Online Appendix for “The Kumaraswamy Distribution: Median-Dispersion ReParameterizations for Regression Modeling and Simulation-Based Estimation” Pablo A. Mitnik and Sunyoung Baek 1. Alternative dispersion orders weaker than the dispersive order In the main text we claim that neither the right-spread order nor the dilation order are relevant for examining whether the second parameter of re-parameterized Kumaraswamy distributions with the same median and support establishes a dispersion order among them. In the case of the right-spread order (also called “excess-wealth order”) this is simply because nonidentical distributions with the same finite support are not related in terms of this order (Sordo 2009:Corollary 7). This is not the case, however, for the dilation order – that is, it is not the case that any two distributions with the same finite support can only be related in terms of the dilation order if they are identical (Muñoz-Perez and Sanchez-Gomez 1990: Theorem 3). Therefore, showing why this order is not relevant for the problem at hand requires a more detailed analysis. The definition of the dilation order involves the notion of “convex order.” Using the same notation as in the main text, Z is greater than π in the convex order if πΈ[π(π)] ≥ πΈ[π(π)] for all convex functions π: β → β, provided these expectations exist. In turn, the random variable π is greater than the random variable π in the dilation order if X−πΈ(π) is greater than π − πΈ(π) in the convex order (Shaked and Shanthikumar 2007:Chapter 3). A necessary and sufficient condition for two distributions to be ordered by dilation, which we use below, is the following (Shaked and Shanthikumar 2007:Theorem 3.A.8). If two random 1 variables X and Y have finite expectations then X is larger than Y in the dilation order if and only if 1 1 1 −1 (π’) −1 (π’) π(π£) = ∫ πΊ −πΉ ππ’ − ∫ πΊ −1 (π’) − πΉ −1 (π’) ππ’ ≤ 0, 1−π£ π£ 0 for all π£ ∈ [0,1). Although it is not the case that any two distributions with the same finite support can only be related in terms of the dilation order if they are identical, this does not imply that Kumaraswamy distributions – and in particular those with the same median – are related in term of this order. Determining whether this is the case analytically is very difficult, due to the lack of a tractable-enough expression for the expectation of Kumaraswamy-distributed variables. In contexts like this, numerical evidence may be helpful. Hence, we conducted a numerical analysis in which we evaluated π(π£) at π£1 = 0.00000000001 and π£2 = 0.99999999999 using pairs of Kumaraswamy-distributed variables Y and X with the same median (π = 0.1, 0.3, 0.5, 0.7 or 0.9) and the following values of the second parameter (for the first re-parameterization): πππ¦ = {0.3, 0.8, … , 7.8} πππ₯ = {πππ¦ + 0.5, πππ¦ + 1, … , 8.3}, where πππ¦ and πππ₯ denote the parameter ππ for variables Y and X, respectively. Hence, in total, we evaluated π(π£1 ) and π(π£2 ) for 680 pairs of Kumaraswamy-distributed variables with a common median. Table A1 includes a small but pattern-revealing sample of our results, showing only the signs of π(π£1 ) and π(π£2 ). Table A2 includes the complete set of results, with actual figures instead of just signs. This table also shows the values of πππ¦ and πππ₯ corresponding 2 to the values of πππ¦ and πππ₯ employed in our analysis. Whenever π(π£1 ) and π(π£2 ) have different signs, the pairs of distributions in question are not ordered by dilation. Using this criterion we found evidence that Kumaraswamy-distributed variables with a common median are not ordered by dilation if πππ¦ > 1, regardless of the values of π and πππ₯ ; or if π > 0.5, regardless of the values of πππ¦ and πππ₯ . In addition, we found out that, in a substantial number of cases in which π ≤ 0.5 and πππ¦ < 1, Kumaraswamy-distributed variables are not ordered by dilation either. For the remaining cases – only 48 of our 680 pairs of distributions – evaluating π(π£) at π£1 and π£2 did not provide evidence that the distributions in each pair are not ordered by dilation. We cannot determine whether this happened because these distributions are actually ordered by dilation or rather because pairs of values of π£ even closer to 0 and 1, respectively, would be needed to show that they are not dilation-ordered. Nevertheless, the evidence from the numerical analysis indicates that the dilation order should not be employed, either, to assess whether the second parameter in the re-parameterizations advanced in the main text is a dispersion parameter. [Tables A1 and A2 about here] 2. Proofs of propositions 4.1 to 4.3 Proposition 4.1. When π → π (π → π), the re-parameterized Kumaraswamy distributions tend to the degenerate distribution with parameter π = π (π = π). Proposition 4.2. When ππ → 0, the Kumaraswamy distribution tends to the degenerate distribution with parameter π = π, for π = π, π. As the proofs of these propositions are rather simple and very similar for both reparameterizations, we only include here the proof for the first re-parameterization. Let X ~ Kp (π, ππ , π, π), with quantile function π»π . The quantile-spread of X is then: 3 ππ πππ»π (π’) = (π − π) {[1 − π’ π(ππ ,π,π,π) ] ππ − [1 − (1 − π’) π(ππ ,π,π,π) ] }, 1 π−π π where π(ππ , π, π, π) = ln [1 − ( π − π ) π ] (ln 0.5)−1 . Taking limit for π → π, π → π, and ππ → 0 we get: lim πππ»π (π’) = (π − π){[1 − π’0 ]ππ − [1 − (1 − π’)0 ]ππ } = (π − π)(0 − 0) = 0 π→ π lim πππ»π (π’) = (π − π){[1 − π’∞ ]ππ − [1 − (1 − π’)∞ ]ππ } = (π − π)(1 − 1) = 0 π→ π lim πππ»π (π’) = (π − π){[1 − π’∞ ]0 − [1 − (1 − π’)∞ ]0 } = (π − π)(1 − 1) = 0. ππ → 0 1/2 It is easy to see that πΏ2 (π) can be written as πΏ2 (π) = ∫0 πππ»π (π’) ππ’. Hence, given that πππ»π (π’) is bounded and that in the three cases above πππ»π (π’) → 0 for any 0 < π’ < 0.5, it follows, using Lebesgue’s dominated convergence theorem, that πΏ2 (π) = 0 when π→ π , π→ π or ππ → 0. As the degenerate distribution is the only distribution with πΏ2 (π) = 0, in all three cases the Kumaraswamy distribution tends to the degenerate distribution with location parameter equal to its median, as stated in Propositions 4.1 and 4.2. Proposition 4.3. When ππ → ∞, the Kumaraswamy distribution tends to the discrete uniform distribution with possible values c and b, for π = π, π. Corollary: When π = 0, π = 1 and ππ → ∞, the Kumaraswamy distribution tends to the Bernoulli distribution with parameter π = 0.5, for π = π, π. The proof of this proposition uses the following lemma. Lemma. If Y ~ Kp (π, ππ , 0, 1), then lim ππ′ (π) = lim ππ →∞ ππ →∞ ln(1 ln 0.5 −1 − ωππ ) π΅ (1 + πππ , 4 ln 0.5 −1 ln(1 − ωππ ) ) = 0.5, where ππ′ (π) is the rth moment around zero of Kp (π, ππ , 0, 1). The proof of this lemma follows. Substituting (6) in the expression for the rth moment around zero of K(p, q, 0, 1) used in the introduction of the main text, the rth moment around zero of Kp (π, ππ , 0, 1) is ππ′ (π) = πΏ1 (ππ ) π΅ (πΏ2 (ππ ), πΏ1 (ππ )), where −1 1 ππ πΏ1 (ππ ) = ln 0.5 [ln (1 − π )] πΏ2 (ππ ) = [1 + π ππ ]. Γ (π£+π₯) Now, from the asymptotic expansion of a ratio of Gamma functions, π£ π₯−π¦ [1 + (π₯−π¦)(π₯+π¦−1) 2π£ Γ (π£+π¦) ∝ 1 + πͺ (π£2 )] with |π£| → ∞ (Tricomi and Erdélyi 1951:133), where π = πͺ(π‘) indicates that |π| < π΄ π‘ for some constant A, we can derive the following asymptotic expansion for the Beta function by using π΅ (πΌ, π½) = π΅(πΌ, π½) ∝ Γ(π½) {1 − π½(π½−1) 2πΌ π€(πΌ)π€(π½) π€(πΌ+π½) : 1 + πͺ (πΌ2 )} πΌ −π½ , with |πΌ| → ∞. This expansion can be employed to represent the limit of ππ′ (π) for ππ → ∞ as: lim ππ′ (π) ∝ ππ →∞ lim Γ (1 + πΏ1 (ππ )) {1 − πΏ1 (ππ )[πΏ1 (ππ ) − 1] ππ →∞ 2 πΏ2 (ππ ) +πͺ( 1 πΏ2 (ππ ) 2 )} πΏ2 (ππ ) where we have used the property π£ Γ(π£) = Γ(π£ + 1). As lim πΏ1 (ππ ) = 0 ππ →∞ and lim πΏ2 (ππ ) = ∞, we then have ππ →∞ 0 1 lim ππ′ (π) ∝ Γ(1) {1 − ∞ + πͺ (∞)} lim πΏ2 (ππ ) ππ →∞ ππ →∞ Using Γ(1) = 1, this entails 5 −πΏ1 (ππ ) . −πΏ1 (ππ ) , lim ππ′ (π) = lim πΏ2 (ππ ) ππ →∞ −πΏ1 (ππ ) ππ →∞ lim ππ′ (π) = lim exp[−πΏ1 (ππ ) ln πΏ2 (ππ )]. ππ →∞ ππ →∞ Replacing πΏ1 (ππ ) and πΏ2 (ππ ) by their definitions, we obtain lim ππ′ (π) = exp [(−1) ln 0.5 lim ππ →∞ ππ →∞ 1 ln (1+π ππ ) 1 ]. (A1) ln(1− πππ ) π₯−1 2π−1 Using the series representation ln π₯ = 2 ∑∞ π=1 2π−1 (π₯+1) with |π₯| > 0 (Gradshteyn and Ryzhik 2007:53), the limit in the hand right side of (A1) can be expressed as 1 ππ lim ln (1 + π ππ ) [ln (1 − π )] ππ →∞ 2π−1 ∞ ∑ −1 = 1 ππ ∞ π ππ 1 [ lim ] π=1 2π − 1 ππ →∞ 2 + π ππ ∑ { 1 −π [ lim 1] π=1 2π − 1 ππ →∞ 2 − π ππ −1 ∞ 1 1 2π−1 2π−1 (1) (−1) =∑ [∑ ] = −1. π=1 2π − 1 π=1 2π − 1 2π−1 −1 } ∞ (A2) Substituting (A2) in (A1), we get lim ππ′ (π) = exp[(−1)ln 0.5 (−1)] = 0.5 ππ →∞ which concludes the proof of the lemma. We can now proceed with the proof of proposition 4.3. ′ (π) = π π΅ (1 + First re-parameterization. Substituting ππ−π π−π π , π) in (4), and then π using π΅(1, π½) = π½ −1 and ( ) = 1, we obtain: π ππ′ (π) = (π − π)π ∑ π π=0 π π π−π π (π ) ( ) π π΅ (1 + , π) π−π π 6 = (π − π)π ∑ π−1 π=0 π π π−π π π π π (π) ( ) π π΅ (1 + , π) + (π − π)π ( ) ( ) π π΅(1, π) π π−π π−π π = (π − π)π ∑ π−1 π=0 π π π−π π (π ) ( ) π π΅ (1 + , π) + π π . π−π π 1 Substituting (6) and ππ = π in the last expression and taking limit for ππ → ∞, we get lim ππ′ (π) = (π − π)π ∑ ππ →∞ π−1 π=0 π π π (π ) ( ) lim πΏ(ππ , π − π) + π π π − π ππ →∞ where πΏ(ππ , π − π) = with ω Μ= π−π π−π ln 0.5 −1 ln(1 − ω Μ ππ ) π΅ (1 + (π − π)ππ , ln 0.5 −1 ln(1 − ω Μ ππ ) ), . As πΏ(ππ , π − π) is the (π − π)th moment around zero of a variable with distribution Kp (π, ππ , 0, 1), the Lemma applies and lim πΏ(ππ , π − π) = 0.5. Hence, ππ →∞ lim ππ′ (π) = 0.5 (π − π)π ∑ ππ →∞ π−1 π=0 π π π (π ) ( ) + ππ. π−π π π π π π π π Now, applying the binomial theorem, (π−π) = (1 + π−π) = ∑ππ=0 ( π ) (π−π) . Therefore, π π π π π π π ∑π−1 π=0 ( π ) (π−π ) = (π−π ) − (π−π ) . We thus have lim ππ′ (π) = 0.5 (π − π)π [( ππ →∞ π π π π ) −( ) ] + π π = 0.5 (π π + π π ). π−π π−π (A3) The rth moment of a discrete uniform distribution with possible values πΌ and π½ is πΈ(π π ) = ∑π=πΌ,π½ π π Pr(π) = πΌπ +π½ π 2 .Therefore, from (A3) and the fact that both the Kumaraswamy distribution and the discrete uniform distribution have bounded support, it follows that when ππ → ∞ the Kumaraswamy distribution converges to the discrete uniform distribution with possible values c and b, as stated in Proposition 4.3. 7 Second re-parameterization. Given the one-to-one mapping between the dispersion parameters of the first and second re-parameterization introduced in the proof of Proposition 3.1, it is the case, using (7), that lim ππ = lim π·(ππ ; ω) = lim ππ → ∞ ππ → ∞ Μ ln ω ππ → ∞ ln(1−0.5ππ ) = ∞. From this and the proof of Proposition 4.3 for the first re-parameterization, the corresponding proof for the second re-parameterization follows immediately. References Gradshteyn, I. and I. Ryzhik (2007). Table of Integrals, Series and Products. New York, Academic Press. Muñoz-Perez, J. and A. Sanchez-Gomez (1990). "Dispersive ordering by Dilation." Journal of Applied Probability 27(2): 440-444. Shaked, M. and G. Shanthikumar (2007). Stochastic Orders. New York, Springer Sordo, M. (2009). "On the Relationship of Location-Independent Riskier Order to the Usual Stochastic Order." Statistics and Probability Letters 79: 155-157. Tricomi, F. G. and A. Erdélyi (1951). "The Asymptotic Expansion of a Ratio of Gamma Functions." Pacific Journal of Mathematics 1: 133-142. 8