Abstract
An inverted U is not established by naming a scatterplot or finding a significant squared term. The shape, the location of its turning point, and the economics behind the relationship all require scrutiny.
A scatterplot is evidence, not an explanation. A fitted straight line summarizes a linear association; it does not turn a cloud of observations into a Kuznets curve. Nor does every curved scatterplot deserve that celebrated name.
Simon Kuznets (1955) considered how the personal distribution of income might change during economic development. His conjecture is commonly represented as an inverted U: inequality initially rises with development and subsequently falls. The conjecture concerns a particular economic relationship, not a generic mathematical shape.
From a straight line to a quadratic
A sensible starting point is the linear conditional-mean specification:
\quad (1)\qquad Y_i=\alpha+\beta X_i+\varepsilon_i, \qquad E(\varepsilon_i\mid X_i)=0.
Here Y_i denotes a measure of income inequality and X_i a measure of economic development, such as real income per capita. A straight line has a constant slope, \beta, and thus cannot generate a turning point.
A familiar empirical representation of the inverted U is:
\quad (2)\qquad Y_i=\alpha+\beta X_i+\gamma X_i^2+\varepsilon_i, \qquad E(\varepsilon_i\mid X_i)=0.
Its slope is \beta+2\gamma X. In the usual case of nonnegative income, \beta>0 and \gamma<0 produce a rising segment followed by a falling segment, provided the turning point falls within the observed domain. A negative estimate of \gamma alone is not sufficient.
The proposed curve is therefore a testable functional form. It is not proof of Kuznets’s economic mechanism. Omitted variables, changes in measurement, heterogeneous countries, and time trends may produce curvature without demonstrating a common developmental path. A cross-section of countries is not automatically a time series of one country’s development.
What the data must show
My preferred sequence is to inspect the scatterplot; fit the straight line; examine residuals and a LOWESS smooth; and then estimate the quadratic if the pattern warrants it. Examine the estimated slopes at the lower and upper bounds of the observed X range. For an inverted U, the lower-end slope should be positive and the upper-end slope negative. Assess uncertainty jointly rather than infer the shape from separate coefficient t-tests. Lind and Mehlum (2010) explain why testing only the quadratic term is inadequate for establishing a U-shaped relationship.
The estimated maximum must also lie strictly within the sample’s range. If it lies outside, the observed relationship can remain monotonic even though the algebraic parabola turns elsewhere. In addition, confidence intervals for a turning point can be broad or unbounded when estimated curvature is weak.
No empirical Kuznets curve should be declared merely because a computer draws an appealing arc. Mathematics identifies the turning point; economic reasoning must explain why the relationship should turn.
A political economy aside
Could public stupidity in choosing political leaders also follow a quadratic curve? Does it rise with prosperity, reach a maximum, and then decline—or have we simply not observed enough data to locate the turning point?
That is satire, not an estimated result. To turn the provocation into research, one would first need defensible measures of political judgment and prosperity, a defined population and period, and a research design capable of distinguishing association from causation. Until then, the question is a gadfly’s invitation to think, not a statistical conclusion.
Appendix: Derivation and inference for an interior turning point
Consider the estimated conditional mean
\quad (3)\qquad \widehat m(X)=\hat\alpha+\hat\beta X+\hat\gamma X^2.
Differentiating with respect to X yields
\quad
(4)\qquad
\frac{d\widehat m(X)}{dX}=\hat\beta+2\hat\gamma XAt an interior stationary point, the first-order condition is
\quad (5)\qquad \hat\beta+2\hat\gamma X^*=0.
For \hat\gamma\ne 0, rearranging gives
\quad
(6)\qquad
X^*=-\frac{\hat\beta}{2\hat\gamma}.The second derivative is
\quad
(7)\qquad
\frac{d^2\widehat m(X)}{dX^2}=2\hat\gamma.Thus \hat\gamma<0 implies that the stationary point is a strict maximum; \hat\gamma>0 implies a strict minimum. For an interior observed maximum one must also have
\quad
(8)\qquad
X_{\min}<X^*<X_{\max}.Equivalently, when \hat\gamma<0, the derivative is positive at X_{\min} and negative at X_{\max}:
\quad
(9)\qquad
\hat\beta+2\hat\gamma X_{\min}>0,\qquad
\hat\beta+2\hat\gamma X_{\max}<0.These endpoint inequalities are the economically pertinent shape restrictions over the data. If X takes negative values, the frequently stated coefficient sign requirement \hat\beta>0 is not necessary; the endpoint slope conditions remain decisive.
Sampling uncertainty. Write g(\beta,\gamma)=-\beta/(2\gamma). Its gradient, in the coefficient order (\beta,\gamma), is
\quad
(10)\qquad
\nabla g=\begin{pmatrix}-1/(2\gamma)\\[2pt]\beta/(2\gamma^2)\end{pmatrix}.If \widehat V is the estimated covariance matrix of (\hat\beta,\hat\gamma), the delta-method approximation is
\quad
(11)\qquad
\widehat{\operatorname{Var}}(X^*)\ \approx\
\nabla g(\hat\beta,\hat\gamma)^{\!\mathsf T}
\widehat V\,
\nabla g(\hat\beta,\hat\gamma).This local approximation can be unreliable when \hat\gamma is near zero. Ratio-based (Fieller-type) methods may yield unbounded confidence sets in that case. Statistical evidence of an inverted U should therefore account for uncertainty in the endpoint slopes as well as the location of X^*.
References
Kuznets, Simon. 1955. “Economic Growth and Income Inequality.” American Economic Review 45 (1): 1–28. https://www.jstor.org/stable/1811581.
Lind, Jo Thori, and Halvor Mehlum. 2010. “With or Without U? The Appropriate Test for a U-Shaped Relationship.” Oxford Bulletin of Economics and Statistics 72 (1): 109–118. https://doi.org/10.1111/j.1468-0084.2009.00569.x.