Thread by @federicolois

1/ Very good thread about all the things that most modelers fail at, which is predicting outcome. As shown your assumptions are everything. For example, our model only estimate immunity through simplified cohort dynamics. Its good for that, it just happen it is also very useful.

https://twitter.com/theotherphilipp/status/1355908107248021504

2/ Language is a model that is useful to communicate ideas. In the same way as language, models shape reality. If your model simplifies an specific aspect, your reality also becomes less rich. You always have to be careful what model you accept as correct and why.

3/ but the most important realization you have to develop is that models are useful, until they arent. The more disconnected from reality, the more effects you are explaining without fully understanding them.

4/ For example, in our model fitting is everything, you don't start from an idealized situation; you start trying to derive what happened and the constraints that reality impose on it. We use serology, asymptomatic measurements, deaths series, prevalence of mutations, etc.

5/ if the model is not able to predict the mechanics of real life you need immediately to figure out if the approximation is gonna have a first, second or third order contribution to the phenomena. That's where the most interesting science happens.

6/ We found out that many things thought to have impact on the abstract don't have sizeable contributions to estimate immunity. For example, not long ago I had a discussion with an epidemiologist that believes that RtoD parameter is important. In the abstract it is.

7/ but the devil is in the details. On our model it can be shown that the actual contribution to the immunity is negligible, to the point that I doubled it, and you can look at it and know that it is wrong, but the outcome of roughly the same.

8/ we estimated it to be 11, this is what happens when you done it to 22. See how the model doesn't allow you to fit properly? But interestingly, under normal life there is a small difference, which for policy purposes is minimal.

9/ So, that begs the question: does it make sense to have complex distributions for the parameter when the mean can get you this with minimal impact in the end result?

10/ I still don't understand why epidemiologist insist in using low capacity predictive models where there are much better ways to do so. And it is not that difficult. We learned what we needed with @LDjaparidze to model epidemiological models in 6 months.