Read on Twitter
SA COVID-19 model updated:
I have not tweeted about it in some time. So as a reminder the model is calibrated to 90% of excess deaths from the MRC (up to 16 January).
https://lrossouw.github.io/covid-19/modelling_covid-19_in_south_africa_at_a_provincial_level.html
Attack rate and predicted deaths shown in charts below (blue - model)
I have not tweeted about it in some time. So as a reminder the model is calibrated to 90% of excess deaths from the MRC (up to 16 January).
https://lrossouw.github.io/covid-19/modelling_covid-19_in_south_africa_at_a_provincial_level.html
Attack rate and predicted deaths shown in charts below (blue - model)
This is not a short thread as the model needs lots of caveats firstly. Modelling COVID-19 even with a "simple" model is not easy. All models are wrong, but some give you useful insights (adapted from George Box).
The model uses an assumed infection fatality rate (IFR) from international research with some adjustments for age profile and HIV prevalence in provinces in SA. This is a key assumptions and is not calibrated from local data. I have tried to calibrate cases too but no luck.
If IFR is higher (due to say hospitals running out of capacity) then the model would overestimate the spread of the epidemic and underestimate future deaths. This might well be the case in this second surge. And hence you need to think about that.
The model was not fit to excess deaths data released today. It takes a couple of days to update. The model also tends to predict forward the status quo so during rising wave (as measured by excess deaths) it will continue to predicting a worsening situation.
So the model predicts for 25 January an attack rate of 68.9% [58.1%-77.1%] and COVID-19 deaths of 130 211 [111 483 - 150 007]. Clearly given the results of 125k excess deaths to 23 Jan the model is overpredicting due to extrapolating the trend. [ ] = 95% confidence interval
The good news is that the model is predicting, somewhat in line with information released today, peaks in most provinces. For example in Gauteng deaths are expected to peak in the next week or two (also see SA at the top of this thread):
https://lrossouw.github.io/covid-19/modelling_covid-19_in_south_africa_at_a_provincial_level.html#98_Gauteng
https://lrossouw.github.io/covid-19/modelling_covid-19_in_south_africa_at_a_provincial_level.html#98_Gauteng
There are I believe two main reasons the projections are turning. The first being that reductions in mobility (associated with Level 3) has resulted in a reduced R over last couple of weeks. You can see that reduction in mobility below.
The second reason is that due the rapid surge in the last month or so we have seen many deaths and these are linked with a large number of infections. This means the attack rate is modelled to be high. 68.9% [58.1%-77.1%].
This implies the spread of the virus may be slowing down due to fewer people to infect. I focus on the lower range of that interval as we are seeing the model is overpredicting (at the peak of the wave) but at 58.1% is still a large portion of the population.
Thus, at least in the model, the virus is slowing down to larger portion of the population having being modelled as having had the disease. But, remember the model works "backwards" from deaths so that IFR assumption is key.
If the IFR is wrong the peak may be mainly due to the Level 3 rules and not high proportions of infected. Which would mean the country responded better than the model is predicting. So also some good news.
The worst case scenario projects what would happen if we 100% stopped caring about protecting ourselves. The prediction is 86.4% [82.5%-90.5%]infected and 220 029 [203 879-238 617] deaths by end of June. This is an underestimate because the variant is more transmissible.
One can see the increased R due to the variant in the model. The "error" term contains an indication of a rise in R that could not be explained by other means. Here is the error term for Eastern Cape:
The error terms adjusts the R. It increased from 0.75 (reducing the R by 25%) in August to about say roughly 1.25 (increasing R by 25%) now. If we assume that at least some of this rise is caused by the variant that does provide some hint of increased transmission.
So the worst case is probably worse than the model is predicting because it assumes that the error term goes back to 1 (resulting in a R of roughly 3). It also means that we are not sure what the herd immunity figure really is due to variant. It's probably higher than before.
It should also be noted that infections and deaths do not stop at the herd immunity and continue well past it. So we still need to be safe and we do need vaccinations to save lives (now and in the future).
The IFR is possibly wrong and the model is overpredicting a bit but the news about reaching the peak of the second wave is good news.
The main risks are the model is very wrong, and of course more new variants may emerge before meaningful vaccination can reach enough of the pop.
