AI and programming classes in Barrow-in-Furness

The learner fetches two Census 2021 tables from Nomis for the 246 output areas of the Barrow-in-Furness district. Their rows add to 31,263 households, against a published district total of 31,258; small census counts are adjusted, so the two differ slightly. From one comes the share of households with exactly three people. From the other comes the share of homes that are detached. Each area is ranked from 1 to 246 on both. The rank correlation between them is 0.04, which is as good as none: knowing how many detached homes an area has tells you nothing about its three-person households.

Now imagine someone preparing data for a study who wants areas that are "interesting" on these measures. They add the two ranks and keep the top quarter, 62 areas. It sounds harmless. Inside that shortlist the correlation is -0.70.

Nothing about Barrow changed between the first row and the second. The filter did it. To get onto the shortlist an area needs a high total, so an area that is low on one measure must be very high on the other, and the shortlist fills with exactly those lopsided cases. A gentler-sounding rule, keeping any area in the top quarter on either measure, gives 114 areas and a correlation of -0.59. Even the rejected areas show -0.26, because they were selected too, by failing the test.

To check this was not chance, the program drew 2,000 random shortlists of 62. Their correlations centred on 0.04, 95% of them fell between -0.18 and 0.26, and the lowest of all was -0.31. No random draw came near -0.70. One thing a careful learner notes: three-person households and detached homes were chosen because they happened to be unrelated here, and nothing in this exercise says why.

That question cannot be delegated to a chatbot, which is part of why programming and statistics still repay a Barrow-in-Furness teenager in 2026. The longer argument is in why teenagers should still learn to code in 2026.