Correlation vs. causation in retail data

Two lines moving together on a chart prove only that they moved together. That's the heart of correlation vs causation in retail data. Correlation means two measures tend to rise and fall at the same time. Causation means one actually drives the other. Scan data shows the first easily and the second only with careful design.

The distinction sounds academic until a budget depends on it. A brand that credits a display for a sales rise, when the rise came from something else, will spend more on displays and wonder why nothing happens.

Why does correlation show up so often in retail data?

Because many retail measures share the same drivers. Season, the calendar, local events and weather push many categories in the same direction at once. Cold drinks and bagged ice rise together in summer. Neither causes the other. Both respond to the season.

Distribution is another common shared driver. When an item gains stores, its sales rise, and so do sales of items merchandised next to it, and so does the number of promotions it appears in. Line those up on a chart and everything correlates with everything.

There's also chance. Look at enough pairs of measures and some will move together purely by coincidence. The more comparisons you run, the more of these accidental matches you'll find.

What are the usual confounders?

A confounder is a third factor that moves both measures, creating a link between them that isn't direct. In retail data, a few come up constantly:

1. Seasonality, which lifts or lowers whole groups of categories together.
2. Changes in store count, which move totals for everything in the panel.
3. Price changes, which often coincide with promotions, displays and advertising.
4. Assortment changes, where a new item arrives at the same time as a new shelf layout.

When you see two measures moving together, ask which of these could explain both. If one could, the apparent relationship may disappear once you account for it.

How do you get closer to causation?

By comparing like with like. The strongest approach is a controlled test: apply a change to one group of stores and not to a similar group, then compare the difference. This is the logic of A/B testing, adapted to physical stores.

Suppose a brand puts a new display in 20 stores and leaves 20 matched stores alone. If both groups had similar sales before, and the display stores pull ahead afterward while the others hold steady, the display is a reasonable explanation. The season, the calendar and the price affected both groups equally.

When a test isn't possible, look for natural comparisons: stores that changed and stores that didn't, over the same period. Before-and-after comparisons with a matched baseline are weaker than true tests but far stronger than a single trend line.

What should you say when you can't prove it?

Say exactly that. "Sales rose in the weeks the display ran" is accurate. "The display raised sales" is a claim that needs a comparison behind it. Precise language protects the analysis and the person presenting it.

It also helps to name the alternatives you checked. "Stores without the display didn't show the same rise" is a much stronger statement than silence. The article on correlation and causation covers the logic in more depth. For aggregated category trends, the NRS Insights monthly reports describe what changed, and the news page lists each release.