Public apologies due?

Only a month ago, a model by Mass General, Harvard, and Georgia Tech predicted that the opening of GA from 27th of April on-wards would result in a rapid increase of deaths over the summer. The predicted number was "over 23,000 deaths."

The image above is from the revised model which now puts the expected or predicted number of deaths at around 15,000 deaths by August. This still seems very high given that right now, GA has totalled just under 2,000 deaths.

Why are the models so inaccurate?

There is a fundamental flaw in the epidemiological models. Modelling responsive human behaviour is no easy task. Without empirical testing, how does one know the rate of compliance, and the actual effect of compliance? The simple answer is, you don't, so you plug in assumptions that are varying degrees of wrong. That's not a condemnation - it's just a reality of modelling.

Culture Matters.

Not only does one model's assumptions - if reasonably accurate for a specific case - not necessarily transfer well to other countries, they may not transfer well to different regions of a diverse country, and they may not transfer well across time within a particular place.

For example, political scientists are still coming to grips with the increasing partisanship in the US, that has diminished the amount of agreement on any politicised issue, only recently realising that it's not a blip but the new normal. On top of that - related, but not synonymous - we have begun to become aware how strong the anti-elite/anti-expert zeitgeist has become.

How are epidemiologists (or behavioural scientists in general) supposed to model that?

And that's assuming they know the empirical effects of X% of people wearing masks, or shutting down colleges, or prohibiting people from going to the park, etc., all of which is quite dubious. The number of different possible combinations of policy rules also creates an impossible - or at least forbiddingly difficult - to model complexity.

For better or for worse (policy)

Epidemiologists could be the smartest of us and still not be able to develop a model that captures everything we need to know for making policy.

Are they at least better than nothing? Not necessarily. What's missing or has had very bad assumptions applied may be crucial. Models may be consistently biased in a particular direction. Their conclusions could lead to policies that are better or worse.

Right now, with Covid-19, the models seem to be consistently biased towards high numbers. That's beneficial to nervous politicians who hate to be seen as doing too little when people are dying. That makes them politically useful. It doesn't mean they're leading to good policies.