Why does a cubic dose term give an absurd prediction for a dose above anything in the training data?

## Least squares says nothing about where there was no data A fitted model is a function defined over the whole real line, but the fitting procedure only ever looked at the range the training inputs covered. Every coefficient was chosen to reduce squared error on those rows. Outside that range the curve is not a prediction supported by evidence; it is whatever shape the chosen basis happens to continue into. That is bearable for a straight line, whose continuation is at least gentle and inspectable. It is not bearable for a polynomial. ## Why polynomial tails explode Consider a drug dose-response fitted with dose, dose squared and dose cubed. The measured response rises with dose and then flattens into a plateau at the top of the tested range. A cubic can imitate that plateau, but only by arranging its terms to cancel: a large positive linear term is offset by a negative quadratic and a positive cubic, and the three roughly balance over the tested doses. The balance is a numerical accident of the fitting range, not a property of the biology. Step outside that range and the cancellation collapses, because the terms grow at different rates. At twice the maximum tested dose the cubic term is eight times larger than it was at the boundary while the linear term is only twice as large. The highest power always wins eventually, so the fitted response either rockets upward or plunges below zero within a short distance of the boundary. Predicting a 340% response, or a negative one, is the visible symptom of a structural fact about the basis. ## What a natural spline changes A cubic spline divides the range at knots and fits a separate cubic on each segment, joined so that value, slope and curvature match. Beyond the outermost knots it is still cubic, so it shares the polynomial's tail problem — and this is why plain cubic splines show very wide uncertainty at the edges of the data. A **natural** cubic spline adds boundary constraints: the fitted function is forced to be **linear beyond the boundary knots**. Those constraints remove degrees of freedom relative to an unconstrained cubic spline on the same knots, and they buy two things. Inside the range, the fit is stabilised near the edges where data is thinnest. Outside it, the continuation is a straight line taking off at the boundary slope, which is a bounded, inspectable guess rather than an explosion. So for the dose-response with a plateau, the two bases give visibly different tails: the cubic terms diverge, the natural spline continues at roughly the plateau's slope. Neither has any evidence behind it. ## Production practice Because no basis makes extrapolation trustworthy, the engineering answer is to control it rather than model it. **Record the domain of validity.** At training time, store the minimum and maximum of every input the model was fitted on, and ship those bounds with the model. **Guard at scoring time.** When a request arrives with a dose above the recorded maximum, the service should do something deliberate: reject it, clamp the input to the boundary and mark the response as clamped, or serve it flagged as an extrapolation. Silently returning the number is the failure. Which of the three is right depends on the cost of a wrong answer, and for anything clinical or financial that decision is a policy question, not an implementation detail. **Monitor the input distribution.** The situation almost never announces itself as one exotic request; it arrives as drift, with the scored population creeping past the range the model was fitted on over months. **Encode known shape instead of trusting the fit.** If the pharmacology says the response saturates, a basis or constraint that saturates — a monotone fit that levels off, or a spline whose tail is deliberately flat — represents the knowledge directly instead of hoping a flexible curve behaves at the boundary. ## The interview point The answer that lands is not 'the model overfitted'. It is that basis expansion buys in-range flexibility and says nothing whatever about out-of-range behaviour, that different bases fail differently out there, and that the professional response is to know your model's valid domain and defend it at the boundary rather than to chase a basis that extrapolates well.