To an alarming degree, science is not self-correcting

It's not just the subsidies, but how they are awarded - right on with the "structure of the academic enterprise".

Consider: To get promoted to professorship, you have to selected by slighly hoarier professors; to be selected, you have to have published works, which are (anonymously) reviewed by other professors, and be awarded grants with (anonymous) review committees of other professors. You probably went to grad school under the tutelage of professors in the same circle... And at least in the past scientists were sufficiently generalist that you had a wide pool of fellows who were reviewing your work at all stages; now we have siloed subdisciplines (like "chemical biology" - which, mind you is not the same as "biological chemistry" or "biochemistry"), and the emergence of "interdisciplinary research" which somehow instead of encouraging generalism, instead promoted dilettantes who couldn't hack it in either of their parent fields...

Is there any wonder why research is increasingly unreliable?

This is an excellent overview article on an important topic. It mentions many of the most influential authors on the topic of accuracy of scientific publications.

Hacker News readers may enjoy "Warning Signs in Experimental Design and Interpretation"[1] by Peter Norvig, a LISP hacker who is now director of research at Google, on how to interpret scientific research. Norvig's essay is my all-time favorite link to share in a Hacker News comment. That's because we see submissions here every day of preliminary studies that can be analyzed by Norvig's checklist on research issues to look for when reading about a scientific finding.

I discuss psychology research weekly with a group of psychologists who study human behavior genetics in a "journal club" (graduate seminar course). Those researchers have told me about other researchers who are trying to clean up the published literature in psychology, for example Jelte Wicherts, whose article "Letting the daylight in: reviewing the reviewers and other ways to maximize transparency in science"[2] in an open-access journal suggests general procedures to improve scientific publishing, for example by changing the incentive structure around reviewing papers submitted for publication. Another helpful researcher on statistical tests to verify results is Uri Simonsohn. The papers he and his colleagues produce[3] are thought-provoking, pointed, and sometimes laugh-out-loud funny.

[1] http://norvig.com/experiment-design.html

[2] Jelte M. Wicherts, Rogier A. Kievit, Marjan Bakker and Denny Borsboom. Letting the daylight in: reviewing the reviewers and other ways to maximize transparency in science. Front. Comput. Neurosci., 03 April 2012 doi: 10.3389/fncom.2012.00020

http://www.frontiersin.org/Computational_Neuroscience/10.338...

[3] http://opim.wharton.upenn.edu/~uws/

This article re-hashes much of the content of Ben Goldacre's book 'Bad Pharma', however it does not propose any solutions (Ben Goldacre does). There is plenty that can be done with legislation, as a customer of healthcare and in clinical trials. If all trials - good results or bad - were published then we would be half way there.

http://www.badscience.net/2013/10/why-and-how-i-wrote-bad-ph...

The article mentions the company 'Amgen' and how they were not able to reproduce some earlier trial results. Alarm bells went off for me right there at the word 'Amgen'. They were the company that had the wonder drug EPO, as in of Lance Armstrong fame. That story is truly fascinating and best pieced together from the book 'Blood Medicine':

http://www.bloodmedicine.info/

...and from the cycling scandal books that have came out recently, e.g. Tyler Hamilton's.

I'm not a fan of Kuhn, largely because he can be read as saying that, indeed, scientific models that get "overthrown" in a paradigm shift are therefore "invalid". (In the extreme, this leads to complete relativism, the idea that there is no such thing as a "valid" model at all. Kuhn himself has waffled on this point: sometimes he says he didn't really mean to take that extreme view, but other times he says things that are really hard to make sense of in any other way.) That leads people to believe that, for example, Newtonian physics must be "invalid" because it was "overthrown" by relativity. But it's really hard to square that claim with the fact that Newtonian physics (along with many other supposedly "invalid" models) gets used every day to make accurate predictions.

I think the bad science described in the article is, at least in part, a reflection of our losing sight of that ultimate objective: we build scientific models of the world in order to make accurate predictions. It's not enough to say, well, I used all the right statistical techniques, I used all the right double-blind control procedures, my results have been replicated. Those things are all necessary, but they're not sufficient; they're not the goal, they're only the starting point. The goal is to make accurate predictions. I think we get a lot of bad science because we don't enforce that requirement enough.

In my experience, when most engineering-oriented people use the word truth, they (wittingly or unwittling) mean the pragmatic definition of "allows us to make stronger-than-previous predictive models about something". Most of the argument in philosophy of science circles strikes me as meta-interesting, not eminently relevant to the practice of science.

Especially when dealing with psychology, a lot of those "what, exactly, is an electron" sort of deeply epistemological statements are not really asked, because we have enough trouble with the simplistic models we have.

So yeah, my recommendation (for what it is worth) is to always assume "most predictive model available" when someone says truth, and be aware that you are making that assumption and that simplification.