Pooled testing combines several people's samples into one assay and retests individuals only when the pool comes back positive. It saves reagents when infections are rare, and a recent line of work pushes it further by using a risk model to decide who gets pooled with whom. The cost is hidden in the chemistry. Mixing a sample with $k-1$ others divides its viral concentration by roughly $k$, so a person with a low viral load can fall below the assay's detection limit and be missed. We argue that risk-sorted pooling does not spread this dilution error evenly. It concentrates the error on exactly the people it places in "safe" pools, namely those with low predicted risk, few symptoms, and often low viral load. These people almost always sit alone in their pool as the only positive, so they suffer the full dilution. They never get "rescued" by a high-load pool-mate whose signal would flip the pool positive and send everyone to an undiluted individual retest.
The proposal treats the makeup of each pool as a design choice with two competing goals. The first is efficiency: fewest tests per person, which favors sorting people by risk. The second is equitable sensitivity: no identifiable group should face a much higher false-negative rate, which favors mixing some higher-risk people into low-risk pools. We will measure the trade-off frontier between the two, and test whether deliberate mixing beats the obvious fix of simply using smaller pools for low-risk people. We also test a second claim. When pooled and individually tested routes have unequal sensitivity, people have a reason to exaggerate symptoms to get routed to individual testing. Past a critical fraction of such misreporters, the efficiency gain of informed pooling collapses. Equalizing sensitivity should remove that incentive and move the collapse point.
Core Claim
In two-stage (Dorfman) pooling, where positive pools trigger individual retests, an infected person's chance of detection depends on their own load and also on who shares their pool. Write $v_j$ for person $j$'s viral concentration, $L$ for the assay detection limit, and $S$ for the pool. Under a simple mass-balance model, an infected person $i$ is detected with probability
$$\mathrm{Sens}_i(S) = \Pr\Big(\tfrac{1}{|S|}\textstyle\sum_{j \in S} v_j > L\Big)\cdot \Pr\big(v_i > L\big).$$
The first factor is where composition enters. A high-load co-infected pool-mate raises the pool mean above $L$ even when $v_i/|S|$ alone would not. Our design objective picks the pool sizes and assignment of people to pools. It minimizes expected tests per person $\mathbb{E}[T]/N$ while controlling the sensitivity of groups $g$, which are defined by triage stratum and, where recorded, symptom status:
$$\min \; \frac{\mathbb{E}[T]}{N} + \lambda \Big(\sum_g (1-\mathrm{Sens}_g)^p\Big)^{1/p}.$$
Here $\mathrm{Sens}_g$ is the average detection probability of infected members of group $g$. The weight $\lambda$ sets the trade-off. The exponent $p$ moves from penalizing average miss rates ($p=1$) to penalizing the worst group ($p\to\infty$).
This interpolation borrows two ideas from multiwinner voting. The additive-versus-max distinction separates "excellence" rules such as $k$-Borda from "diversity" rules such as Chamberlin–Courant, as mapped in Properties of Multiwinner Voting Rules. The $L_p$ norm construction comes from Manipulation-Robust Selection of Citizens' Assemblies. That paper uses $L_p$ norms over selection probabilities to limit how much any one volunteer can gain by misreporting. We reuse the norm family for a different purpose: aggregating group miss rates. We do not claim any voting-theoretic guarantee transfers.
Where This Sits
Three papers motivate the gap, and each stops short of it.
Density-Dependent Group Testing replaces the idealized "any positive makes the pool positive" rule with a response curve $f(\rho)$ in the fraction $\rho$ of positives in the pool. It proves that the number of tests needed is governed by how $f$ behaves as $\rho\to 0$. When $f(\rho)\approx\rho^\alpha$ with $\alpha\ge 1$, pooling loses its advantage entirely. That paper studies non-adaptive designs, i.i.d. infections, and average recovery. It does not ask how miss rates distribute across covariate-defined groups under adaptive triage.
Smart pooling uses a gradient-boosting risk model to send high-risk patients to individual testing. This keeps pooling efficient up to 50% prevalence, against roughly 30% for random pooling. It reports a wet-lab false-negative rate of about 5.4% at pool size 10. That figure is an aggregate. The risk-sorting mechanism is precisely what we predict skews the error toward low-risk pools.
The closest prior work is Correlation Improves Group Testing: Modeling Concentration-Dependent Test Errors. It shows that correlated concentrations within pools reduce dilution-induced errors, and that concentration-aware grouping beats random mixing. We build on its mechanism but change the question in three ways. We ask who bears the remaining error, not only how large it is on average. We design assignments from pre-test covariates under a triage policy, where the natural efficiency-maximizing choice (grouping low-risk people together) is exactly the one that removes the rescue effect for them. We also ask how the design behaves when those covariates are self-reported and can be gamed. If that paper's grouping already turns out to equalize group sensitivity, our frontier will show it, and we will report it as a positive finding for their approach.
Two other lines constrain the design rather than motivate it. Optimal Dorfman Group Testing for Symmetric Distributions and Community-Aware Group Testing show that correlated infection changes the optimal pool size. The community-aware paper reports 55–75% fewer tests than community-agnostic binary splitting. Any mixing we do must therefore be weighed against the efficiency these clustered designs capture. Combinatorial Group Testing with Selfish Agents shows that strategic participants can push group testing from logarithmic to linear cost when the number of defectives is unknown. That model is noiseless and ignores dilution. Our manipulation study is narrower and more applied: people misreport symptoms, not test participation.
We also keep the manipulation claim modest. The $\sqrt{n}$ transition of A Smooth Transition from Powerlessness to Absolute Power arises because i.i.d. votes fluctuate at order $\sqrt{n}$. In triage, misreports shift the design's inputs directly, so we expect a threshold in the misreporting fraction rather than in $\sqrt{n}$. We treat any resemblance to that paper as an empirical question.
Research Program
The first component is a calibrated simulator. It samples viral loads per stratum from distributions spanning several orders of magnitude, following the regime studied in Semi-quantitative group testing for efficient and accurate qPCR screening of pathogens with a wide range of loads. Loads are allowed to correlate with symptom status. Pool concentration is converted to test outcomes through both a hard detection limit and a smooth $f(\rho)$ with background noise $f(0)>0$, following the density-dependent model. We calibrate against the two wet-lab anchors in the file: about 5.4% false negatives at pool size 10 from Smart Pooling, and 96.7% sensitivity at pool size 6 from Exploring Chlamydia trachomatis screening in Shenzhen.
The second component is the assignment optimizer. It takes risk scores and group labels as input and outputs per-stratum pool sizes plus a mixing matrix: how many members of each risk stratum go into each pool type. We solve it by local search over mixing matrices. Sensitivity is evaluated by Monte Carlo, and the $\lambda$ and $p$ sweep traces the frontier.
The third component is a theory note. It derives closed-form $\mathbb{E}[T]/N$ and $\mathrm{Sens}_g$ for two strata under mass balance, which gives the frontier's shape. It then states the condition under which mixing strictly beats shrinking the low-risk pool size at equal cost. We expect this condition to depend on the load gap between strata and on high-risk prevalence. This formalizes the coupling of dilution and prevalence that the file lists as open.
The fourth component is a manipulation study. A fraction $\mu$ of low-risk infected and uninfected people report symptoms that push them above the triage cutoff. We measure how $\mathbb{E}[T]/N$ rises with $\mu$ and locate the $\mu^*$ at which cost reaches 1.0, the cost of testing everyone individually. We also measure each design's incentive gap: how much sensitivity a person gains by being routed to individual testing.
Baselines
The naive reference points are individual testing, at 1.0 tests per person with no dilution loss, and classical Dorfman with the pool size optimized for known prevalence under random assignment. These bracket cost and sensitivity. A proposal that cannot beat Dorfman on worst-group sensitivity at equal cost has no reason to exist.
The strongest existing method is Smart Pooling's risk triage with risk-sorted pools, reimplemented on the same risk scores. Its reported operating point is 0.36 expected tests per specimen at 5% prevalence, with an operational threshold of up to 50% prevalence (Smart pooling). Nobody has reported a per-group sensitivity result for it; no reported result found on Lacuna for subgroup sensitivity under risk-sorted pooling. We will report both. Our design beats it if, at matched $\mathbb{E}[T]/N$ within bootstrap confidence intervals, it raises worst-group sensitivity with non-overlapping intervals. Equivalently, it beats Smart Pooling if it reaches Smart Pooling's worst-group sensitivity at lower cost.
We also include the dilution-compensated fixed pool size of 6 from the Shenzhen study, balanced hypergraph-factorization designs from Group testing via hypergraph factorization applied to COVID-19, and the community-aware two-stage algorithm on its own synthetic topologies. The key ablation is stratified Dorfman: each risk stratum gets its own optimized pool size, and strata are never mixed. It isolates whether composition matters beyond size. Every design uses the same risk model, load simulator, prevalence grid, and Monte Carlo budget. Hyperparameters are tuned on a held-out seed set with an equal search budget per design.
Why not just…?
Why not just use smaller pools for everyone, as the Shenzhen study did? Uniformly smaller pools raise everyone's sensitivity. They do so by paying the cost in every stratum, including high-risk strata that do not need it. The frontier comparison shows directly whether uniform shrinking is dominated. If uniform shrinking sits on our frontier, we will report that composition adds nothing.
Why not just give the low-risk stratum its own smaller pool size? This is the real threat, and it is why stratified Dorfman is our central ablation rather than a side baseline. At low prevalence, small pools are expensive per positive found. Mixing buys sensitivity through the rescue effect instead. The theory note predicts when each approach wins. If stratified sizing matches mixing across the tested prevalence and load-gap range, we narrow the claim to an auditing result: risk-sorted pooling hides a group sensitivity gap, and per-stratum sizing fixes it cheaply.
Why not just decode with cycle-threshold values or a Bayesian dilution model? Semi-quantitative decoding (Semi-quantitative group testing) and Bayesian adaptive testing with dilution (BAYESIAN GROUP TESTING WITH DILUTION EFFECTS) improve inference once pools are formed. They cannot recover a positive whose pool fell below the detection limit and read negative. We will run each design under both a hard-threshold decoder and a smooth-$f$ probabilistic decoder, to check whether composition effects survive better decoding.
Why not just stop using self-reported symptoms? Triage could rely only on coarse, hard-to-game metadata such as location and time, as in the Bogotá test-center data. We will run this arm. We expect it to reduce manipulation but also to raise the cost of informed pooling. The comparison quantifies what equalized designs retain relative to it.
Proposed Experiments
We use the Bogotá patient cohort (2,068 samples with symptom metadata) and the test-center cohort (7,162 samples) from Smart Pooling if access can be obtained. If cycle-threshold values are absent or the data are unavailable, we fall back to the calibrated simulator. Around these we run the Density-Dependent synthetic setting ($n=5{,}000$, $k=13$ and $166$) and the community-aware topologies (200 families of 5, 20 classes of 50). Prevalence is swept from 1% to 50%. We vary the load gap between symptomatic and asymptomatic strata.
The metrics are $\mathbb{E}[T]/N$, overall sensitivity, worst-group sensitivity, the incentive gap, and the collapse fraction $\mu^*$.
Three checks are decisive. First, under risk-sorted pooling, the low-risk stratum's sensitivity should fall measurably below the high-risk stratum's, while the aggregate false-negative rate matches the reported 5.4% anchor at pool size 10. Second, at matched cost, mixed designs should raise worst-group sensitivity beyond stratified sizing in the regime the theory predicts, and fail to do so outside it. Third, designs with smaller incentive gaps should show a larger $\mu^*$. If the first check fails, the gap is not real and the program stops there. If the second fails, the contribution reduces to an audit plus per-stratum sizing, which we would still publish.
References
- Committee Scoring Rules: Axiomatic Classification and Hierarchy
- Manipulation-Robust Selection of Citizens' Assemblies
- Density-Dependent Group Testing
- Smart pooling: AI-powered COVID-19 informative group testing
- Correlation Improves Group Testing: Modeling Concentration-Dependent Test Errors
- Optimal Dorfman Group Testing for Symmetric Distributions
- Community-Aware Group Testing
- Combinatorial Group Testing with Selfish Agents
- A Smooth Transition from Powerlessness to Absolute Power
- Semi-quantitative group testing for efficient and accurate qPCR screening of pathogens with a wide range of loads
- Exploring Chlamydia trachomatis screening in Shenzhen: a cost-effective sample pooling strategy for pre-marital and pre-pregnancy populations
- Group testing via hypergraph factorization applied to COVID-19
- Bayesian group testing with dilution effects