Risk-Sensitive Foraging
Risk-Sensitive Foraging is a stochastic behavioral ecology framework predicting that an animal's preference between a constant food reward and a variable reward of equal mean depends on whether its expected energy budget lies above or below its physiological starvation threshold.
Risk-Sensitive Foraging is a stochastic optimization framework in behavioral ecology and optimal foraging theory that explains how foraging animals incorporate reward variance (σ2)—in addition to mean expected energy intake (μ)—into their patch and prey choices. Pioneered experimentally by Thomas Caraco, Steven Martindale, and Thomas S. Whittam in 1980 (Animal Behaviour) and formalized mathematically as the Energy-Budget Rule (or Z-Score Model) by David W. Stephens in 1981 (Animal Behaviour), risk-sensitive foraging theory addresses a fundamental limitation of deterministic foraging models (such as Charnov's Marginal Value Theorem): in nature, two foraging options with the exact same average rate of energy return (μconstant = μvariable) can yield radically different probabilities of survival when one option provides a predictable, steady stream of food (σ2 ≈ 0) and the other provides a high-variance "boom-or-bust" gamble (σ2 ≫ 0).
Note that in the technical vocabulary of microeconomics and classic behavioral ecology, "risk" refers specifically to probabilistic variance in resource reward (σ2), whereas the danger of being killed by a predator is termed "predation hazard" (which is analyzed via Giving-Up Density and the Landscape of Fear, though modern state-dependent models integrate both). Grounded in Jensen's Inequality and nonlinear fitness utility curves, Risk-Sensitive Foraging proves that animals dynamically switch between risk-averse behavior (preferring certainty) on a positive energy budget and risk-prone behavior (preferring a high-variance gamble) when facing imminent starvation on a negative energy budget.
Caraco's (1980) Junco Experiments and the Energy-Budget Rule
Experimental Discovery in Yellow-Eyed Juncos (Junco phaeonotus)
In 1980, Thomas Caraco, Steven Martindale, and Thomas S. Whittam tested whether small granivorous passerines—yellow-eyed juncos (Junco phaeonotus, ~20 g) wintering in Arizona—were sensitive to variance in seed rewards when the arithmetic mean reward was held strictly equal. Inside an aviary choice arena, individual juncos chose between two feeding stations:
- The Constant (Safe) Station: Delivered exactly 3 millet seeds on every visit (μ = 3.0, σ2 = 0).
- The Variable (Risky) Station: Delivered either 0 seeds (with probability 0.5) or 6 seeds (with probability 0.5)—yielding the exact same mean of 3.0 millet seeds per visit (μ = 3.0), but with high variance (σ2 = 9.0).
Caraco et al. manipulated each bird's physiological energy budget prior to testing by varying either the pre-trial fasting duration (1 hour vs. 4 hours of deprivation) or the ambient air temperature (which alters thermogenic metabolic demand; see Hesse's Rule and allometric macroecology). The juncos exhibited a striking, reversible behavioral switch:
- On a Positive Energy Budget (Satiated / Warm Conditions): When the rate of energy intake from the constant 3-seed feeder was sufficient to meet the bird's daily metabolic needs and survive the coming night, juncos strongly preferred the constant 3-seed station (Risk-Averse Foraging, choosing certainty 75%–85% of the time).
- On a Negative Energy Budget (Food-Deprived / Cold Stress): When the bird was starved or exposed to cold temperatures so that the constant 3-seed intake rate was insufficient to accumulate the fat reserves required to survive the night, the exact same juncos flipped their preference to the variable 0-or-6-seed station (Risk-Prone Foraging, gambling on the high-variance option 70%–85% of the time)!
| Forager Energetic State | Relationship Between Mean Intake (μ) and Survival Requirement (R) | Shape of Fitness Utility Function U(x) Around μ | Jensen's Inequality Prediction (E[U(X)] vs. U(μ)) | Optimal Behavioral Choice | Empirical Taxonomic Examples |
|---|---|---|---|---|---|
| Positive Energy Budget | Expected intake exceeds threshold (μ > R) | Concave downward (U″(x) < 0; diminishing marginal fitness returns) | E[U(Xvar)] < U(μconst) | Risk-Averse (Prefer constant, low-variance patch) | Satiated juncos (Junco), bumblebees (Bombus), hummingbirds, shrews |
| Balanced Energy Budget | Expected intake equals threshold (μ = R) | Inflection point (U″(x) = 0; locally linear utility) | E[U(Xvar)] = U(μconst) | Risk-Neutral (Indifferent to variance; 50:50 choice) | Foragers right at daily energetic break-even |
| Negative Energy Budget | Expected intake falls below threshold (μ < R) | Convex upward (U″(x) > 0; accelerating marginal fitness returns) | E[U(Xvar)] > U(μconst) | Risk-Prone (Prefer variable, high-variance gamble) | Starving winter juncos, cold-stressed shrews (Sorex), founding queens |
| Variance in Handling / Delay Time | All energy budgets (hyperbolic discounting of time E / T) | Convex upward in denominator time T (f(T) = 1/T, f″(T) > 0) | E[E / Tvar] > E / μT | Universal Risk-Prone for Time (Prefer variable delay) | European starlings (Sturnus), pigeons (Columba), rats |
Mathematical Derivation: Stephens's (1981) Z-Score Starvation Model and Jensen's Inequality
1. Stephens's (1981) Overnight Survival Threshold (R) Model
In 1981, David W. Stephens derived the Energy-Budget Rule from first principles of probability theory for a small diurnal endotherm (such as a wintering songbird or shrew) that must accumulate at least R kilojoules of fat reserves by dusk (t = Tday) to avoid freezing or starving to death overnight. Suppose that at the end of the foraging day, a bird choosing patch i (with mean daily intake μi and standard deviation σi) has a total energy reserve X distributed normally as X ~ N(μi, σi2).
The probability that the bird's dusk energy reserve falls below the lethal threshold R (the probability of overnight starvation) is given by the cumulative standard normal distribution Φ(z):
Pr(Starvation: X < R) = Φ( [R − μi] / σi ) = Φ(zi)
where zi = (R − μi) / σi is the standardized Z-score of the energetic shortfall. Because Φ(z) is a monotonically increasing function of z, natural selection favors the foraging option that minimizes zi = (R − μi) / σi:
- Case 1: Positive Energy Budget (μ > R, so R − μ < 0 is negative): Because the numerator (R − μ) is negative, making the standard deviation σi as small as possible (σi → 0) drives zi → −∞, reducing the probability of starvation to zero. If the bird can already survive the night on the guaranteed mean μ, gambling on a high-σ patch only introduces the risk of an unlucky streak that drops X below R. Therefore, when μ > R, the optimal forager is strictly Risk-Averse!
- Case 2: Negative Energy Budget (μ < R, so R − μ > 0 is positive): Because the numerator (R − μ) is positive, choosing a safe, low-variance patch (σi → 0) drives zi → +∞, guaranteeing 100% certainty of death by starvation overnight! Conversely, choosing a high-variance patch (large σi) shrinks the positive Z-score toward zero (zi → 0+), raising the probability of survival toward 50% (because a lucky streak of 6-seed rewards can push X above R). Therefore, when μ < R, the optimal forager must switch to Risk-Prone behavior!
2. Generalization via Jensen's Inequality on Sigmoidal Fitness Utility U(x)
As generalized by Leslie A. Real (1980), Alasdair I. Houston, and John M. McNamara (1982, 1999), Stephens's step-function threshold is a limiting case of Jensen's Inequality applied to any smooth sigmoidal fitness utility function U(x) relating energy reserves x to reproductive value or survival:
- For any concave-downward function (U″(x) < 0, where each additional gram of fat yields diminishing marginal fitness returns), Jensen's Inequality states that the expected utility of a random variable X is strictly less than the utility of its mean: E[U(X)] ≈ U(μ) + ½ σ2 U″(μ) < U(μ) → Risk Aversion.
- For any convex-upward function (U″(x) > 0, such as the lower tail of a sigmoidal survival curve below the inflection point R, or an accelerating egg-production threshold in colonial spider or wasp queens), Jensen's Inequality states that E[U(X)] ≈ U(μ) + ½ σ2 U″(μ) > U(μ) → Risk Proneness.
The Variance-Discounting Model and why Animals Are Risk-Averse for Amount but Risk-Prone for Time Delay
In 1996, Alejandro Kacelnik and Moreno Brito e Abreu demonstrated a striking empirical dichotomy across vertebrates and invertebrates when variance occurs in reward amount (E) versus handling/waiting time delay (T):
- When reward amount (E) varies (e.g., 1 vs. 5 seeds after a fixed 10-second delay), animals on positive budgets are almost universally risk-averse.
- When time delay (T) varies (e.g., a fixed 3-seed reward delivered after either 2 seconds or 18 seconds, mean μT = 10 s, versus a constant 10-second delay), the exact same animals are universally risk-prone (preferring the variable delay)!
Why? Because foraging currency is a ratio of energy to time (R = E / T)—or because animals hyperbolically discount delayed rewards as V = E / (1 + kT)—the utility function with respect to the numerator (E) is linear or concave, whereas the function f(T) = E / T with respect to the denominator (T) is strictly convex upward (∂2(E/T)/∂T2 = +2E / T3 > 0)! By Jensen's Inequality:
½ (E / 2 + E / 18) = 0.278 E > E / 10 = 0.100 E
Thus, receiving a reward after 2 seconds half the time yields a nearly threefold higher average rate of intake than waiting 10 seconds every time!
Ecological Extensions: Group Foraging, Spider Webs, and Plant Root Foraging
Risk-sensitive foraging extends far beyond laboratory seed feeders to shape community ecology across the biosphere:
- Group Foraging and the Producer–Scrounger Game as a Variance-Reduction Mechanism: As shown by Thomas Caraco (1981) and Luc-Alain Giraldeau, foraging in a flock and playing the scrounger strategy (joining patches discovered by others) lowers the variance (σ2) in daily food discovery compared to solitary producer searching. Accordingly, subordinate birds join flocks and scrounge when on a positive energy budget (risk-averse variance reduction), but break away to forage solitarily as producers when starving on a negative energy budget (risk-prone gambling for an unshared bonanza patch).
- Colonial vs. Solitary Orb-Weaving Spiders: George W. Uetz showed that in colonial orb-weaving spiders (Metepeira incrassata) in Mexico, solitary web-building is a high-variance (risk-prone) strategy, whereas joining a communal web colony reduces prey-capture variance (σ2) by 50%, explaining why spiders shift from solitary to colonial webs across prey-abundance gradients.
- Plant Root Foraging (Dener et al., 2016): Remarkably, in 2016, Efrat Dener, Alex Kacelnik, and Hagai Shemesh (Current Biology) grew pea plants (Pisum sativum) with their roots split between two pots—one with constant nutrient supply and one with variable nutrient supply of equal mean. Just like Caraco's juncos, pea plants proliferated roots into the constant pot (risk-averse) when mean nutrient levels were high, and into the variable pot (risk-prone) when mean nutrient levels were below the threshold needed to set seed!
Frequently Asked Questions
How does "risk" in Risk-Sensitive Foraging differ from "predation risk" in the Landscape of Fear?
In classic decision theory and Risk-Sensitive Foraging (Caraco, 1980; Stephens, 1981), "risk" means statistical variance (σ2) or unpredictability in food reward, whereas the danger of being eaten by a predator is termed predation hazard (quantified by Joel Brown's Giving-Up Density). However, John M. McNamara and Alasdair I. Houston (1986, 1992) unified both concepts via State-Dependent Dynamic Programming, showing that a starving animal (μ < R) simultaneously becomes both risk-prone for food variance and hazard-prone against predators (accepting higher predation danger at exposed feeders) because certain death by starvation outweighs a probabilistic chance of predation.
Why are large herbivores less likely to switch to risk-prone foraging on a daily timescale than shrews or hummingbirds?
In accordance with allometric macroecology and the Jarman–Bell Principle, fat storage capacity scales as M1.0 while metabolic rate scales as M0.75, so time to starvation scales as tstarve ∝ M+0.25. A 3 g hummingbird or masked shrew (Sorex cinereus) can starve to death in 2 to 4 hours, meaning it confronts the lethal threshold R every single afternoon. A 500 kg moose or bighorn sheep has weeks of adipose and rumen reserves, buffering daily foraging variance so its short-term utility curve remains concave (risk-averse) except during late-winter blizzards or pre-rut reproductive thresholds.
What is Weber's Law (Scalar Expectancy Theory) as an alternative explanation for risk sensitivity?
Proposed by John Gibbon, Mark Bateson, and Alejandro Kacelnik (1995), the mechanistic psychophysical (Weber's Law) model notes that animals perceive reward magnitudes and time intervals with logarithmic or scalar error (standard deviation proportional to the mean: σperceived = k μ). Because logarithmic compression (ln x) is intrinsically concave downward across all positive values, psychophysical perception naturally produces risk aversion for reward amounts and risk proneness for time delays as a baseline cognitive rule, upon which physiological energy-budget shifts are superimposed.
References
- Caraco, T., Martindale, S., & Whittam, T. S. (1980). An empirical demonstration of risk-sensitive foraging preferences. Animal Behaviour, 28(3), 820–830. https://doi.org/10.1016/S0003-3472(80)80142-4
- Stephens, D. W. (1981). The logic of risk-sensitive foraging preferences. Animal Behaviour, 29(2), 628–629. https://doi.org/10.1016/S0003-3472(81)80128-5
- Kacelnik, A., & Bateson, M. (1996). Risky theories—the effects of variance on foraging decisions. American Zoologist, 36(4), 402–434. https://doi.org/10.1093/icb/36.4.402
- Dener, E., Kacelnik, A., & Shemesh, H. (2016). Pea plants show risk sensitivity. Current Biology, 26(13), 1763–1767. https://doi.org/10.1016/j.cub.2016.05.008