Polynomial Regression Explained: How to Run, Read and Report It
Polynomial regression fits a curved relationship by adding powers of the predictor (x², x³, …) to a linear model. It is still a linear regression, because the model is linear in its coefficients, so the standard least-squares machinery applies. This guide explains when a curve is needed, how to test it, and how to interpret a quadratic model with a fully worked example.
What is polynomial regression?
A quadratic model is y = b0 + b1x + b2x² + ε, and a cubic adds b3x³. The sign of b2 gives the curvature: negative for an inverted U, positive for a U shape. For a quadratic, the turning point (maximum or minimum) is at:
x* = −b1 / (2b2)
The slope of the curve is not constant: it equals b1 + 2b2x, so b1 alone is the slope only at x = 0 (or at the mean if x is centred).
When should you use it?
Use it when residuals from a straight-line fit curve systematically, or when theory predicts a threshold or optimum: dose-response with a peak, performance versus arousal, growth that levels off. Prefer the lowest degree that fits adequately; quadratic is often enough, and degrees above three are rarely justified.
Testing whether the curve is needed
Compare the quadratic with the linear model using the F-test for the added term:
F = (R²quad − R²lin) / [(1 − R²quad)/(n − 3)], df = (1, n − 3)
This is equivalent to the t-test for b2 (F = t²). Centring x before squaring reduces the correlation between x and x², which stabilises the coefficients but does not change the fitted curve or the test of b2 (Cohen et al., 2003).
Assumptions
The assumptions are those of linear regression: independent errors, constant variance and approximately normal residuals, plus a correct functional form. High-degree polynomials can oscillate wildly and extrapolate poorly beyond the data, and they are sensitive to outliers at the extremes.
Original example (illustrative)
Question. How does plant growth vary with fertiliser dose, and is there an optimal dose? Growth (cm) at doses 1 to 14 units (invented): 1→28.7; 2→35.5; 3→43.3; 4→47.4; 5→50.2; 6→54.9; 7→58.7; 8→58.7; 9→55.0; 10→52.5; 11→51.2; 12→48.9; 13→39.4; 14→37.8.
| Model | R² | Adj. R² | Residual SE |
|---|---|---|---|
| Linear | 0.063 | −0.015 | 9.20 |
| Quadratic | 0.970 | 0.965 | 1.71 |
The quadratic term improves R² by ΔR² = 0.907, F(1, 11) = 335.31, p < .001. The raw-scale quadratic coefficients are:
| Term | b | SE | t | p |
|---|---|---|---|---|
| Intercept | 19.937 | 1.595 | 12.50 | < .001 |
| x | 9.265 | 0.489 | 18.93 | < .001 |
| x² | −0.581 | 0.032 | −18.31 | < .001 |
The turning point is x* = −9.265/(2 × −0.581) = 7.97 units, where predicted growth peaks at about 56.9 cm. Centring dose at its mean reduces the correlation between the linear and quadratic terms from r = 0.97 to r = 0.00.
Interpretation. Growth rises with dose up to roughly 8 units and then declines, so a straight line misrepresents the relationship. The turning point is an estimate with its own uncertainty, and it should not be extrapolated beyond the observed range of 1 to 14. These numbers are invented for teaching. In a real study, check that the decline is not driven by a few points at the high end.
How to run polynomial regression with StatFAS
Open the Polynomial Regression Calculator. Arrange two numeric columns, predictor and outcome.
Understanding the output
- Coefficients with tests: the highest-order term determines whether the added curvature is needed.
- R² and adjusted R²: adjusted R² penalises extra terms, and R² always increases with degree.
- Fitted curve and residual plot: look for remaining patterns and edge effects.
- Turning point: for a quadratic, compute −b1/(2b2) and, if possible, its confidence interval.
APA 7 reporting example
A quadratic regression fit the data significantly better than a linear model, ΔR² = .907, F(1, 11) = 335.31, p < .001. The quadratic term was negative (b = −0.581, 95% CI [−0.651, −0.511]), indicating an inverted-U relationship with predicted maximum growth at a dose of approximately 8.0 units.
Common mistakes and troubleshooting
- Using a high degree to chase every wiggle. Overfitting yields a curve that fits this sample and predicts poorly. Choose the degree from theory and parsimony.
- Extrapolating beyond the data, where polynomial curves can turn sharply.
- Dropping the lower-order terms (for example, keeping x² without x), which makes the model depend on the origin of x.
- Interpreting b1 as the overall slope in a curved model.
- Ignoring collinearity. Centre x to stabilise coefficients.
Limitations and alternatives
Polynomials are global: changing data at one end alters the fit everywhere. Splines, generalised additive models and local regression handle flexible shapes more gracefully. For saturating or sigmoidal patterns, a non-linear model such as a logistic growth curve may match the mechanism better. If your outcome is binary, a polynomial term can also be added to a logistic regression. For several predictors see multiple linear regression.
Conclusion
Polynomial regression is a simple way to capture curvature. Test whether the extra term improves fit, centre your predictor, interpret the curve and its turning point rather than a single slope, and stay within the range of your data.
1.
statfas-polynomial-regression-interface.png, alt: "StatFAS Polynomial Regression Calculator interface"2.
statfas-polynomial-regression-example-data.png, alt: "Example data entered in the StatFAS Polynomial Regression Calculator"3.
statfas-polynomial-regression-settings.png, alt: "Polynomial Regression Calculator settings in StatFAS"4.
statfas-polynomial-regression-result.png, alt: "Polynomial Regression Calculator results generated in StatFAS"Frequently asked questions
Is polynomial regression linear or non-linear?
It is a linear regression model because it is linear in the coefficients, even though the relationship between x and y is curved. Ordinary least squares can be used to fit it.
How do I choose the polynomial degree?
Start from theory and plots, choose the lowest degree that captures the pattern, and compare models with an F-test for the added term, adjusted R² or information criteria. Avoid degrees above three without strong justification.
Why should I centre the predictor?
Centring reduces the correlation between x and x², which stabilises the coefficient estimates. It does not change the fitted curve or the test for the highest-order term.
References
- Cohen, J., Cohen, P., West, S. G., & Aiken, L. S. (2003). Applied multiple regression/correlation analysis for the behavioral sciences (3rd ed.). Erlbaum.
- Draper, N. R., & Smith, H. (1998). Applied regression analysis (3rd ed.). Wiley.