A formal statistical critique, by two spatial-statistics researchers, of the geospatial method in Worobey et al. 2022. Argues (a) that using the centroid of early-case locations (or a similarly simple constructed point) as a proxy for an epidemic’s true point of origin is statistically unjustified — there is no theory establishing that correspondence — and (b) that the Monte Carlo procedure Worobey et al. used to conclude no other location could be the origin is methodologically flawed (wrong reference distribution/multiple-comparisons handling). Concludes the “market was the epicenter” claim does not follow from the geospatial data as analyzed, without asserting an alternative origin.
relevance_note: the canonical formal-statistics rebuttal of the market geospatial-clustering claim, targeting Worobey 2022’s core inferential step directly.
Data and limitations (Section 2)
O-3 - The 155-case December-2019 dataset is incomplete and coarse (vs ~257 documented cases, no onset dates, shared addresses)
Stoyan & Chiu list four data-quality limitations of the case set before proceeding with it: (1) precise coordinates were unavailable and were extracted from a map with ⇐50 m noise; (2) seven cases shared one residential address yet were treated as distinct points; (3) no symptom-onset dates were recorded, so the temporal ordering is unknown; (4) the set of 155 is incomplete relative to ~257 December-2019 cases documented by Demaneuf (2022). They flag these but state their critique targets the applied statistics rather than data quality.
Link to original
Results (Section 3)
O-1 - Recomputed Monte Carlo p-values for the market as case center are marginal and weaken toward non-significance as the case sample shrinks
Stoyan & Chiu reconstruct Worobey et al.’s Monte Carlo procedure (1,000 inhomogeneous-Poisson simulations with intensity proportional to residential population density) and compute the p-value for the market being the “center” of the early-case cloud under three center definitions and four case subset sizes.
Center definition n=155 n=150 n=100 n=80 Centroid (mean) 0.030 0.024 0.060 0.122 Center-point 0.008 0.009 0.043 0.068 Mode 0.070 0.072 0.155 0.200 Significance is at best marginal at the full sample and disappears as the sample is trimmed, showing the market-as-center result is not robust. (Worobey’s own reported median market-to-case distance for the 155 cases was 16.11 km.)
Link to original
O-2 - Bootstrap places the market at the edge of the case-centroid cloud with Wanda Plaza an equally admissible center
Bootstrapping the 155 case locations, the market “can hardly be considered a part of the clouds of centroids”, sitting at the periphery. The Wanda Plaza shopping/hotel complex is “clearly located in the central part of the point clouds formed by centroids” and lies within the confidence regions of all three kinds of “center” when n=80. So the geospatial data do not single out the market over other central-Wuhan landmarks as the case cloud’s center.
Link to original
Arguments (Sections 3-4)
A-1 - Rejecting a population-density (non-contagious) null cannot single out the market as the origin
Worobey’s test compares the observed cases’ median distance-to-market against 1,000 inhomogeneous-Poisson patterns whose intensity is proportional to residential population density — a null in which case locations are scattered like the resident population, with no person-to-person contagion. A contagious respiratory disease necessarily produces spatial clustering tighter than that population-proportional scatter. Hence the observed cases will look “too concentrated” relative to this null around essentially any central reference point, not uniquely around the market. Stoyan & Chiu make this explicit: substituting another landmark (e.g. Wanda Plaza) for the market in the same test reproduces the rejection. The alternative hypothesis actually being tested (“cases are not distributed proportionally to population density”) is therefore mismatched to the conclusion drawn (“the market is the origin”): rejecting the former licenses only the statement that transmission clusters cases, not any statement locating the source. A low p-value is thus uninformative about origin.
Validity verdict
status: approved; reason_if_not_false: checked.
Reconstructed step: the test’s alternative hypothesis is “cases are not distributed proportionally to residential population density”; a contagious respiratory disease clusters tighter than population scatter, so this null is rejected around any sufficiently central reference point, not uniquely the market → therefore rejecting it does not discriminate the market from other central landmarks as the source. Traced myself: locating a point source requires a test that discriminates the candidate location against alternative central points; a null of “no contagion / population-proportional scatter” cannot do this because its rejection is entailed by contagion alone, wherever the source. The empirical claim that substituting Wanda Plaza reproduces the rejection is here a premise (its truth is priced at step 8); conditional on it, no undercutting defeater survives. The inference holds exactly as stated.
Link to original
A-2 - The geometric center of early cases is not a justified estimator of the epidemic origin (centrality does not imply causality)
There is no epidemiological or statistical result guaranteeing that the geometric center of case residences equals the source location. For a person-to-person contagious disease the spatial distribution of early cases is generated by (a) where susceptible hosts live and (b) the transmission process spreading outward and via commuting — so the “center” of the case cloud tracks the residential population’s geography, not the point of introduction. Two further weaknesses compound this: the specific estimator Worobey used, the coordinate-wise median, is not rotation-invariant, so the computed “center” depends on the arbitrary choice of coordinate axes; and Stoyan & Chiu’s bootstrap shows the market sits at the periphery of the centroid cloud while other landmarks (Wanda Plaza) are equally admissible centers. Therefore, even granting for argument that the market were the geometric center of the cases, that centrality would not imply it is the causal origin — centrality does not imply causality.
Validity verdict
status: approved; reason_if_not_false: checked.
The load-bearing move is a non-implication claim: “market is the geometric center of early cases” does not deductively entail “market is the causal origin.” Traced myself: the case cloud is generated by host residence geography plus the transmission process, none of which is pinned to the introduction point; there is no theorem mapping the centroid of an outward-spreading contagious process back onto its seed, so the entailment fails. The estimator points (non-rotation-invariance of the coordinate-wise median; market at the periphery of the bootstrap centroid cloud) are supporting, not load-bearing. The conclusion is deliberately scoped to “could not be inferred” (deduced), and does not overreach to “carries zero probabilistic evidence” — so no undercutting defeater applies. Holds as stated.
Link to original