Hacker News Reader

Comment by oasisbob | Hacker News Reader

> all of Worobey's or Pekar's reasoning starts from the assumption that there was no significant cryptic (unsampled) human spread.

They don't assume that. See the discussion in Pekar. ("The high extinction rate [...] Failed introductions of intermediate haplotypes are also possible.") The whole point of their epidemiological modeling is to disentangle the likelihood of this happening. (See: "Separate introductions of lineages A and B")

The tMRCA doesn't allow enough time for any deep cryptic spread. The authors show this. It's not assumed.

> We show that it is highly unlikely that SARS-CoV-2 circulated widely in humans earlier than November 2019 and that there was limited cryptic spread, with, at most, dozens of SARS-CoV-2 infections in the weeks leading up to the inferred tMRCA, but likely far fewer.

View on HN

The details of that epidemiological model start around page 7 of

https://www.science.org/doi/suppl/10.1126/science.abp8337/su...

I haven't fully studied all the layers yet, but as far as I can tell it's numerical sludge, in the same general direction as econometrics, with so many free parameters or arbitrary choices in the model that you could plausibly get almost any result you wanted. Similar models got their chance to make falsifiable predictions of the case count in the actual pandemic, and they failed repeatedly, often by huge margins, in both directions.

For a specific criticism, the authors report that they're robust against variation in doubling time, but nothing about robustness to the extent of the overdispersion of SARS-CoV-2's spread (i.e., the superspreader characteristic, where most patients infect very few people but a minority infect a lot). They appear to model that only through the connectivity of the infection network, which appears to be fixed. All other things being equal, more stochastic early spread will make it harder to confidently reach any conclusion.

I'm not saying this work isn't interesting at all, or that I could do better; but it's many layers of stacked uncertainty. I believe the confidence that the authors express in this paper is already excessive, and Worobey's statements to the media go considerably beyond that.

Numerical sludge. Got it.

Enjoy your criticism.

Sorry if that phrasing was too mean; I've described many models that I judged to be uselessly complex as "numerical sludge" in the past, including my own less successful efforts. Do you at least understand my specific objection above? You're almost certainly a better biologist than me, but differential equation models get a lot closer to my real expertise.

The output of their model is the structure of the resulting phylogenetic tree. I believe the primary determinant of that will be the extent of overdispersion of transmission, i.e. the shape of the pdf of how many people you expect each patient to infect. For example, let's say R0 = 2 and each patient always infects exactly two others. Then all lineages will survive (at least initially, until a significant fraction of the population is recovered and immune). The phylogenetic tree will be as perfect a binary tree as the mutation rate and our sampling permit. If we cut off the beginning, then we'll see many descendant lineages.

Instead, let's say the spread is absurdly overdetermined, so that each patient infects either zero others (with p = 0.99) or 200 (with p = 0.01). Then (after trying the simulation enough times for the pandemic to finally get started) our tree will just be a succession of super-spreading events, with far fewer lineages.

I believe that by manipulating that pdf, we can reshape that phylogenetic tree to almost anything we want. I see no description from the authors of why they believe the pdf resulting from their choice of a contact network is the right one, nor any study of robustness to changes in the shape of that pdf. Am I missing something?

Reply on news.ycombinator.com