Turning Points in Quadratic Regression

Why a Quadratic Turning Point Matters

A quadratic regression describes a curved relationship between a predictor and a fitted outcome. If the curve bends downward, it has a highest point; if it bends upward, it has a lowest point. That point is the turning point, or vertex, of the fitted curve. It can summarize where the fitted outcome changes direction, but its practical meaning depends on whether the predictor value is covered by the data.

For example, a fitted curve might predict that an outcome increases as a predictor rises, then decreases after a particular value. The turning point locates that change in the fitted relationship. It does not by itself establish that the relationship is causal, that an individual outcome will equal the fitted value, or that the turning point is reliably estimated.

The Quadratic Model and Its Vertex

Let \(x\) be a quantitative predictor and \(Y\) an outcome. A quadratic regression represents the conditional mean of the outcome with a quadratic function of \(x\). Its fitted curve uses estimated coefficients, which we write as \(b_0\), \(b_1\), and \(b_2\).

The corresponding regression model for observed values is \(Y_i=\beta_0+\beta_1x_i+\beta_2x_i^2+\varepsilon_i\). Here \(\beta_0+\beta_1x_i+\beta_2x_i^2\) is the conditional mean model, and \(\varepsilon_i\) represents the difference between an observed outcome and that mean. The intended interpretation requires the quadratic mean function to be a reasonable description of the relationship over the predictor values being considered. Standard regression fitting also relies on suitable error conditions, such as independent errors with mean zero and constant variance; normality is relevant to some inferential procedures, not to calculating the vertex of a fitted curve. This tutorial concerns the fitted curve’s geometry, not uncertainty in the estimated turning point.

The units help keep the two coordinates distinct. If \(x\) is measured in hours and \(Y\) in milligrams, then \(x_v\) is in hours and \(\hat{y}_v\) is in milligrams. The coefficients have corresponding units: \(b_1\) is outcome units per predictor unit, while \(b_2\) is outcome units per squared predictor unit.

Deriving the Turning Point

There are two direct ways to identify the vertex. The slope of the fitted quadratic at \(x\) is \(b_1+2b_2x\). Setting that slope equal to zero and solving gives the predictor coordinate. Substituting that coordinate into the fitted curve gives the fitted outcome.

To simplify the fitted outcome, substitute \(x_v\) into the equation and combine the terms:

Equivalently, completing the square rewrites the fitted curve as \(\hat{y}=b_0-b_1^2/(4b_2)+b_2(x+b_1/(2b_2))^2\). The squared term is zero at \(x=-b_1/(2b_2)\), confirming both vertex coordinates. This form also makes the classification clear: if \(b_2<0\), moving away from the vertex makes the squared-term contribution negative, so the vertex is a maximum. If \(b_2>0\), moving away makes it positive, so the vertex is a minimum. This agrees with the second derivative, \(2b_2\).

If \(b_2=0\), the fitted function is linear (or constant), not a quadratic curve with a unique turning point. A calculated vertex should therefore not be reported by dividing by zero or by treating a nearly zero coefficient as automatically meaningful.

Check the Observed Predictor Range

Let \(x_{\min}\) and \(x_{\max}\) be the smallest and largest predictor values in the data used to fit the curve. The vertex is within the observed range when \(x_{\min}\le x_v\le x_{\max}\). This is a useful first check: if the turning point is outside the range, it is an extrapolation beyond the observed predictor values.

When the vertex is outside the range, the fitted curve does not turn within the observed interval. Its slope does not change sign there, so the curve is increasing throughout that interval or decreasing throughout it. If the practical question is which fitted value is highest or lowest within the observed interval, check the fitted values at the interval’s endpoints rather than automatically using the vertex.

An in-range vertex is not necessarily well supported by the data. The smallest and largest observations can enclose a wide gap with few or no observations near the turning point. The range check answers whether the vertex is between the extremes; it does not measure how densely the data cover its neighborhood, validate the quadratic form, or quantify uncertainty in the vertex.

Common Mistakes and Practical Meaning

• Using the wrong sign in the vertex formula. The predictor coordinate is \(-b_1/(2b_2)\), not \(b_1/(2b_2)\). Substituting into the derivative, \(b_1+2b_2x=0\), provides a quick sign check.
• Calling every vertex a maximum. The sign of \(b_2\) determines the shape: negative means a maximum, positive means a minimum.
• Reporting only the predictor coordinate. A turning point has two coordinates. Calculate \(\hat{y}_v\) as well, and state the units of both.
• Assuming the vertex is observed just because it is in range. An in-range value may lie in a sparsely observed gap. Check where the predictor values actually occur, not only their minimum and maximum.
• Treating the fitted value as a guaranteed individual outcome. The vertex gives a feature of the fitted curve. Individual outcomes can differ from fitted values, and the calculation does not establish causation.

A turning point can give a compact description of a quadratic fit: where its direction changes, whether that change is upward or downward, and what outcome the curve predicts there. The range check is essential to that description. A vertex outside the range is an extrapolated feature; a vertex inside the range is still only a preliminary indication of support, not proof that the curve is appropriate or that the turning point is precise.