Hacker News Reader

Comment by nstart | Hacker News Reader

I understand that sentiment. But at the same time, if it ain't misinformation, I'm not sure what's wrong here. Sometimes the way to get started is by writing stuff while you are still too arrogant to not doubt (said with much affection). From code to essays, people's pasts are littered with stuff they cringe at years later. But it is what gets them started. So yea. Maybe we don't put anyone on a pedestal for shiny charts. But maybe we don't tear them down either :)

(also I say this fully appreciating that you were explaining the parent poster's view and you may not necessarily support it)

View on HN

The issue is that making a mistake here isn't like making a mistake about, say, movie review sentiment analysis. Already there have been a few "I know stats!" posts that didn't do the right controls but still got sent around the internet as proof of someone's argument (see the "Viral advertising exec now knows epidemiology" post that medium took down). These data are confusing and difficult. It turns out there's a bunch of people who have studied this field for years and can help us avoid those issues (epidemiologists). We should listen to them. That's not to say we shouldn't open the data, but handle with extreme care, which isn't the modality of most early data scientists. A 4 week bootcamp doesn't lead to the most rigorous science, usually that's not a big issue (Pareto principle and all that), but when it's lives and active problems, it can become a big issue.

I say this as a graduate of Insight which is a kind of DS bootcamp.

No! We should rigorously scrutinise epidemiologists and do lots and lots of competing analysis.

What makes you think epidemiologists are so great? The one who has guided the UK's response has been severely criticised by fellow academics in the past for bogus low quality modelling. The Imperial College paper has basic errors and flawed assumptions that are obvious even to untrained laymen.

I think you're wildly over-estimating how much statistical and logical training most academics get. This is one of the basic underlying problems that the replication crisis has exposed: an unending flood of academic papers, especially those doing modelling, that fall apart when examined by people with statistical and mathematical training.

You're right that a 4 week bootcamp won't make someone a flawless handling of data. But it might still be a more rigorous form of training than epidemiologists get.

I agree with points from the parent comment, but more so from you. It's good to have experts who've rigorously studied historical practices and their outcomes; we can learn many things about which practices work best, and what mathematical relationships they reveal. But history is not an accurate predictor of the future, and theory is only a stepping stone for better testing (such as Great Britain's theory for herd immunity, which was a ground breaking theoretical approach to dealing with a pandemic).

I think the most effective approach, which adds to your argument, is ensemble modelling from numerous, independent researchers from many fields. This is akin to how many people guessing at the number of gumballs in a container at the fair will average to the correct amount, while any single guess will not. There is a Freakonomics episode, Superpredictors, which demonstrates the use of this statistical approach for anticipating the outcome of voting in foreign politics.

Reply on news.ycombinator.com