Originally Posted by
lazygun37
OK, I'll admit it -- at this point I'm sort of enjoying both of your inability to admit even the most obvious mistakes, at least to me. I mean, come on Rob: you typed over 400 angry words in response to me without once acknowledging your were wrong -- and then you hilariously say this in your very next post (in response to somebody else):
You're got to admit that's pretty funny.
What's less funny is that I think you still don't understand *why* you were wrong, because you then make the same mistake again literally in the very next sentence:
So, yeah, I'm telling you that's also wrong. I mean, even wiigelec understands by now that reporting delays of up to 1-2 weeks may cause slight problems if you're trying to draw conclusions based on data for the last 5 days. Which is why nobody serious would do that. (To be fair, wiigelec also still doesn't seem to understand the difference between totals *known* on a given day and totals *up to* a given day. But whatever.)
I live to serve.
In fact, I'm going to seriously address your point about my supposed "inconsistency" and my supposed "flip-flopping" between requiring perfect data from you while using imperfect data and approximations myself. If either of you could one day respond equally seriously, that would be lovely.
So: the problem I have with you and Rip is not at all that you're using imperfect or imprecise data. It's that you are often using *wrong* data (or, equivalently, misinterpreting data) to arrive at incorrect or unsubstantiated conclusions.
Statistics is all about helping us make decisions under uncertainty. All data is imprecise and inaccurate at some level. But in order to make optimal decisions based on imprecise data we need to understand the data we are using. If I use a measurement of some quantity -- like death counts -- I need to first understand how imprecise it is (because of purely random small number statistics), I need to understand how inaccurate it is (perhaps because of delayed reporting), and I need to understand whether it even measures the quantity I would like it to measure (e.g. if I'm trying to measure daily death counts, I can't rely on measurements of totals known on a given day). With this understanding, you can draw meaningful conclusions even from highly imprecise and/or inaccurate data. Without this understanding, you're virtually certain to draw the wrong conclusions and possibly mislead people.
Let me take a specific example. The average number of flu deaths per year in the US is around 30,000. The range around that is pretty large -- from something like 18,000 to 51,000 deaths in the last 7 years with complete data (and up to 61,000 if we include the last two years with incomplete data). So a statement like "22,000 deaths is pretty typical" is perfectly sensible in this context. And just to prove that admitting faults doesn't kill you: I do wish I'd said "not atypical" instead of "pretty typical". But it makes zero difference to the argument in this case, which was simply that the number of deaths was already in the realm of the seasonal and still rising. By contrast, you were using your incorrect daily death counts to draw the conclusion that the peak may well have been a week ago -- but that conclusion is completely unsupported by the correct data. I hope you can see the difference.
By the same token, all models are wrong at some level. That's true of Newtonian physics, of special and general relativity, of Maxwell's equations and whatever else you want to name. The question is not whether a model is "perfect". It isn't -- if it was, it wouldn't be a model. The question is whether it captures the essential parts of the problems you are trying to capture. And, in interpreting such models, it's critical to also account for "modelling error" -- i.e. the uncertainty that's associated with the fact that your model is imperfect.
Don't get me wrong -- of course models can be flat out wrong. What that means is that they fail to capture aspects of the data that they *should* capture if they were right. But this is why it's just disingenuous to say things like "the Imperial College model is BS; it predicted 2.2M deaths, and now we're going to have only 60,000". That prediction was intended to capture the case of an unchecked outbreak, so comparing it to the actual numbers for an outbreak with interventions doesn't make sense. And given that their model does predict deaths in the tens of thousands for scenarios with interventions, it's actually doing pretty well. Not perfect, obviously -- but well enough to help in decision making.