Curvature and the Second Derivative

Why Curvature Needs Its Own Description

A quadratic regression curve does not have one unchanging slope: its slope varies with the predictor. The second derivative describes how that slope changes as the predictor increases. This provides a precise way to say whether a fitted curve bends upward or downward, and how rapidly it bends.

Curvature is not the same as whether the outcome is increasing or decreasing. A curve can bend upward while its slope is still negative, or bend downward while its slope is still positive. The second derivative describes the change in slope; the first derivative describes the slope at a particular predictor value.

Quadratic Mean Function and Curvature

For a quantitative predictor \(x\) and outcome \(Y\), write the quadratic regression model as \(Y_i=\beta_0+\beta_1x_i+\beta_2x_i^2+\varepsilon_i\). Here \(\beta_0\), \(\beta_1\), and \(\beta_2\) are model coefficients, and \(\varepsilon_i\) is the deviation of observation \(i\) from the model’s conditional mean. For interpreting the curve as the conditional mean, the key assumption is \(E(\varepsilon_i\mid x_i)=0\), so that \(m(x)=E(Y\mid x)=\beta_0+\beta_1x+\beta_2x^2\). The quadratic form should be a reasonable description over the predictor range being interpreted.

The fitted curve replaces the unknown coefficients with their least-squares estimates, conventionally written \(b_0\), \(b_1\), and \(b_2\): \(\hat{m}(x)=b_0+b_1x+b_2x^2\). Differentiating this polynomial twice gives its estimated curvature.

If \(\beta_2>0\), then \(m''(x)>0\), so the slope increases as \(x\) increases and the curve bends upward. If \(\beta_2<0\), then \(m''(x)<0\), so the slope decreases and the curve bends downward. If \(\beta_2=0\), the second derivative is zero and the mean function is linear rather than curved.

The first derivative makes this interpretation explicit: \(m'(x)=\beta_1+2\beta_2x\). Increasing \(x\) by one unit changes this slope by \(2\beta_2\) outcome units per predictor unit. Thus, the second derivative is a rate of change in slope, not an outcome change by itself.

Why the Factor of Two Appears

The quadratic term’s coefficient is not itself the second derivative. Differentiating \(x^2\) once gives \(2x\); differentiating again gives \(2\). The constant and linear terms have second derivatives of zero. Consequently, the second derivative is twice the coefficient of the squared term.

There is also a useful interpretation in terms of equal steps along the predictor scale. For any positive step size \(h\), compare the mean at \(x+h\), at \(x\), and at \(x-h\). Expanding each quadratic term gives \(m(x+h)=\beta_0+\beta_1(x+h)+\beta_2(x+h)^2\) and \(m(x-h)=\beta_0+\beta_1(x-h)+\beta_2(x-h)^2\). Subtracting twice the middle value cancels the constant and linear terms. The remaining quadratic terms are \(2\beta_2h^2\). Therefore:

This identity shows why positive curvature corresponds to the middle value lying below the average of the two equally spaced outer values, while negative curvature corresponds to the middle value lying above that average. It also confirms that the second derivative describes a consistent pattern across the quadratic curve, not just at one selected \(x\).

Interpreting Curvature with Units

Units help keep curvature distinct from slope. If \(Y\) is measured in outcome units and \(x\) in predictor units, then the slope has units of outcome units per predictor unit. The second derivative has units of outcome units per predictor unit squared. For example, if the outcome is a concentration in milligrams per litre and the predictor is a dose in milligrams, curvature is measured in \((\text{milligrams per litre})/\text{milligram}^2\).

The coefficient \(b_2\) has the same units as curvature, because it multiplies \(x^2\) to produce an outcome in the model. The coefficient is half the curvature: \(\hat{m}''=2b_2\). Reporting the coefficient without multiplying by two, or describing it as a change in outcome per predictor unit, gives the wrong interpretation.

Curvature Describes Shape, Not Causation

A fitted second derivative describes the shape of the model’s estimated conditional mean over the predictor range in question. It is an association-based description unless the study design and assumptions support a causal interpretation. A curved regression fit by itself does not show that changing the predictor would cause the outcome to change.

The quadratic model has constant curvature, but that mathematical feature does not guarantee that a quadratic form is suitable at every possible predictor value. Interpret the curve over a range where the data provide a reasonable basis for the model; far outside that range, the polynomial may give implausible predictions or a shape unsupported by observations. Curvature alone also does not tell you where the slope is zero.

The calculation \(2b_2\) is a description of the fitted curve. How uncertain that estimate is, and whether the data provide evidence of nonzero curvature, are separate inferential questions. The sign of a point estimate should not be treated as proof that the underlying mean curve bends in that direction.

Common Mistakes

• Using \(b_2\) instead of \(2b_2\). The squared term’s coefficient is half the second derivative. Differentiate \(b_2x^2\) twice to obtain \(2b_2\).
• Calling curvature the slope. Curvature is the rate at which slope changes. The slope itself is \(b_1+2b_2x\), which depends on \(x\).
• Reading the sign as the direction of change in the outcome. A positive second derivative means the slope is increasing, not necessarily that the outcome is increasing. Check the slope separately if that question matters.
• Leaving out squared predictor units. Curvature is measured in outcome units per squared predictor unit, not in outcome units per predictor unit.
• Assuming the fitted shape establishes a causal effect. A regression curve summarizes a modeled relationship; causal conclusions require appropriate design and assumptions.