Gist of this paper:
The authors look for circumstantial evidence of community transmission of COVID-19 within Wuhan for dates substantially prior to the seafood market outbreak. They do this by attempting to triangulate* web search volume data with satellite mapping of Wuhan hospital parking lots.
*suggestively gesture back and forth between
In the parking lots, they're looking for more vehicles than expected based on prior years' imagery.
In the web search trends, they're looking for "disease signals".
Specifically, they're looking at two terms: a symptom of sufficient specificity to be useful for differential diagnosis of COVID-19 vs. influenza ("diarrhea"), and a more generic term representing influenza-like illness ("cough").
Here's the teardown:
1. "We developed a comprehensive list of hospitals in Wuhan (using Google Maps, Wikipedia and PubMed)."
List not provided.
2. Search terms are given in English even though they were actually input in Mandarin.
You can verify this by trying to search the English terms yourself. By process of elimination and the shape of their graphs, the terms had to have been:
"cough" → 咳嗽
"diarrhea" → 腹泻
Live link below. Requires signup. Make sure to select Hubei → Wuhan (湖北 → 武汉).
http://index.baidu.com/v2/main/index.html#/trend/咳嗽?words=咳嗽...
3. "We extracted the relative search volumes of the terms “cough” and “diarrhea” using WebPlotDigitizer, v4.215."
What? The values are provided on tooltip hover. This data is written into the page. You can just pull the real data out of the html.
Based on a cursory glance at WebPlotDigitizer, the vague method quote above appears to be tantamount to saying "we made up the data from screenshots".
4. They're looking at seasonal trends.
This is clearly visible from basic inspection. This is not interesting.
5. Their Relative Search Volume scale doesn't agree with Baidu.
Their numbers are wrong. Just look at Figure 2b's y-axis.
https://files.catbox.moe/v05oss.png
Maxima: It shouldn't go to 600. The actual values peak on Baidu at 711.
Minima: None of their data should be plotting near 0. The minimum volume in their date range is 76.
6. The uptick for "diarrhea" (腹泻) that they claim exists doesn't.
Again, this is Figure 2b. The red line is the problem here.
This appears to be deliberate fabrication. No competent researcher can miss the discrepancy between what this graph depicts and what the source depicts.
咳嗽 ("cough" / blue line on harvard preprint) was following a completely normal-looking seasonal cycle during the period under discussion.
腹泻 ("diarrhea" / red line on harvard preprint) was, again, following a completely normal-looking seasonal cycle, with a slightly elevated baseline.
The red line is mis-plotted to trend up in a sudden surge right where the researchers would need this to occur for their argument to hold any value at all. Then, rather than declining substantially prior to the January 2020 cluster, this plot depicts the trend holding through to 2020.
This information is simply false. That's not what Baidu reports.
Here's a side-by-side of the discrepancy:
https://files.catbox.moe/5s6i7o.png
Here's a superimposed version, for charitable reading of Figure 2b. I've stretched it to fit. Red should overlay green.
https://files.catbox.moe/eubuci.png
7. Screw the sat mapping, frankly.
If you can't be trusted to screenshot a graph without us having to arrive at this point, there's no point even assessing the content of your claims about something orders of magnitude more complex.
The sat mapping doesn't constitute evidence so much as it constitutes having paid for the visuals needed to launder this preprint into mainstream news. The only thing that can currently be done with this information is a data dump for other researchers to use.
8. Proofread your paper. 2019 → 2020 below:
> Both search query terms show a large increase approximately 3 weeks preceding the large spike of confirmed COVID-19 cases in early 2019 (Figure 2, c, purple). There is a large decrease in hospital volume and search query data following the public health lockdown of Wuhan on January 23, 2019 (Figure 2, third annotation).
9. No stats anywhere.
Can't do time series analysis without math. Spinning a yarn / telling a just-so story shouldn't be the level of reporting expected of academics in this context.
10. No attempt to build a keyword list or other reference information of continuing utility.
So the analysis dies with this one paper, essentially. There's nowhere to go from here.