betaFeedback

WorksMathematics

A rank-associated measure of population heterogeneity based on order statistics

Abstract

This paper introduces a rank-associated measure of population heterogeneity that quantifies how population differences are reflected in the relative positions of observations within a pooled sample. The measure is defined as a weighted average of Kolmogorov-Smirnov distances between conditional and unconditional distributions of order statistics, conditional on the population of origin at each rank. We establish its main properties, including boundedness, invariance under common strictly monotone transformations, and a characterization of its null value through independence between order statistics and population indicators. We derive bounds relating the measure to weighted pairwise distributional distances and population diversity across ranks, and examine the contributions of the minimum and maximum, obtaining a simple expression for two populations. Monte Carlo simulations with four populations assess its finite-sample behavior and asymptotic convergence. The empirical application reveals a shift in heterogeneity toward higher ranks in the germ-free group.

Introduction

Comparing heterogeneous populations is an important problem in statistics, with applications in reliability and survival analysis, economics, environmental studies, and experimental research. A common approach is to compare their marginal distributions. For continuous distributions F1,…,FmF_1,\ldots,F_m, one possibility is the average Kolmogorov–Smirnov distance,

Mm:=2m(m−1)∑1≤i<j≤mdKS(Fi,Fj)M_m := \frac{2}{m(m-1)} \sum_{1 \le i < j \le m} d_{\mathrm{KS}}(F_i,F_j)

where dKS(F,G)=sup⁡x∈R∣F(x)−G(x)∣d_{\mathrm{KS}}(F,G) = \sup_{x \in\mathbb{R}} \lvert F(x)-G(x)\rvert denotes the Kolmogorov–Smirnov distance [9, 4]. Although useful for comparing marginal distributions, such measures do not account for how population identities are reflected in the relative positions of observations in the pooled sample.

In this paper, we approach population heterogeneity from this perspective. Consider independent random variables X1,…,XmX_1,\ldots,X_m, where XiX_i represents an observation from population ii, and let X1:m≤⋯≤Xm:mX_{1:m} \leq\cdots\leq X_{m:m} denote their order statistics [3]. The indicator δk,i=1{Xi=Xk:m}\delta_{k,i} = 1\{X_i=X_{k:m}\} identifies whether the observation from population ii occupies rank kk. The relationship between order statistics and their associated rank indicators has been studied in several contexts. [1] established characterizations under proportional hazard rates, while [8] investigated related results through order structures. [5] considered the independence between time and cause. More recently, [12] developed general characterizations of independence between order statistics and rank indicators, with particular attention to the minimum and maximum in right- and left-censoring models, respectively.

Motivated by this line of research, we introduce a rank-associated measure of population heterogeneity that considers all order statistics and population indicators. The measure is constructed by comparing the conditional distribution of each order statistic, given the population supplying the corresponding observation, with its unconditional distribution. We aggregate the resulting Kolmogorov–Smirnov distances across ranks and populations, weighting each contribution by pk,i=P(δk,i=1)p_{k,i} = \mathbb{P}(\delta_{k,i}=1). In this way, the measure captures how population identities are distributed across the pooled ordering, complementing comparisons based solely on marginal distributions.

We establish the main properties of the proposed measure, including boundedness, invariance under common strictly monotone transformations, and a characterization of its null value in terms of independence between order statistics and population indicators. We also derive bounds connecting it to weighted pairwise distributional distances and to the diversity of population identities across ranks. Finally, we examine its decomposition into contributions from the minimum, maximum, and intermediate order statistics.

The remainder of the paper is organized as follows. Section 2 introduces the measure and establishes its main properties. Section 3 considers two populations through an example involving independent exponential random variables. Section 4 compares the proposed measure with marginal heterogeneity using a lognormal example. Section 5 presents simulation results and an application to real data. Finally, Section 6 discusses the interpretation and limitations of the proposed measure, summarizes the main findings, and outlines directions for future research.

A rank-associated measure of heterogeneity

Let X1,…,XmX_1,\ldots,X_m be independent, absolutely continuous random variables, with respective distribution functions F1,…,FmF_1,\ldots,F_m and densities f1,…,fmf_1,\ldots,f_m. Throughout this section, assume that ties occur with probability zero; that is, P(Xi=Xj)=0\mathbb{P}(X_i=X_j)=0 for all i≠ji\ne j.

For k,i∈{1,…,m}k,i\in\{1,\ldots,m\}, define the rank indicator δk,i:=1{Xi=Xk:m}\delta_{k,i}:=1\{X_i=X_{k:m}\}, which equals one if the observation from population ii occupies rank kk and zero otherwise. Let

pk,i=P(δk,i=1).(1)p_{k,i}=\mathbb{P}(\delta_{k,i}=1). \tag*{(1)}

For each rank kk, these probabilities satisfy ∑i=1mpk,i=1\sum_{i=1}^{m}p_{k,i}=1. Moreover, for each population ii, ∑k=1mpk,i=1\sum_{k=1}^{m}p_{k,i}=1.

Whenever pk,i>0p_{k,i}>0, define the conditional distribution function

Fk,i∗(x):=P(Xi≤x∣δk,i=1).(2)F_{k,i}^{*}(x):=\mathbb{P}(X_i\leq x\mid\delta_{k,i}=1). \tag*{(2)}

This is the distribution of an observation from population ii conditional on its occupying rank kk in the pooled ordering. This conditional representation is closely related to the characterizations of independence between order statistics and rank indicators considered in Vila and Almeida [12].

The distribution function of the kkth order statistic is denoted by Fk:mF_{k:m}. Since exactly one population supplies the observation occupying rank kk, the law of total probability gives

Fk:m(x)=∑i=1mpk,iFk,i∗(x),(3)F_{k:m}(x) = \sum_{i=1}^{m} p_{k,i}F^*_{k,i}(x), \tag*{(3)}

where pk,ip_{k,i} is as given in (1).

Thus, the distribution of an order statistic is a mixture of the conditional distributions associated with the possible population identities, with mixing probabilities given by the rank probabilities.

Definition 2.1. The rank-associated heterogeneity measure is defined as

Hm:=Hm(X1,…,Xm)=1m∑k=1m∑i=1mpk,idKS(Fk,i∗,Fk:m),(4)H_m := H_m(X_1,\ldots,X_m) = \frac{1}{m}\sum_{k=1}^{m}\sum_{i=1}^{m}p_{k,i}d_{\mathrm{KS}}(F^*_{k,i},F_{k:m}), \tag*{(4)}

where terms with pk,i=0p_{k,i}=0 are defined to be zero.

The measure averages, over all ranks, the discrepancies between the conditional distribution associated with each population and the unconditional distribution of the corresponding order statistic. The factor 1/m1/m gives equal weight to the different ranks, while pk,ip_{k,i} reflects the probability that population ii occupies rank kk.

The measure has a direct probabilistic interpretation. A large value of HmH_m indicates that knowing the population identity of an observation occupying a particular rank can substantially change the distribution of the corresponding order statistic. A small value indicates that the conditional distributions are close to their unconditional counterparts.

Proposition 2.1. The measure HmH_m satisfies 0≤Hm≤10 \le H_m \le1.

Furthermore, Hm=0H_m=0 if and only if Fk,i∗=Fk:mF^*_{k,i}=F_{k:m} for every pair (k,i)(k,i) such that pk,i>0p_{k,i}>0.

Proof. Since the Kolmogorov–Smirnov distance between two distribution functions belongs to [0,1][0,1], and since ∑i=1mpk,i=1\sum_{i=1}^{m}p_{k,i}=1, we obtain

0≤∑i=1mpk,idKS(Fk,i∗,Fk:m)≤1.0 \le\sum_{i=1}^{m}p_{k,i}d_{\mathrm{KS}}(F^*_{k,i},F_{k:m}) \le1.

Averaging over the mm ranks yields 0≤Hm≤10 \le H_m \le1.

Since all terms in (4) are nonnegative, Hm=0H_m=0 if and only if every term with positive weight is zero. This is equivalent to the stated equality of distribution functions. ■

The preceding result establishes that the proposed measure is zero precisely when the population identity does not affect the distribution of the order statistic, conditional on the rank being specified.

This interpretation can be expressed in terms of independence.

Proposition 2.2. For any kk and ii such that pk,i>0p_{k,i}>0, Xk:mX_{k:m} and δk,i\delta_{k,i} are independent if and only if Fk,i∗=Fk:mF^*_{k,i}=F_{k:m}.

Consequently, Hm=0H_m=0 if and only if Xk:mX_{k:m} is independent of δk,i\delta_{k,i} for every pair (k,i)(k,i) with pk,i>0p_{k,i}>0.

Proof. The first equivalence follows from Vila and Almeida [12], which establishes that the independence between an order statistic and its corresponding rank indicator is equivalent to the equality of the associated conditional and unconditional distributions.

For the second assertion, since all terms defining HmH_m are nonnegative, Hm=0H_m = 0 if and only if Fk,i∗=Fk:mF^{*}_{k,i} = F_{k:m} for every pair (k,i)(k,i) such that pk,i>0p_{k,i} > 0. The result then follows directly from the first equivalence.

An immediate consequence is that Hm=0H_m = 0 when the random variables are independent and identically distributed (i.i.d.) with a continuous distribution. In this case, the population identity is independent of the rank occupied by the observation, so the conditional distribution of the corresponding order statistic is the same regardless of the population from which the observation originates.

Proposition 2.3. For each k=1,…,mk = 1,\ldots,m, let

Dk:=1−∑i=1mpk,i2.D_k := 1 - \sum_{i=1}^{m} p_{k,i}^{2}.

Then

Hm≤1m∑k=1mDk=1−1m∑k=1m∑i=1mpk,i2.H_m \le\frac{1}{m}\sum_{k=1}^{m}D_k = 1 - \frac{1}{m}\sum_{k=1}^{m}\sum_{i=1}^{m}p_{k,i}^{2}.

In particular, Hm≤1−1/mH_m \le1 - 1/m.

Proof. For each k,ik,i, the mixture representation (3) and the triangle inequality imply

dKS(Fk,i∗,Fk:m)≤∑j=1mpk,jdKS(Fk,i∗,Fk,j∗)≤∑j≠ipk,j=1−pk,i.d_{\mathrm{KS}}(F^{*}_{k,i},F_{k:m}) \le\sum_{j=1}^{m}p_{k,j}d_{\mathrm{KS}}(F^{*}_{k,i},F^{*}_{k,j}) \le\sum_{j\ne i}p_{k,j}=1-p_{k,i}.

Consequently,

Hm=1m∑k=1m∑i=1mpk,idKS(Fk,i∗,Fk:m)≤1m∑k=1m∑i=1mpk,i(1−pk,i)=1m∑k=1m(1−∑i=1mpk,i2).H_m = \frac{1}{m}\sum_{k=1}^{m}\sum_{i=1}^{m}p_{k,i}d_{\mathrm{KS}}(F^{*}_{k,i},F_{k:m}) \le\frac{1}{m}\sum_{k=1}^{m}\sum_{i=1}^{m}p_{k,i}(1-p_{k,i}) = \frac{1}{m}\sum_{k=1}^{m}\left(1-\sum_{i=1}^{m}p_{k,i}^{2}\right).

Finally, applying the Cauchy–Schwarz inequality, we have (∑i=1mpk,i)2≤m∑i=1mpk,i2\left(\sum_{i=1}^{m}p_{k,i}\right)^{2}\le m\sum_{i=1}^{m}p_{k,i}^{2}. Since ∑i=1mpk,i=1\sum_{i=1}^{m}p_{k,i}=1, this immediately implies ∑i=1mpk,i2≥1/m\sum_{i=1}^{m}p_{k,i}^{2}\ge1/m, thereby establishing the stated universal bound.

The quantity DkD_k in Proposition 2.3 is the Gini–Simpson diversity index ([6]) associated with the population identity of the observation occupying the kkth rank. Thus, the preceding result shows that the contribution of each rank to HmH_m is bounded by the diversity of the corresponding rank assignment. In particular, the upper bound depends not only on the number of populations but also on how evenly the populations are represented at each rank.

A representation in terms of the marginal distributions

The conditional distributions in (4) can be expressed directly in terms of the marginal distribution functions and densities.

For each ii, define

Wk,i(x):=P(∑j≠i1{Xj<x}=k−1).W_{k,i}(x) := \mathbb{P}\left(\sum_{j\ne i}\mathbf{1}_{\{X_j<x\}}=k-1\right).

This is the probability that exactly k−1k-1 observations from populations other than ii are smaller than xx. Since

{∑j≠i1{Xj<x}=k−1}=⋃A⊆{1,…,m}∖{i}∣A∣=k−1[(⋂j∈A{Xj<x})∩(⋂j∉Aj≠i{Xj≥x})].\left\{\sum_{j\ne i} 1_{\{X_j<x\}}=k-1\right\} = \bigcup_{\substack{A\subseteq\{1,\ldots,m\}\setminus\{i\}\\ |A|=k-1}} \left[ \left(\bigcap_{j\in A}\{X_j<x\}\right) \cap \left(\bigcap_{\substack{j\notin A\\ j\ne i}}\{X_j\ge x\}\right) \right].

by independence, the probability Wk,i(x)W_{k,i}(x) can be written as

Wk,i(x)=∑A⊆{1,…,m}∖{i}∣A∣=k−1∏j∈AFj(x)∏j∉Aj≠i[1−Fj(x)].W_{k,i}(x) = \sum_{\substack{A\subseteq\{1,\ldots,m\}\setminus\{i\}\\ |A|=k-1}} \prod_{j\in A}F_j(x) \prod_{\substack{j\notin A\\ j\ne i}}\left[1-F_j(x)\right].

By definition, δk,i=1{Xi=Xk:m}\delta_{k,i}=1_{\{X_i=X_{k:m}\}}. Since the random variables are independent and continuously distributed, conditioning on Xi=xX_i=x gives

P(δk,i=1∣Xi=x)=P(Xi=Xk:m∣Xi=x)=P(∑j≠i1{Xj<x}=k−1)=Wk,i(x).\mathbb{P}(\delta_{k,i}=1\mid X_i=x) = \mathbb{P}(X_i=X_{k:m}\mid X_i=x) = \mathbb{P}\left(\sum_{j\ne i}1_{\{X_j<x\}}=k-1\right) = W_{k,i}(x).

Consequently, for any measurable set A⊆RA\subseteq\mathbb{R}, the law of total probability yields

P(Xi∈A,δk,i=1)=∫AP(δk,i=1∣Xi=x)fi(x) dx=∫Afi(x)Wk,i(x) dx.\mathbb{P}(X_i\in A,\delta_{k,i}=1) = \int_A\mathbb{P}(\delta_{k,i}=1\mid X_i=x)f_i(x)\,\mathrm{d}x = \int_A f_i(x)W_{k,i}(x)\,\mathrm{d}x.

Taking A=RA=\mathbb{R}, we obtain the probability that XiX_i occupies rank kk:

pk,i=P(δk,i=1)=∫−∞∞fi(x)Wk,i(x) dx.p_{k,i}=\mathbb{P}(\delta_{k,i}=1) = \int_{-\infty}^{\infty}f_i(x)W_{k,i}(x)\,\mathrm{d}x.

Furthermore, provided that pk,i>0p_{k,i}>0, the conditional distribution of XiX_i given δk,i=1\delta_{k,i}=1 is absolutely continuous, with density

fk,i∗(x)=fXi∣δk,i=1(x)=fi(x)Wk,i(x)pk,i.(5)f_{k,i}^{*}(x)=f_{X_i\mid\delta_{k,i}=1}(x) = \frac{f_i(x)W_{k,i}(x)}{p_{k,i}}. \tag*{(5)}

Thus, the rank probabilities are given by

pk,i=∫−∞∞fi(x)Wk,i(x) dx.(6)p_{k,i} = \int_{-\infty}^{\infty}f_i(x)W_{k,i}(x)\,\mathrm{d}x. \tag*{(6)}

By combining (3) and (5), the density of the kkth order statistic is given by

fk:m(x)=∑i=1mfi(x)Wk,i(x).(7)f_{k:m}(x)=\sum_{i=1}^{m}f_i(x)W_{k,i}(x). \tag*{(7)}

The expressions in (6) and (7) follow from the standard density representation of order statistics for independent, not necessarily identically distributed, random variables; see David and Nagaraja [3].

For the minimum and maximum, these expressions simplify considerably. In particular, W1,i(x)=∏j≠i[1−Fj(x)]W_{1,i}(x)=\prod_{j\ne i}[1-F_j(x)] and Wm,i(x)=∏j≠iFj(x)W_{m,i}(x)=\prod_{j\ne i}F_j(x). It follows that

f1,i∗(x)=fi(x)∏j≠i[1−Fj(x)]p1,i,whereasfm,i∗(x)=fi(x)∏j≠iFj(x)pm,i.(8)f_{1,i}^{*}(x)=\frac{f_i(x)\prod_{j\ne i}[1-F_j(x)]}{p_{1,i}}, \qquad \text{whereas}\qquad f_{m,i}^{*}(x)=\frac{f_i(x)\prod_{j\ne i}F_j(x)}{p_{m,i}}. \tag*{(8)}

These two expressions reveal an important distinction between the lower and upper extremes. For the minimum, the observation from population ii must be smaller than all the remaining observations. For the maximum, it must be larger than all the remaining observations. Thus, the conditional distributions associated with the two extremes reflect different aspects of the marginal distributions, particularly their lower- and upper-tail behavior.

Connection with pairwise distributional dispersion

The proposed measure HmH_m in (4) can also be related to a weighted average of pairwise distances between the conditional distributions.

For each rank kk, define

Ak,m:=∑i=1mpk,idKS(Fk,i∗,Fk:m)andPk,m:=∑i=1m∑j=1mpk,ipk,jdKS(Fk,i∗,Fk,j∗),A_{k,m} := \sum_{i=1}^{m} p_{k,i}d_{\mathrm{KS}}(F^*_{k,i}, F_{k:m}) \quad\text{and} \quad P_{k,m} := \sum_{i=1}^{m}\sum_{j=1}^{m} p_{k,i}p_{k,j}d_{\mathrm{KS}}(F^*_{k,i}, F^*_{k,j}),

where pk,ip_{k,i} is as given in (1).

Here, Ak,mA_{k,m} measures the weighted dispersion of the conditional distributions around their mixture, whereas Pk,mP_{k,m} measures their weighted pairwise dispersion.

Proposition 2.4. For every kk,

12Pk,m≤Ak,m≤Pk,m.(9)\frac{1}{2}P_{k,m} \leq A_{k,m} \leq P_{k,m}. \tag*{(9)}

Consequently, defining

Pm:=1m∑k=1mPk,m,P_m := \frac{1}{m}\sum_{k=1}^{m}P_{k,m},

we have

12Pm≤Hm≤Pm.(10)\frac{1}{2}P_m \leq H_m \leq P_m. \tag*{(10)}

Proof. By the mixture representation (3) and the triangle inequality,

dKS(Fk,i∗,Fk:m)=dKS(Fk,i∗,∑j=1mpk,jFk,j∗)≤∑j=1mpk,jdKS(Fk,i∗,Fk,j∗).d_{\mathrm{KS}}(F^*_{k,i}, F_{k:m}) = d_{\mathrm{KS}}\left(F^*_{k,i}, \sum_{j=1}^{m}p_{k,j}F^*_{k,j}\right) \leq\sum_{j=1}^{m}p_{k,j}d_{\mathrm{KS}}(F^*_{k,i}, F^*_{k,j}).

Multiplying by pk,ip_{k,i} and summing over ii gives the upper bound.

For the lower bound, the triangle inequality yields

dKS(Fk,i∗,Fk,j∗)≤dKS(Fk,i∗,Fk:m)+dKS(Fk:m,Fk,j∗).d_{\mathrm{KS}}(F^*_{k,i}, F^*_{k,j}) \leq d_{\mathrm{KS}}(F^*_{k,i}, F_{k:m}) + d_{\mathrm{KS}}(F_{k:m}, F^*_{k,j}).

Multiplying by pk,ipk,jp_{k,i}p_{k,j} and summing over ii and jj, we obtain Pk,m≤2Ak,mP_{k,m} \leq2A_{k,m}. This proves (9). Averaging over the ranks gives (10). ■

This result establishes a direct mathematical connection between the proposed measure and a weighted pairwise distributional dispersion. The distinction is that HmH_m measures deviations from the natural mixture distribution Fk:mF_{k:m}, rather than comparing all conditional distributions directly.

Heterogeneity associated with the minimum and maximum

The proposed measure HmH_m in (4) aggregates the contributions of all order statistics. However, the minimum and maximum deserve particular attention because they represent the two extreme positions in the pooled ordering.

Define

Hmmin⁡:=∑i=1mp1,idKS(F1,i∗,F1:m)andHmmax⁡:=∑i=1mpm,idKS(Fm,i∗,Fm:m),H_m^{\min} := \sum_{i=1}^{m}p_{1,i}d_{\mathrm{KS}}(F^*_{1,i}, F_{1:m}) \quad\text{and} \quad H_m^{\max} := \sum_{i=1}^{m}p_{m,i}d_{\mathrm{KS}}(F^*_{m,i}, F_{m:m}),

where pk,ip_{k,i} is as given in (1).

The contribution of the intermediate order statistics is

Hmint:=∑k=2m−1∑i=1mpk,idKS(Fk,i∗,Fk:m),H_m^{\mathrm{int}} := \sum_{k=2}^{m-1} \sum_{i=1}^{m} p_{k,i}d_{\mathrm{KS}}(F_{k,i}^{*}, F_{k:m}),

where Hmint=0H_m^{\mathrm{int}}=0 when m=2m=2. The overall measure can then be decomposed as

Hm=1m(Hmmin⁡+Hmmax⁡+Hmint).(11)H_m = \frac{1}{m}\left(H_m^{\min} + H_m^{\max} + H_m^{\mathrm{int}}\right). \tag*{(11)}

The decomposition in (11) makes it possible to distinguish heterogeneity associated with the lower extreme from heterogeneity associated with the upper extreme. In particular, Hmmin⁡H_m^{\min} measures the extent to which the distribution of the minimum depends on the population supplying that minimum, whereas Hmmax⁡H_m^{\max} measures the corresponding dependence for the maximum.

For two populations, there are no intermediate order statistics, and therefore

H2=12(H2min⁡+H2max⁡).(12)H_2 = \frac{1}{2}\left(H_2^{\min} + H_2^{\max}\right). \tag*{(12)}

The decomposition in (12) has two components that also admit a simple pairwise representation. Let q=P(X1<X2)q = \mathbb{P}(X_1 < X_2). Since, by (3), F1:2=qF1,1∗+(1−q)F1,2∗F_{1:2}=qF_{1,1}^{*}+(1-q)F_{1,2}^{*} and F2:2=(1−q)F2,1∗+qF2,2∗F_{2:2}=(1-q)F_{2,1}^{*}+qF_{2,2}^{*}, we obtain H2min⁡=2q(1−q)dKS(F1,1∗,F1,2∗)H_2^{\min}=2q(1-q)d_{\mathrm{KS}}(F_{1,1}^{*},F_{1,2}^{*}) and H2max⁡=2q(1−q)dKS(F2,1∗,F2,2∗)H_2^{\max}=2q(1-q)d_{\mathrm{KS}}(F_{2,1}^{*},F_{2,2}^{*}). Therefore,

H2=q(1−q)[dKS(F1,1∗,F1,2∗)+dKS(F2,1∗,F2,2∗)].(13)H_2=q(1-q)\left[d_{\mathrm{KS}}(F_{1,1}^{*},F_{1,2}^{*})+d_{\mathrm{KS}}(F_{2,1}^{*},F_{2,2}^{*})\right]. \tag*{(13)}

The representation in (13) shows explicitly that the overall heterogeneity measure for two populations is determined by the discrepancies between the conditional distributions of the minimum and maximum, weighted by the probability of the two possible population orderings.

The relative contribution of the two extremes can also be summarized by

Am:=Hmmax⁡−Hmmin⁡Hmmax⁡+Hmmin⁡,(14)A_m := \frac{H_m^{\max}-H_m^{\min}}{H_m^{\max}+H_m^{\min}}, \tag*{(14)}

provided that Hmmax⁡+Hmmin⁡>0H_m^{\max}+H_m^{\min}>0. It follows from (14) that −1≤Am≤1-1 \le A_m \le1. Positive values indicate that the maximum contributes more than the minimum, whereas negative values indicate the opposite. A value of zero means that the two extreme contributions are equal. This quantity complements HmH_m by describing the direction of the heterogeneity associated with the extremes.

Invariance properties

The proposed measure HmH_m in (4) is invariant under a common strictly increasing transformation of all the observations.

Proposition 2.5. Let X1,…,XmX_1,\ldots,X_m be independent, absolutely continuous random variables, and let gg be a strictly monotone function. Then

Hm(g(X1),…,g(Xm))=Hm(X1,…,Xm).H_m(g(X_1),\ldots,g(X_m))=H_m(X_1,\ldots,X_m).

Moreover, if gg is strictly decreasing, then

Hk,m(g(X1),…,g(Xm))=Hm−k+1,m(X1,…,Xm),H_{k,m}(g(X_1),\ldots,g(X_m))=H_{m-k+1,m}(X_1,\ldots,X_m),

for k=1,…,mk=1,\ldots,m.

Proof. If gg is strictly increasing, the result follows from the invariance of the Kolmogorov–Smirnov distance under strictly increasing transformations and the preservation of the ranks. Suppose now that gg is strictly decreasing. In this case, the ordering of the observations is reversed. In particular, the observation occupying rank kk after transformation is the one that occupied rank m−k+1m-k+1 before transformation. Thus, the rank probabilities satisfy pk,i(g)=pm−k+1,ip^{(g)}_{k,i}=p_{m-k+1,i}.

Furthermore, the conditional distributions associated with rank kk after transformation are the images, under gg, of the conditional distributions associated with rank m−k+1m-k+1 before transformation. Since the Kolmogorov–Smirnov distance is invariant under a common strictly monotone transformation, we obtain

Hk,m(g(X1),…,g(Xm))=Hm−k+1,m(X1,…,Xm).H_{k,m}\bigl(g(X_1),\ldots,g(X_m)\bigr)=H_{m-k+1,m}(X_1,\ldots,X_m).

Summing over kk and using the fact that k↦m−k+1k\mapsto m-k+1 is a permutation of {1,…,m}\{1,\ldots,m\} gives

Hm(g(X1),…,g(Xm))=Hm(X1,…,Xm).H_m\bigl(g(X_1),\ldots,g(X_m)\bigr)=H_m(X_1,\ldots,X_m).

This completes the proof.

The preceding result shows that the proposed measure is invariant under both increasing and decreasing changes of scale. Although an increasing transformation preserves the ordering of the observations, a decreasing transformation reverses the rank-specific contributions. Consequently, the overall heterogeneity measure remains unchanged, whereas its rank profile is reflected around the central rank.

The case of two populations

The case m=2m=2 provides a particularly transparent interpretation of the proposed measure. Let X1X_1 and X2X_2 be independent, absolutely continuous random variables with distribution functions F1F_1 and F2F_2, respectively. Define q=P(X1<X2)q=\mathbb{P}(X_1<X_2). Then p1,1=qp_{1,1}=q, p1,2=1−qp_{1,2}=1-q, and p2,1=1−qp_{2,1}=1-q, p2,2=qp_{2,2}=q.

From (8) the conditional densities associated with the minimum are

f1,1∗(x)=f1(x)[1−F2(x)]q,f1,2∗(x)=f2(x)[1−F1(x)]1−q.(15)f^*_{1,1}(x)=\frac{f_1(x)\left[1-F_2(x)\right]}{q},\qquad f^*_{1,2}(x)=\frac{f_2(x)\left[1-F_1(x)\right]}{1-q}. \tag*{(15)}

and those associated with the maximum are

f2,1∗(x)=f1(x)F2(x)1−q,f2,2∗(x)=f2(x)F1(x)q.(16)f^*_{2,1}(x)=\frac{f_1(x)F_2(x)}{1-q},\qquad f^*_{2,2}(x)=\frac{f_2(x)F_1(x)}{q}. \tag*{(16)}

These expressions are useful for both theoretical investigations and numerical evaluation; see also David and Nagaraja [3].

An illustrative example: exponential populations

The distinction between the minimum and maximum contributions becomes particularly clear when the two populations follow exponential distributions. Let X1∼exp⁡(θ1)X_1\sim\exp(\theta_1), X2∼exp⁡(θ2)X_2\sim\exp(\theta_2), independently, where θ1,θ2>0\theta_1,\theta_2>0 are the rate parameters. The probability that population 1 supplies the minimum is

q=P(X1<X2)=θ1θ1+θ2.q = \mathbb{P}(X_1 < X_2) = \frac{\theta_1}{\theta_1+\theta_2}.

The conditional distribution of the minimum given that it comes from population 1 is exponential with rate θ1+θ2\theta_1+\theta_2. The same distribution is obtained when the minimum comes from population 2. Hence, F1,1∗=F1,2∗F^*_{1,1}=F^*_{1,2}, which implies

dKS(F1,1∗,F1,2∗)=0andH2min⁡=0.(17)d_{\mathrm{KS}}(F^*_{1,1},F^*_{1,2})=0 \quad\text{and}\quad H_2^{\min}=0. \tag*{(17)}

The situation is different for the maximum. Its conditional survival functions are

S2,1∗(x)=θ1+θ2θ2e−θ1x−θ1θ2e−(θ1+θ2)xandS2,2∗(x)=θ1+θ2θ1e−θ2x−θ2θ1e−(θ1+θ2)x.S^*_{2,1}(x)=\frac{\theta_1+\theta_2}{\theta_2}e^{-\theta_1x}-\frac{\theta_1}{\theta_2}e^{-(\theta_1+\theta_2)x} \quad\text{and}\quad S^*_{2,2}(x)=\frac{\theta_1+\theta_2}{\theta_1}e^{-\theta_2x}-\frac{\theta_2}{\theta_1}e^{-(\theta_1+\theta_2)x}.

These distributions are different whenever θ1≠θ2\theta_1\ne\theta_2. Therefore,

H2max⁡=2q(1−q)dKS(F2,1∗,F2,2∗)>0H_2^{\max}=2q(1-q)d_{\mathrm{KS}}(F^*_{2,1},F^*_{2,2})>0

for unequal rates. Combining this result with (17), we obtain

H2min⁡=0,H2max⁡>0if θ1≠θ2.(18)H_2^{\min}=0,\qquad H_2^{\max}>0\quad\text{if }\theta_1\ne\theta_2. \tag*{(18)}

Thus, by (18), for two exponential populations with different rates, the proposed measure detects heterogeneity through the maximum but not through the minimum. This behavior is a consequence of the memoryless property of the exponential distribution: conditional on being the first observation, the minimum has the same exponential distribution regardless of which population supplies it; see Ross [10].

This example illustrates why separating the minimum and maximum contributions can be informative. Although both extremes are derived from the same two marginal distributions, their conditional distributions may exhibit markedly different behavior. In this case, the minimum does not reveal the difference between the populations through the proposed conditional comparison, whereas the maximum does.

Comparison with marginal heterogeneity

We now investigate whether the proposed measure HmH_m in (4) provides information that is not captured by a conventional marginal heterogeneity measure.

For two populations, consider the marginal Kolmogorov–Smirnov measure

M2=dKS(F1,F2).(19)M_2=d_{\mathrm{KS}}(F_1,F_2). \tag*{(19)}

Unlike H2H_2, the measure in (19) compares the original marginal distributions directly. It does not explicitly account for the conditional distributions associated with the minimum and maximum.

The following example demonstrates that two population configurations can have the same value of M2M_2 but substantially different values of H2H_2.

Let LN⁡(μ,σ2)\operatorname{LN}(\mu,\sigma^2) denote the lognormal distribution with parameters μ\mu and σ2\sigma^2, so that its logarithm follows a normal distribution with mean μ\mu and variance σ2\sigma^2.

Consider the following two configurations.

Configuration A. X1∼LN(0,0.52)X_1 \sim\mathrm{LN}(0, 0.5^2), X2∼LN(0.49585,0.52)X_2 \sim\mathrm{LN}(0.49585, 0.5^2).

Configuration B. X1∼LN(0,0.52)X_1 \sim\mathrm{LN}(0, 0.5^2), X2∼LN(0.54935,1.52)X_2 \sim\mathrm{LN}(0.54935, 1.5^2).

In both configurations, the populations are independent. The parameters of the second population were selected so that the marginal Kolmogorov–Smirnov distances are approximately equal. By using densities in (15) and (16), numerical integration gives the results reported in Table 1.

QuantityConfiguration AConfiguration B
M2M_20.38000.3800
q=P(X1<X2)q = \mathop{\mathrm{P}}(X_1 < X_2)0.75840.6359
dKS(F1,1∗,F1,2∗)d_{\mathrm{KS}}(F^*_{1,1}, F^*_{1,2})0.17580.4469
dKS(F2,1∗,F2,2∗)d_{\mathrm{KS}}(F^*_{2,1}, F^*_{2,2})0.17580.5850
H2H_20.06440.2389

Table 1. Marginal and rank-associated heterogeneity measures for the two lognormal configurations.

The values of H2H_2 are obtained from (13). For Configuration A, H2≈(0.7584)(0.2416)(0.1758+0.1758)≈0.0644H_2 \approx(0.7584)(0.2416)(0.1758 + 0.1758) \approx0.0644. For Configuration B, H2≈(0.6359)(0.3641)(0.4469+0.5850)≈0.2389H_2 \approx(0.6359)(0.3641)(0.4469 + 0.5850) \approx0.2389.

Thus, although the two configurations have essentially the same marginal heterogeneity, their rank-associated heterogeneity measures differ considerably. In fact, H2(B)/H2(A)≈3.71H_2^{(B)}/H_2^{(A)} \approx3.71.

This difference illustrates the additional information provided by the proposed measure.

In Configuration A, the two populations have the same log-scale standard deviation. Their difference is primarily associated with a shift in location. The conditional distributions associated with the minimum and maximum are relatively close. In Configuration B, the populations differ in both location and scale. In particular, the second population has a substantially larger log-scale standard deviation. Although the marginal Kolmogorov–Smirnov distance remains approximately the same as in Configuration A, the conditional distributions associated with the ranks are much more different.

The example highlights an important distinction. The marginal measure M2M_2 quantifies how different the two original distributions are, whereas H2H_2 quantifies how different the rank-conditioned distributions are from the corresponding unconditional order-statistic distributions. The two measures therefore describe related but distinct aspects of population heterogeneity.

The separate contributions of the minimum and maximum also provide useful information. The two configurations can be compared not only through the overall value of H2H_2, but also through the relative importance of H2min⁡H_2^{\min} and H2max⁡H_2^{\max}. This distinction helps identify whether the observed rank-associated heterogeneity is primarily related to the lower or upper extreme.

Simulation results and application

This section presents a numerical evaluation of the asymptotic and finite-sample properties of the proposed heterogeneity measures, using a comprehensive simulation study and an application to real data. Specifically, we analyze the performance of the global and local measures under diverse structural configurations that encompass a broad range of location and scale heterogeneity among populations.

Simulation results

For illustrative purposes, the observations in this section were generated from log-normal distributions. In each scenario, four random variables, X1X_1, X2X_2, X3X_3, X4X_4, were simultaneously drawn from log-normal distributions to represent the (m=4m = 4) populations. To ensure consistency across sample sizes and minimize the impact of sampling variability on the asymptotic convergence analysis, we first constructed a master dataset comprising n=500,000n = 500,000 observations. Smaller sample sizes, ranging from (1,000) to (250,000), were then obtained by subsampling from this master dataset in such a way that all observations in smaller samples were nested within those of the larger samples.

Conditional density and cumulative distribution functions across ranks and simulation scenarios

Figure 1. Conditional density, fk,i∗(x)f^{\ast}_{k,i}(x), and cumulative distribution, Fk,i∗(x)F^{\ast}_{k,i}(x), functions across ranks (k=1,…,4k = 1, \ldots, 4) and simulation scenarios, along with the corresponding estimates of H4H_4.

Asymptotic convergence of the global heterogeneity index across sample sizes

Figure 2. Asymptotic convergence of the global heterogeneity index H4H_4 across sample sizes under the four simulation scenarios is demonstrated.

The study considers four distinct structural scenarios to evaluate the effects of varying forms and degrees of heterogeneity among populations. In the first scenario (Independence), the four random variables are independent and identically distributed, with μj=0\mu_j = 0 and σj=0.5\sigma_j = 0.5 for all j=1,2,3,4j = 1, 2, 3, 4, resulting in H3≈0H_3 \approx0. The second scenario (Position) introduces progressive differences exclusively in the location parameter, μj\mu_j, while keeping the scale parameter fixed (σj=0.5\sigma_j = 0.5); specifically, μj+1=μj+0.2\mu_{j+1} = \mu_j + 0.2 with μ1=0\mu_1 = 0. The third scenario (Scale) varies only the scale parameter, holding the location parameter constant, so that μj=0\mu_j = 0 and σj+1=σj+0.3\sigma_{j+1} = \sigma_j + 0.3 with σ1=0.4\sigma_1 = 0.4. Finally, the fourth scenario (Position and Scale) represents a more complex setting in which both location and scale parameters vary simultaneously across populations, allowing for the investigation of the proposed procedure under combined sources of distributional heterogeneity: μj+1=μj+0.2\mu_{j+1} = \mu_j + 0.2 and σj+1=σj+0.3\sigma_{j+1} = \sigma_j + 0.3 for all j=1,2,3,4j = 1, 2, 3, 4.

Figures 1 and 2 illustrate the statistical stability and asymptotic convergence of the global estimator H4H_4 with increasing sample size (n). Specifically, H4H_4 approaches zero under the independence scenario (Scenario 1), whereas it converges to nonzero values reflecting the degree of structural divergence in the alternative scenarios (Scenarios 2–4). Moreover, the conditional density and cumulative distribution functions (fk,i∗f^*_{k,i} and Fk,i∗F^*_{k,i}, respectively) exhibit strong discriminatory power across ranks. The effects of location, scale, and their combination are clearly reflected in the corresponding curves, particularly at higher ranks (k=3k = 3 and k=4k = 4).

Application results

The Hoel [7] dataset comprises two groups: Control (n=99n = 99) and Germ-free (n=82n = 82). The three causes of failure (m=3m = 3) were treated as competing populations. In the Control group, the proportions were 22.22% (n=22n = 22) for thymic lymphoma, 38.38% (n=38n = 38) for reticulum cell sarcoma, and 39.39% (n=39n = 39) for other causes. In the Germ-free group, the corresponding proportions were 35.37% (n=29n = 29), 18.29% (n=15n = 15), and 46.34% (n=38n = 38), respectively. The analysis was stratified by experimental group, with rank-conditioned survival times modeled using the log-normal distribution and parameters estimated by maximum likelihood. This stratification prevents potential distortions in H3H_3 that could arise from aggregating groups with different survival time scales, while enabling the assessment of heterogeneity in failure-time distributions across the three competing causes within each environmental condition.

The comparative stratified analysis of the Hoel [7] data (see Figure 3) demonstrates that the germ-free environment not only extends animal longevity but also alters the competitive dynamics among risks. The global heterogeneity measures indicate similar structural levels between groups (Control: H3=0.2342H_3 = 0.2342; Germ-free: H3=0.2296H_3 = 0.2296), reflecting comparable overall dispersion across failure cause ranks. In contrast, the asymmetry measure exhibits a pronounced sign inversion (Control: A3=−0.2340A_3 = -0.2340; Germ-free: A3=0.1250A_3 = 0.1250), suggesting that the absence of microbiota shifts heterogeneity from early-stage failures to higher ranks and reduces the rigidity of temporal separation between pathologies. Notably, for the extreme ranks (k=1k = 1 and k=3k = 3), only two curves are shown instead of three, due to insufficient occurrences (nki<2n_{ki} < 2, where nkin_{ki} denotes the number of observations experiencing cause ii at rank kk) for certain causes to serve as the absolute minimum or maximum failure time.

Rank-conditioned functions for competing risks data

Figure 3. Rank-conditioned functions for competing risks data (m=3m = 3 causes), stratified by group (Control and Germ-free). The left and right panels display the functions for the Control and Germ-free groups, respectively.

Consequently, rank-conditioned log-normal parametric modeling demonstrates that the absence of microbiota systematically alters the hierarchy of competing causes. The environmental factor serves both as a global retarding mechanism for failure time and as a modulator of temporal dispersion throughout the organisms’ lifespan.

Discussion and concluding remarks

This paper introduced a rank-associated measure of population heterogeneity based on the conditional distributions of order statistics. Unlike conventional measures of marginal heterogeneity, the proposed measure accounts for the association between population identity and rank in the pooled sample. In particular, while MmM_m compares the original population distributions, HmH_m quantifies the discrepancies between the conditional and unconditional distributions of the order statistics, thereby capturing heterogeneity associated with pooled rankings.

The construction is related to the measures of departure from independence studied in Vila and Almeida [12], which focus on the minimum and maximum in right- and left-censoring settings, respectively. The present approach extends this perspective to all order statistics and population indicators, providing an aggregate assessment of rank-associated heterogeneity. We established that Hm=0H_m = 0 if and only if each order statistic is independent of its corresponding population indicator δk,i\delta_{k,i} whenever pk,i>0p_{k,i} > 0. This characterization connects the proposed measure to the literature on independence and characterizations involving order statistics and their associated indicators [1, 8, 5, 12]. We also established boundedness, invariance under common strictly monotone transformations, and bounds relating HmH_m to weighted pairwise distributional distances and the diversity of population identities across ranks.

The use of the Kolmogorov–Smirnov distance allows nonparametric comparisons of entire distribution functions without requiring finite moments [9, 4]. Moreover, the probabilities pk,ip_{k,i} provide natural weights and determine an upper bound for HmH_m through the diversity of population identities across ranks. The decomposition into minimum, maximum, and intermediate contributions also helps identify where rank-associated heterogeneity arises. For example, in the exponential case, the minimum contribution vanishes, whereas the maximum contribution is positive when the populations have different rates. Thus, heterogeneity may vary substantially across order statistics.

The numerical results reveal differences in rank-associated heterogeneity even when marginal heterogeneity is identical. Monte Carlo simulations support the stability and asymptotic convergence of the proposed measure, while the empirical application illustrates its ability to identify structural differences in competing risks. In particular, the results indicate a shift in failure heterogeneity toward higher ranks in the absence of microbiota, a pattern that may not be apparent from marginal distributional comparisons alone. These findings highlight the complementary information provided by incorporating population identity and rank into the assessment of heterogeneity.

Some limitations of the proposed measure should be acknowledged. The Kolmogorov–Smirnov distance captures the largest absolute discrepancy between distribution functions and may therefore overlook differences in moments or absolute scale. Additionally, probability weighting can reduce the influence of population–rank combinations with small probabilities. Consequently, HmH_m is best viewed as a complementary measure of rank-associated population heterogeneity rather than a universal measure of distributional differences.

Future research could investigate alternative probability distances, particularly Wasserstein-based measures, as considered in Vila and Almeida [12] for the minimum and maximum. Such alternatives may provide complementary perspectives on distributional discrepancies, with different sensitivities to location, scale, and tail behavior. Their properties and relationships with marginal heterogeneity and population diversity across ranks remain open questions [4, 11].

Declarations

Ethics Approval Not applicable.

Disclosure statement There are no conflicts of interest to disclose.

Author Contributions Statement All authors contributed equally to the conception and design of the study, data analysis, interpretation of the results, manuscript preparation, and revision. All authors read and approved the final version of the manuscript.

Data availability The dataset analyzed in this study is publicly available at https://raw.githubusercontent.com/IowaBiostat/data-sets/main/Hoel1972/Hoel1972, as described in Hoel [7].

References

  1. [1]Bapat, R. B. and Kochar, S. C. (1993). Characterizations of identically distributed independent random variables using order statistics. Statistics & Probability Letters, 17:225–230.DOI
  2. [3]David, H. A. and Nagaraja, H. N. (2003). Order Statistics, 3rd ed. Wiley, Hoboken, New Jersey.
  3. [4]Gibbs, A. L. and Su, F. E. (2002). On choosing and bounding probability metrics. International Statistical Review, 70(3), 419–435.
  4. [5]Kella, O. (2024). On independence of time and cause. Statistics & Probability Letters, 204:109944.arxiv.org/abs/2207.05328
  5. [6]Jost, L. (2006). Entropy and diversity. Oikos, 113(2), 363–375. https://doi.org/10.1111/j.2006.0030-1299.14714.x
  6. [7]Hoel, David G. (1972). A representation of mortality data by competing risks. Biometrics, 28(2), 475–488.pubmed.ncbi.nlm.nih.gov/4556703
  7. [8]Malov, S. V. (1996). On characterizations of independent identically distributed random variables via order structures. Statistics & Probability Letters, 26:97–104.DOI
  8. [9]Massey, F. J. (1951). The Kolmogorov–Smirnov test for goodness of fit. Journal of the American Statistical Association, 46(253), 68–78.DOI
  9. [10]Ross, S. M. (2014). Introduction to Probability Models, 11th ed. Academic Press.
  10. [11]Székely, G. J. and Rizzo, M. L. (2013). Energy statistics: A class of statistics based on distances. Journal of Statistical Planning and Inference, 143(8), 1249–1272. https://doi.org/10.1016/j.jspi.2013.03.018.
  11. [12]Vila, R. and Almeida, F. (2026). Characterizations of independence between order statistics and rank indicators. Preprint. Avaliable at https://arxiv.org/abs/2608.05428.

Paper details

vlnc:W_wej3tjwq
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

Not recorded for this paper

View statement
Looking for related papers…
Contents