Posted on 04/02/2020 9:05:29 AM PDT by Willgamer
Of course, you are correct. I dont know analytics. It just looks to me that if you try to dit a curve to # new deaths without forcing it into a polynomial function you are just going to get an exponential curve. But, of course, I dont know analytics. How would you do it?
"All models are approximations.
Essentially, all models are wrong, but some are useful.
However, the approximate nature of the model must always be borne in mind."
George E. P. Box, Statistician
Some simple examples ...
1. Motor vehicle crashes. The physics of every motor vehicle crash is unique, and the reaction of a human body to an impact in a crash will vary wildly. Fatality rates on a highway can also be influenced by things like EMT response time, underlying medical issues for victims, etc. And yet that doesn't stop me (in my profession) from using models to make predictions about the costs and benefits of highway projects where roadway geometry is changed, speed limits are raised or lowered, or other changes are made in the design and operation of a highway.
2. Hospital capacity forecasts. This type of modeling is done all over the place. In general terms, a population in a region will have X hospital beds for each 10,000 people. This can be divided even further into subsets like X1 (emergency room capacity), X2 (top-level trauma center capacity), X3 (infectious disease handling), X4 (cardiovascular treatment capacity), X5 (orthopedic treatment capacity), X6 (neurosurgery capacity), etc., etc. There are tremendous variations among different populations that will drive inaccuracies in this type of modeling, but that doesn't stop the medical industry and public health officials from using this approach to forecast changing health facility needs over time.
Sobering note here ... I am aware of at least one state government that uses third grade reading test scores as a primary variable in forecasting prison cell needs ten years down the road. Think of all the things that can change in ten years with human beings ... and yet the reading test scores of a large group of 8 year-old kids turns out to be a very good indicator of felony crimes committed by 18 year-olds a decade later.
My sense here is that the wide (effectively unlimited) range of variables in biological modeling presents less of a challenge than the SPEED at which these models must be re-calibrated in many circumstances. A viral outbreak is a classic example of this. This is not like tracking lung cancer rates and tobacco use over years and decades. It's more like trying to model the structural damage in a building while it's burning to the ground in a matter of minutes or hours.
In the real world we also got South Korea and Singapore. Very different results from Italy and New York, I’d say.
A huge problem is these models are being produced by academics who are no longer selected by merit but by race and sex.
“Sobering note here ... I am aware of at least one state government that uses third grade reading test scores as a primary variable in forecasting prison cell needs ten years down the road. Think of all the things that can change in ten years with human beings ... and yet the reading test scores of a large group of 8 year-old kids turns out to be a very good indicator of felony crimes committed by 18 year-olds a decade later.”
Yup. It’s a great big state on the West coast, no?
“.... a limited number of factors,while the real world has a huge number of them. .”
And its sometimes very difficult to impossible to know how to weight these factors in how they contribute to the reality you are trying to model. Often its a linear model trying to simulate something thats non-linear which means your model will only be accurate (sort-of!) within a very narrow range of parameters. So you guess, hopefully and not be too wrong, but accept that you likely are and don’t fall in love with your model.
I’ll try to help you see-
re #1- the physics of every motor vehicle crash is unique... NO, the physics algorithms are well known, the data is varies wildly.
re #2- this is queuing modeling, a well described subject.
In the cases you cite, the models are built with a reasonable knowledge of the inputs, variables, and algorithms.
Modeling real world biological events requires identifying inputs, variables, and algorithms that we only know very incompletely, much more incompletely than in the above cases.
The proof is the utter failure of both climate change and epidemiological models to predict the future (except with the occasional blind squirrel finding a nut).
Again, nature in the “wild”, as opposed to in vitro testing, is unimaginably complex... well beyond our current understanding.
You can update these models with the SPEED of real time inputs, but it will offer little improvement. This is easily demonstrated by taking the model, say any of the Wuhan ones currently being used, and plug in the complete information from a month ago and run the model forward. Does it correctly give today’s numbers. NO.
Excellent post.
Disclaimer: Opinions posted on Free Republic are those of the individual posters and do not necessarily represent the opinion of Free Republic or its management. All materials posted herein are protected by copyright law and the exemption for fair use of copyrighted works.