A friend of mine has been playing with simple models of the spread of SARS2 in Japan since the beginning. It was always interesting but nothing mind-blowing. Recently he added vaccination into it and things got very interesting.
This thread will cover implications of the model and speculation about Japan’s future epidemic the technical details of the model speculation about why such a simple model is so good (in Japan)
Implications: The model strongly implies that the recent massive drop is due to vax. This bottom chart shows the 2 components, mobility and vax. Vax has been a massive negative contribution in the last few months, outweighing mobility since late August.
The vaccination effect gave us a lot of headroom to increase mobility without getting back into exponential growth but it looks like we're already at the edge of growth (according to the model and the public test results).
As vaccination-induced immunity wanes, the numbers could rise sharply but vaccination is still ongoing (at a lower rate than before), so that might be a while. The uncertainty about waning makes predictions very hard.
In general, forecasting in Japan has not been great. E.g. the recent massive drop seems to have blind-sided everyone. This simple model doesn’t really help to forecast case numbers because we can’t predict the G-Mobility data.
What this simple model can do is tell us how much we can add to mobility before entering growth. This could give us some control over our fate by triggering changes in behaviour or vaccination schedules.
Depending on what interventions remain in place, it might be possible for Japan to have a very normal life and stay out of growth all or most of the year with well-timed vaccinations.
This article makes a very similar point. I cannot find any details of Prof Osawa’s model.
https://www.japantimes.co.jp/news/2021/10/09/national/japan-booster-shot-plan/
Meta-implications: G-Mobility and vax are enough to account for almost all variation in the rate of spread. That implies that, in Japan, G-Mobility is an excellent proxy for interpersonal infection risk. Any model without something like that is probably not a good model.
Forecasting vaccination rates is pretty easy. Forecasting Google Mobility is hard. It is likely influenced by
- weather
- govt policies
- headlines about people dying untreated in ambulances
- backlash against holding the Olympics
but some of these are somewhat predictable.
The model: It’s relatively simple. It assumes that the rate on a given day (R_t) is some combination of
- 6 G-Mobility indices
- number of people fully-vaxxed
- number of people who were fully-vaxxed over 168 days ago (to model waning effect)
- fraction of delta vs non-delta
With just these, it does a remarkable job. Here’s the modelled vs the observed R_t for Tokyo (middle chart). It captures the trends very well - the various waves, the massive drop and the recent reversal of this as people get out and about.
It assumes that R_t is a weighted linear sum of these 9 things. It chooses the weights using standard linear regression against the historical data. That’s all there is.
You can read a bit more about it here
https://github.com/graphy-covidjp/graphy-covidjp.github.io/blob/master/blog/2021-11-01.md
You can find lots more graphs here
https://graphy-covidjp.github.io/#predictions
It’s updated daily.
Why it works so well? I think part of the reason is that the interventions in Japan fall into 2 categories masks, ventilation etc. These have been pretty constant from early on. govt mandates. These have all been of the type that show up clearly in the Google mobility data.
In other countries there are varying mandates for masks in public/school/etc. sometimes varying a lot by town/county. Tracking these and integrating them into a model adds much more complexity.
People in Japan rarely visit friends’ houses. In other countries, when closures are enforced, people go to their friends’ houses. This is invisible in G-Mobility. In Japan, “residential” almost certainly means “in your own home”.
I have concerns about the modelling of waning in that we don’t really have a lot of people in the 200+ days group, so its influence may have been underestimated and the 200 days threshold is somewhat arbitrary.
