New Stanford Institute To Target Bad Science

ananyo writes "John Ioannidis, the epidemiologist who published an infamous paper entitled 'Why most published research findings are false', has co-founded an institute dedicated to combating sloppy medical studies. The new institute is to focus on irreproducibility, waste in science and publication bias. The institute, called the Meta-Research Innovation Centre or METRICS, will, the Economist reports, 'create a "journal watch" to monitor scientific publishers' work and to shame laggards into better behaviour. And they will spread the message to policymakers, governments and other interested parties, in an effort to stop them making decisions on the basis of flaky studies. All this in the name of the centre's nerdishly valiant mission statement: "Identifying and minimising persistent threats to medical-research quality."'"

Comments

Re:Don’t throw a wet blanket on science

[gstoddart]

For instance, there are the innumerable parallel computing papers that use O(N^2) algorithms to show a speedup on a GPU or supercomputer where there exists a serial O(log N) algorithm that runs faster on a PC. (No joke.)

Except that while there might be some problems which have O(log N) solutions as well as O(N^2) solutions, there are still things which still only have O(N^2) solutions, correct?

So if you can learn how to solve a known O(N^2) problem better (even if there is a known O(log N) solution), what you learn is still applicable to to other O(N^2) problems for which there isn't a known O(log N) solution.

I'm not sure what you're describing is evidence of malfeasance, or that they're working on solving a class of solution, and not necessarily that specific problem.

To me it sounds more like they're probably aware of the O(log N) solution, but that's irrelevant because they're looking at how to use parallelism to address things which are O(N^2), because there's many many of those.

So much of math comes down to solving an equivalent problem you already know how to solve.

Maybe they're figuring out how to address a problem which is O(N^2) by one method, so that once they know how to solve it faster with parallelism, they can learn how to solve other problems which nobody has an O(log N) solution for.

It may not be all about solving that particular problem, but that class of problem. Because mostly it seems like we've never figured out how to do real parallelism except for things which are classed as 'embarassingly parallel' because it already lends itself to breaking it up -- like SETI@Home.