Saturday, March 28, 2020

On the Problems With the COVID-19 Modeling That was Screaming Alarms

At the post, Now He Tells Us: The Latest From Nutjob Ferguson, Kratoklastes comments:

They used a SEIR (Susceptible -> Exposed -> Infected -> Recovered) model, where the changes over time in each of S, E, I and R are governed by a set of very basic differential equations.

About a month ago I built a discrete-time version of a SEIR model, because it seemed to me that there were far too few deaths for covid19 to be worth worrying about.

I picked what seemed to be sensible parameters - and lo, when ICL released their latest model, it turned out I had picked very nearly the same values as they had (I had a longer 'infectious' period).

The model is horribly unrealistic on a bunch of levels; first, there's no genuine constraint on the number of infected - such constraint as there is, happens as a result of a stupidly unrealistic mechanism for the transition equations for 'Susceptible'.

I replaced the stupid mechanism with one that was based on behaviour (i.e., the number of interactions that infected people have per day, and the probability that an interaction infects someone else).

Anyhow... suffice it to say that if a single asymptomatic individual was introduced into the US population in December 2019, by mid-Feb 2020 there would have been millions of asymptomatic-infected wandering around infecting others. Well before the 'lockdowns' and what-not. By the time governments started to respond, SEIR says it was already too late.

I'm mounting the model on a webserver this weekend, so that people can see for themselves how the model behaves when they 'tweak' the parameters - including allowing them to change the date for 'Patient Zero', and then see how the modelled infection happens over time.

As it happens, my opinion of the SEIR model is that it is amateurish garbage... that's irrelevant: it's the model whose results are being bandied about as if it produces 'hard numbers'.

My background is in large-scale economic and econometric modeling - specifically in performing systematic sensitivity analysis to get some idea about the statistical properties of forecast 'manifolds' (as opposed to 'lines' or 'points').

A lot of people are very hostile towards economic and econometric modeling - mostly because it tells them what they don't want to hear (e.g., that the policy they're infatuated with, won't do what they calim). By contrast, medical statisticians (including epidemiologists) get a 'halo effect' by being part of an authoritarian profession with government imprimatur... despite the fact that medical error is the 3r or 4th highest cause of deaths every year.

Lastly: Consider these two things together, and ask yourself how it affects the model.

① There is a very high probability that Patient Zero exhibited no symptoms, and was simply waved through border controls: prior to March 15th, nobody was bothering to test (or quarantine) travellers who had no symptoms. Furthermore, 30-50% of people infected downstream of Patient Zero will have been asymptomatic.

② There won't have been a single 'Patient Zero' for the whole US, either: there will have been several on the same day, at every port of entry into the US. Many of them will never have developed symptoms, and will now be clear of the virus.

If there were 3, or 5, or 20 original 'Patient Zero' in late December, then there would be the same number of SEIR processes happening simultaneously.

(This is what makes me certain it's a nothingburger: otherwise there should be ~1000 people a day dying in the US at this stage).



  1. There were 401 deaths yesterday. The rate of increase is pretty high right now, and it would just take a 20% daily increase (below what we saw last week) to have death rates in the ~1000 per day by next weekend.


    - Are new deaths being accurately recorded? Should people who die with a positive coronavirus test but died coincidentally, still count?

    - If there's an increase in effective drug treatment (which is so political now it likely can't happen soon enough), that would slow the rate of growth.

    - Most importantly, if we're so far along that the inflection point arrives this week, then the rate of growth of total deaths will keep us below 1000/day.

    Wait and see!

    In the meantime let's get ourselves ready for the post coronovirus counter attack on the sweetheart cronyism between governments, hospitals and medical suppliers. It's up to us to let the people know all of their sacrifices were made to save these cronies from the deals they have made to reduce supply.

    We had to flatten the curve because they flattened the supply curve. A good libertarian book waiting to happen.


    1. He didn't say there wouldn't be deaths of a 1,000 per day but that the models said they would already be happening.

    2. Yep and I agree. My point is that if 1000 per day starts happening in a week, most people will think the models were pretty close and won't agree with the statement "this is a big nothingburger".

      I'm not sure what I think on that right now.


  2. I can relate to the Contractor's observations. I particularly like his suggestions about a "counter attack". However, I must say again that statistics is not the right tool in analyzing the effect of viruses on humans. Nor is it helpful in forecasting infection or death rates. Kraktoklastes admits as much when he states "I picked what seemed to be sensible parameters..." When reality showed these to be wrong he changed them and complained that the SEIR model was "garbage". There is an old saying in statistics called GIGO (garbage in, garbage out). If your parameters and assumptions (part of your inputs) are inaccurate your results (the outputs) will be inaccurate.

    To avoid GIGO you must analyze an operation that repeats in exactly the same way over and over again. Manufacturing production lines are like this and statistics is very useful in quality control. The interaction between viruses and humans is the opposite of this. The interactions change from time to time and from individual to individual as the virus and the human immune system responds and changes. Epstein hinted at this clumsily in his comment about evolution. My point is, because of the nature of the interaction process, you can find statistics that forecast high death rates and statistics that forecast low death rates. Statistics is simply not helpful and the data reported by the CDC is actually unhelpful.

    However, humans have thousands of years of experience with viruses and our own immune system. This knowledge tells us that healthy humans have little to fear from viruses. And it tells us voluntary individualized action is the best defense. Wash hands, cover mouth when coughing, stay at home when feeling ill or if you want to avoid contact with others who you believe to be ill. No coercive mandates necessary and thus no collateral damage.

    Again, first principles would be the best guide in reacting to any health threat. First do no harm. Unfortunately bureaucrats and politicians are not incented or rewarded for doing nothing.

    1. and to maintain and fortify your immune system with Vitamin D3 (80% from sunshine/20% from food& supplements); Vitamin C; Zinc (; avoiding sugar and simple carbohydrates. Don't just be allopathic. Now that telecommuting is the rage, consider relocating to a less urban, sun drenched locale.