Eisenhauer’s Relative Dispersion Coefficient (CRD)

• Eisenhauer’s Relative Dispersion Coefficient (CRD) is a unit-free metric that quantifies variability by comparing standard deviation to half the data range.
• Its invariance under linear transformations ensures that CRD remains consistent regardless of changes in scale or origin, unlike Pearson’s CV.
• The normalized CRD (CRD₍c₎) offers bounded comparisons across different sample sizes, enhancing cross-study reproducibility in statistical analyses.

Eisenhauer’s Relative Dispersion Coefficient (CRD) is a dimensionless, unit-invariant statistic for quantifying the relative dispersion in a data sample. Unlike traditional measures such as Pearson’s coefficient of variation (CV), CRD is derived solely from canonical measures of spread, remains invariant under linear transformations of scale or origin, and can be normalized for meaningful interpretation across sample sizes. CRD’s theoretical and practical advantages stem from its mathematical construction, statistical properties, and empirical performance in reproducible workflows using modern statistical software (Silveira et al., 2021).

1. Formal Definition

Given a sample $x_{1},\dots,x_{n}$, with the sample mean $\bar x = \frac{1}{n}\sum_{i=1}^{n} x_i$, sample standard deviation $$s = \sqrt{\frac{1}{n-1} \sum_{i=1}^n (x_i-\bar x)^2}$$, and sample range $r = \max_i x_i - \min_i x_i$, the uncorrected CRD is defined as:

$CRD = \frac{s}{r/2} = \frac{2s}{r}$

Both $s$ and $r$ retain the units of the original data, but their ratio, $CRD$, is inherently dimensionless. This construct makes CRD a direct measure of the standard deviation relative to the half-spread of observed values, capturing how much variability exists within the bounded span of the data (Silveira et al., 2021).

2. Invariance to Linear Transformations

A principal criterion for measures of relative variability is invariance under changes of units or origin. If each datum is transformed according to $y_i = a x_i + b$, with $a \neq 0$ and $\bar x = \frac{1}{n}\sum_{i=1}^{n} x_i$0 arbitrary, then

• $\bar x = \frac{1}{n}\sum_{i=1}^{n} x_i$1
• $\bar x = \frac{1}{n}\sum_{i=1}^{n} x_i$2

So:

$\bar x = \frac{1}{n}\sum_{i=1}^{n} x_i$3

Thus, CRD is unaffected by shifts in location ($\bar x = \frac{1}{n}\sum_{i=1}^{n} x_i$4) or scaling and unit changes ($\bar x = \frac{1}{n}\sum_{i=1}^{n} x_i$5), a property not enjoyed by estimators like Pearson’s CV, which can change arbitrarily under such transformations. This invariance ensures that CRD reflects properties of dispersion that are intrinsic to the data’s shape, not contingent on external measurement conventions (Silveira et al., 2021).

3. Normalization and Sample-Size Correction

The raw $\bar x = \frac{1}{n}\sum_{i=1}^{n} x_i$6 does not necessarily reside within $\bar x = \frac{1}{n}\sum_{i=1}^{n} x_i$7 and its extremal possible values depend on $\bar x = \frac{1}{n}\sum_{i=1}^{n} x_i$8. Eisenhauer (1993) introduced a normalized, sample size–corrected form:

$\bar x = \frac{1}{n}\sum_{i=1}^{n} x_i$9

The minimum $$s = \sqrt{\frac{1}{n-1} \sum_{i=1}^n (x_i-\bar x)^2}$$0—approaching $$s = \sqrt{\frac{1}{n-1} \sum_{i=1}^n (x_i-\bar x)^2}$$1—occurs when all values cluster at two extremes. The maximum, $$s = \sqrt{\frac{1}{n-1} \sum_{i=1}^n (x_i-\bar x)^2}$$2, is attained when values are spread equally between two points. $$s = \sqrt{\frac{1}{n-1} \sum_{i=1}^n (x_i-\bar x)^2}$$3 indicates effectively no relative dispersion; $$s = \sqrt{\frac{1}{n-1} \sum_{i=1}^n (x_i-\bar x)^2}$$4 indicates maximal dispersion for sample size $$s = \sqrt{\frac{1}{n-1} \sum_{i=1}^n (x_i-\bar x)^2}$$5. This normalization enables rigorous comparison across datasets of varying size, facilitating generalization and interpretability not possible with unbounded statistics (Silveira et al., 2021).

4. Comparison with Pearson’s Coefficient of Variation

Pearson’s coefficient of variation is defined as:

$$s = \sqrt{\frac{1}{n-1} \sum_{i=1}^n (x_i-\bar x)^2}$$6

Although frequently deployed to estimate relative variability, CV exhibits critical flaws:

• Non-invariance to additive shifts: $$s = \sqrt{\frac{1}{n-1} \sum_{i=1}^n (x_i-\bar x)^2}$$7 changes if a constant is added to all data, since $$s = \sqrt{\frac{1}{n-1} \sum_{i=1}^n (x_i-\bar x)^2}$$8 but $$s = \sqrt{\frac{1}{n-1} \sum_{i=1}^n (x_i-\bar x)^2}$$9, so $r = \max_i x_i - \min_i x_i$0.
• Mixing location and scale: $r = \max_i x_i - \min_i x_i$1 employs a central tendency (mean) in its denominator, which confounds location with spread.
• Lack of boundedness: $r = \max_i x_i - \min_i x_i$2 can exceed $r = \max_i x_i - \min_i x_i$3 arbitrarily, impeding intuitive interpretation.
• Instability near zero mean: $r = \max_i x_i - \min_i x_i$4 becomes infinite or undefined if $r = \max_i x_i - \min_i x_i$5—an issue in many practical scenarios.

An adjusted, bounded CV normalized by Kirby (1974) is:

$r = \max_i x_i - \min_i x_i$6

Despite this, CV retains its other limitations. In contrast, $r = \max_i x_i - \min_i x_i$7 and $r = \max_i x_i - \min_i x_i$8 are strictly unit-free, shift- and scale-invariant, bounded, and depend only on pure measures of spread—standard deviation and range—making them robust alternatives for quantifying relative dispersion (Silveira et al., 2021).

5. Practical Computation in R

Direct computation of $r = \max_i x_i - \min_i x_i$9 and $CRD = \frac{s}{r/2} = \frac{2s}{r}$0 can be easily embedded in statistical workflows. For a numeric vector $CRD = \frac{s}{r/2} = \frac{2s}{r}$1:

$s$5

For example, with a sample of temperatures in Celsius:

$s$6

This exemplifies invariance to unit transformations: the CRD is identical for Celsius and Fahrenheit. Conversely, $CRD = \frac{s}{r/2} = \frac{2s}{r}$2 differs between these scales (Silveira et al., 2021).

6. Interpretation and Empirical Guidance

A raw $CRD = \frac{s}{r/2} = \frac{2s}{r}$3 near zero denotes tight clustering relative to range; a value near unity indicates standard deviation approaching half the observed range, signifying maximal internal dispersion. The normalized $CRD = \frac{s}{r/2} = \frac{2s}{r}$4, restricted to $CRD = \frac{s}{r/2} = \frac{2s}{r}$5, allows calibrated comparison:

• $CRD = \frac{s}{r/2} = \frac{2s}{r}$6: low dispersion
• $CRD = \frac{s}{r/2} = \frac{2s}{r}$7: moderate dispersion
• $CRD = \frac{s}{r/2} = \frac{2s}{r}$8: extreme relative dispersion for sample size $CRD = \frac{s}{r/2} = \frac{2s}{r}$9

Empirical application pairs $s$0 values with density estimates; broader, flatter empirical densities correspond to higher $s$1 values, while tall, narrow densities pair with lower $s$2. These features make $s$3 particularly apt for unit-free assessment of variability across disparate measurement scales, research contexts, and sample sizes (Silveira et al., 2021).

7. Context within Statistical Methodology

Eisenhauer’s CRD and its sample size–corrected counterpart represent advances in descriptive statistics, resolving several interpretive and computational defects in more established metrics such as CV. Their theoretical justifications, invariance properties, and boundedness provide stronger guarantees for reproducibility and cross-dataset interpretation. These features make CRD suitable for association with density-based summary plots, robust aggregation in meta-analytic settings, and consistent quantification of dispersion irrespective of measurement scales or origins. The adoption of $s$4 provides a more rigorous, theoretically grounded foundation for comparative studies of variability in quantitative data analysis (Silveira et al., 2021).