Apologies to those of you who’ve been waiting for a model update: I’ve been slowed down a bit by work, start-of-term chaos with the kids, and by trying to organise an U13 girls rugby team. But it’s finally here in its glorious 25-tweet thread detail. Hope you enjoy…. 1/25
The July iteration of my model did an OK job of predicting (at least in “ballpark” terms) the level of hospital admissions over the last couple of months – in fact it’s almost spot-on right now, albeit maybe in the same way that a stopped clock is correct twice a day. 2/25
But that model won’t be a good guide to what happens over the autumn and winter, because it’s missing two significant drivers: waning immunity, and booster vaccinations. So I’ve upgraded the model to include those factors, and am ready to give you the results. 3/25
My 10 high-level conclusions are:
1.Without any waning of immunity, we’d probably be facing a relatively benign autumn/winter. In fact, even if behaviour returned to a pre-pandemic normal, we could see relatively few hospitalisations and deaths after November: 4/25
2. But with waning immunity, the outlook is considerably less favourable, with the potential for sizeable waves both in the autumn of 2021, and in early 2022. The outcome depends (unsurprisingly) on the speed of waning – and the faster it is, the worse things get. 5/25
3. The outcomes are worse if there is an age gradient on waning, such that older people’s immunity wanes faster than youngsters’. But if there is no waning of severe-disease protection, outcomes are significantly better (but still not entirely benign). 6/25
4. In this model, vaccinating 12-15s makes little difference to population-level outcomes, but there may be other reasons for doing this, and the model isn’t fully age-stratified so may miss some effects. Hence I’d be cautious in using this analysis for policy-making. 7/25
5. Booster vaccinations for older / vulnerable people (I’ve used the old JCVI groups 1-4 here) improves the situation, with a 20-30% reduction in total deaths and hospitalisations, although at the cost of a slightly higher peak in January/February. 8/25
6. However, the case for boosters in other groups is less strong – and it could even be counter-productive. Here we see giving boosters quickly to 50-70s and then to 18-50s before Christmas actually tends to magnify and delay the peak, with worse overall results. 9/25
But we can do better. For example, if we give boosters quickly to the 70+, but then slow down to a more sustainable rate (~1m doses per week in England) then that tends to suppress the peak, not magnify it. There may be even better strategies I haven’t found yet. 10/25
(Note I’m conscious there is a lot of uncertainty in this analysis, and so again would be cautious before using it directly for policy-making. It would be helpful to see the results of other models, e.g. from SPI-M, to check if they give similar predictions). 11/25
(And please note I’m not addressing here the ethical question of whether the gains from booster vaccinations in the UK are sufficient to outweigh the case for sending those vaccines to other countries – I’m just giving you some analysis to inform that debate!) 12/25
7. So far I’ve assumed that the immunity from booster vaccinations wanes in the same way as for 2nd doses; however if booster immunity is “stickier” (as suggested by @andrew_croxford here ), then the outlook for 2022 is much better. 13/25
8. I’ve also assumed so far that behaviour remains fixed at its September levels, with some ongoing WFH, use of masks, etc. But if this caution falls away, reverting closer to a pre-pandemic “normal” over the winter, we could see much larger waves. 14/25
9. But even if we did get a large wave that needed “flattening” to protect NHS capacity, I don’t see a case for a full-on lockdown; moving back one step along the roadmap (i.e. to the position in England before July 19th) should be sufficient to bring R<1 for a period. 15/25
(please note I’m not advocating doing this – I’m hoping it won’t be needed, but it makes sense to have a contingency plan. Also, while it would “flatten the curve” by splitting a large peak into two smaller ones, it could actually increase the total number of admissions). 16/25
10.Overall, the number of covid-related deaths predicted by the model, across a plausible set of scenarios, ranges between 10k and 50k, for the period of Sept 2021 to June 2022. This is in addition to the ~130k already recorded to date in England. 17/25
PLEASE NOTE that I have relatively low confidence in using the model to make forecasts over the next few months – there are a lot of factors that are hard to predict (e.g. human behaviour) or for which we currently have relatively sparse data (e.g. immunity waning), and… 18/25
…which can make a big difference to outcomes. So rather than reading numerical predictions from the model, I would suggest using it to reach directional conclusions (e.g. faster waning makes things worse, boosters improve things – but not if delivered too fast, etc.) 19/25
I would also note that the model excludes some factors that may become relevant over a longer time horizon e.g. births, deaths and aging. And lacking a full age-stratification or regional breakdown, it may miss some important effects driven by those factors. 20/25
For that reason and others, it would be good to see a new set of SPI-M models released by SAGE, particularly if these are being used to inform decisions by JCVI / CMOs / government around booster vaccinations and for 12-15 year olds. 21/25
In the meantime, I can compare my new “mid waning” base case to my old model, and the previous set of SPI-M models. Note the dotted line is not really a forecast, because it doesn’t include any impact from booster vax – the reality should be (on average) lower than this. 22/25
One note: you can see the new model doesn’t match hospitalisations perfectly in July. This is because it isn’t great at fitting admissions when there are large swings in the age mix of cases, as there were in the summer. Please bear in this in mind for future forecasts! 23/25
Also to answer an obvious question; no, the model isn’t published (sorry) – maybe when I have a bit more time to spend on this. But I do try to be transparent with the assumptions, which are documented in the attached tables – with changes and additions flagged in green. 24/25
In particular I’ve updated the model to the latest vaccine take-up / supply / effectiveness data, and tweaked a few other parameters (e.g. generation time for delta is now 4 days, not 5 days). As always, questions and suggestions for improvement are very welcome. 25 & END.

• • •

Missing some Tweet in this thread? You can try to force a refresh
 

Keep Current with James Ward

James Ward Profile picture

Stay in touch and get notified when new unrolls are available from this author!

Read all threads

This Thread may be Removed Anytime!

PDF

Twitter may remove this content at anytime! Save it as PDF for later use!

Try unrolling a thread yourself!

how to unroll video
  1. Follow @ThreadReaderApp to mention us!

  2. From a Twitter thread mention us with a keyword "unroll"
@threadreaderapp unroll

Practice here first or read more on our help page!

More from @JamesWard73

12 Sep
One of the questions I often get in response to my modelling threads (such as last Thursday’s below) is: what does this imply for levels of hospital occupancy with covid? Unfortunately I don’t have a good model of hospital stay dynamics, but helpfully...
…I know a man who does, and @nicfreeman1209 very generously offered to convert my various scenarios for weekly admissions into a corresponding occupancy forecast. So what follows is very much a collaboration on the analysis, but the policy commentary is all mine.
Essentially, Nic’s model uses the known data for hospital admissions and occupancy to estimate a distribution of how long people stay in hospital with covid – here’s some discussion on an earlier version:
Read 13 tweets
9 Sep
I know, I’m meant to be doing that modelling thread (it will come later this evening!). But first I got distracted by today’s release of data from @PHE_uk – thanks to the amazing @kallmemeg and the team there. The main point of interest in today’s report has been …
… the release of data on cases, hospitalisations and deaths split by vaccine status. In particular Table 4 of the report has caused some consternation as it suggests case rates are higher in vaccinated groups than unvaccinated, for age groups 40-79.
There are some real reasons why we might expect vaccine effectiveness to have declined – including the impact of delta, and waning of immunity over time. But there are also a number of potential confounders and distortions here, including:
Read 12 tweets
30 Aug
I spent a bit of time yesterday building immunity waning into the model, which is one of those bits of code I hope no-one ever sees (because it’s a massive hack), but it seems to be working OK so 🤷‍♂️. Now all I need is some numbers to go in the waning rate assumptions! 🧵
And that’s where I could use some help. The way the model works is as follows: as well as the classic compartments for fully susceptible (S) and recovered and immune via prior infection (R), I also have states for immune by vaccination (V) and partially immune (P).
People in state P are able to get infected, and to pass the disease on, but are unable to get severe disease (so won’t be hospitalised, or die). So in terms of waning, I’ve set things up with three different waning processes: R to P, V to P, and P to S, as illustrated below:
Read 8 tweets
28 Aug
England case data is looking moderately encouraging, with case ratios (compared to same day last week) dipping below 1 even on my adjusted curve which removes the Boardmasters spike – see orange line below. Note that last day will be adjusted up, but is unlikely to go over 1. 🧵
As a result, my estimate of R in England also dips below 1 (but only just… to 0.99) for the first time since early August.
I’m conscious that the case rates in over-40s are still growing, which isn’t ideal, and may mean that hospital admissions continue to grow as well. The hope is that, as in previous cycles, it’s the younger groups that move first, and then drag the other ones along.
Read 5 tweets
26 Aug
Three points as a postscript to the @boardmasters saga:
1. My final estimate of the number of reported covid cases from infections at Boardmasters is just under 9,000 (note this is restricted to the 15-19yo age group, as that’s where there’s a clear signal at regional level) 🧵 Image
2. While the Boardmasters spike was artificially increasing reported weekly case ratios for England last week, it is now artificially depressing them (as last week’s spike is now the baseline we’re measuring against) – note the adjusted orange line is now above the blue. Image
Without the spike, the case ratio for England would have been almost exactly flat for specimen dates 17th-20th Aug, suggesting an underlying trend with R = 1. So if people tell you English cases are falling, they may be literally correct, but I’d not celebrate too fast.
Read 10 tweets
24 Aug
I’ve been promising to update my comparison of how the July model projections are going. So here it is. As you can see, the model (orange line) did quite well in predicting admissions for the first couple of weeks (in mid/late July), but missed the sudden decline in cases… 🧵
…that happened after the Euros, and so the actuals fell below the projection (good news). More recently, the gap has been closing, but that’s for the (bad) reason that hospitalisations have started to rise again, while my model was projecting a gentle decline.
A couple of notes on the other models shown here:
1. I think we may have done LSHTM a slight favour by selecting their non-waning model for comparison, when their (more conservative) waning model was probably more their real central case.
Read 13 tweets

Did Thread Reader help you today?

Support us! We are indie developers!


This site is made by just two indie developers on a laptop doing marketing, support and development! Read more about the story.

Become a Premium Member ($3/month or $30/year) and get exclusive features!

Become Premium

Too expensive? Make a small donation by buying us coffee ($5) or help with server cost ($10)

Donate via Paypal Become our Patreon

Thank you for your support!

Follow Us on Twitter!

:(