Back to how I work

Data into decisions

Looking beneath the dashboard

A hypothetical brand-affinity drilldown showing why raw co-purchases can tell a different story from base-rate-adjusted affinity.

Basket analysis · Anchor brand ASynthetic data
10,000 baskets · 2,000 with APaired baskets
Everyday
1,100
Seasonal
600
Specialty
500

The next question

Everyday leads on raw co-purchases. But it is present in 50% of all baskets in this slice. Is the pairing special, or is the brand simply everywhere?

Lift = P(B | A) ÷ P(B). Above 1 means more frequent pairing than the baseline.

Start with the denominator

My work at Central Retail included product co-purchase analysis and an interactive brand-affinity view to support assortment and cross-selling discussions. This case is an illustrative extension of that type of work, not a claim about a particular retail finding.

In the synthetic example, Everyday has the most baskets paired with anchor brand A. It is also present in half of all baskets. A high raw count can therefore reflect a popular brand rather than a distinctive relationship.

Adjust for the baseline

Lift compares the chance of finding B in a basket containing A with the chance of finding B in any basket.

Lift = P(B given A) / P(B).

Specialty has fewer co-purchases, but its overall lift is 2.50. The pair occurs two and a half times as often as the brand’s baseline presence would suggest.

Then ask whether promotion is doing the work

Select Adjusted affinity, then Regular baskets. Specialty retains 3.00× affinity. Seasonal falls to 0.75× even though its overall lift was above 1.

This suggests a more specific investigation: is Seasonal’s pairing concentrated in promoted baskets, while Specialty’s association persists outside promotions?

What I would check before acting

Basket size, sample sufficiency, store and customer mix, product availability, seasonality, and repeat purchasing could all change the interpretation. Association alone does not establish incremental sales or a causal cross-sell effect.

A sensible next step would be a controlled test with a clear outcome measure. The demo is designed to show how a better question emerges from a drilldown, not to prescribe a real assortment decision.