Read on Twitter
A riff on data, models and intervention in a COVID world.
When you don't have data, you absolutely need models - it helps you understand what could happen, good and bad scenarios and what you need to measure to understand more thing. With no - or little - data, models are king.
When you have data, the role of models change. In fast moving epidemic even working out what is happening *right now* is complex; your data is coming in different streams, it has biases (which change over time), technical issues (also time dependent) >>
<< and the things you want to measure/know about also change over time - eg, in this case, infections. Here models have three roles.
1. To help do the fusion of the different data streams. This is in some sense "old school" modelling and data fusion; eg Kalman filters from the 1960s is totally in zone here. Here models of reality (how infections become cases become hospitalisations) bind data streams together
You can learn the binding, or enforce the binding; you can have time varying parameters integrated over or not, but your data hopefully constrains everything enough for this to be as much "de-noising" data streams as it is "modelling".
2. Using data and models to do "nowcasts" and short term "forecasts". Forecasts here really do mean "we believe the world will be like this in X days time". Weather forecasts are the best analogy - measurements today to predict 1, 2, 7 days out.
Because of the long chain from infections -> active cases -> hospitalisations -> deaths plus reporting lags one can be pretty accurate given one has enough data in the other areas. The (sad) deaths happening now are about infection events 4-6 weeks ago mainly.
3. The third sort is back to the original scenario models. These can be crude, but useful - no human is good at understanding exponential growth. They can be sophisticated, and thus fragile to assumptions (eg, that control measures will or wont change).
Scenario models are not forecasts. They are there to understand the options we have on the table and critical parameters for this.
We have moved into a COVID world with lots more data (hats off to the ONS survey and the REACT survey in the UK; genuinely a good move here. I don't think a similar thing is present in FR or ES). We have far deeper testing across most European countries (all credit).
Wastewater testing has proved its worth in a variety of settings in Europe and US; we should aim for *every* urban area to have regular COVID testing on their wastewater. Finally the use of hospitalisation data as a key tracking parameter is widespread of course.
As such much of the decision space is constrained by this data. We know largely where we are *now* (or at least, for very sure, last week) in terms of infection levels and can forecast well the next bit.
...BUT... the presence of data, with good, appropriate modelling in any of the 3 ways above does not imply action! To change infection and hospitalisation we don't just have to know about it, estimate it, and plan - we need to *intervene*.
Interventions are many - the simplest, and the most important is isolation (aka "hyper-localised lockdown of individuals"). This is the most important thing to get right as our ability to see the infection more clearly comes into focus.
Other things include the routine but important ... washing hands, wearing masks.
But everything - including Trace - loops back to the key intervention of isolation when we know people (a) have the virus or (b) are at high risk of having the virus.
Vaccines - assumming they work as well as stated for Pfizer and more of the same from the other vaccines would be great - will really really help - not least because we can target the at risk groups.
