john_horton

The study (which strangely isn't available on SSRN or the author's website) appears to be about recruiting law and b-school grads.

I wouldn't be that surprised if her results are true. Whenever there are lots of applicants and lots of employer uncertainty about applicant quality, screening on things like credentials that correlate (however weakly) with average ability can still be "rational."

To give an example, assume that GMAT score is a) unobservable by employers and b) perfectly predictive of future performance in the firm. If HBS grads have a mean 1 point advantage over Sloan grads, it makes sense to lexically prefer HBS to Sloan grads, even though slightly less than half the time the employer will make the wrong (ex post) hiring choice.

This dynamic also tends to be self-reinforcing. If credentials are all that matter to future employers (and all schools cost the same), then all students will try to attend the most prestigious school they can. If admissions committees are not pure coin flips, then it makes sense for employers to believe that the HBS/Sloan "distinction" means something.

I wonder if you would even need to pay people---someone (google would seem like an obvious candidate) could write a browser extension with a "total crap" button. I know I'm sufficiently annoyed when I land on some spammy garbage that I'd like to punish them with a consequential down vote.

Hopefully you could get enough legitimate uptake to make paying people to fraudulently vote wouldn't be worth it.

Interesting - do you have a sample chapter you can make public?

(if there already is on the site you linked to, sorry for being obtuse).

I see your point---my point is that this article is not a good jumping off point for discussions about philosophy of science or epistemology (bold title notwithstanding). It's about bad statistical practice. More importantly, it's about a bad practice that is well-known (at least among statisticians) and has remedies.

I don't think this is what the article is about. At least for physical and biological phenomena, there is some objective truth. For example, Seroquel does or does not reduce schizophrenic symptoms---the reason an original finding is "temporary" and wears off is not because the efficacy has changed over time---it's because it was never efficacious in the first place (and we were mistaken because of bad research practices).

Andrew Gelman has a great blot post on this:

http://www.stat.columbia.edu/~cook/movabletype/archives/2010...

In a nutshell, the problem arises from trying to estimate small effects and then screening results on statistical significance.

Strange that it was banned. Maybe it had to do with the vignettes---some of the examples are from political science (e.g., election results, polling) and another deals with arsenic in drinking water...or maybe Hu Jintao feels the way to deal with hierarchal data is to cluster standard errors.

What would be really wonderful (and what I'm hoping will get built) is way to generate web-friendly visualizations directly from the amazing ggplot2 package.

Depending on your needs, MTurk workers can give you decent feedback for very little money.

I'm skeptical that (a) the personnel system is a huge driver of attrition and that (b) the proposed solution would work well.

On (a), deployments are generally pretty terrible, esp. if you have a family. Most officers I know that got out didn't say it explicitly, but they made a comparison of life in the military vs. life outside the military. In this comparison, deployments are such a huge factor that the other stuff is pretty secondary.

On (b), officers have very few objective measures of performance. Those that exist have lots of noise. Good platoon sergeants can make incompetent platoon leaders still look pretty good. You can pencil-whip maintenance records. You get good or bad draws of soldiers & assignments etc. This encourages politicking and focusing on appearance over substance. Adding more "flexibility" to such a system could have some bad second-order effects, by ramping up the incentives to appear good to your commanders, without actually being good.

Many of these would be at home in the "what not to do" section of Tufte's book.

Also - anyone notice on the White House graphic it says that the pope visited "President Cart"?

I was at NIPS and talked to one of the authors. I thought the paper was interesting, but I think the "you're not paying enough" critique is spot on. Humans clearly can be better at this task---you just can't give them strong incentives to cut corners on quality, which happens with a low piece-rate and a task that takes on the order of 3 ~ 4 minutes to do properly.

Agree about beauty---but I think the real genius of ggplot2 is the grammar. After using it, I think it becomes easier to think about data graphically and then quickly generate the visualization you have in your head. ggplot2 alone is a reason to switch to R.

And then the World Bank will write-off that debt under their Heavily Indebted Poor Countries Initiative (HIPC): http://goo.gl/iaQfJ

The fact that the US has a lot of control over the World Bank seems reasonable given that they put up the lion's share of the capital. As far as dictating terms and conditions, regardless of their merit, there aren't many banks that will give million dollar loans to bad credit risks w/o strings attached. What's the alternative scenario? No bank? They just give out money?

Agreed - relative inequality definitely matters.

Income is more concentrated today compared to the post-war period, but I don't think that makes the mean meaningless---it depends upon how much weight you put on a dollar held by different parts of the distribution. That being said, I think you'd be hard pressed to show that any part of the income distribution is worse-off in real terms than they were in 1950.

Incidentally, we only know about these changes in the income distribution because empirically oriented economists have been collecting and analyzing this data for many years.

There are lots of different moments & quantiles of the income distribution that are interesting and useful to look at, but its not like the mean is terribly misleading here. In the plot you provide, median income in real dollars has increased from 25K in 1950 to 40K in 1990---a 70% increase in real income.

A few points:

(1) Many if not most economists are empirical. Recent John Bates Clark medals---a good reflection of what academic economists think is the best work being done right now---have gone to economists doing empirical work. Recent examples include what works for development (evaluated with large randomized controlled trials), what are the causes of growing income inequality, how should we structure auctions etc. http://en.wikipedia.org/wiki/John_Bates_Clark_Medal . To say that most economists have no contact or interest in empiricism is incorrect.

(2) The point about simple models is not that economists fully believe them---it's that they force everyone to be explicit about assumptions. They also try to identify what are the key features of admittedly very complex phenomena. This is what models do in most scientific endeavor. That the models often perform poorly is a sign of an immature science---not fraud or dishonesty. You'd get eternal fame and become fantastically rich if you came up with a financial model with great predictive powers---do economists not create such models because they are stupid and corrupt? That seems unlikely.

(3) The transistor example completely misses the key distinction that every presentation of the simple supply and demand model always makes - i.e., there is a difference between the short term and long term. If demand doubled for transistors tomorrow, the price wouldn't rise?

(4) You confuse absolute and comparative advantage.

(5) "In the 1950s the US had a great economy. Except for some points of information, biomedical, and materials technology, it's not clear that our standard of living is as high now."

This is where I stopped reading - this is so obviously wrong http://www.wolframalpha.com/input/?i=GDP/capita+real+US that it's hard to take seriously any conclusions on the relative merits of different industrial policies the follow from this "fact."

I think the relative supply and demand are going to determine prices here, just like any other market. A price floor could be imposed, but this would likely price some work out of the market or lead to other work-arounds as trading partners try to avoid constraints.

As far as protecting workers, in a survey I ran last year, there wasn't any evidence that workers on MTurk feel more exploited by their online bosses compared to their offline bosses:

http://www.technologyreview.com/blog/arxiv/24646/

Though obviously those that feel particularly mistreated could be selecting out of the market.

In some sense, this is what TurKit already is:

http://groups.csail.mit.edu/uid/turkit/

And I think TurKit is part of the back-end for Soylent.

Thanks for the mention (I made mechanical proust) - I agree that there's a lot of dross, but some responses are poignant, funny or interesting. It would be interesting to play around w/ how questions get posed.