Precautionary Saving against Correlation under Risk and Ambiguity
What this paper finds — and why it matters
Layer 1: Overview
Research question and motivation: How much to save is a central household financial decision, and uncertainty drives the “precautionary saving motive.” The precautionary-saving literature has mostly studied one-dimensional (single-attribute) risk, yet households face multidimensional risk: both wealth and health conditions matter for saving. Because wealth and health are plausibly related, the authors argue the correlation between two risky attributes should be incorporated into precautionary-saving analysis. They further note that correlation between two attributes is harder to quantify than a single attribute’s risk (less experience, fewer observations), so they also introduce ambiguity about the correlation. The paper’s purpose is to characterize how the correlation between two risky attributes (wealth and health) affects optimal savings under multivariate preferences, both when correlation is known (risk) and when it is ambiguous.
Model setup: A purely theoretical two-date model (t=0, t=1). The individual has time-separable lifetime utility from a bivariate utility function u(x,y) over wealth x and health y, increasing and concave in both (u^(1,0)>=0, u^(0,1)>=0, u^(2,0)<=0, u^(0,2)<=0); the sign of the cross derivative u^(1,1) is left unrestricted. The risk-free interest rate is zero and there is no time discounting, so the analysis isolates the effect of risk on saving. At t=1 the individual faces “good” and “bad” income risks (epsilon_G, epsilon_B occurring with probabilities 1-p, p) and “good”/“bad” health risks (delta_G, delta_B with probabilities 1-q, q), all four mutually independent. Correlation between income and health risk is captured by a parameter k: the probability of simultaneous bad income and bad health is kpq. When k=1 the risks are independent (joint probability = pq); k>1 (k<1) indicates positive (negative) correlation; correlation increases in k. The individual chooses saving s to maximize lifetime utility (equation 1). “Good” vs “bad” risks are ranked by stochastic dominance (FSD, Nth-order NSD, and Ekern’s Nth-degree risk increase).
Main findings (theoretical propositions, no estimated magnitudes): (1) Proposition 1 — when income risk is ranked by Nth-order and health risk by Mth-order stochastic dominance, optimal savings increase (decrease) in correlation k if (-1)^(n+m) u^(n+1,m)(x,y) >= (<=) 0 for n=1..N, m=1..M. This condition defines “mixed correlation aversion (seeking).” In the special case N=M=1, optimal savings increase in k if u^(2,1)>=0, i.e., the individual is “cross prudent” (decrease if cross imprudent, u^(2,1)<=0). Intuition: cross-prudent individuals dislike the simultaneous occurrence of bad income and bad health, which becomes more likely as k rises, so they save more. (2) Proposition 2 (ambiguous correlation, smooth ambiguity model of Klibanoff et al. 2005, 2009) — if the second-order utility phi exhibits decreasing absolute ambiguity aversion (DAAA) and u exhibits mixed correlation aversion or seeking, then ambiguous correlation raises the optimal amount of savings relative to the risky benchmark with correlation k_O = sum q_theta k_theta. The result combines a “timing of uncertainty effect” (governed by beta(s_O)>=1 iff phi exhibits DAAA) and the sign of a covariance term. (3) Proposition 3 extends the same result to Nth-/Mth-degree risk increases: under DAAA and (-1)^(N+M) u^(N,M)>=(<=)0 and (-1)^(N+M) u^(N+1,M)>=(<=)0, ambiguous correlation raises savings.
Implications: Whether correlation raises or lowers precautionary saving depends entirely on the signs of higher-order cross derivatives of utility, and under ambiguity additionally on the absolute-ambiguity-aversion coefficient. The authors link results to experimental evidence (Attema et al. 2019 find both cross prudence and imprudence; correlation aversion in gains, seekingness in losses) and to empirical work on public health systems, which by changing the wealth-health correlation affect precautionary saving (e.g., Rosen and Wu 2004; Atella et al. 2012; Chou et al. 2003; Jappelli et al. 2007), broadly consistent with cross prudence.
Layer 2: Deep Dive
What is the core mechanism linking correlation to saving, and how is it formalized?
Correlation between income and health risk is parameterized by a single scalar k that scales the joint probability of the simultaneous bad outcome to kpq (with k=1 = independence, k>1 = positive correlation, k<1 = negative correlation), following the representation of Doherty and Schlesinger (1990). The derivative of expected period-1 utility with respect to k reduces (Lemma 1) to pq times [E[f(eps_B,del_B)] - E[f(eps_G,del_B)] - E[f(eps_B,del_G)] + E[f(eps_G,del_G)]], so the sign of the response to correlation is governed by a cross-difference whose sign maps directly onto the signs of higher-order cross derivatives of u. As k rises, the simultaneous occurrence of two bad outcomes becomes more likely; agents who dislike that combination (mixed correlation averse / cross prudent) save more to protect against it.
What exactly is ‘mixed correlation aversion (seeking)’ and how does it relate to correlation aversion and cross prudence?
An individual is mixed correlation averse (seeking) if (-1)^(n+m+1) u^(n,m)(x,y) >= (<=) 0 for all n=1..N, m=1..M. It is a bivariate extension of Caballe and Pomansky’s (1996) univariate mixed risk aversion, and generalizes Epstein and Tanny’s (1980) correlation aversion (which corresponds to u^(1,1)<=0). Cross prudence (u^(2,1)>=0, per Eeckhoudt et al. 2007) is the third-order version of correlation aversion. The paper’s saving conditions use mixed correlation aversion (seekingness) excluding the second-order correlation-aversion term, expressed via the derivative pattern (-1)^(n+m) u^(n+1,m) >= (<=) 0.
How is the ‘good’ vs ‘bad’ ranking of risks made rigorous?
Through stochastic dominance. eps_G dominates eps_B in the sense of Nth-order stochastic dominance (NSD) iff E[u(w+eps_G,h)]>=E[u(w+eps_B,h)] for all u with (-1)^(n+1) u^(n,0)>=0, n=1..N (mixed risk aversion in wealth); analogously for health via Mth-order dominance (MSD). FSD corresponds to N=M=1. The paper also uses Ekern’s (1980) Nth-degree risk increase, where the first N-1 moments coincide (e.g., a 2nd-degree increase is a Rothschild-Stiglitz mean-preserving spread; a 3rd-degree increase is an increase in downside risk per Menezes et al. 1980).
How is ambiguity about correlation modeled, and what drives the ambiguity result?
The individual perceives a finite set of possible correlations {k_1<…<k_Theta} with subjective second-order probabilities q_theta, and evaluates them via the recursive smooth ambiguity model of Klibanoff et al. (2005, 2009) using an increasing, concave, thrice-differentiable second-order utility phi (concavity = ambiguity aversion). Evaluating the FOC at the benchmark s_O (the optimum under the mean correlation k_O = sum q_theta k_theta) decomposes the effect into a ’timing of uncertainty effect’ (Osaki and Schlesinger 2014), captured by beta(s_O) which is >=1 iff phi exhibits decreasing absolute ambiguity aversion (DAAA), plus a covariance term Cov(phi’(v), v_s). Under mixed correlation aversion/seeking, v(s,k) and v_s(s,k) move in opposite directions in k (Lemma 3), so because phi’ is decreasing the covariance is positive; combined with DAAA this yields higher savings (Proposition 2).
What is the role of decreasing absolute ambiguity aversion (DAAA)?
DAAA (lambda(z) = -phi’’(z)/phi’(z) decreasing in z) is the ambiguity analogue of decreasing absolute risk aversion. The Appendix proves (following Osaki and Schlesinger 2014) that beta(s)>=1 iff the ambiguity precautionary premium Psi_A >= the ambiguity premium pi_A, which is equivalent to DAAA. DAAA ensures the timing-of-uncertainty effect pushes toward more saving. The authors caution that empirical/experimental evidence on the sign of absolute ambiguity aversion is thin; Berger and Bosetti (2020) is cited as an exception finding evidence for DAAA, and the authors say more evidence is needed.
How do the theoretical predictions connect to experimental and empirical observations?
Experimentally, Attema et al. (2019) measure multivariate risk preferences (wealth and longevity as a health proxy) and observe both cross prudence and cross imprudence, and correlation aversion in the gain domain with correlation seekingness in the loss domain. So the model implies savings can rise or fall with correlation depending on the individual. Empirically, the wealth-health correlation is shaped by public health systems: a more protective system separates wealth and health risk (lowers correlation). Rosen and Wu (2004) find poor health leads to safer investment (consistent with cross prudence); Atella et al. (2012) find households invest more in risky assets when health risk is mitigated by a protective national health system; Chou et al. (2003, Taiwan) find public health insurance reduced precautionary saving (a correlation decrease); Jappelli et al. (2007, Italy) find higher precautionary saving where health care quality is lower (a correlation increase); Ayyagari and He (2017) and Christelis et al. (2020) find Medicare/Medicare Part D increased risky investment. These are described as consistent with cross prudence.
How does this paper differ from the closest prior work?
Versus Eeckhoudt and Schlesinger (2008), which studies how risky shifts in future income affect saving via higher-order stochastic dominance, this paper adds correlation between two attributes and multivariate preferences. Versus Courbage and Rey (2007), who compare a certain-health vs risky-health setting, this paper compares two settings where health is risky in both but the income-health correlation differs, using the simpler Doherty-Schlesinger (1990) correlation representation. Versus Osaki and Schlesinger (2014) and Gierlinger and Gollier (2017), who study ambiguity in future income, this paper introduces ambiguity into the correlation rather than into income itself. The mixed-correlation-aversion concept builds on Jokung (2011) and Eeckhoudt et al. (2007, 2009).
What are the policy implications and their scope conditions?
Because public health systems alter the correlation between wealth and health (e.g., medical-expense coverage separates the two risks, lowering correlation), they affect precautionary saving. The directional prediction is conditional: under cross prudence, lower correlation (more generous public health coverage) reduces precautionary saving and a positive wealth-health correlation raises saving above the independence benchmark; under cross imprudence the signs reverse. Under ambiguity the prediction additionally requires DAAA plus the relevant cross-derivative sign pattern. The authors stress that because experimental evidence shows both cross prudence and imprudence, no unconditional policy prediction follows – e.g., for cross-imprudent individuals ambiguous correlation might lower savings.
What are the main caveats and directions for future research?
The results are sufficiency conditions tied to signs of higher-order cross derivatives, which are hard to interpret and whose empirical signs are not firmly established (experimental evidence is insufficient). The model is a stylized two-date setup with zero interest rate, no time discounting, additive time-separable utility, interior unique optimum, and a single scalar correlation parameter. The authors note the framework extends straightforwardly to multi-period models and suggest studying settings where the value and uncertainty of correlation change over time.