Standardized Shape Invariants

Why Standardize Shape Summaries?

A measurement can change units or origin without changing the distributional shape we want to describe. For example, converting a measurement from one unit to another multiplies it by a positive constant, while resetting the zero point adds a constant. A useful shape summary should not change just because of these recodings. Standardized moments and ratios of quantile differences are two ways to construct such summaries.

Throughout, consider a real-valued random variable \(X\) and a positive location-scale transformation \(Y=a+bX\), where \(a\in\mathbb{R}\) and \(b>0\). The shift \(a\) changes location, and \(b\) changes scale. This tutorial concerns summaries of distributional shape that are unchanged by both operations.

Standardized Skewness and Excess Kurtosis

The powers of \(\sigma\) remove the units: \(\mu_3\) has units cubed and \(\sigma^3\) has the same units; likewise, \(\mu_4\) and \(\sigma^4\) have matching units. Skewness records a signed third-moment summary, while excess kurtosis compares the standardized fourth moment with the normal benchmark of \(3\). These descriptions require the relevant moments to exist; they are not defined for every distribution.

Proof. Write \(\mu_X=E[X]\) and \(\sigma_X^2=\operatorname{Var}(X)\). Since \(Y=a+bX\), \(\mu_Y=a+b\mu_X\). Therefore \(Y-\mu_Y=b(X-\mu_X)\). It follows that \(\operatorname{Var}(Y)=b^2\sigma_X^2\), so \(\sigma_Y=b\sigma_X\), using \(b>0\). For \(r=3,4\), whenever the corresponding absolute moment is finite, \(\mu_r(Y)=E[(Y-\mu_Y)^r]=b^r\mu_r(X)\). Thus \[ \frac{\mu_3(Y)}{\sigma_Y^3} =\frac{b^3\mu_3(X)}{(b\sigma_X)^3} =\frac{\mu_3(X)}{\sigma_X^3}, \qquad \frac{\mu_4(Y)}{\sigma_Y^4} =\frac{b^4\mu_4(X)}{(b\sigma_X)^4} =\frac{\mu_4(X)}{\sigma_X^4}. \] Subtracting \(3\) from the second standardized moment gives the same excess kurtosis for both variables. \(\square\)

The positivity condition matters for skewness: if \(b<0\), then the third central moment changes sign while the standard deviation remains positive, so skewness changes sign. The fourth standardized moment is unchanged for any nonzero \(b\), but this tutorial uses positive transformations so that quantiles retain their ordering as well.

Quantile-Difference Ratios

For \(0<p<1\), define the quantile \(Q_X(p)\) by the generalized inverse \(Q_X(p)=\inf\{x\in\mathbb{R}:F_X(x)\ge p\}\), where \(F_X\) is the distribution function of \(X\). This convention also applies to discrete distributions, where the distribution function can jump.

Proof. The distribution functions satisfy \(F_Y(y)=P(a+bX\le y)=F_X((y-a)/b)\), since \(b>0\). For any real \(x\), \(F_Y(a+bx)=F_X(x)\). Equivalently, the set of \(y\) for which \(F_Y(y)\ge p\) is the set of values \(a+bx\) for which \(F_X(x)\ge p\). Because \(x\mapsto a+bx\) is increasing, taking the infimum of the first set gives \(a+b\) times the infimum of the second, plus \(a\). Thus \(Q_Y(p)=a+bQ_X(p)\). Each quantile difference therefore satisfies \(Q_Y(p_j)-Q_Y(p_i)=b[Q_X(p_j)-Q_X(p_i)]\). In a ratio of such differences the common factor \(b\) cancels, provided the denominator is nonzero. \(\square\)

A familiar three-quantile shape summary is Bowley’s quartile skewness: \[ B_X=\frac{Q_X(0.75)+Q_X(0.25)-2Q_X(0.5)} {Q_X(0.75)-Q_X(0.25)}, \] when the denominator is positive. Its numerator is the difference between the upper and lower quartile gaps. The quantile transformation rule shows that both numerator and denominator are multiplied by \(b\), so \(B_Y=B_X\). A different invariant, the ratio of the upper and lower quartile gaps, describes their relative sizes rather than their signed difference.

Moment Summaries and Quantile Summaries in Practice

Both kinds of summaries discard information, but they do so differently. Standardized moments use all observations through expectations of powers of distance from the mean. This makes them sensitive to tail behavior; it also means a summary may not exist when the relevant moment is infinite. Quantile summaries depend on selected points of the distribution rather than on powers of extreme values. They can be useful when moment assumptions are not appropriate, but a few quantiles still cannot describe an entire distribution.

Common Pitfalls and Interpretation

• Assuming a moment summary always exists. Finite variance alone does not guarantee finite third or fourth moments. Check the required moment condition before reporting skewness or kurtosis.
• Forgetting the direction of a transformation. A negative multiplier reverses quantile order and changes the sign of skewness; the positive-scale invariance statements above do not apply unchanged.
• Using a ratio with a zero denominator. Quantile differences can be zero, especially for discrete distributions with ties. Check the denominator before defining the ratio.
• Treating an invariant as a complete description. Standardized skewness, excess kurtosis, and a selected quantile ratio each summarize only part of a distribution. Agreement on one or several summaries does not establish that two distributions have the same shape.
• Confusing invariance with resistance to unusual observations. Standardization removes units and shifts; it does not make moments insensitive to extreme values. Quantile summaries avoid averaging powers of extremes but depend on which quantiles are selected.