Replicating no-intervention results

[andyljones]

I'd like to adapt your work to answer some additional questions I have, and as a starting point I wanted to check that I can replicate the results from the paper. A good place to start seemed like the no-intervention curves in Fig 8 & 9:

To my understanding, making these changes to the script and running

python simulate_agepolicies.py 5 5 --index 0

should get me a deaths csv that matches the figure, but at least for my five runs (vs your 100) it ends at 300k deaths rather than 400k - and they're pretty tightly clustered

Presumably there's some other parameter I need to alter that I'm missing. Do you have any ideas what it might be?

[andyljones]

Incidentally, I also noticed the lockdown factor in the script is 2, while in the paper it's reported as 10? Which one is used in the figures?

[CharpignonML]

Regarding your second question: Figure 8 and Figure 9 are simulations of scenarios, and as such they are not necessarily representative of all values used to calibrate the parameters in the previous section, i.e., the set (probability of infection given contact, mortality multiplier, start of outbreak/patient zero date). To choose the set that fits the Lombardy data best (the actual curve for the number of deaths being at the 50th percentile of our set of simulations), a lockdown factor of 10 was used.

In Figure 8, the lockdown factor is taken equal to 1 (as if there was not any social distancing measure in place for the agents who are not already isolating themselves - which they: (1) can do voluntarily with a mean time to isolation of 4.6 days if not having severe/critical symptoms, (2) are assigned to do immediately - via quarantine - if they present severe/critical symptoms, or (3) are simulated to do - starting at lockdown time - if they belong to the considered age category), whereas in Figure 9 the lockdown factor is taken equal to 2.

Does this help? Does that change the results of your simulations?

[andyljones]

Aha! My mistake was not making the connection between 'physical distancing' and 'lockdown', which is daft of me. I understand now, thanks very much!

Unfortunately, It doesn't change the results of my simulations. The curves I'm trying to replicate are the baseline no-intervention ones, where the lockdown never arrives.

[CharpignonML]

Hi Andy!

Sorry for the delay in response on our side -- we were finishing up the manuscript for submission.

• Yes, our physical distancing is ensured via several mechanisms: isolation within the household for agents who shelter in place (and therefore, contacts with agents outside the household do not exist), quarantine for severely and critically infected agents, and reduction in the number of daily contacts as imposed by the lockdown factor that modifies the country-specific contact matrix after a certain point in time (representing a policy intervention). Let us know if you have any further questions.

• We have now updated the release code to reflect the sample_households.py version of the code that we used to generate the policy intervention simulations in Lombardy, Italy, as well as the correct .pickle file for the Italian regional population (of 10 million agents) that accounts for the (hypertension, diabetes) comorbidity distribution. You are correct that the plateau value for the no intervention scenario should be about 360k deaths. If our paper is getting accepted, this will be the threshold value that will be appear in the manuscript.

Thanks a lot for sharing these pieces of feedback, and do not hesitate to send any other remarks and/or suggestions for improvement our way!

Marie.

[andyljones]

That's great to hear, thanks Marie.

Assuming I'm running the simulation correctly then, may I ask if your revised manuscript includes figures for age threshold lockdowns? My original interest in your work was that I was curious about the effect of locking down, say, everyone over 30. The results in the draft manuscript only show the outcomes for locking down ten-year age brackets independently though, and running the sim with age threshold policies instead gives more substantial effects:

• No intervention: 300k deaths
• Quarantine over 50s: 100k deaths
• Quarantine over 30s: 50k deaths
• Quarantine everyone: 30k deaths

To be clear, I get these by setting frac_stay_home = [0., 0., 1., 1., 1.] to lock down everyone over 30 say.

Incidentally, I really do appreciate your openness and friendliness here. Publishing code in sync with - or ahead of! - publishing the manuscript is still not a common thing, and it's great to see here.