Coding and AI classes in Gosport

The learner downloads Census 2021 table TS045 from the Nomis API for all 279 output areas in Gosport. Across the whole borough, 7,420 of 35,912 households have no car or van: 20.66%. That is the truth the surveys are trying to find. Each pretend survey picks a few output areas at random and estimates the share from those alone. To say how sure it is, the program uses a technique called the bootstrap: it resamples the chosen areas 1,000 times, recalculates the share each time, and takes the middle 95% of those results as its range. One survey of 20 areas, for example, estimated 21.6% with a range of 14.9% to 29.0%. That range does contain 20.66%. The real question is how often that happens.

Every one of these ranges was labelled 95%. With only five areas, just 80.8% of them caught the truth: the method was over-confident, drawing ranges too narrow for the little evidence it had. Checking a stated confidence against reality like this is called calibration, and a well-calibrated method would score close to 95% at every size. Around 20 to 40 areas the bootstrap gets close, at 93.6% and 93.7%. At 80 areas it overshoots to 96.8%, for a reason worth understanding: 80 areas is more than a quarter of the whole borough, and the bootstrap assumes the areas were picked from an endless supply, so it overstates how much a larger survey could vary.

The ranges also shrink as the survey grows, from 17.11 points wide with five areas to 5.35 with eighty, and the typical miss of the estimate falls from 3.59 points to 0.80. But the most useful lesson is the first row. A confident-sounding range built from very little data can be wrong one time in five while still calling itself 95%.

A Gosport teenager who can test whether confidence is earned will get far more from AI, which makes learning to code in 2026 well worthwhile. The longer argument is in why teenagers should still learn to code in 2026.