> That would miss a lot of factors that are important if you're trying to predict epidemics at the society-wide level, but OK, so be it. You need a firm footing of the basics before you can progress to more complex scenarios.
That's how EE/CS stuff usually works (at least outside ML), building complex systems hierarchically out of well-understood primitives. The life sciences are different. There's almost nothing there we understand well enough to build like that, so almost all results of practical importance (a novel antibiotic, a vaccine, a cultivar of wheat, etc.) are produced by experiment and iteration on the complete system of interest, guided to some extent by our limited theoretical understanding.
This discrepancy has been noted many times; it's just a completely different way of working and thinking. If you haven't, then you might read "Can a biologist fix a radio?".
> Microbiologists seem to manage it?
A grad student in microbiology can grow millions of test organisms in a few days, at the cost of a few dollars, and get all the usual benefits of the central limit theorem. A grad student in epidemiology absolutely can't, since their test organisms are necessarily people. So you're quite correct that it's basically a social science, since it depends on aggregate human behavior in the same way e.g. that economics does, and is therefore just as dismal. Unfortunately it's also the best and only science capable of answering questions of significant practical importance, like whether the hospitals are about to be overrun. I'd tend to agree that stuff like Imperial College's CovidSim has so many parameters and so little ground truth as to have almost no predictive value. R0 seems fine to me though, and usefully well-defined, in the same way that the CAGR of a country's GDP seems fine.
In the life sciences, it's often possible to design an experiment under artificial conditions that will get a repeatable answer, like the growth rate of a plant in a certain controlled environment. It's much more difficult to use the result of such a repeatable experiment for any practical purpose; consider, for example, the steep falloff in drug candidates as they move from in vitro screens (cheap and repeatable, but only weakly predictive) to human trials (predictive by definition, but expensive and noisy). I'm absolutely not a life scientist myself, in part because I think I'd find that maddening; but essentially all results of practical benefit there came from researchers working in that way.
That's how EE/CS stuff usually works (at least outside ML), building complex systems hierarchically out of well-understood primitives. The life sciences are different. There's almost nothing there we understand well enough to build like that, so almost all results of practical importance (a novel antibiotic, a vaccine, a cultivar of wheat, etc.) are produced by experiment and iteration on the complete system of interest, guided to some extent by our limited theoretical understanding.
This discrepancy has been noted many times; it's just a completely different way of working and thinking. If you haven't, then you might read "Can a biologist fix a radio?".
> Microbiologists seem to manage it?
A grad student in microbiology can grow millions of test organisms in a few days, at the cost of a few dollars, and get all the usual benefits of the central limit theorem. A grad student in epidemiology absolutely can't, since their test organisms are necessarily people. So you're quite correct that it's basically a social science, since it depends on aggregate human behavior in the same way e.g. that economics does, and is therefore just as dismal. Unfortunately it's also the best and only science capable of answering questions of significant practical importance, like whether the hospitals are about to be overrun. I'd tend to agree that stuff like Imperial College's CovidSim has so many parameters and so little ground truth as to have almost no predictive value. R0 seems fine to me though, and usefully well-defined, in the same way that the CAGR of a country's GDP seems fine.
In the life sciences, it's often possible to design an experiment under artificial conditions that will get a repeatable answer, like the growth rate of a plant in a certain controlled environment. It's much more difficult to use the result of such a repeatable experiment for any practical purpose; consider, for example, the steep falloff in drug candidates as they move from in vitro screens (cheap and repeatable, but only weakly predictive) to human trials (predictive by definition, but expensive and noisy). I'm absolutely not a life scientist myself, in part because I think I'd find that maddening; but essentially all results of practical benefit there came from researchers working in that way.