One Calibration Rule for Pollution Indices and Polygenic Scores

Two fields that never cite each other share one statistical habit and one failure. Geochemists judge whether soil is polluted by comparing a measured metal level against a "background" level expected from local rock. The standard tool, the geo-accumulation index, is $I_{geo} = \log_2(C/1.5B)$. Here $C$ is the measured concentration, $B$ is the background, and the constant 1.5 is a fixed allowance for natural variation. Geneticists judge whether a person is at high inherited risk by comparing their polygenic risk score against other people's. A polygenic risk score is a weighted sum of many small genetic effects, almost always learned from European-ancestry cohorts. In both fields the verdict depends on which reference group you compare against and how much natural spread you tolerate before raising a flag. In both fields the reference and the tolerance are usually fixed when they should depend on context: on rock type in one case, on genetic ancestry in the other.

We propose to treat both indices as one object, a score measured against a context-dependent reference distribution, and to fix them with one method. The method has two parts. The first is an audit: among samples known to be "unexceptional" (pristine background soils, or people from the reference distribution), the fraction that the index flags should be the same at every value of the context. The second is a repair. We estimate the full conditional distribution of the reference measurement given continuous context, pooling information across related contexts. We then replace both the fixed 1.5 allowance and fixed percentile cutoffs with quantiles of that distribution. The central claim is that the arbitrary constant 1.5 and the ancestry-blind risk decile are the same mistake, and that it is measurable. Correcting the spread and the tails, not only the center, is what removes the misclassification. The mean-shift corrections that genetics has already adopted only partly do this, and geochemistry has not adopted any.

Core Claim

Write $x$ for the raw measurement on a log scale: $\log_2 C$ for a metal, or the raw score for a genome. Write $z$ for continuous context, such as lithology and parent-material descriptors for a soil sample, or a person's genetic similarity to reference populations. Take $Q_{1-\alpha}(z)$ to be the $(1-\alpha)$ quantile of $x$ among reference samples with context $z$. The calibrated index is

$$I^{*} = x - Q_{1-\alpha}(z),$$

measured in bits of excess for metals. Integer class boundaries are kept so the output reads like the familiar 0–6 contamination scale. For risk scores the analogous quantity is the conditional percentile $F_z(x)$, and "top decile" means $F_z(x) > 0.9$.

The classical index is the special case $Q_{1-\alpha}(z) = \log_2 B + \log_2 1.5$. That is a constant center plus a tolerance of about 0.58 bits, the same everywhere. That special case can be flag-rate invariant only if natural log-scale variability is identical across all rock types, which the geochemistry literature itself lists as an open problem. The corresponding assumption for risk scores is that the score's spread, not only its mean, is the same across ancestries.

The testable prediction is this. On held-out reference samples, $I^{*}$ keeps the flag rate within sampling error of $\alpha$ across context. The classical index and mean-only calibration do not, and the gap grows with distance from the context where the original reference was fixed.

Where the Evidence Comes From

The genetics side supplies both the motivation and much of the method. Ancestry Calibration of Polygenic Risk Scores Improves Risk Stratification and Effect Estimation in African American Adults calibrates raw scores against continuous genetic similarity to a Yoruba reference in 8,841 REGARDS participants. After calibration, 33.6% of the top type-2-diabetes decile and 55.0% of the top height decile were reassigned. This is direct evidence that fixed cutoffs misplace people. We reuse its continuous-ancestry framing, and we ask whether modeling ancestry-dependent spread and tails adds anything beyond what it reports. The same paper also constrains our design. It shows that entering ancestry as a regression covariate next to an uncalibrated score creates spurious associations it calls "statistical ghosts." We therefore calibrate before any downstream modeling rather than adjusting afterward.

On cross-ancestry cancer polygenic risk scores shows that European-trained breast and prostate cancer scores shift in absolute level across UK Biobank ancestry groups, yet keep within-group ranking. We use it as the case where within-group percentiles are known to work, so any gain there must come from continuous rather than categorical context.

Multi-ancestry polygenic risk scores using phylogenetic regularization (DendroPRS) lets effect sizes drift along an ancestry tree. Sparsity penalties on the root and on branch offsets let related populations borrow strength. We borrow this sharing structure for a different quantity: the reference distribution rather than the effect sizes. For soils it becomes a hierarchy running from a national background to a region to a lithological unit.

Optimal strategies for learning multi-ancestry polygenic scores vary across traits is a caution against pooling. For traits driven by locally selected loci, adding European data hurt African prediction. Our hierarchical sharing must therefore be able to switch off, and we test that.

Polygenic risk score portability for common diseases across genetically diverse populations quantifies the problem. European-optimized scores fall to 39.4% of their European accuracy in African-ancestry test samples. Precision diagnosis for monogenic diabetes requires ethnicity specific criteria for genetic testing shows the clinical cost of a single threshold. Ethnicity-specific BMI cutoffs raised detection from 8.8% to 16%. That study used hand-tuned categorical cutoffs, and we aim to replace them with learned continuous ones.

The geochemistry side supplies the target and the fixed-constant problem. Application of Different Indices for Soil Heavy Metal Pollution Risk Assessment Comparison and Uncertainty: A Case Study of a Copper Mine Tailing Site shows how badly the indices disagree on the same cores near a copper tailing site in Bayannur. The single-pollution index flagged 18.14% of the area as highly contaminated, while the Nemerow index called 61.36% safe. That study built its background from 15 points within 2 km and sampled three depth layers, noting surface enrichment for every metal except chromium. This gives us a natural internal check.

Ecological and Health Risk Assessment of Heavy Metals in Farmland in the South of Zhangbei County, Hebei Province, China applies $I_{geo}$ with a regional background to 53 farmland samples. Research on the Characteristics of Heavy Metal Pollution in Lake and Reservoir Sediments in China Based on Meta-Analysis compiles six metals across 71 lakes over 2003–2022, which gives a cross-regional view of how background varies. None of these studies asks whether their index flags pristine material at the same rate in different geological settings. That is the gap.

The research file found no study bridging the two fields. The closest work is Vargas et al.'s continuous ancestry calibration, which we extend to spread and tails and carry into geochemistry.

Research Program

The work has three parts.

First, we build the conditional reference model. We fit quantile regression of $x$ on $z$ among reference samples only. For soils, the reference samples are designated background sites and, as an independent check, deep-layer samples below the depth of anthropogenic enrichment. For genomes, the reference is the full cohort, or controls for disease traits. Coefficients are shared hierarchically in the DendroPRS style, with a penalty on each context's offset from its parent. Cross-validation can choose zero sharing, which answers the Lehman et al. concern.

Second, we define the flag-rate invariance audit. We bin held-out reference samples by context and compute, in each bin, the fraction flagged at the index's alert level. The summary statistic is the largest deviation from the nominal rate across bins, with bootstrap intervals. The audit is index-agnostic, so it can score the classical $I_{geo}$, the Nemerow index, raw decile cutoffs, and every calibrated variant on the same footing.

Third, we check the covariate trap in the environmental setting. In simulation, we generate outcomes that depend on true excess contamination but not on lithology. We then show whether regressing the outcome on uncalibrated $I_{geo}$ plus a lithology covariate manufactures a lithology effect, the geochemical version of Vargas et al.'s ghost, and whether $I^{*}$ avoids it. We simulate because the file contains no cohort that links measured soil metals to genotyped individuals. We do not claim a real gene-by-environment result.

Baselines

The naive reference point is the index as practitioners use it today. For soils that is $I_{geo}$ with the published background and the 1.5 constant, computed exactly as in the Bayannur and Zhangbei studies. For genomes it is the raw score with cohort-wide decile cutoffs. It is run on identical samples, and beating it means a smaller maximum flag-rate deviation on held-out reference samples. For the risk-score side, the relevant numbers in the file are Vargas et al.'s: the top-versus-bottom T2D decile odds ratio rises from 7.97 [6.31–10.13] uncalibrated to 10.77 [8.41–13.91] calibrated, and the height correlation rises from 0.24 to 0.32. For $I_{geo}$ flag-rate invariance across lithologies, no reported result was found on Lacuna.

The strongest existing method is Vargas et al.'s continuous ancestry calibration, and its geochemical analogue is a locally re-estimated mean background. It is the strongest competitor because it already addresses shifts in the center using continuous context, which is likely most of the error. We reimplement it on the same reference samples, with the same context variables and the same cross-validation folds, and give it the same hyperparameter budget as our model. On REGARDS-like data we should at least reproduce its reported 10.77 odds ratio. Beating it means two things. The first is a statistically smaller flag-rate deviation in the tails, meaning the top decile or the alert class. The second is a top-decile odds ratio or correlation no worse than its reported values. A tie on the center with a win only in the tails is the expected outcome and would still count.

DendroPRS is a baseline only on the score-quality question. Its reported non-White BMI correlation of 0.14 is the bar for whether calibration on top of a better score still helps.

Two ablations isolate the key ingredient. The mean-plus-constant-tolerance ablation keeps our learned center but restores a fixed 0.58-bit allowance. This separates learning the center from learning the tolerance. The no-sharing ablation fits each context independently, which shows whether the hierarchy helps when background samples are scarce, as with the 15 Bayannur points.

Why not just…?

Why not just re-estimate a local background and keep the 1.5 constant? This is what careful geochemists already do, as Bayannur's 2 km background shows, and it fixes the center. It still imposes the same tolerance in every setting. Our mean-plus-constant ablation is exactly this method. If its flag-rate deviation on deep-layer and background samples matches ours, then the 1.5 constant is harmless in practice, and we will report that. The file's own open problems suggest otherwise: natural variability differs with element mobility and rock mineralogy. With only 15 background points per site, a local estimate is also noisy, which is where hierarchical sharing should matter.

Why not just use Vargas et al.'s calibration, or z-score within ancestry groups? For the center of the distribution this is a real threat, and on some traits it may be as good as our method. Within-group z-scoring uses categorical labels, which the file identifies as a failure for admixed individuals. Mean-only calibration assumes the score's spread is constant across ancestry. The decile-reclassification numbers do not tell us whether that assumption holds. We settle the question by stratifying the audit by genetic-similarity bins. If spread is constant, our method reduces to theirs, and on the genetics side our contribution is the audit alone.

Why not just add ancestry principal components or lithology as regression covariates? Vargas et al. showed directly that this creates spurious ancestry effects next to uncalibrated scores. Our simulation tests whether the same thing happens with lithology and $I_{geo}$. Covariate adjustment also leaves individual classifications unchanged, and those classifications are what screening and remediation decisions use.

Proposed Experiments

On the genetics side we use UK Biobank ancestry groups (African, East Asian, South Asian, admixed American, Middle Eastern), with continuous genetic similarity computed against 1000 Genomes references. The traits are height, BMI and type-2 diabetes, plus breast and prostate cancer to replicate Koprulu et al.'s within-group results. Where access allows, we use REGARDS for a direct comparison with Vargas et al.

On the geochemistry side we use three data sources. The first is the Bayannur depth profiles, where 30–50 cm samples serve as held-out lithogenic reference and surface samples serve as test. The second is the Zhangbei farmland set. The third is the 71-lake meta-analysis, for cross-regional background variation. The Dehui land-use survey provides a sanity check: among farmland under different uses on similar parent material, $I^{*}$ should separate land-use classes no worse than $I_{geo}$ does. The geochemical data are small and available through the source papers rather than as public benchmarks. We say plainly that this side is a smaller, proof-of-concept test.

The decisive checks are three. First, on held-out reference samples, $I^{*}$ keeps the flag rate within bootstrap error of nominal across context bins, while classical $I_{geo}$ and raw deciles drift. Second, the learned-tolerance model beats the mean-plus-constant ablation in the tails. Third, in Bayannur, deep-layer samples are flagged at the nominal rate, surface copper and lead remain flagged near the tailing source, and chromium, which the source paper reports is not surface-enriched, is not newly flagged. If the second check fails in both fields, the claim that tolerance matters is false. What remains is a transferable audit plus the finding that mean calibration suffices.