The Covid-19 Files (4): Comments on “Impact of NPIs…” Paper of 16th March

The UK’s coronavirus response, in particular the transition over the last week from a “mitigation” policy to “suppression”, has been informed by computer modelling carried out by a team led by Imperial College, London, whose findings are documented in the paper Impact of non-pharmaceutical interventions (NPIs) to reduce COVID-19 mortality and healthcare demand (pdf), dated 16th March.

The paper suggests a series of interventions that would need to be in place for 5 months to reduce the scale of the UK epidemic by 75%. I wondered whether the model would indicate a shorter period of more rapid suppression of the epidemic with more drastic interventions (as in China, Italy and other countries).

The paper provides a correspondence email address (Professor Neil Ferguson, who is himself recovering from Covid-19!), so I’ve just now emailed my comments (please excuse the copy-paste artefacts):



I hope you’re now recovered from your bout of Covid-19!  I know you’re very busy so I’ll get straight to the point.

Page and table numbers below refer to your paper “Impact of NPIs…” of 16th March. 

1. You note (p.3) that the impact of NPIs “depends critically on how people respond to their introduction” [my stress].  It’s obviously also very important how strictly NPIs are enforced, e.g. by mandated closure of bars, cafes, venues, non-essential shops and other sites of contact, regulation of use of supermarkets and pharmacies etc.  Have you therefore carried out sensitivity analyses of the parameters given in Table 2 (and any relevant others) with a view to providing government with policy options that could potentially bring down infection rates more rapidly and to a lower-level? 

E.g. for SD (and similarly SDO) what would be the effect of “lockdown” measures aimed at reducing workplace contacts by say 90% (i.e. to just the say 10% who are health and other key workers) and for “contact outside household, school or workplace” by, say 95% rather than 75%?  Similarly for CI and HQ the reduction in non-household contacts could be 95% with 95% compliance.  For PC, I understand essentially all universities are or soon will be closed (you have 25% remaining open), though some are housing overseas students trapped in the UK (and some no doubt carrying out essential research work related to the epidemic, such as modelling!).

In the real world, of course, an “x% reduction in contacts” resulting from a work at home (and furlough) policy may represent part of the population reducing their workplace contacts by 100% and the rest (health and other key workers) by somewhat less.  Is this captured by the model?

I’m sure you’re in a better position than I am (in particular since I don’t have information as to how the parameters interact) to devise sets of parameters representing at least one additional scenario of stricter suppression of the epidemic.  

2. Similarly, what would be the effect of reducing infection rates during those contacts that continue?  There must be parameters in the model for this.  (Or perhaps it may be expedient to model this by further reducing the number of contacts, though that would make calibration with the real world more difficult).  In London at present, it seems very few people are wearing face masks, e.g. on public transport (it might be particularly important for store checkout staff to do so); there is apparently some crowding on buses and tubes (I advocate running full services to reduce this, advising passengers to sit apart, not cutting back services); food stores are being allowed to become crowded; and touch screens, keypads and basket and trolley handles are not being disinfected between customers.  And that’s just some of what I’ve personally noticed!    

3. I’m not sure effective contact-tracing is currently taking place in the UK, but it needs to be re-established.  If contact-tracing is employed then not all contacts are equally dangerous and I wonder if the model can capture this.  At the risk of echoing Rumsfeld, there are known contacts (e.g. meeting attendees) and unknown contacts (e.g. those sharing public transport).  The point is that your model allows for the quarantine of contacts within the household (HQ), but not (at least explicitly) for the quarantine of other contacts, even when we know who they are. 

The advantage of treating known and unknown contacts differently would be to model the effect of different kinds of intervention, of course.  In the meantime, though, you could introduce a CQ (contact quarantine) intervention and model the extent to which that reduces infection rates (i.e a % of contacts of known cases who self-isolate).  This would indicate the effectiveness of contact-tracing and isolation measures, but, even if there are now insufficient public contact-tracing resources, one step the government could take is to urge those who have symptoms and are self-isolating at home to contact their recent close contacts and let them know.  For example, it seems I was at a meeting earlier this month with someone who has since developed symptoms, yet I only found out about this by chance.

4. On p.4 you describe seeding occurring “at an exponentially growing rate”.  For what period is this assumed to continue?  The reason I ask is because – although seeding may well be growing at an exponential rate now, with apparently uncontrolled repatriation to the UK from (other) infected areas – once the first wave of the epidemic is suppressed (and, to be clear, I would advocate something more like “eradication” as in China, bringing the epidemic down to a lower level and more rapidly than modelled in your suppression scenario) it would be useful to model the effect of a testing and quarantine process for arrivals to the UK from those remaining affected areas (or of all arrivals). 

It’s not the priority right now, of course, but it would be highly desirable to model the strict travel protocols and exhaustive contact-tracing procedures that could potentially prevent a second wave of the epidemic, once the current wave has been suppressed.

5. Is there a “learning” element in the model?  I’m sure much of the following is included, but, to spell it out, the mortality rate will be a function of the number of cases occurring at a given time and in the preceding period (because of pressure on health services), which should be adjusted for the time of year (since there is more pressure on health services in winter months), but also of the cumulative number of cases (because of learning by health professionals in the UK) and of time (because of shared learning by health professionals globally, i.e. identification of the most effective treatment regimes such as the use of antiviral drugs). 

Obviously we await a vaccine (when the model can incorporate vaccination rates), but much can be done before it arrives, cf. AIDS which is now manageable.  Another way in which time is on our side is thus in the availability of health workers – if we can suppress the epidemic until there is even limited availability of a vaccine we would be immensely better off, since the available supplies could be used to immunise health-workers. 

Incidentally, even absent any vaccine, infection rates might be high among health workers, creating individual and even some herd immunity in that subset of the total population (and improving their effectiveness because of the reduced need for personal protection).  This would effectively increase resources to deal with any second wave of the epidemic, further suggesting the imperative of drastic action now to suppress the first wave.

Btw, I don’t suppose you model health workers separately from the rest of the population?  If not, maybe that’s a suggestion for future work.       

6. Have you tried to validate your model by inputting parameters representing the NPIs implemented in other countries, e.g. China, Italy?

I’m deeply concerned that the controls implemented so far in the UK and in particular London are insufficient to reduce R even to below 1.  If I’m correct (and policy remains unchanged) this will only become apparent in a week to 10 days when curves of serious cases (we’re not counting mild cases) and, later, mortality do not begin to level off or even inflect.  At that point the incubation period will mean case numbers will continue to increase for at least another 7-10 days. 

Furthermore, there are practical difficulties in maintaining infection controls for 5 months as modelled, so government needs an option that would reduce R to well below 1, to bring infection rates down to a very low level in a month or two, as in China. 

My concerns about current UK policy are shared by a number of very analytically talented individuals for whom I have a great deal of respect.


Tim Joslin