3 ms·
People complain, but a Gaussian is good enough for a rough approximation. The real curve is something like the derivative of a logistic, convoluted with somethi
by gus_massa 7y ago
People complain, but a Gaussian is good enough for a rough approximation. The real curve is something like the derivative of a logistic, convoluted with something like a 15 day interval[1], and some noise, a lot of noise. Take a look at the SIR models and similar https://en.wikipedia.org/wiki/Compartmental_models_in_epidemiology https://en.wikipedia.org/wiki/Compartmental_models_in_epidem... , that are betters models. They are close enough to a Gaussian, but the Gaussian has a faster decay, so a Gaussian approximation underestimate the length of the tail.
I don't want to be mean, but the curve for Argentina is fucking inaccurate. [Hi from Argentina!] https://covid-gauss.site/country/argentina https://covid-gauss.site/country/argentina The expected peak is in the middle of May, the epidemy is not reducing here :(. Moreover, we made a huge mistake that effectively broke the quarantine last week, so I expect the peak to be earlier and higher.
Perhaps your fit is good enough for post-peak countries, but not for pre-peak countries.
[1] The important part is not how many people is diagnosed, but how many people is in the hospital. So it is necessary to model that the bad cases will need something like 2 weeks / 1 month of hospitalization.
- eatthatpie 7y agoThanks for sharing the link about SIR :) So far I was thinking about changing the model for a given country if the Gaussian approach would cause unacceptable inaccuracy. If the data for confirmed cases is not that accurate, can you recommend a more valuable data source? Best wishes for people there in Argentina :)
- gus_massa 7y agoThe Gaussian is good enough for a graphic, but I'm not sure how much systematic error adds when it is used to fit the data. I don't have a good intuition or theoretical knowledge about that. I use the phrase "Everything is a Gaussian" all the time in the sense of https://en.wikipedia.org/wiki/Central_limit_theorem https://en.wikipedia.org/wiki/Central_limit_theorem but I'm not sure if it produce a good fit when you have only the left tail. The problem with Argentina is that we had the usual exponential grow, and then we started a quarantine so the grow changed to a slower exponential. The automatic fit confuses the change in the shape with the region near the top of the Gaussian, so it produces a bad result. Sometimes it is tricky to fit a function [1] Looking at the data, they look weird. I expected something like this image: https://imgur.com/a/eO8VyQ3 https://imgur.com/a/eO8VyQ3 But it doesn't look like the data at all. In particular I don't remember the big peaks and the zeros that are in the data, but perhaps I was not paying attention. Best wishes for you too. [1] Have you ever fit a oscillating function? It is a mess. I only got good results if I pick the initial point by hand.