Public health budgets are usually split by risk factor. One program targets tobacco, another alcohol, another workplace exposures. Each program measures its own share of preventable disease and chases it. This proposal asks how much of the jointly preventable burden such separate programs can be guaranteed to capture, compared with a single planner who coordinates all of them. The answer we propose comes from a tool in algorithmic game theory called smoothness. Smoothness is a way to bound how far self-interested agents fall short of the best joint outcome without solving for their equilibrium. Our claim is that the needed bound can be read off a quantity epidemiologists already report: how a set of single-factor attributable fractions compares with their joint attributable fraction.
The opening is concrete. A population attributable fraction (PAF) is the share of cases that would disappear if an exposure were moved to its lowest-risk level. When the PAFs of several risks overlap, so that each factor's marginal gain shrinks once the others are removed, the burden function shows diminishing returns. That is exactly the structure under which Sessa et al. prove that independently learning agents reach at least $1/(1+\alpha)$ of the optimum. Here $\alpha$ is the curvature, a number in $[0,1]$ that measures how much the factors overlap. When risks are synergistic, so that the joint gain exceeds the sum of the parts, the guarantee disappears and coordination becomes necessary. The result is a cheap "decentralization certificate" for comparative risk assessment, along with a sharp statement of when it fails.
Core Claim
Treat each of $K$ prevention programs as a player. Player $k$ chooses $a_k \in [0,1]$, the fraction of the population it moves into the target (minimum-risk) range of exposure $k$, subject to a budget. Let $W(a)$ be the expected burden averted relative to current exposure. Each program is rewarded with its current marginal attributable contribution, $u_k(a)=W(a_k,a_{-k})-W(0,a_{-k})$. This is what a program measures if it re-estimates its PAF on the population as it stands now, rather than at an old baseline.
Under this reward scheme the setting is a valid utility game in the sense used by Sessa et al.. If $W$ is monotone and shows diminishing returns along these coordinates, any no-regret dynamic reaches time-averaged welfare of at least $W(a^*)/(1+\alpha)$. Here $a^*$ is the coordinated optimum, and
$$\alpha = 1-\min_k \frac{W(\mathbf{1})-W(\mathbf{1}_{-k})}{W(e_k)},$$
which compares each factor's marginal gain once all other factors are handled, $W(\mathbf{1})-W(\mathbf{1}_{-k})$, with its stand-alone gain $W(e_k)$. The stand-alone gain is the single-factor PAF. The marginal gain follows from the joint PAF with and without factor $k$.
Two observations make this more than a relabeling of the field. The first concerns the non-monotonicity barrier that the research file flags against Sessa et al. U-shaped dose-response curves, such as those for bone-mineral markers in the MBD-5D case-cohort study, are non-monotone in raw exposure. They become monotone in the "fraction moved into the target range" coordinate, provided moving someone into range never raises their joint risk. That condition can be tested. It fails only when risks act against each other. The second observation is that the multiplicative-independence assumption standard in Global Burden of Disease aggregation forces diminishing returns. Under that assumption, removing one relative risk helps less when the others are already removed. Departures from independence change $\alpha$ and can push the game into the region where the bound breaks.
A back-of-envelope reading of published numbers shows that the certificate is informative rather than trivial. The illustrations below assume independence and ignore budgets.
In MBD-5D, the narrow PTH target alone has a PAF of 8.8%, the calcium-plus-phosphorus block 8.8%, and all three together 16.8% (MBD-5D). This matches $1-(0.912)^2$ almost exactly. Treating the two blocks as players gives $\alpha \approx 0.09$, so separate targeting would be guaranteed about 92% of the coordinated gain.
For laryngeal cancer, Zhou et al. attribute 66.46% of deaths to tobacco, 12.4% to alcohol and 5.92% to occupational risks. Under independence, alcohol's marginal value once tobacco and occupation are handled is only about 32% of its stand-alone value. That gives $\alpha\approx 0.68$ and a guarantee near 60%. A dominant risk makes the smaller programs' self-measured shares badly overstate their contribution once the dominant risk has been addressed.
Research Program
The first part is theory. We will state the game precisely, with budget-constrained continuous actions and a burden function averaged over the exposure distribution. We will prove the $1+\alpha$ guarantee in "fraction-to-target" coordinates. We will extend it to an expected version in which each program learns from noisy, re-estimated PAFs. Noisy estimation of this kind, using methods such as targeted maximum likelihood estimation (TMLE), is how Mertens et al. actually estimate population effects. The extension argument of Roughgarden carries pointwise smoothness to time-averaged no-regret play, so the main new work is the estimation-noise term.
We will also construct explicit counterexamples. When the joint risk is synergistic, the coordinated optimum can concentrate budget on a pair of factors whose joint benefit only appears together, and independent programs can get stuck far below it. When the risks act against each other, monotonicity fails. These counterexamples define the "do not decentralize" region. They play the same role as the non-smooth network formation games in Roughgarden. Kesselheim and Kodric also constrain the design: smoothness bounds can flip under a change in the payoff model. We therefore fix the welfare objective as expected burden averted, and we treat risk-averse program objectives only as a sensitivity analysis.
The second part is estimation. We need $\alpha$ with uncertainty. We need single and leave-one-out joint PAFs computed on the same individual-level data, with a check of whether the multiplicative-independence form fits. Where individual data exist, we estimate interactions directly rather than assuming independence.
Baselines
The naive reference point is open-loop targeting. Each program allocates its budget once, using its baseline single-factor PAF, and never re-estimates. This is how per-factor PAF rankings are used in practice. It is run on the same fitted burden model and budgets as every other method. Our proposed dynamic beats it if it achieves a higher averted burden in the large-$\alpha$ settings. We expect no gain where $\alpha$ is near zero.
The upper reference is the coordinated planner. It directly maximizes $W(a)$ under the joint budget using projected gradient ascent with many restarts. It is the strongest competitor because it is the outcome decentralization is measured against. We also report continuous greedy as the standard centralized approximation, with its $1-1/e\approx 0.63$ guarantee.
The closest existing empirical result for decentralized learning in this class of games is the approximation ratio of about 0.71 that Sessa et al. report for distributed no-regret learning in sensor coverage. Yao et al. report empirical Price of Anarchy (the ratio of the best joint outcome to the outcome agents actually reach) as low as 1.10 on MovieLens-1M. We found no reported result on Lacuna for decentralized multi-risk prevention programs on any health dataset, so the health comparisons are new.
The key ablation removes the interaction structure. We refit the burden model under forced multiplicative independence and recompute both $\alpha$ and the simulated gap. If the certificate's predictions are unchanged when interactions are removed, the fitted interactions add nothing.
All methods share the same data, burden model, budgets, number of iterations and random seeds.
Why not just…?
Why not just compare the sum of single PAFs with the joint PAF, as epidemiologists already do? This is the most serious threat, and we concede that $\alpha$ is close kin to this familiar overlap measure. What the overlap ratio does not supply is an operational guarantee. It does not say that programs re-estimating and reallocating under budgets will reach a stated fraction of the coordinated optimum, and it does not say when that statement becomes false. The experiments test whether $1/(1+\alpha)$ predicts the simulated decentralized ratio better than the raw overlap ratio does. If the raw ratio predicts just as well, we will report that the theory adds a guarantee but no predictive power.
Why not just simulate the decentralized programs directly? For three markers, direct simulation is cheap and exact, and we use it as ground truth. The certificate needs only routinely published single and joint PAFs, while simulation needs a fully specified joint risk model for every population. That makes the certificate useful at the scale of 204 countries in the Global Burden of Disease data. If the bound is too loose to rank populations by coordination need, that is a real negative result and we will report it.
Why not model individual behavioral responses with an agent-based simulation? That addresses a different gap: people adapting to interventions, which the research file names as the static-counterfactual limitation. It needs behavioral data none of these cohorts contain. We deliberately place the strategic agents at the program level, where the objectives are documented and measurable.
Proposed Experiments
The first experiment uses MBD-5D (case-cohort study; Population-level associations of achievement of targets for bone-mineral markers with survival in haemodialysis patients with mildly elevated intact-PTH levels: a case-cohort study). We will request the individual data, or rebuild the burden surface from the published splines. We will treat calcium, phosphorus and PTH control as three programs, fit the joint risk with interaction terms, and test whether moving a patient into range on one marker can raise risk given the others. We then compare the certified ratio with the simulated no-regret ratio.
The second experiment uses the individual participant data pool of Voerman et al.. Preconception BMI and gestational weight gain map naturally onto separate care programs, which lets us test the claimed link between interaction and curvature on two risks with very different odds ratios.
The third experiment uses the ki cohorts analyzed by Mertens et al.. Welfare here is the mean length-for-age Z-score gain rather than a case count. This tests whether the certificate carries over to continuous population effects estimated with TMLE noise.
The fourth experiment applies the certificate across Global Burden of Disease risk clusters, using laryngeal cancer from Zhou et al. and dementia from Xu et al.. The point is to map where per-factor programs are certifiably adequate. We state clearly that these joint fractions inherit the independence assumption.
The decisive checks are three. The first is the rank correlation between certified and simulated ratios across all settings. The second is a constructed synergistic population in which the theory predicts decentralized failure, which we must then observe in simulation. The third is the ablation showing whether fitted interactions change any conclusion that independence would have given.
References
- Bounding Inefficiency of Equilibria in Continuous Actions Games using Submodularity and Curvature
- Population-level associations of achievement of targets for bone-mineral markers with survival in haemodialysis patients with mildly elevated intact-PTH levels: a case-cohort study
- Global trends and risk factors of laryngeal cancer: a systematic analysis for the Global Burden of Disease Study (1990–2021)
- Causes and consequences of child growth faltering in low-resource settings
- Intrinsic robustness of the price of anarchy
- Price of Anarchy for Mechanisms with Risk-Averse Agents
- How Bad is Top-$K$ Recommendation under Competing Content Creators?
- Maternal body mass index, gestational weight gain, and the risk of overweight and obesity across childhood: An individual participant data meta-analysis
- Epidemiological and sociodemographic transitions in the global burden and risk factors for Alzheimer's disease and other dementias: a secondary analysis of GBD 2021