Quadratic regression is a statistical method that fits a curved line to your data instead of a straight one. While linear regression draws the best straight line through a set of points, quadratic regression draws the best parabola, a U-shaped or inverted-U curve described by the equation ŷ = ax² + bx + c. It’s the go-to technique when your data clearly curves rather than following a straight path.
The Quadratic Equation and Its Parts
The equation ŷ = ax² + bx + c has three coefficients, and each one controls a different aspect of the curve’s shape:
- a (the quadratic coefficient) controls how much the curve bends and in which direction. A positive value of a creates a U-shape that opens upward, meaning the data has a minimum point. A negative value creates an inverted U that opens downward, meaning the data has a maximum point. The larger the absolute value, the tighter the curve.
- b (the linear coefficient) influences the slope and shifts where the peak or valley of the curve falls along the horizontal axis.
- c (the constant) is the y-intercept, the predicted value of y when x equals zero.
If you’ve worked with linear regression (y = mx + b), quadratic regression simply adds the ax² term. That single addition lets the model capture relationships where the effect of x on y speeds up, slows down, or reverses direction.
When You Need a Curve Instead of a Line
The clearest signal that you need quadratic regression comes from a residual plot. When you fit a linear model to your data, you can plot the residuals (the differences between predicted and actual values) against x. If those residuals scatter randomly above and below zero, the linear model is a reasonable fit. But if the residuals form a U-shape or an arc, that pattern tells you there’s curvature in the data that a straight line is missing. A quadratic model can capture that curvature.
You can also spot the need for it just by graphing your raw data. If the points trace a curve rather than a line, trying to force a straight line through them will systematically overpredict in some regions and underpredict in others. Quadratic regression solves that by bending the prediction line to follow the data’s natural shape.
How the Coefficients Are Calculated
Quadratic regression uses the same core principle as linear regression: the method of least squares. The goal is to find values of a, b, and c that minimize the total squared distance between each actual data point and the curve. In other words, the algorithm tries every possible parabola and selects the one where the combined prediction errors are as small as possible.
For a straight line, least squares solves a relatively simple system of two equations. For a quadratic curve, it solves a system of three equations (one for each coefficient). You don’t need to do this by hand. Spreadsheet software, graphing calculators, and statistical tools all have built-in functions that compute the three coefficients instantly from your data set.
Finding the Peak or Valley
One of the most useful features of a quadratic regression model is the vertex, the highest or lowest point on the curve. In practical terms, this is the point where the trend reverses. If you’re modeling something like crop yield versus fertilizer amount, the vertex tells you exactly how much fertilizer produces the maximum yield before adding more starts to hurt.
The x-coordinate of the vertex is calculated with the formula x = −b / (2a). Once you have that x value, plug it back into the full equation to get the corresponding y value. Together, those two numbers give you the turning point of your data. This also defines the axis of symmetry, a vertical line at x = −b / (2a) where the parabola mirrors itself on either side.
Whether the vertex represents a maximum or minimum depends entirely on the sign of a. A negative a means the parabola opens downward, so the vertex is a peak. A positive a means it opens upward, so the vertex is a valley.
Common Real-World Applications
Quadratic regression shows up anywhere a relationship naturally rises and then falls (or falls and then rises). Projectile motion is the classic example. When you throw a ball or launch a projectile from a catapult, the height over time follows a parabolic arc. Researchers at the Yale-New Haven Teachers Institute used catapult experiments to show students how adjustments to arm length, release angle, and spring tension change the trajectory, all modeled with quadratic regression.
In business and economics, quadratic models often describe the relationship between price and revenue. Revenue tends to increase as price rises, but only up to a point. Beyond that, customers stop buying and revenue drops. The vertex of the fitted parabola identifies the price point that maximizes revenue.
Biology and agriculture use quadratic regression to model growth responses. Plant growth versus temperature, enzyme activity versus pH, and crop yield versus water supply all tend to follow inverted-U patterns. There’s an optimal level in the middle, with diminishing or negative returns on either side. A quadratic model captures that shape and pinpoints the optimum.
Checking Whether the Model Fits
Fitting a quadratic curve to your data doesn’t automatically mean it’s the right model. You need to verify two things: that the overall model explains the data well, and that the quadratic term itself is actually necessary.
R-squared tells you the first part. It measures what proportion of the variation in y is explained by the model, on a scale from 0 to 1. An R-squared of 0.95 means the curve accounts for 95% of the variation in your data. For calibration curves in laboratory settings, an R-squared above 0.99 is typically expected.
The second check is whether the quadratic term (the ax² part) is statistically significant. Your software will report a p-value for each coefficient in the model. If the p-value for the x² term is very small (conventionally below 0.05), that tells you the curvature is real and not just noise. If the p-value is large, the data may not actually curve, and you might be better off with a simple linear model. In one example from UCLA’s statistical methods group, the quadratic term had a t-statistic of 21.94 and a p-value near zero, leaving no doubt that the curvature was meaningful.
Quadratic vs. Linear Regression
Linear regression is simpler, easier to interpret, and should be your starting point. It works well when the relationship between your variables is roughly constant: every unit increase in x produces about the same change in y. Quadratic regression becomes necessary when that rate of change itself is changing. If each additional unit of x has a larger (or smaller) effect than the last, you’re looking at a curved relationship.
A quadratic model will always produce an R-squared at least as high as the linear version, simply because it has an extra coefficient to work with. That doesn’t mean it’s always better. Adding the quadratic term only makes sense when the data genuinely curves and the x² coefficient is statistically significant. Otherwise, you’re adding complexity without gaining real predictive power.
It’s also worth noting that quadratic regression is the simplest form of polynomial regression. If your data has more complex shapes, with multiple bends, you could fit cubic (x³) or higher-order polynomials. But quadratic models handle the vast majority of real-world curved relationships, and they remain straightforward to interpret because the vertex gives you a single, meaningful turning point.