Abstract
We study comparative statics of Nth-degree risk increases within a large class of problems that involve bidimensional payoffs and additive or multiplicative risks. We establish necessary and sufficient conditions for unambiguous impact of Nth-degree risk increases on optimal decision making. We develop a simple and intuitive approach to interpret these conditions : novel notions of directional Nth-degree risk aversion that are characterized via preferences over lotteries.
Similar content being viewed by others
Notes
It is more common to define Nth-degree risk increases in terms of a utility function u (α) as follows: ᾶ 2 is an increase in Nth-degree risk over ᾶ 1 if and only if E[u(ᾶ 2)] < E[u(ᾶ 1)] for all N times continuously differentiable real valued utility function u such that (−1) (N+1) u (N) > 0. Our equivalent characterization is somewhat more natural when dealing with the effect of risk increases on optimally chosen variables (i.e., on marginal utility).
Ekern’s (1980) result, but not Lemma 1, also implies that ᾶ 2≽ N ᾶ 1 is necessary to have unanimity of ranking within the class of utility functions q such that (−1)N q (N) > 0, so the information provided by the two results is somewhat different. We thank an anonymous referee for pointing this out.
Denuit et al.’s (1999) result is expressed in terms of maximal generators for s-convex orders and established for non-necessarily regular functions by the introduction of the concept of Nth degree divided difference. The authors remark that for regular functions the conditions on Nth-degree divided difference are equivalent to conditions on the Nth derivative. In Appendix A we propose a direct and simpler proof of the result when the functions are assumed to be regular.
The canonical portfolio problem arises as a special case of this framework in which the attributes are perfect substitutes (setting U(y, z) = v(y + z) and redefining the variables), making it, in essence, a univariate problem. Assumption A2 is always satisfied in this setting.
Note that the condition R N(y, z) ≤ N is coupled with ( − 1)N U (0, N)(y, z) ≥ 0 which, for N = 2, excludes risk averse consumers. In the special case that α = 0 this additional condition is no longer necessary. Therefore, assuming risk aversion, ( − 1)2 U (0, 2)(y, z) < 0, leads to a change in the direction of the inequality of the first condition, which becomes R 2(y, z) ≥ 2, as established by Rothschild and Stiglitz (1971) and Eeckhoudt and Schlesinger (2008) in the context of the 2-date saving problem with rate-of-return risk.
Suppose, for example, that relative risk aversion is constant: v(z) = (1 − γ)− 1 z 1 − γ. Then, if γ < 1 an Nth-degree risk increase in the rate-of-return will decrease savings. For γ > 1, an Nth-degree risk increase in the rate-of-return will increase savings for some wealth levels and it will decrease savings for other wealth levels.
Eeckhoudt et al. (2009) showed that if \(\widetilde {\alpha }_{2}\succcurlyeq _{N}\widetilde {\alpha }_{1}\) and \(\tilde {x}_{1}\succcurlyeq _{M}\tilde {x} _{2} \), then the 50-50 lottery [\(\widetilde {\alpha }_{2}+ \tilde {x}_{1}, \widetilde {\alpha }_{1}+ \tilde {x}_{2}\)] is an (N+M)th-degree risk increase over [\(\widetilde {\alpha }_{2}+ \tilde {x}_{2}, \widetilde { \alpha }_{1}+ \tilde {x}_{1}\)] in the sense of Ekern (1980), i.e., the second lottery is preferred by all decision makers with ( − 1)N + M U (0, N + M)(y, z) ≤ 0. When ρ y = 0 and ρ z > 0, and for a fixed value of y, our Proposition deals with the special case in which M = 1. Denuit et al. (2010a) proved a multivariate version of Eeckhoudt et al. (2009)’s results.
The case with ρ y ≤ 0 and ρ z ≥ 0 is the relevant scenario to interpret the conditions found in Section 2.
References
Baiardi, D., & Menegatti, M. (2011). Pigouvian tax, abatement policies and uncertainty on the environment. Journal of Economics, 103(3), 221–251.
Block, M.K., & Heineke, J.M. (1973). The allocation of effort under uncertainty: the case of risk-averse behavior. Journal of Political Economy, 81, 376–385.
Chiu, W.H., & Eeckhoudt, L. (2010). The effects of stochastic wages and non-labor income on labor supply: update and extensions. Journal of Economics, 100, 69–83.
Chiu, W.H., Eeckhoudt, L., Rey, B. (2012). On relative and partial risk attitudes: theory and implications. Economic Theory, 50, 151–167.
Courbage, C., & Rey, B. (2012). Priority setting in health care and higher order degree change in risk. Journal of Health Economics, 31, 484–489.
Dardanoni, V. (1988). Optimal choices under uncertainty: the case of two-argument utility functions. Economic Journal, 98, 429–450.
Dardanoni, V., & Wagstaff, A. (1990). Uncertainty and the demand for medical care. Journal of Health Economics, 9, 23–38.
Denuit, M., De Vylder, E., Lefevre, C. (1999). Extremal generators and extremal distributions for the continuous s-convex stochastic orderings. Insurance: Mathematics and Economics, 24, 201–217.
Denuit, M., Eeckhoudt, L., Tsetlin, I., Winkler, R.L. (2010a). Multivariate concave and convex stochastic dominance. INSEAD working paper, 2010/29.
Denuit, M., Eeckhoudt, L., Rey, B. (2010b). Some consequences of correlation aversion in decision science. Annals of Operations Research, 176, 259–269.
Denuit, M., Eeckhoudt, L., Menegatti, M. (2011). Correlated risks, bivariate utility and optimal choices. Economic Theory, 46, 39–54.
Eeckhoudt, L., & Schlesinger, H. (2006). Putting risk in its proper place. American Economic Review, 96, 280–289.
Eeckhoudt, L., & Schlesinger, H. (2008). Changes in risk and the demand for saving. Journal of Monetary Economics, 55, 1329–1336.
Eeckhoudt, L., Schlesinger, H., Tselin, I. (2009). Apportioning of risks via stochastic dominance. Journal of Economic Theory, 144, 994–1003.
Eeckhoudt, L., Rey, B., Schlesinger, H. (2007). A good sign for multivariate risk taking. Management Science, 53, 117–124.
Ekern, S. (1980). Increasing Nth-degree risk. Economics Letters, 6, 329–333.
Gollier, C. (2001). The economics of risk and time. MIT Press.
Hadar, J., & Seo, T. K. (1990). The effects of shifts in a return distribution on optimal portfolios. International Economic Review, 31, 721–736.
Jean, W.H. (1980). The geometric mean and stochastic dominance. Journal of Finance, 35, 151–158.
Keenan, D.C., Kim, I., Warren, R.S. (2006). The private provision of public goods under uncertainty: a symmetric-equilibrium approach. Journal of Public Economic Theory, 8, 863–873.
Kimball, M. (1990). Precautionary saving in the small and in the large. Econometrica, 58, 58–73.
Leland, H.E. (1968). Saving and uncertainty: the precautionary demand for saving. Quarterly Journal of Economics, 82, 465–73.
Menezes, C., Geiss, C., Tressler, J. (1980). Increasing downside risk. American Economic Review, 70, 921–932.
Menezes, C., & Wang, X.H. (2005). Increasing outer risk. Journal of Mathematical Economics, 41, 875–886.
Rothschild, M., & Stiglitz, J. (1970). Increasing risk I: a definition. Journal of Economic Theory, 2, 225–243.
Rothschild, M., & Stiglitz, J. (1971). Increasing risk II: its economic consequences. Journal of Economic Theory, 3, 66–84.
Sandler, T., Sterbenz, F.P., Posnett, J. (1987). Free riding and uncertainty. European Economic Review, 31, 1605–1617.
Sandmo, A. (1970). The effect of uncertainty on saving decisions. Review of Economic Studies, 37, 353–60.
Tressler, J.H., & Menezes, C.F. (1980). Labor supply and wage rate uncertainty. Journal of Economic Theory, 23, 425–436.
Author information
Authors and Affiliations
Corresponding author
Additional information
The authors thank the participants of the conference “Risk and choice: A conference in honor of Louis Eeckhoudt” (Toulouse, July, 2012) as well as those of EGRIE conference (Palma, September, 2012) and of the Risk Seminar (Paris-Dauphine, January 2013) and in particular Christophe Courbage, Harris Schlesinger and Bertrand Villeneuve. Special thanks to Louis Eeckhoudt who discussed the paper and provided very useful hints and comments as well as to two anonymous referees who permitted us to greatly improve the paper in particular by bringing our attention to the possibility to generalize the increasing risk results to a stochastic dominance framework. The first two authors gratefully acknowledge the financial support of the ANR (RISK project) as well as of the Risk Foundation (Groupama Chair).
Appendices
Appendix A: Increases in Nth-degree risk
Proof of Lemma 1
The fact that 2. implies 1. results directly from Ekern (1980). For the sake of completeness we rederive it. Let \((\widetilde {\alpha }_{1},\widetilde { \alpha }_{2})\) be such that \(\widetilde {\alpha }_{2}\succcurlyeq _{N} \widetilde {\alpha }_{1}.\) We have
Since q and its N first derivatives are continuous on \(\mathbb {R}_{+}\), they are bounded on all compact subsets and all our integrals are well defined. □
By definition, all the terms in the sum are equal to 0 and \(F_{\widetilde { \alpha }_{2}}^{[ N] }(x)-F_{\widetilde {\alpha }_{1}}^{\left [ N \right ] }(x)\geq 0\) on [0, B]. By 2., the integral is then nonnegative and \(E[ q(\widetilde {\alpha }_{2})] \geq E\left [ q( \widetilde {\alpha }_{1})\right ] \).
Let us now prove that 1. implies 2. Let q be an N times continuously differentiable function on \(\mathbb {R}_{+}\) such that \(E\left [ q(\widetilde { \alpha }_{2})\right ] \geq E[ q(\widetilde {\alpha }_{1})] \) for all pair \((\widetilde {\alpha }_{1},\widetilde {\alpha }_{2})\) such that \( \widetilde {\alpha }_{2}\succcurlyeq _{N}\widetilde {\alpha }_{1}.\) Let, if it exists, α be a nonnegative real number such that q (N)(α) ≠ 0 and let ε denote a positive real number such that q (N)(t) ≠ 0 for t ∈ [α, α + ε]. The sign of q (N)remains then constant on [α, α + ε]. Let \(\widetilde { \beta }_{1}\) and \(\widetilde {\beta }_{2}\) be two nonnegative bounded above random variables such that \(\widetilde {\beta }_{2}\succcurlyeq _{N} \widetilde {\beta }_{1}.\) Let B be a common upper bound for \(\widetilde { \beta }_{1}\) and \(\widetilde {\beta }_{2}\) and let \(\widetilde {\alpha }_{1}= \frac {\varepsilon }{B}\widetilde {\beta }_{1}+\alpha \) and \(\widetilde {\alpha }_{2}=\frac {\varepsilon }{B}\widetilde {\beta }_{2}+\alpha .\) The random variables \(\widetilde {\alpha }_{1}\) and \(\widetilde {\alpha }_{2}\) take their values in [α, α + ε] and it is easy to check that \(F_{\widetilde {\alpha }_{i}}^{[ k] }( t) =\left ( \frac {\varepsilon }{B}\right )^{k-1}F_{\widetilde {\beta }_{i}}^{[ k] }\left ( \frac {t-\alpha }{\varepsilon }B\right ) \) for k = 1, 2. . . and i = 1, 2. Therefore, \(\widetilde {\alpha }_{2}\succcurlyeq _{N} \widetilde {\alpha }_{1}\) which implies, by 1., that \(E\left [ q(\widetilde { \alpha }_{2})\right ] \geq E[ q(\widetilde {\alpha }_{1})] .\) We have
By definition, we have \(F_{\widetilde {\alpha }_{2}}^{[ k] }(\alpha +\varepsilon )=F_{\widetilde {\alpha }_{1}}^{[ k] }(\alpha +\varepsilon )\) for k = 1, . . . , N and \(F_{\widetilde {\alpha }_{2}}^{[ N] }(t)-F_{\widetilde {\alpha }_{1}}^{[ N] }(t)\geq 0\) and the inequality is strict for some t in [α, α + ε], and even on a neighborhood of t by continuity of \( F_{\widetilde {\alpha }_{1}}^{[ N] }\) and \(F_{\widetilde {\alpha } _{2}}^{[ N] }.\) Since the sign of q (N)remains constant on [α, α + ε], this gives that ( − 1)N q (N)(t) > 0 on α, α + ε and ( − 1)N q (N)(x) > 0 for all x such that q (N)(x) ≠ 0 which completes the proof.
1.1 Proof of Proposition 1
Let us prove that 2. implies 1. Let \(\left ( \widetilde {\alpha }_{1}, \widetilde {\alpha }_{2}\right ) \) be such that \(\widetilde {\alpha } _{2}\succcurlyeq _{N}\widetilde {\alpha }_{1}\) and let \(q(\alpha )=g(x_{1}^{\ast },\alpha ,\mu ).\) We have \(\frac {\partial ^{N}g}{\partial \alpha ^{N}}(x_{1}^{\ast },\alpha ,\mu )=-pU^{( 1,N) }\left ( K-x_{1}^{\ast }p,x_{1}^{\ast }\mu +\alpha \right ) +\mu U^{\left ( 0,N+1\right ) }( K-x_{1}^{\ast }p,x_{1}^{\ast }\mu +\alpha ) .\) By 2., we have \(( -1)^{N}\frac {\partial ^{N}g}{\partial \alpha ^{N}} (x_{1}^{\ast },\alpha ,\mu )\geq 0\) and ( − 1)N q (N)(α) ≥ 0. Since \(\widetilde {\alpha }_{i}\) is bounded above ( i = 1, 2), \( x_{1}^{\ast }\mu +\widetilde {\alpha }_{i}\) is bounded and bounded away from zero and U and its N + 1 first derivatives are bounded on the convex hull of the set of values taken by \(\left ( K-x_{1}^{\ast }p,x_{1}^{\ast }\mu + \widetilde {\alpha }_{i}\right ) \). The same applies to q on the convex hull of the set of values taken by \(\widetilde {\alpha }_{i}\), i = 1, 2. By Lemma 1, this leads to \(E[ q(\widetilde {\alpha }_{1})] \leq E [ q(\widetilde {\alpha }_{2})] .\) By definition, we have \(E\left [ q(\widetilde {\alpha }_{1})\right ] =0,\) which gives \(E\left [ q(\widetilde { \alpha }_{2})\right ] \geq 0\) or \(E\left [ g(x_{1}^{\ast },\widetilde {\alpha } _{2},\mu )\right ] \geq 0.\) By concavity of U, it is easy to check that \(E [ g(x,\widetilde {\alpha }_{2},\mu )] \) is a decreasing function of x. Since \(x_{2}^{\ast }\) is characterized by \(E\left [ g(x_{2}^{\ast }, \widetilde {\alpha }_{2},\mu )\right ] =0,\) we obtain that \(x_{2}^{\ast }\geq x_{1}^{\ast }.\)
Let us prove 1. implies 2. As in the proof of Lemma 1, we consider \(\widetilde {\beta }_{1}\) and \(\widetilde {\beta }_{2}\) two nonnegative random variables with a common upper bound B such that \(\widetilde {\beta } _{2}\succcurlyeq _{N}\widetilde {\beta }_{1}.\) As above, we introduce the random variables \(\widetilde {\alpha }_{1,\varepsilon }=\frac {\varepsilon }{B} \widetilde {\beta }_{1}+\alpha ^{\ast }\) and \(\widetilde {\alpha } _{2,\varepsilon }=\frac {\varepsilon }{B}\widetilde {\beta }_{2}+\alpha ^{\ast }\) for some α ∗ > 0 and some ε > 0. The random variables \(\widetilde {\alpha }_{1,\varepsilon }\) and \(\widetilde {\alpha } _{2,\varepsilon }\) take their values in [α ∗, α ∗ + ε] and \(\widetilde {\alpha }_{2,\varepsilon }\succcurlyeq _{N}\widetilde {\alpha }_{1,\varepsilon }.\) Let us consider a given real number γ and let us define the random variables \( \widetilde {\gamma }_{i,\varepsilon }\), i = 1, 2, as lotteries giving \( \widetilde {\alpha }_{i,\varepsilon }\) with probability ε and γ with probability 1 − ε. The constant γ provides an additional degree of freedom that will prove useful in order to control the pair \(( K-x_{1}^{\ast }p,x_{1}^{\ast }\mu +\alpha ) .\) We have \(\widetilde {\gamma }_{2,\varepsilon }\succcurlyeq _{N}\widetilde {\gamma }_{1,\varepsilon }\) as an immediate consequence of the stability of this order relation under probability mixtures. Let \(x_{1,\varepsilon }^{\ast }\) and \(x_{2,\varepsilon }^{\ast }\) be respectively the solutions of \(P_{U,K,p}( \widetilde {\gamma }_{1,\varepsilon },\mu )\) and \(P_{U,K,p}(\widetilde {\gamma }_{2,\varepsilon },\mu ).\) By 1., we have \(x_{2,\varepsilon }^{\ast }\geq x_{1,\varepsilon }^{\ast }.\) By definition, we have \(E\left [ g(x_{1,\varepsilon }^{\ast },\widetilde {\gamma }_{1,\varepsilon },\mu ) \right ] =0\) and \(E\left [ g(x_{2,\varepsilon }^{\ast },\widetilde {\gamma } _{2,\varepsilon },\mu )\right ] =0.\) By concavity of U, g is a decreasing function of x and we have \(E\left [ g(x_{1,\varepsilon }^{\ast },\widetilde { \gamma }_{2,\varepsilon },\mu )\right ] \geq 0.\) Let q be defined by \( q(\alpha )=g(x_{1,\varepsilon }^{\ast },\alpha ,\mu ).\) We have
By construction, the left side of the equality is nonnegative, all the terms in the sum are equal to zero and \(F_{\widetilde {\alpha }_{2},P}^{\left [ N \right ] }(t)-F_{\widetilde {\alpha }_{1},P}^{[ N] }(t)\) is nonnegative and nonzero. Note that \(\widetilde {\gamma }_{i,\varepsilon }\) is bounded and bounded away from zero for i = 1, 2. Therefore q and its first N derivatives are bounded on the convex hull of the set of values taken by \( \widetilde {\gamma }_{i,\varepsilon }\), i = 1, 2. Therefore, ( − 1)N q (N)(t) is nonnegative at least on a given subinterval of α ∗, α ∗ + ε. Let \(\alpha _{\varepsilon }^{\ast }\ \)be in α ∗, α ∗ + ε such that \((-1)^{N}q^{(N)}(\alpha _{\varepsilon }^{\ast })\geq 0.\) We have then
where \(x_{1,\varepsilon }^{\ast }\) satisfies
or
Remark that, until now, α ∗, K and γ have been arbitrarily chosen. Let us now choose them carefully in order to derive our result. Let (Y, Z) be arbitrary in \(( \mathbb {R}_{+}^{\ast })^{2}\). By our Inada condition, \(\lim _{z\rightarrow 0}\frac {U^{\left ( 0,1\right ) }( Y,z) }{U^{( 1,0) }( Y,z) } =\infty \) and \(\lim _{z\rightarrow \infty }\frac {U^{( 0,1) }\left ( Y,z\right ) }{U^{( 1,0) }( Y,z) }=0\) which gives that there exists some z ∗ > 0 such that \(\frac {U^{( 0,1) }( Y,z^{\ast }) }{U^{( 1,0) }( Y,z^{\ast }) }=\frac {p}{\mu }\) or − p U (1, 0)(Y, z ∗) + μ U (0, 1)(Y, z ∗) = 0. We choose x ∗ > 0 such that μ x ∗ < inf(Z, z ∗) and α ∗ and K are taken such that α ∗ = Z − μ x ∗ > 0 and K = Y + p x ∗. Let γ be given by γ = z ∗ − μ x ∗ > 0. We have
The solution of Eq. 7 for ε = 0 is then given by x ∗. In a well chosen neighborhood of x ∗, x ↦ K − x p and \(x\longmapsto x\mu +\frac {\varepsilon }{B}\widetilde {\beta }_{1}+\alpha ^{\ast }\) are bounded and bounded away from 0. The functions U 1, 0 and U (0, 1) being continuously differentiable, the function \((x,\varepsilon )\longmapsto \varepsilon E\left [ g(x,\frac {\varepsilon }{B}\widetilde {\beta }_{1}+\alpha ^{\ast },\mu )\right ] +(1-\varepsilon )g(x,\widetilde {\gamma },\mu )\) is then differentiable with respect to x at (x ∗, 0). Furthermore, the derivative of this last function with respect to x at (x ∗, 0) is nonzero (concavity of U ). The solution \(x_{1,\varepsilon }^{\ast }\) of Eq. 7 is then continuous with respect to ε in a neighborhood of 0 which gives \(\lim _{\varepsilon \rightarrow 0}x_{1,\varepsilon }^{\ast }=x^{\ast }.\) Furthermore, we clearly have \(\lim _{\varepsilon \rightarrow 0}\alpha _{\varepsilon }^{\ast }=\alpha ^{\ast }.\) Taking the limit in Eq. 6 when ε tends to 0, we obtain
or, by construction
1.2 Proof of Proposition 2
Let us prove that 2. implies 1. Let \(\left (\widetilde {\mu }_{1},\widetilde { \mu }_{2}\right ) \) be such that \(\widetilde {\mu }_{2}\succcurlyeq _{N} \widetilde {\mu }_{1}\) and let \(q(\mu )=g(x_{1}^{\ast },\alpha ,\mu ).\) We have \(q^{(N)}(\mu )=-p(x_{1}^{\ast })^{N}U^{\left ( 1,N\right )}(K-x_{1}^{\ast }p,x_{1}^{\ast }\mu +\alpha )+\mu (x_{1}^{\ast })^{N}U^{(0,N+1)}\left (K-x_{1}^{\ast }p,x_{1}^{\ast }\mu +\alpha \right )+N(x_{1}^{\ast })^{N-1}U^{(0,N)}\left (K-x_{1}^{\ast }p,x_{1}^{\ast }\mu +\alpha \right ) \). Since we assumed that \(x_{1}^{\ast }\geq 0\), by 2, we have
\((-1)^{N}p(x_{1}^{\ast })^{N}U^{(1,N) }(K-x_{1}^{\ast }p,x_{1}^{\ast }\mu +\alpha )\leq 0\) and
which gives ( − 1)N q (N)(μ) ≥ 0. Since \( \widetilde {\mu }_{i}\) is bounded above ( i = 1, 2), \(x_{1}^{\ast }\widetilde { \mu }_{i}+\alpha \) is bounded and bounded away from zero and U and its N + 1 first derivatives are bounded on the convex hull of the set of values taken by \(\left ( K-x_{1}^{\ast }p,x_{1}^{\ast }\widetilde {\mu }_{i}+\alpha \right ) \). The same applies to q on the convex hull of the set of values taken by \(\widetilde {\mu }_{i}\), i = 1, 2. By Lemma 1, this leads to \(E[ q(\widetilde {\mu }_{1})] \leq E\left [ q(\widetilde {\mu } _{2})\right ] .\) By definition, we have \(E\left [ q(\widetilde {\mu }_{1}) \right ] =0\) which gives \(E[ q(\widetilde {\mu }_{2})] =E\left [ g(x_{1}^{\ast },\alpha ,\widetilde {\mu }_{2})\right ] \geq 0.\) By concavity of U, it is easy to check that g(x, α, μ) is a decreasing function of x. From there we derive that \(x_{2}^{\ast }\geq x_{1}^{\ast }.\)
Let us prove that 1. implies 2. As in the proof of Lemma 1 and Proposition 1, we consider \(\widetilde {\beta }_{1}\) and \( \widetilde {\beta }_{2}\) two nonnegative random variables with a common upper bound B such that \(\widetilde {\beta }_{2}\succcurlyeq _{N}\widetilde { \beta }_{1}.\) As above, we introduce the random variables \(\widetilde {\mu } _{1,\varepsilon }=\frac {\varepsilon }{B}\widetilde {\beta }_{1}+\mu ^{\ast }\) and \(\widetilde {\mu }_{2,\varepsilon }=\frac {\varepsilon }{B}\widetilde { \beta }_{2}+\mu ^{\ast }\) for some μ ∗ > 0 and some ε > 0. The random variables \(\widetilde {\mu }_{1,\varepsilon }\) and \( \widetilde {\mu }_{2,\varepsilon }\) take their values in [μ ∗, μ ∗ + ε] and \(\widetilde {\mu }_{2,\varepsilon }\succcurlyeq _{N}\widetilde {\mu }_{1,\varepsilon }.\) Let us consider a given γ and let us define the random variables \(\widetilde {\gamma } _{i,\varepsilon }\), i = 1, 2, as a lottery that takes the value \(\widetilde { \mu }_{i,\varepsilon }(\omega )\) with probability ε and the value γ with probability 1 − ε. As previously, we have \( \widetilde {\gamma }_{2,\varepsilon }\succcurlyeq _{N}\widetilde {\gamma } _{1,\varepsilon }.\) Let \(x_{1,\varepsilon }^{\ast }\) and \(x_{2,\varepsilon }^{\ast }\) be respectively the solutions of \(P_{U,K,p}(\alpha ,\widetilde { \gamma }_{1,\varepsilon })\) and \(P_{U,K,p}(\alpha ,\widetilde {\gamma } _{2,\varepsilon }).\) If \(x_{1,\varepsilon }^{\ast }\) is nonnegative then, by 1., we have \(x_{2,\varepsilon }^{\ast }\geq x_{1,\varepsilon }^{\ast }.\) By definition, we have \(E\left [ g(x_{1,\varepsilon }^{\ast },\alpha ,\widetilde { \gamma }_{1,\varepsilon })\right ] =0\) and \(E\left [ g(x_{2,\varepsilon }^{\ast },\alpha ,\widetilde {\gamma }_{2,\varepsilon })\right ] =0.\) By concavity of U, g is a decreasing function of x and we have \(E\left [ g(x_{1,\varepsilon }^{\ast },\alpha ,\widetilde {\gamma }_{2,\varepsilon }) \right ] \geq 0.\) Let q be defined by \(q(\mu )=g(x_{1,\varepsilon }^{\ast },\alpha ,\mu ).\) We have
By construction, the left side of the equality is nonnegative, all the terms in the sum are equal to zero and \(F_{\widetilde {\mu }_{2},P}^{[ N] }(t)-F_{\widetilde {\mu }_{1},P}^{[ N] }(t)\) is nonnegative and nonzero. Note that \(\widetilde {\gamma }_{i,\varepsilon }\) is bounded and bounded away from zero for i = 1, 2. Therefore q and its first N derivatives are bounded on the convex hull of the set of values taken by \( \widetilde {\gamma }_{i,\varepsilon }\), i = 1, 2. Therefore, ( − 1)N q (N) is nonnegative at least on a given subinterval of μ ∗, μ ∗ + ε. Let \(\mu _{\varepsilon }^{\ast }\ \)be in μ ∗, μ ∗ + ε such that \( (-1)^{N}q^{(N)}(\mu _{\varepsilon }^{\ast })\geq 0.\) We have then
where \(x_{1,\varepsilon }^{\ast }\) satisfies
or
Remark that, until now p, μ ∗, K, α and γ have been arbitrarily chosen. Let us now choose them carefully in order to derive our result. We assume first that γ is equal to 1. Let x ∗ > 0 be given and let (M, Y, Z) in \(( \mathbb {R}_{+})^{3} \) such that M < Z. We take \(p=\frac {U^{( 0,1) }(Y,x^{\ast }+Z-M)}{U^{( 1,0) }(Y,x^{\ast }+Z-M)}\), K = Y + p x ∗, α = Z − M and \(\mu ^{\ast }=\frac {M}{x^{\ast }}.\) By construction, we have α > 0 and we have
The solution of Eq. 9 for ε = 0 is then given by x ∗. In a well chosen neighborhood of x ∗, x ↦ K − x p and \(x\longmapsto x\left (\mu ^{\ast }+\frac {\varepsilon }{B}\widetilde { \beta }_{1}\right ) +\alpha \) are bounded and bounded away from 0. The functions U (1, 0) and U (0, 1) being continuously differentiable, the function \((x,\varepsilon )\longmapsto \varepsilon E\left [ g(x,\alpha ,\frac {\varepsilon }{B}\widetilde {\beta }_{1}+\mu ^{\ast })\right ] +(1-\varepsilon )g(x,\alpha ,\gamma )\) is then differentiable with respect to x at (x ∗, 0). Furthermore, the derivative with respect to x at (x ∗, 0) is nonzero. The solution \( x_{1,\varepsilon }^{\ast }\) of Eq. 9 is then continuous with respect to ε in a neighborhood of 0 which gives \( \lim _{\varepsilon \rightarrow 0}x_{1,\varepsilon }^{\ast }=x^{\ast }\) and guarantees that \(x_{1,\varepsilon }^{\ast }>0\) for ε small enough . Since we clearly have \(\lim _{\varepsilon \rightarrow 0}\mu _{\varepsilon }^{\ast }=\mu ^{\ast },\) taking the limit in Eq. 8 when ε tends to 0, we obtain
or
This result being true for all (Y, Z) in \(\left (\mathbb {R}_{+}^{\ast }\right )^{2},\) all M ∈ (0, Z) all x ∗ > 0 and for \(p=\frac {U^{(0,1)}(Y,x^{\ast }+Z-M)}{U^{(1,0)}(Y,x^{\ast }+Z-M)}.\) When x ∗goes to 0, Y, M and Z being fixed, \(x^{\ast }\frac {U^{(0,1)}\left (Y,x^{\ast }+Z-M\right )}{U^{( 1,0) }\left (Y,x^{\ast }+Z-M\right )}\) goes to 0 and we obtain that
for all (Y, Z) in \(( \mathbb {R}_{+}^{\ast })^{2}\) and all M ∈ (0, Z) or equivalently
We assumed that \(z\frac {U^{( 0,1) }(Y,z)}{U^{( 1,0)}(Y,z)}\) is unbounded. We then have either \(\lim _{z\rightarrow 0}z\frac { U^{( 0,1) }(Y,z)}{U^{( 1,0) }(Y,z)}=\infty \) or \( \lim _{z\rightarrow \infty }z\frac {U^{( 0,1) }(Y,z)}{U^{\left ( 1,0\right ) }(Y,z)}=\infty \). If \(\lim _{z\rightarrow 0}z\frac {U^{\left ( 0,1\right ) }(Y,z)}{U^{( 1,0) }(Y,z)}=\infty \), it suffices to take \(x^{\ast }=\frac {\zeta }{2}\) and \(M=Z-\frac {\zeta }{2}\) for ζ arbitrarily small to make the quantity \(px^{\ast }=\frac {U^{\left ( 0,1\right ) }(Y,x^{\ast }+Z-M)}{U^{( 1,0) }(Y,x^{\ast }+Z-M)} x^{\ast }=\frac {1}{2}\zeta \frac {U^{( 0,1) }(Y,\zeta )}{U^{\left ( 1,0\right ) }(Y,\zeta )}\) arbitrarily large. Since M, Y and Z are bounded, Eq. 10 gives then ( − 1)N U (1, N)(Y, Z) ≤ 0. If \(\lim _{z\rightarrow \infty }z\frac {U^{\left ( 0,1\right ) }(Y,z)}{U^{( 1,0) }(Y,z)}=\infty ,\) it suffices to take x ∗sufficiently large to make the quantity x ∗ + Z − M sufficiently large and \(\frac {U^{( 0,1) }(Y,x^{\ast }+Z-M)}{ U^{( 1,0) }(Y,x^{\ast }+Z-M)}( x^{\ast }+Z-M) \) arbitrarily large. Since Z is kept fixed, \(\frac {x^{\ast }}{x^{\ast }+Z-M}\) is arbitrarily close to 1 and p x ∗arbitrarily large. Since M, Y and Z are bounded, Eq. 10 gives then ( − 1)N U (1, N)(Y, Z) ≤ 0.The Proof of Corollary 1 follows directly by reversing the inequalities above.
1.3 Proof of Proposition 3
In an expected utility framework, Nth-degree risk aversion in the direction of (ρ y , ρ z ) is equivalent to
for all \(( y,z,x_{1},x_{2}) \in \mathbb {R}_{+}^{4}\) with x 2 > x 1. The previous inequation is satisfied for all x 2 > x 1 if and only if \(E\left [ U\left ( y+t\rho _{y},z+t\rho _{z}+\widetilde {\alpha } _{2}\right ) \right ] -E\left [ U\left ( y+t\rho _{y},z+t\rho _{z}+\widetilde { \alpha }_{1}\right ) \right ] \) is increasing in t or, equivalently, if and only if \(E\left [ f\left ( y+t\rho _{y},z+t\rho _{z}+\widetilde {\alpha } _{2}\right ) \right ] \) is larger than \(E\left [ f\left ( y+t\rho _{y},z+t\rho _{z}+\widetilde {\alpha }_{1}\right ) \right ] \) for all \((y,z,t)\in \mathbb {R} _{+}^{3}\). This is satisfied if and only if \(E\left [ f\left ( y,z+\widetilde { \alpha }_{2}\right ) \right ] \) is larger than \(E\left [ f\left ( y,z+\widetilde { \alpha }_{1}\right ) \right ] \) for all \((y,z)\in \mathbb {R}_{+}^{2}.\) Since we want this inequality to be true for all \((\widetilde {\alpha }_{1}, \widetilde {\alpha }_{2})\) such that \(\widetilde {\alpha }_{2}\) is an increase in Nth-degree risk over \(\widetilde {\alpha }_{1},\) this inequality is equivalent, by Lemma 1, to ( − 1)N f (0, N) ≥ 0.
1.4 Proof of Proposition 4
In an expected utility model, Nth-degree multiplicative-risk attraction in the direction of (ρ y , ρ z ) is equivalent to
for all \((\widetilde {\alpha }_{1},\widetilde {\alpha }_{2})\) such that \( \widetilde {\alpha }_{1}\preccurlyeq _{N}\widetilde {\alpha }_{2}\) and all (x 1, x 2) such that 0 < x 1 < x 2 and y + x i ρ y ≥ 0, i = 1, 2. This inequality is satisfied for all x 2 > x 1 if and only if \(E \left [ U( y+t\rho _{y},z+t\rho _{z}\widetilde {\alpha }_{2}) \right ] -E\left [ U\left ( y+t\rho _{y},z+t\rho _{z}\widetilde {\alpha } _{1}\right ) \right ] \) is increasing on \(T(y)=\left \{t\in \mathbb {R} _{+}:y+t\rho _{y}\geq 0\right \} \) or, equivalently, if and only if \(E\left [ h(\widetilde {\alpha }_{2},y,z,t) \right ] \) is larger than \(E [ h( \tilde {\alpha }_{1},y,z,t) ] \) for all \((y,z)\in \mathbb {R}_{+}^{2}\) and for t ∈ T(y) where h(α, y, z, t) = ρ y U (1, 0) (y + t ρ y , z + t ρ z α) + α ρ z U (0, 1) (y + t ρ y , z + t ρ z α). By Lemma 1, this last inequality is satisfied for every pair \((\widetilde {\alpha }_{1},\widetilde {\alpha }_{2})\) such that \(\widetilde {\alpha }_{2}\) is an increase in Nth-degree risk over \( \widetilde {\alpha }_{1}\) if and only if \((-1)^{N}\frac {\partial ^{N}h}{\partial \alpha ^{N}}\geq 0.\)
Simple calculus gives \(\frac {\partial ^{N}h}{\partial \alpha ^{N}}=t^{N}\rho _{y}\rho _{z}^{N}U^{(1,N)}+\alpha t^{N}\rho _{z}^{N+1}U^{\left ( 0,N+1\right ) }+Nt^{N-1}\rho _{z}^{N}U^{( 0,N) }\) where derivatives of U are taken at (y + t ρ y , z + α t ρ z ). Since t ≥ 0 and ρ z ≥ 0, our necessary and sufficient condition can then be rewritten as ( − 1)N t ρ y U (1, N) + α t ρ z U (0, N + 1) + N U (0, N) ≥ 0 for all \((\alpha ,y,z)\in \mathbb {R}_{+}^{3}\ \)and for t ∈ T(y). Denoting y + t ρ y by Y and z + α t ρ z by Z, our necessary and sufficient condition is equivalent to ( − 1)N(Y − y)U (1, N) + (Z − z)U (0, N + 1) + N U (0, N) ≥ 0 for all \((Y,Z)\in \mathbb {R}_{+}^{2}\) and all (y, z) such that y ≥ Y and 0 ≤ z ≤ Z and where the derivatives are taken at (Y, Z). Taking successively y = Y and z = Z, y = Y and z = 0 we obtain that (− 1)N U (0, N) ≥ 0 and ( − 1)N Z U (0, N + 1) + N U (0, N) ≥ 0. Letting y go to ∞, we obtain ( − 1)N U (1, N) ≤ 0. Conversely, if these 3 conditions are satisfied and for a given y ≥ Y, we have ( − 1)N(Y − y)U (1, N) ≥ 0 and
which gives that our necessary and sufficient condition is satisfied.
A far as Nth-degree multiplicative-risk aversion is concerned, it is characterized by the fact that \(E\left [ U\left ( y+t\rho _{y},z+t\rho _{z} \widetilde {\alpha }_{2}\right ) \right ] -E\left [ U\left ( y+t\rho _{y},z+t\rho _{z}\widetilde {\alpha }_{1}\right ) \right ] \) is decreasing on \(T(y)=\left \{ t\in \mathbb {R}_{+}:y+t\rho _{y}\geq 0\right \} \) or, equivalently, by the fact that \(E[ -h( \widetilde {\alpha }_{2},y,z,t) ] \) is larger than \(E[ -h( \tilde {\alpha }_{1},y,z,t) ] \) for all \((y,z)\in \mathbb {R}_{+}^{2}\) and for t ∈ T(y) or, finally, by \( ( -1)^{N+1}\frac {\partial ^{N}h}{\partial \alpha ^{N}}\geq 0.\) The rest of the proof is then directly adapted from the multiplicative-risk attraction setting.
Appendix B: Nth-degree stochastic dominance
In this appendix we generalize Lemma 1, Propositions 1 and 2, and Corollary 1. We begin by formalizing the concept of Nth-degree stochastic dominance.
Definition 4 (Jean)
Let \(\widetilde {\alpha }_{1}\) and \(\widetilde {\alpha }_{2}\) denote two random variables with values in [0, B]. We say that \( \widetilde {\alpha }_{2}\) is dominated in the sense of Nth-degree stochastic dominance by \(\widetilde {\alpha }_{1},\) and we denote it by \(\widetilde { \alpha }_{2}\succcurlyeq _{NSD}\widetilde {\alpha }_{1},\) if \(F_{\widetilde { \alpha }_{2}}^{[ N] }(x)\geq F_{\widetilde {\alpha }_{1}}^{\left [ N \right ] }(x)\) for all x ∈ [0, B] where the inequality is strict for some x and \(F_{\widetilde {\alpha }_{2}}^{[ k] }(B)\geq F_{\widetilde {\alpha }_{1}}^{[ k] }(B)\) for k = 1, . . . , N − 1.
Jean (1980) characterizes Nth-degree stochastic dominance: He establishes that \(\widetilde {\alpha }_{2}\) is dominated in the sense of Nth-degree stochastic dominance by \(\widetilde {\alpha }_{1}\) if and only if \(E\left [ q( \widetilde {\alpha }_{2})\right ] >E[ q(\widetilde {\alpha }_{1})] \) for all N times continuously differentiable real valued function q such that ( − 1)k q (k) > 0 for k = 1, . . . , N where \(q^{(k)}=\frac { d^{k}q}{d\alpha ^{k}}.\) The following Lemma characterizes the set of N times continuously differentiable functions for which \(E\left [ q(\widetilde { \alpha }_{2})\right ] \geq E[ q(\widetilde {\alpha }_{1})] \) for all pair \((\widetilde {\alpha }_{1},\widetilde {\alpha }_{2})\) where \(\widetilde {\alpha }_{2}\succcurlyeq _{NSD}\widetilde {\alpha }_{1}.\)
Lemma 2
Let q be a given real valued function that is N times continuously differentiable on ℝ+. The following are equivalent.
-
1.
For all pair \((\widetilde {\alpha }_{1},\widetilde {\alpha }_{2})\) such that \(\widetilde {\alpha }_{2}\succcurlyeq _{NSD}\widetilde {\alpha }_{1},\) we have \(E[ q(\widetilde {\alpha }_{2})] \geq E\left [ q(\widetilde { \alpha }_{1})\right ] .\)
-
2.
For all x ≥ 0, we have ( − 1)k q (k) ≥ 0 for k = 1, . . . , N.
Proof
The fact that 2. implies 1. results directly from Jean (1980). Let us now prove that 1. implies 2. Let q be an N times continuously differentiable function on \(\mathbb {R}_{+}\) such that \(E\left [ q(\widetilde {\alpha }_{2})\right ] \geq E\left [ q(\widetilde {\alpha }_{1}) \right ] \) for all pair \((\widetilde {\alpha }_{1},\widetilde {\alpha }_{2})\) such that \(\widetilde {\alpha }_{2}\succcurlyeq _{NSD}\widetilde {\alpha } _{1}. \) Let k be in {1, . . . , N} and let \((\widetilde {\alpha }_{1},\widetilde {\alpha }_{2})\) such that \(\widetilde {\alpha } _{2}\succcurlyeq _{k}\widetilde {\alpha }_{1}.\) By definition, we have \(F_{ \widetilde {\alpha }_{2}}^{[ k] }(x)\geq F_{\widetilde {\alpha } _{1}}^{[ k] }(x)\) for all x ∈ [0, B] where the inequality is strict for some x and \(F_{\widetilde {\alpha }_{2}}^{\left [ i \right ] }(B)=F_{\widetilde {\alpha }_{1}}^{[ i] }(B)\) for i = 1, . . . , k − 1. By integration, we obtain \(F_{\widetilde {\alpha }_{2}}^{\left [ j\right ] }(x)\geq F_{\widetilde {\alpha }_{1}}^{[ j] }(x)\) for all j ≥ k and all x ∈ [0, B] and, in particular, \( \widetilde {\alpha }_{2}\succcurlyeq _{NSD}\widetilde {\alpha }_{1}\) hence \(E [ q(\widetilde {\alpha }_{2})] \geq E\left [ q(\widetilde {\alpha } _{1})\right ] .\) By Lemma 2 we have ( − 1)k q (k) ≥ 0 on 0, B for k = 1, . . . , N. □
The following Proposition establishes the effect on optimal choice of an Nth-degree stochastically dominated shift on the initial endowment α.
Proposition 5
Let U be a given increasing, strictly concave and infinitely differentiable utility function satisfying Assumption A1. Let us consider p and μ as given. The following properties are equivalent:
-
1.
For all initial endowment \((K,\widetilde {\alpha }_{1})\), a switch from \(\widetilde {\alpha }_{1}\) to \(\widetilde {\alpha }_{2}\) such that \(\widetilde {\alpha }_{2}\succcurlyeq _{NSD}\widetilde {\alpha }_{1}\) increases the optimal level of the choice variable, i.e., \(x_{2}^{\ast }\geq x_{1}^{\ast }\).
-
2.
For all (y, z), we have ( − 1)k (− p U (1, k)(y, z) + μ U (0, k + 1)(y, z)) ≥ 0 for k = 1, . . . , N.
Proof
Let us prove that 2. implies 1. Let \(\left ( \widetilde { \alpha }_{1},\widetilde {\alpha }_{2}\right ) \) be such that \(\widetilde { \alpha }_{2}\succcurlyeq _{NSD}\widetilde {\alpha }_{1}\) and let \(q(\alpha )=g(x_{1}^{\ast },\alpha ,\mu ).\) By 2., we have ( − 1)k q (k)(α) ≥ 0 for k = 1, . . . , N. By Lemma 2, this leads to \(E[ q(\widetilde {\alpha }_{1})] \leq E\left [ q( \widetilde {\alpha }_{2})\right ] .\) By definition, we have \(E\left [ q( \widetilde {\alpha }_{1})\right ] =0,\) which gives \(E\left [ q(\widetilde { \alpha }_{2})\right ] \geq 0\) or \(E\left [ g(x_{1}^{\ast },\widetilde {\alpha } _{2},\mu )\right ] \geq 0.\) By concavity of U, it is easy to check that \(E [ g(x,\widetilde {\alpha }_{2},\mu )] \) is a decreasing function of x. Since \(x_{2}^{\ast }\) is characterized by \(E\left [ g(x_{2}^{\ast }, \widetilde {\alpha }_{2},\mu )\right ] =0,\) we obtain that \(x_{2}^{\ast }\geq x_{1}^{\ast }.\) □
Let us prove 1. implies 2. As in the proof of the Lemma 2, it suffices to remark that, for k = 1, . . . , N, \(\widetilde {\alpha } _{2}\succcurlyeq _{k}\widetilde {\alpha }_{1}\) implies \(\widetilde {\alpha } _{2}\succcurlyeq _{NSD}\widetilde {\alpha }_{1}.\)
Finally, the following Proposition establishes the effect on optimal choice of an Nth-degree stochastically dominated shift on the multiplicative variable μ.
Proposition 6
Let U be a given increasing, strictly concave and infinitely differentiable utility function satisfying Assumption A2. The following properties are equivalent:
-
1.
For all initial endowment (K, α) and all asset cost and payoff \( (p,\widetilde {\mu }_{1})\) such that \(x_{1}^{\ast }\geq 0\), and any switch from \(\widetilde {\mu }_{1}\) to \(\widetilde {\mu }_{2}\) with \(\widetilde {\mu } _{2}\succeq _{NSD}\widetilde {\mu }_{1}\) increases (decreases) the optimal level of the choice variable, i.e., \(x_{2}^{\ast }\geq x_{1}^{\ast }\).
-
2.
For all (y, z), we have ( − 1)k U 1, k(y, z) ≤ 0(r e s p. ≥ 0)R k(y, z) ≤ N and( − 1)k U 0, k(y, z) ≥ 0(r e s p. ≤ 0) for k = 1, . . . , N.
Proof
Follows the proof of Proposition 5. □
Rights and permissions
About this article
Cite this article
Jouini, E., Napp, C. & Nocetti, D. Economic consequences of Nth-degree risk increases and Nth-degree risk attitudes. J Risk Uncertain 47, 199–224 (2013). https://doi.org/10.1007/s11166-013-9176-6
Published:
Issue Date:
DOI: https://doi.org/10.1007/s11166-013-9176-6