Three confounds undercut the whole-dataset correlation. (1) Selection design: on 12 January every wildlife-stall sample was sequenced regardless of qPCR status, whereas the other sequenced samples were predominantly qPCR-positive and mostly from non-wildlife locations; wildlife species are therefore over-represented among the sequenced negatives, mechanically driving a spurious negative correlation between wildlife mtDNA and SARS-CoV-2 (their Figure S4 simulation). (2) Non-wildlife stall positives were very likely human-shed, so including them in a univariate correlation contaminates the signal. (3) Timing/decay: in a zoonotic scenario animal shedding precedes human shedding, and most wildlife stalls were sampled 11 days after the case-linked stalls, giving animal-derived RNA more time to decay — again weakening any positive animal correlation. When the analysis is restricted to the balanced n=70 wildlife-stall set, human mtDNA does not correlate with the virus while some wildlife species (porcupine, marmot) do. Hence the negative-correlation result relied on by skeptics is an artifact of pooling and design, not a robust biological signal; a correlation on these post-outbreak samples is in any case a weak instrument for identifying the infected host.
Step 6 verdict — approved (checked)
Reconstructed step: a selection-bias (collider-style) argument. Because the 12 January protocol sequenced every wildlife-stall sample regardless of qPCR status while the other sequenced samples were predominantly qPCR-positive, wildlife species are mechanically over-represented among sequenced negatives; conditioning the correlation on a selection variable correlated with both mtDNA source and viral status manufactures a spurious negative association independent of any true biological relationship. Hence the pooled negative correlation is not evidence against an infected-animal source. This is a valid statistical inference, and conditional on the premises (the Fig S4 simulation reproducing the artifact; the balanced n=70 reanalysis where human mtDNA is uncorrelated and porcupine/marmot are positively correlated) it holds. Probed for a defeater — could the negative correlation be a real signal the balanced set fails to capture? — but the simulation directly demonstrates the design generates it, so the “artifact” reading survives. The conclusion is appropriately limited to “not evidence against,” not “evidence for.” Traced from the premises, so checked.