Suppression, minimum cell sizes, and small numbers
When a figure is computed from very few stores, two problems arise at once. It is statistically unreliable, because a single store dominates it. And it can be identifying, because a figure from three stores in one area effectively reports on those stores.
Suppression is the standard response, and encountering it in a report is a sign of a provider taking both problems seriously.
Why are small cells unreliable?
Because the estimate is dominated by individual variation. A market estimate built from four stores moves when one of them has an unusual week, and the movement looks like a market signal.
Wide uncertainty is the honest description. Publishing a point estimate from a handful of stores without that uncertainty implies a precision that does not exist.
Why are small cells identifying?
Because aggregation only protects when there is something to aggregate. A category total for a postcode with two stores is close to a statement about each of them, and both the stores and their competitors can work out what it says.
Can aggregation ever fail to protect? Yes, when the cells get small, which is exactly why minimum thresholds for publication exist.
What thresholds are used?
They vary by provider and by context, and the principle is consistent: a minimum number of contributing units, sometimes with an additional rule about concentration so that no single unit dominates a cell.
The concentration rule matters. Ten stores where one accounts for most of the volume is not meaningfully protected by the count alone.
What should a buyer do with a suppressed cell?
Aggregate upward. Combine regions, combine weeks, or combine categories until the cell is large enough to publish, and accept the coarser answer.
What should not happen is pressing a provider to release a suppressed figure. The suppression exists for reasons that protect the stores supplying the data, and a provider that relaxes it on request is one to be careful with generally.
Does this affect trend analysis?
It can, when a cell is above the threshold in some periods and below it in others. A series with gaps caused by suppression must not be read as a series with gaps caused by zero sales.
Providers should distinguish suppressed from zero in the output. If yours does not, that is worth raising, because the two are opposite findings.
How does it relate to privacy generally?
It is one of the mechanical protections in a responsible data operation, alongside aggregation, access control and clear limits on what may be inferred. Store owners contribute data through the systems they run every day, and they are entitled to know how it is handled.
A data business in this space is well served by declining a question rather than by finding a technically permissible version of it.
What should be disclosed about suppression?
The thresholds, the rules, and how suppressed cells are marked. Those three let a reader interpret a gap correctly, which is otherwise impossible.
Frequently asked questions
Is suppression a sign of weak data?
The opposite, usually. It is a sign the provider knows where its data is thin and declines to publish beyond it.
Can I get suppressed figures under a confidentiality agreement?
Sometimes for reliability-based suppression, rarely for identification-based suppression, and providers differ. Ask which kind applies before asking for an exception.
How do I plan around suppression?
By defining the analysis at a level the data can support. Specifying a cut and then discovering it is suppressed wastes a cycle; asking about cell sizes first does not.
Does suppression bias a total?
It should not, if suppressed cells are still included in higher-level aggregates. Ask whether suppression is a display rule or an exclusion, because those have different consequences.
What about very small brands?
A small brand's figures are exactly the ones most likely to hit thresholds. Aggregating across time is usually the most workable response.
Is this the same as rounding?
No. Rounding reduces precision; suppression withholds. Some providers use both, and they should be distinguished in the output.
Do public statistics do this?
Yes, routinely. National statistical agencies suppress small cells for the same two reasons, and their published rules are a reasonable model for what a commercial provider should be able to explain.
What is the key question?
Is this cell suppressed or zero. Everything else follows from the answer.
### How should a buyer design around thresholds?
By specifying the analysis at a level the data supports, which means asking about typical cell sizes before defining the cut. A brand that specifies a region-by-format-by-week analysis and then discovers most cells are suppressed has lost a cycle.
Aggregating time is usually the cheapest way to build a cell up, because it preserves the geographic or format detail that the question was actually about.
### How should a report mark suppression?
With a distinct symbol and a footnote that explains it, never with a blank or a zero. A blank invites the reader to assume nothing happened, which is the opposite of what suppression means.
The same applies to your own internal reporting when you aggregate purchased data. Suppression that is preserved through one report and lost in the next produces a chart that quietly asserts zeros. Public statistical agencies such as the Census Bureau publish their disclosure rules as a model.
Does suppression apply to shares as well as totals?
It should, because a share computed from a suppressed cell is as unreliable and as identifying as the total was. A provider that suppresses totals and publishes shares from the same cells has closed one door and left another open.
Working with retail data? See NRS Insights data solutions or contact NRS Insights.