Explaining the Risks of Extrapolating Nonlinear Curves

Why Extrapolation Needs More Than a Good Fit

A fitted curve can summarize a nonlinear pattern well where observations were collected and still give a misleading prediction beyond them. The further a requested predictor value lies outside the observed range, the less the data can distinguish among possible curve shapes there. As a result, extrapolation depends not only on the observed fit but also on assumptions about how the relationship continues.

This distinction matters whenever a model is used to predict at a new predictor value. A curve may be drawn smoothly across a graph, but a smooth line is not evidence that its shape remains accurate beyond the data. In particular, nonlinear forms can turn, flatten, or accelerate, and those differences can become large outside the range used to fit them.

Definition: Let the observed predictor values range from \(x_{\min}\) to \(x_{\max}\). A prediction at \(x\) is interpolation when \(x_{\min} \leq x \leq x_{\max}\), and extrapolation when \(x < x_{\min}\) or \(x > x_{\max}\). The fitted regression function \(f(x)\) describes the model’s conditional mean response at predictor value \(x\), written \(E(Y\mid X=x)=f(x)\). A fitted value beyond the observed range is a model-based extension, not a direct summary of observations at that value.

For a fitted curve to describe the mean response, a basic model is \(Y_i=f(x_i)+\varepsilon_i\), where \(x_i\) is the predictor value for observation \(i\), and \(\varepsilon_i\) is the error around the mean curve. The mean-curve interpretation requires \(E(\varepsilon_i\mid X_i=x_i)=0\). Usual regression analysis also relies on the observations and their errors being collected and modeled appropriately for the intended analysis. Those conditions do not, by themselves, establish that the same curve continues to describe the relationship outside the observed range.

Extrapolation requires an additional judgment: whether the selected functional form and the process generating the data remain relevant at the new predictor values. A good fit over the observed range helps assess the curve there. It cannot, on its own, verify the curve’s behavior where there are no observations.

How Curves That Agree Can Separate

A short algebraic argument shows why agreement within the data need not imply agreement outside it. Suppose a candidate curve \(f(x)\) is modified by adding \(c(x-a)(x-b)\), where \(a\) and \(b\) are the endpoints of the observed predictor range and \(c\) is a nonzero constant. The added term equals zero at both endpoints, so the modified curve has the same endpoint predictions. Its size inside the range is limited by the range’s width, but outside the range it grows in magnitude as \(x\) moves away from the endpoints.

(x-a)(x-b) = \left(x-\frac{a+b}{2}\right)^2 - \left(\frac{b-a}{2}\right)^2

Inside \([a,b]\), the squared term in this expression is between \(0\) and \(((b-a)/2)^2\). Therefore \(\left|(x-a)(x-b)\right| \leq (b-a)^2/4\) there. Outside the interval, the product is positive and increases as \(x\) moves farther from the endpoints. Thus, two curves can be close across the observed interval yet have a difference that grows beyond it. This is not a technical flaw in fitting: the observed data simply provide limited information about the unobserved continuation.

Worked Example: Two Curves With Similar Fits Inside the Range

Suppose a hypothetical data set has predictor values from \(0\) to \(4\), measured in hours. Two candidate curves for a response measured in units are \(f_1(x)=10+2x\) and \(f_2(x)=10+2x+0.1x(x-4)\). Compare them at observed-range values and at \(x=12\) hours.

At \(x=0\), both predict \(10\) units, since the added term is \(0.1(0)(-4)=0\). At \(x=4\), both predict \(18\) units, since the added term is also zero. At the midpoint \(x=2\), the first curve predicts \(10+2(2)=14\) units, while the second predicts \(14+0.1(2)(-2)=13.6\) units. Their predictions differ by only \(0.4\) units there.

At \(x=12\), which is outside the observed range, the predictions are

f_1(12)=10+2(12)=34,\qquad f_2(12)=10+2(12)+0.1(12)(12-4)=43.6.

The curves differ by \(9.6\) units at \(12\) hours, despite agreeing at both observed-range endpoints and differing by only \(0.4\) units at the midpoint. These calculations show how extensions of candidate forms can separate; they do not establish which prediction is closer to the true mean response at \(12\) hours. The data contain no observations at that value to settle that question.

Polynomial Curves Can Turn or Accelerate

Polynomial regression is especially important to handle carefully outside the range. A polynomial’s terms may balance one another over the observed values, but at sufficiently large positive or negative values the highest-degree term dominates. Unless its behavior is supported by knowledge about the process, that mathematical continuation may be unrealistic. Even a turning point computed exactly from a fitted polynomial is not evidence that the real relationship turns there.

Worked Example: A Quadratic That Eventually Predicts a Decline

Imagine a hypothetical study relating daily dose \(x\), in milligrams, to a mean symptom-relief score \(f(x)\), in points. The observed doses range from \(0\) to \(6\) mg/day. A fitted quadratic is \(f(x)=5+3x-0.2x^2\). Its predictions at the observed endpoints are \(f(0)=5\) points and \(f(6)=5+18-7.2=15.8\) points; intermediate values include \(f(2)=10.2\) and \(f(4)=13.8\). The fitted curve is increasing over this observed range.

To find where its direction changes, differentiate the fitted function: \(f'(x)=3-0.4x\). Setting this equal to zero gives \(3-0.4x=0\), so \(x=7.5\) mg/day. The derivative is positive below \(7.5\) and negative above it. The model therefore predicts a maximum at \(7.5\) mg/day, just beyond the observed range. But the observations up to \(6\) mg/day do not establish that symptom relief actually peaks at \(7.5\) mg/day.

Farther outside the data, the polynomial predicts \(f(12)=5+36-0.2(144)=12.2\) points and \(f(20)=5+60-0.2(400)=-15\) points. If the score is defined so that negative symptom relief is impossible, the prediction at \(20\) mg/day is not credible. The arithmetic is correct for the fitted quadratic; what is unsupported is treating that quadratic as a reliable description at \(20\) mg/day. A different curve, or a different relationship at high doses, could fit the observed range comparably while implying a different continuation.

A polynomial’s far-range behavior follows from its leading term. For example, for \(f(x)=\beta_0+\beta_1x+\beta_2x^2\) with \(\beta_2 \neq 0\), the ratio of the lower-degree terms to the quadratic term satisfies

\frac{\beta_0+\beta_1x}{\beta_2x^2} = \frac{\beta_0}{\beta_2x^2} + \frac{\beta_1}{\beta_2x} \longrightarrow 0 \quad\text{as }x\longrightarrow\infty.

Consequently, the quadratic term determines the curve’s eventual direction and unbounded growth or decline. This explains why a polynomial that behaves gently over a narrow interval can become extreme far away. It does not show that the real-world relationship has that extreme behavior.

Extrapolated Conclusions Depend on the Chosen Form

Different nonlinear forms encode different claims about how a relationship continues. A line assumes the mean response changes at a constant rate. A saturating curve assumes that increases become smaller as the predictor grows and that the response approaches a plateau. When the observed range is limited, both forms can sometimes describe the observations reasonably well, yet give very different predictions farther out.

Worked Example: Linear Increase Versus Saturation

Consider hypothetical measurements of a continuous biomarker response, in response units, at doses from \(2\) to \(6\) mg/day. Suppose two candidate mean curves are \(g(x)=20x/(5+x)\), a saturating form, and \(h(x)=2.7+1.4x\), a linear form. Compare their predictions at three observed-range values, then at \(20\) mg/day.

At \(x=2\), \(g(2)=40/7\approx5.714\) response units and \(h(2)=2.7+2.8=5.5\) units. At \(x=4\), \(g(4)=80/9\approx8.889\) units and \(h(4)=2.7+5.6=8.3\) units. At \(x=6\), \(g(6)=120/11\approx10.91\) units and \(h(6)=2.7+8.4=11.1\) units. These predictions are relatively close over the observed dose range.

At \(x=20\), the predictions are \(g(20)=400/25=16\) response units and \(h(20)=2.7+28=30.7\) units. The difference is \(30.7-16=14.7\) units. The saturating curve assumes a plateau near \(20\) units, while the line assumes that the same rate of increase continues. The observed-range comparisons alone do not decide which assumption is more credible at \(20\) mg/day. If relevant scientific knowledge supports a plateau, that knowledge may favor the saturating model; if not, the saturation claim also requires justification. Neither extrapolated value becomes verified simply because its curve fits the observed range.

An extrapolated statement should therefore identify both the numerical prediction and the assumption behind it. “The fitted curve predicts a mean response of 16 units at \(20\) mg/day, if the saturating form remains appropriate there” is more informative than “the response will be 16 units.” It distinguishes a model’s consequence from a fact established by observed data. It also avoids treating a predicted conditional mean as a guaranteed value for every individual.

What Must Be Assumed Beyond the Data?

A credible extrapolation needs more than a curve that looks reasonable between observed points. The assumptions become stronger as the prediction moves farther from the data, or into conditions where the process could change. Useful questions include:

Does the functional form have a reason to continue? A mechanistic or scientific reason for saturation, continued growth, or a turning point is stronger support than a convenient fit alone.
Is the prediction in a comparable setting? Changes in population, measurement, treatment, or other conditions can change the relationship represented by the fitted curve.
Are the predicted values meaningful? A curve that yields impossible values, violates known physical constraints, or contradicts established knowledge should not be treated as credible merely because its in-range fit is good.
How far outside the observed range is the target? A small extension beyond the data still requires an assumption about continuation; a distant extension generally asks substantially more of that assumption.

These questions do not guarantee a correct extrapolation. They make the basis for it visible and help distinguish a cautious, qualified model-based prediction from an unsupported claim about what must happen.

Common Mistake: Treating a Smooth Fit as Evidence Outside the Data

A common error is to reason, “The curve fits the observed points well and looks smooth, so its prediction beyond them is reliable.” The correction is that the fit evaluates the curve where data provide evidence; smoothness describes the drawn function, not its verified behavior beyond those data. Another mistake is to interpret a polynomial turning point outside the observed range as a discovered real-world turning point. It is only the location where that fitted formula changes direction. Whether the underlying relationship changes direction there needs stronger support.

Key takeaway: Extrapolation extends a fitted relationship beyond the predictor values that were observed. Nonlinear forms that fit similarly within that range can diverge outside it, so extrapolated conclusions must be presented as conditional on the curve’s continued validity and any relevant scientific assumptions.

Check Your Understanding

Answer each question by distinguishing what the fitted curve calculates from what the observed data support.

A data set has predictor values from \(3\) to \(10\). Is a prediction at \(x=8\) interpolation or extrapolation? What about a prediction at \(x=12\)?
Why can two curves that agree at both endpoints of an observed interval still differ substantially outside it?
A fitted quadratic has a turning point just beyond the largest observed predictor value. What does that turning point establish, and what does it not establish?
A line and a saturating curve fit the observed data similarly but predict different means at a much larger predictor value. What additional information could help assess which continuation is plausible?
Rewrite this claim more cautiously: “The response will decline at high doses because the fitted quadratic turns downward.”