing cells.

The cost is that a coarse cell averages over potentially different effects. A channel may include campaigns with different audiences or creative roles. A geographic group may combine areas with different response patterns. If one segment benefits and another does not, the cell average can be positive without making a safe claim about either subgroup. Aggregation improves statistical stability by spending interpretive resolution.

Fine cells spend that stability to gain specificity. A campaign-level or segment-level report can align more closely with the action a marketer wants to take, but each cell has fewer observations. Noise becomes more influential after division by the cell count, and a minimum reporting threshold can suppress the smallest cells entirely. Looking across many fine cells also requires simultaneous uncertainty control; a result that looks strong when selected from a large collection may not clear a corrected lower bound.

The privacy-robust framework expresses the trade-off through a segment-safe lower bound. For a reported cell g, the lower bound for its weighted average effect must be reduced by a heterogeneity penalty related to the cell’s segment diameter. In the paper’s formulation, a cell supports a safe claim for every segment inside it only when the robust lower bound minus the heterogeneity term exceeds the segment’s business threshold. The exact penalty depends on a modeling condition about how effects vary across the segment space and on an empirical proxy built from nested fine cells.