Clive Best

On March 16th Neil Ferguson’s team published “Report 9” which changed government policy and triggered a “lockdown” a week later. The results of his simulation showed that the number of COVID  patients would soon overwhelm the ~4000 ICU bed capacity in the NHS. The code that produced his results has now been made available and I have spent the last few days struggling to get it working. Here are the first results I get after running Ferguson’s “Report 9″model.

COVID simulation for UK if left “unmitigated”

and in yet more detail.

Predicted “unmitigated” deaths using R0=2.4 IFR = 0.9% as used in report 9.

I was surprised to see the date of the peak predicted by the model because in reality the epidemic  occurred about a month earlier (starting March 1). So it looks like Ferguson originally thought that we had much more time to prepare for this emergency than in reality we did have.

The main impact of this report were the  measures he proposed to suppress the epidemic thereby avoiding “overwhelming the NHS” and “save lives”. I have spent the last 3 days struggling to get his code working and it has been a bit of a nightmare. The released procedure as was published on GITHUB could only really be run on a supercomputer,  while instead I have an iMac! There are 4 types of suppression interventions.

1. PC – School and University closures, restaurants, bars, non-essential shops etc.
2. CI – case isolation (7 days)
3. HQ – Household quarantine (14 days)
4. SD – Social distancing (at various levels)

The newly released “report9” process proposes to run the Covid-Sim model 10 times (multi-threaded) and then take the average. (The main reason to run it 10 times is because you get slightly different values each time).  In addition to this there are an additional 45 combinations of intervention strength and 4 different values of R0  (2.0, 2.2, 2.4, 2.6 ) to run. This makes a grand total of 180  CovidSim batch jobs, which is equivalent to 1800 single threaded runs! This can only really be run on a supercomputer. The full suite of combinations is basically impossible to run on my iMAC. So instead I decided  to restrict all my combinations to a single run (instead of 10) and to use only R0=2.4 because this was used in his original paper.  This produces a more reasonable set of 60 sequential runs which still took me about 2 days to finally finish while getting a headache. Here are the results I get for one  typical 4 level intervention scenario, more or less  corresponding to those shown in report 9.

Impact of 4 suppression scenarios on predicted deaths. The green curve more or less represents the lockdown measures the UK consequently adopted. Note that the dates are  a month later than what actually occurred. The maximum peak in deaths/day reached in the green scenario is ~ 400

Intervention detail. I am not yet quite sure why the second peaks appear yet !

The full lockdown measure finally adopted is shown in green resulting in a smaller peak in deaths after about 28 days followed by a long tail. So how do these prediction compare to what actually happened in reality. UK lockdown measures were introduced on March 23rd a month earlier than envisaged above. Here are the daily deaths in hospitals (excluding care homes) as reported by the NHS.

UK Hospital deaths by actual date

This indeed shows the same shape but twice as many deaths occurred than expected, yet at no time were ICU beds overwhelmed. The outbreak occurred a month earlier than Ferguson predicted. This  seems to be because there were far more infections in the community than were originally thought. It is now estimated that R0 was actually 3 instead of 2.4.

Hindsight is a wonderful thing, but it seems pretty clear that the UK  should probably have locked down a week earlier, and as a result total deaths probably would have been halved.

Update: (27/5) If you compare Sweden with a lockdown imposed on the same date as UK (March 23) then Sweden is apparently doing much better.

Everyone is watching Sweden where the chief epidemiologist Anders Tegnell has resisted lockdown. Cafes, schools and restaurants have all remained open throughout. Their policy of voluntary social distancing measures while protecting the elderly seems to be working.

I ran Ferguson’s model for Sweden after finally working out how he normalises to different populations. Basically he normalises model death predictions to those actually recorded for a specific date. For the UK this date is April 10 (10,000 deaths). This is how the model compares to Sweden if I use the same date (400 deaths). First with lockdown on 23 March.

Swedish deaths are less than predicted under a March23 lockdown scenario

Now with lockdown a week earlier on 16th March

ICL model compared to deaths in Sweden. The blue curve is a UK style lockdown beginning 16th March ( a week earlier than UK). The red curve is unmitigated deaths. The green curve are recorded deaths.

Accumulated deaths then appear to be about 1000 higher than they would have been had they applied a UK style lockdown on the 16th March. However the UK figures also show an overshoot of about 5000 deaths even with the 23rd March lockdown.

Different ICL lockdown timing simulations. The green and cyan graphs follow what actually occurred.

So in general Sweden and the UK are in a similar state currently. However the Swedish trend is showing a smaller decline implying that R is around or slightly above 1. If the goal in Sweden is to reach herd immunity while protecting the elderly, then it seems to be working. Everything will depend on whether a vaccine becomes available in September. If not then Sweden’s strategy could well pay off.

I have been running  the Imperial College Covid-19 model for different scenarios. The published  parameter files (on GITHUB) use an R0 value of 3.0 with an IFR of ~1%, higher than that published in February (2.4). This predicts a total number  of 620,000 deaths in the UK if the disease were allowed to run its course through to “herd immunity” without any mitigation. These are the figures that SAGE must have been  considering in early March, when the UK policy then was to only test arrivals from China and flatten infections  towards an eventual  herd immunity.  On the 13th March however the advice was suddenly changed:

The new advice issued by the Chief Medical Officer is as follows:
Stay at home for 7 days if you have either:

• a high temperature
• a new continuous cough

Do not go to a GP surgery, pharmacy or hospital.

Something dramatic changed very quickly and by March 23rd a full lockdown was imposed – Places (schools, shops, restaurants, pubs, businesses ) were closed, travel restricted  and social distancing measures applied. The new slogan was “stay at home, protect the NHS, save lives”.

I have used Neil Ferguson’s (ICL) model to investigate exactly why this sudden policy change occurred. It seems that SAGE had by then concluded that the current R value  was ~3.0 and infections exploding mainly in London. ICL’s  model was now predicting 620,000 deaths without government intervention. This was way too much to “take it on the chin” by “slowing the curve”. But what would have happened if the decision for lockdown had instead been taken  a week earlier or a week later ? This is what the ICL model says.

Different ICL lockdown timing simulations

Figure 1. How the timing of lockdown measures affected final UK death rates. Note also the  strange effect that the models and data all coincide on day 100 with 10,000 deaths (10th April). This appears to be because Ferguson forces the model to agree with the actual data  on this day.

It would have been far worse to delay the decision a week than to advance it a week! Hindsight is a wonderful luxury for all armchair critics. Fergusson’s model  predicts that if the lockdown had been imposed a week earlier, then it might have saved up to 15000 lives  in the short term. However if it had instead been applied a week afterwards it may instead have cost an extra 40,000 deaths !

The IC model is driven by various parameter files which are very obtuse and difficult to fathom out without any proper documentation. So instead I simply delayed all  intervention start  dates by ~ 1 week to get these results. However this results in the day 100 effect as described above. ATTP has an ad hoc  fix for this but I am not sure if this doesn’t perhaps also bias the result.