The "worst case" models were projecting how things would go if we took no precautions, which led to precautions being taken, which subsequent models took into account.
Social distancing is working.
but nearly every single model has been way off.
It's an apples-to-oranges comparison, though. The ones that projected millions of deaths were assuming that no precautions would be taken. This was done to illustrate how much those precautions were necessary. New models include the measures that we have taken. Some project what happens if relax those measures too soon, others show what happens if we tighten them. It varies by model.
As for why the models have been "off" – this is essentially what's happened:
Models: Hey, look at how bad things are gonna get if we don't do anything!
Us: Oh damn that does look bad, we should do something
Us: [Does something]
Models: Hey good job, you've slowed the spread a bit, which has a huge ripple effect on future infections, so our new curve is much flatter!
The virus is HIGHLY infectious. The initial models projected zero social distancing measures, so every person who is now *not* getting infected due to the measures that we are taking is subsequently preventing hundreds of others from being infected.
The CDC actually traced Chicago's outbreak back to one person who attended two family gatherings, getting at least 14 people infected, three of whom died. Those 14 people then infected even more people, who infected more, etc.
I'm not sure what the actual number would be in terms of average people one sick person would infect, but let's call it 10 (probably a bit high, but the Chicago example is higher, so let's just use that as an example).
- Patient zero passes the virus onto 10 people.
- Those 10 people then each pass it on to another 10
- Those people infect 10 more people, who each infect another 10 people.
That's 10,000 people infected very quickly. But in the second scenario, let's say that 5/10 people practice social distancing, and measures don't change. Here's how the math changes in Scenario B:
- Patient Zero has 50% fewer social interactions, infecting 5 instead of 10
- Those 5 also have 50% fewer interactions, infecting 5
- Those people each infect 5 more people, who each infect 5 more
With 50% social distancing, cases from that one patient would drop from 10,000 down to just 625 – a reduction of 93%.
Let's say social distancing catches on more than 50% halfway through this process. Scenario C:
- Patient Zero infects five people
- Those five infect five more, but then social distancing measures increase by another 50%
- That wave of people on average infects 2.5 more people, who in turn infect 2.5 more each
Same number of waves, but the increasing effectiveness reduces the spread down to 156 new cases, a 75% reduction from Scenario B and a 98.5% reduction from what the spread would be with no measures at all.
Now, if measures RELAX instead of getting stricter:
- Patient Zero infects 5 people
- Those 5 infect 5 more, then social distancing measures get relaxed or we decide to reopen everything
- That wave then infects 10 people each, who infect 10 more
Again, same number of waves, but the relaxed measures lead to 2,500 people getting, a 400% increase over the static social distancing scenario and a 1600% increase in cases over stricter social distancing. And that's just four waves of transmission, which is a very short timeline all things considered.
So, essentially, any preventative action early on will have a HUGE ripple effect, because when one person avoids infection, potentially hundreds of subsequent infections are also avoided as a result. That's why the models have been "off" by so much – and that discrepancy is the direct result of us heeding those models and taking action.
That's not to say that the models are perfect – there are still plenty of unknowns (individuals with antibodies, untested cases, etc) but some of the models try to account for those the best that they can. But hopefully that explains why the initial models had way scarier numbers than what we're experiencing – because we're dealing with exponential spread, so any mitigation has a HUGE impact on future modeling.