Good spot. I think my point stands, that indicating a recent change in a 7-day average by comparing to a 14-day (or 21 or 28 or ...) average is inferior to comparing to the 7-day average from yesterday, or the day before or ...., i.e. looking at the rolling mean.Bird on a Fire wrote: ↑Mon Mar 01, 2021 11:11 pmAh, I think he's not plotting 7- vs 14- (etc-)day rolling averages, but the average of the most recent 7 vs 14 etc days - ie, averaging over the period shown by the lines.
As I've said, I couldn't easily find a description of how or why he made the decisions implicit in his charts - but I'm not skilled in Twitter . But charts like this... ... explicitly refer to "Modelled cases" and the "plumes" look like some kind of uncertainty interval. Any fit or forecast requires a model, whether acknowledged or not, and whether it's called "statistical" or not.Bird on a Fire wrote: ↑Mon Mar 01, 2021 11:11 pmAll his stuff is quite unconventional, though. He models the pandemic trajectory as a random walk with no underlying statistical model whatsoever, just fitting to the data.
His plots do seem to do a good job of fitting, but I can't recall seeing a comparison of say, last month's forecast with the actual observed figures, because the plots he posts each day are based on an updated version of the model.
Even descriptive statistics such as means entail important assumptions. For example, the 7-day arithmetic mean of daily cases has a clear and simple interpretation, the total cases in a week (divided by 7). The geometric mean (or mean of logs) of cases is much more debatable - what is the meaning of the product of daily cases.
On the other hand, the 7 day geometric mean (or mean of logs) of daily ratios is easily interpretable, the ratio over a complete week (7th rooted). Here the arithmetic mean is more debatable - what is the meaning of the arithmetic sum of ratios?