[diognesofsinope]

Same thought -- makes you think it has less do you with masks/mandates and more with city-folk being cautious as delta surges. Maybe the mandates encourage this behavior shift?

Top points if politicians noticed this difference off the bat and create the mask mandate to engineer the illusion that it was their policy that worked, even though it would've happened anyways -- aka action bias.

In any case, I think we can all agree people need to stop (1) using different random X-day moving averages and (2) give us the data.

edit: there's an entire literature on methodologies (1) Regression Discontinuity Design (RDD) and (2) Difference-in-Difference (DiD) that's designed to address this problem in analysis.

[mjevans]

The reason for a 7 day running average is because there is a great deal of noise in the source data, that is how you clarify the picture to remove short term trends in numbers.

Confounding trends might exist for various reasons like: labs want people to have a day or two off, so process less on the weekends or some days of the week in spreading that weekend out.

People might interact in different risk settings over the course of a week. (E.G. For some higher risk on the job and shut in on the weekends. For others time to go out clubbing on the weekends and shut in jobs during the week).

So the 7 day running average just says: we can't disentangle to get at the true raw numbers, this is the best we can do to get useful data out of the data-feed.

PS: The only way to get truly useful data would be sufficient testing capacity to issue fully randomized tests on everyone, in a mandatory way, on a regular basis. A 'well designed survey' would probably still have biases but might approach that. To the best of my knowledge (public news media), currently most tests are probably focused on populations perceived to be at higher risk, who show visible possible symptoms, or who are obviously ill.

[diognesofsinope]

I get why it's used, but I strongly believe people are better off seeing the raw data. Exactly as you said, we can probably infer if there is variation in testing (weekends).

Or at the very least show the 7-period SD so we can see if the confidence intervals overlap.

You can absolutely use averages to hide data points and data quality.

PS: I disagree with your PS -- if we're being wildly theoretical and impractical about this we should just create a second Earth where everyone wears mask and a third Earth where no-one wears masks and compare all the Earths....

And none of what you're saying addresses the core problem I mentioned, which is that the divergence in the series starts before the policy date.