💻 computer science

Scarcity and Predictive Uncertainty: Implications for Societal Resource Allocation

This paper demonstrates that systematic differences in predictive uncertainty across populations fundamentally alter optimal resource allocation strategies—shifting from prioritizing low-uncertainty individuals under scarcity to high-uncertainty ones under abundance—thereby creating new moral dilemmas and efficiency losses in domains like education, healthcare, and social services.

Original authors: Shafkat Farabi, Patrick J. Fowler, Sanmay Das

Published 2026-08-06

📖 7 min read🧠 Deep dive

CC BY 4.0

Original authors: Shafkat Farabi, Patrick J. Fowler, Sanmay Das

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are the captain of a rescue boat in the middle of a stormy ocean. You have a limited number of life jackets, but hundreds of people are in the water. The goal isn't just to save people; it's to save as many as possible. This is the world of "scarce resource allocation," a field where scientists and policymakers try to figure out how to hand out things like hospital beds, tutoring hours, or housing vouchers when there simply aren't enough to go around. Usually, they use computers to make predictions: "If we give this student extra help, will they pass the test?" or "If we give this patient a ventilator, will they survive?"

But here is the twist: not all predictions are created equal. Some predictions are like looking through a clean window; you can see exactly what's coming. Others are like looking through a foggy, scratched-up pane of glass; you might guess the right direction, but you're not sure. This paper explores what happens when the "fog" (predictive uncertainty) is different for different groups of people. It asks a tricky question: If you have a limited supply of life jackets, should you give one to the person you are sure will survive if you help them, or the person you are unsure about, who might need it more? The answer turns out to be surprisingly counterintuitive, flipping like a coin depending on how many life jackets you have.

The Foggy Forecast and the Life Jacket Dilemma

This paper, titled "Scarcity and Predictive Uncertainty," dives into a specific problem in how we use AI and data to make tough choices. The authors, Shafkat Farabi, Patrick J. Fowler, and Sanmay Das, build a mathematical model to see what happens when we try to be "fair" and "efficient" at the same time, but our crystal balls are clouded by different amounts of fog for different people.

In the real world, we often use two main strategies to decide who gets help. The first is Vulnerability First: "Give help to the people who are in the worst shape right now." The second is Maximum Marginal Benefit (MMB): "Give help to the people who will improve the most because of it." Think of MMB like a gardener deciding which wilted plants to water. You might skip the ones that are already dead (too far gone) and the ones that are already healthy (they don't need it). Instead, you water the ones that are just on the edge of dying, because a little water will make them bloom.

The problem arises when the gardener isn't sure how thirsty the plants actually are. Maybe for some plants, the soil is dry and predictable (low uncertainty). For others, the soil is a mystery; it might be wet, or it might be a desert (high uncertainty). The paper asks: If you want to save the most plants, does the "mystery soil" change who you water?

The Great Flip-Flop

The authors ran simulations (mathematical experiments) to see how this plays out. They discovered a fascinating "flip-flop" in how resources should be distributed, depending entirely on how scarce those resources are.

When resources are very scarce (The "Triage" Phase):
Imagine you have only one life jacket for a whole group of swimmers. The paper finds that the smartest move is to give it to the person with the clearer forecast (low uncertainty), even if they are in the exact same trouble as the person with the foggy forecast. Why? Because with a scarce resource, you need to be sure your help works. If you give the life jacket to someone with a "foggy" future, there's a real chance they might drown anyway, or that the jacket wasn't the deciding factor. But if you give it to someone with a "clear" future, you know for a fact that your help pushed them over the edge to safety. In this scenario, the "foggy" people get left behind, not because they are less important, but because the math says their outcome is too unpredictable to guarantee a win.

When resources are abundant (The "Feast" Phase):
Now, imagine you have a huge pile of life jackets. Suddenly, the strategy flips! The paper shows that when you have plenty to go around, you should start targeting the people with the foggy forecasts (high uncertainty). Why? Because when you have enough resources, you can afford to take risks. The people with clear forecasts are likely to survive anyway or don't need as much help to cross the finish line. But the people with the foggy forecasts? They are the ones who might be sitting just below the surface, waiting for a push. If you give them a resource, the "fog" clears, and they might suddenly shoot up to safety. In a world of plenty, the "mystery" people become the best targets for help because they have the most room to improve.

The Cost of Ignoring the Fog

The paper also looked at what happens if the decision-maker doesn't know about the different levels of fog. They call this the "Uncertainty Unaware" approach. It's like a lifeguard who assumes everyone's water is the same depth, even though some parts are deep and murky while others are shallow and clear.

The results showed that ignoring the fog causes a significant loss in efficiency, especially when resources are low. If you treat everyone the same, you end up wasting life jackets on people who didn't need them or missing the people who did. The authors found that in low-resource settings, this mistake can lead to about a 20% drop in the number of people successfully saved compared to a strategy that accounts for the fog.

Real-World Test: The Student Exam

To see if this math holds up in the real world, the authors tested their ideas using data from the PISA (Programme for International Student Assessment), a massive global test of 15-year-old students. They looked at math scores and tried to predict which students would pass a certain proficiency level if they got extra tutoring.

They found that prediction models were more uncertain for students from lower socioeconomic backgrounds (the "foggy" group) compared to those from higher backgrounds (the "clear" group). When they simulated a tutoring program:

Under the "Uncertainty Aware" strategy: If the school had very little money for tutoring, they would focus entirely on the students with clearer predictions (the higher socioeconomic group), because they were the surest bets to pass.
Under the "Uncertainty Unaware" strategy: The school would spread the tutoring out more evenly, which turned out to be less efficient.

However, as the school budget grew, the "Uncertainty Aware" strategy shifted, eventually giving more attention to the students with higher uncertainty, just as the theory predicted.

The Big Question

The paper concludes with a moral puzzle. We have shown that it is mathematically efficient to treat people differently based on how "foggy" their future looks. But is that fair? If two students are in the exact same spot, but one comes from a background where the future is harder to predict, should they get less help just because the math says they are a riskier bet?

The authors don't solve this moral debate, but they highlight a new dilemma for society. They suggest that while ignoring uncertainty leads to wasted resources, using it to decide who gets help might mean abandoning the most vulnerable people when resources are tight. It's a reminder that in the game of life, the "smartest" math move isn't always the "fairest" human move, and the answer changes depending on how many life jackets we have in the boat.

Technical Summary: Scarcity and Predictive Uncertainty

Problem Statement
This paper addresses a critical gap in the literature on automated decision systems (ADS) for societal resource allocation: the impact of heterogeneous predictive uncertainty across different population subgroups. While existing research often focuses on how improved prediction accuracy leads to better social outcomes, this work investigates what occurs when the uncertainty of those predictions varies systematically (e.g., due to demographic differences in model performance or fundamental life variability). The authors examine how this uncertainty interacts with two dominant prioritization strategies in "local justice" contexts: Maximum Marginal Benefit (MMB), which targets individuals most likely to cross a success threshold due to intervention, and Vulnerability-First (VF), which targets those with the lowest baseline welfare. The core problem is determining how knowledge of varying predictive uncertainty ( $\sigma$ ) alters optimal allocation decisions, particularly when resources are scarce, and what the efficiency costs are when decision-makers ignore this heterogeneity.

Methodology
The authors formulate a novel mathematical model of scarce resource allocation involving a population of individuals characterized by a baseline welfare state $x$ and an uncertainty level $\sigma$ .

State Dynamics: An individual's welfare at time $t=1$ is modeled as $x' = x + L + I$ , where $L = \sigma Z$ represents baseline prediction uncertainty (with $Z$ being a symmetric, log-concave random variable) and $I = \xi + \tau W$ represents the intervention (a deterministic drift $\xi$ plus additional uncertainty $\tau W$ ).
Success Definition: Success is defined as a binary outcome where the final state $x'$ exceeds a threshold $\theta$ (e.g., passing a test or surviving).
Policy Formulations:
1. Uncertainty-Aware MMB (UA-MMB): The optimal policy that maximizes the expected population mass above $\theta$ by explicitly accounting for individual $\sigma$ . The authors derive that this policy follows a cutoff rule based on the marginal utility gain function $\Delta(x, \sigma)$ .
2. Uncertainty-Unaware MMB (UU-MMB): A policy that optimizes based on the expected utility gain averaged over all possible uncertainty levels, ignoring individual $\sigma$ .
3. Vulnerability-First (VF): A policy that allocates resources to individuals with the lowest baseline $x$ , ignoring both $\sigma$ and marginal benefit calculations.
Analysis: The paper utilizes convex analysis and properties of log-concave distributions to characterize the behavior of $\Delta(x, \sigma)$ . It proves theorems regarding the existence of crossover points where marginal benefits shift from positive to negative.
Empirical Validation: The theoretical findings are illustrated using a real-world dataset from the Programme for International Student Assessment (PISA) 2022. The authors simulate an intervention to improve math scores, using XGBoost models to predict performance. They demonstrate that prediction uncertainty (measured by RMSE) varies significantly between socioeconomic groups (high vs. low ESCS), allowing them to test the allocation behaviors of UA-MMB versus UU-MMB under varying resource constraints.

Key Contributions and Results

The "Scarcity Flip" in Prioritization: The central finding is that the optimal prioritization under MMB depends non-monotonically on resource availability relative to predictive uncertainty.
- Scarce Resources: When resources are very limited, the UA-MMB policy strictly prioritizes individuals with lower predictive uncertainty, even if they share the same baseline welfare state ( $x$ ) as high-uncertainty individuals. This is because lower uncertainty yields a higher probability of reliably crossing the success threshold $\theta$ .
- Abundant Resources: As resources become more plentiful, the prioritization flips. The policy begins to target individuals with higher predictive uncertainty. This occurs because the allocation boundaries expand, and the high-uncertainty group's "gain curve" (the range of $x$ values benefiting from intervention) expands faster than the low-uncertainty group's, eventually overtaking it.
Efficiency Losses of Ignoring Uncertainty: The paper quantifies the "Relative Efficiency Loss" (REL) of policies that fail to account for heterogeneous uncertainty.
- UU-MMB: Suffers significant efficiency losses (approx. 20% in the PISA simulation) when resources are scarce because it may allocate to low-uncertainty individuals who have negative marginal benefits or miss high-potential high-uncertainty individuals.
- VF: Consistently suffers high efficiency losses across all resource levels because it prioritizes based on baseline state rather than the probability of crossing the threshold, often allocating to individuals who will not cross $\theta$ even with intervention.
Nature of the Utility Gain Function ( $\Delta$ ): The authors prove that under uncertainty, the marginal benefit of intervention is not strictly monotonic with respect to baseline state. There exists a crossover point $x_0 > \theta$ where individuals above the threshold may still benefit from intervention (due to the "smearing" effect of uncertainty), and individuals far above the threshold may face negative marginal gains.

Significance and Claims
The paper claims that its findings have substantial implications for the ethics and mechanics of resource allocation in domains such as medical triage, education, and homelessness services.

Ethical Dilemma: The results raise a new moral question: Is it just to allocate a scarce resource to one person over another solely based on the predictability of their future (predictive uncertainty), even if their current needs are identical? The paper notes that while MMB is often viewed as just in triage, incorporating uncertainty may systematically deprioritize vulnerable groups if those groups also happen to have higher predictive uncertainty (e.g., due to data bias or fundamental life instability).
Policy Design: The authors argue that decision-makers seeking to maximize marginal benefit must account for predictive uncertainty, not just the estimated state. Ignoring this heterogeneity leads to sub-optimal outcomes, particularly in low-resource settings.
Context Dependence: The distributional effects of uncertainty-aware allocation are context-dependent. In the PISA study, higher uncertainty was associated with the disadvantaged group (low ESCS), leading to their deprioritization under scarcity. However, the authors caution that in other contexts, higher uncertainty might align with advantaged groups, meaning empirical assessment is required rather than assuming a fixed direction of bias.

The paper concludes that while uncertainty-aware policies maximize efficiency, they introduce complex distributional trade-offs that challenge traditional notions of local justice, necessitating a careful balance between utilitarian efficiency and equitable treatment of groups with different levels of predictability.