john_horton

This sounds like a case where the problem of collective action would actually work in society's favor: a single vandal would bear the full cost of destroying the sensor (e.g., the risk of getting caught), but the "benefits" of the destroyed sensor would be diffuse, spread out to all would-be shooters.

http://en.wikipedia.org/wiki/Collective_action

I don't disagree with that - a bar chart that doesn't start at zero can be misleading. But for what he was trying to do---compare means---it would have been a perfectly fine choice to do a lineplot/very thin bar plot and start the y-axes somewhere other than zero if he included error bars and clearly labeled things.

http://www.edwardtufte.com/bboard/q-and-a-fetch-msg?msg_id=0...

Agree that 3D graphs are terrible, but it's a canard that y-axes always need to start at 0 and that doing anything else is somehow misleading.

On this point, see http://krugman.blogs.nytimes.com/2011/09/14/axes-of-evil/

Re: oDesk data - what do you want to know?

I regular post data from oDesk on my own blog (I work at oDesk w/ these folks: https://www.odesk.com/info/l/research/): http://onlinelabor.blogspot.com/.

And there's the "oConomy": https://www.odesk.com/oconomy/ There's also lots of info available via the API.

This is interesting - do you have a good academic reference for that summarizes this literature? Maybe like a handbook chapter?

Agreed - but I think there are enough unbalanced forces within most organizations towards the "let's a have meeting" solution to problems that this is making a "ha, ha, only serious" point that resonates.

It's been done (though instead of money, candy and instead of the urban poor, UC Berkeley CS students):

http://hci.berkeley.edu/cs260-fall10/images/b/ba/FinalPaper-...

We'll have to disagree about the axes---I think the choice of where to start axes depends on what you're trying to illustrate. For me, absolute differences in wages was what I wanted highlight.

Re: Standard error bars - you're absolutely right. I should have included them, but it was my first time using the googleVis package & didn't see an easy way to add them. I don't have the numbers right in front of me, but standard deviation for each skill was on the order of ~$7/hour and each skill had to have at least 30 obs., though many where much, much higher. If we say n=50, that's an SE of about $1/hour, so clearly one shouldn't take much stock in very small differences in wages. My goal was just to unearth some interesting data---I probably should tone down the implied recommendation lest I'm responsible for a glut of unemployable lisp hackers in a few years:)

1) The data are what they are - these are the rates people are charging by listed skill. It would be very interesting to look at prices by actual completed projects, but that's just another approach one can take.

2) I make it very clear that there's an association, not a causal claim. I flag the fact, make a lame joke about it and put "might* in the title.

3) There's no Tufte rule about starting charts at 0 - you make choices that make it easier for people to make comparisons. Paul Krugman said it better than anyone on this topic: http://krugman.blogs.nytimes.com/2011/09/14/axes-of-evil/

You're right---there's nothing about the demand side yet. That's my next blog post. I'll show the earned wages by skill and some measures of demand & trade volume in different skills.

So my anecdotal impression is that the most successful freelancers learn to get very picky about who they work with---they want to see a verified payment method (esp. if the client is proposing a fixed-price contract), a history of hires and successful projects at reasonable wages and a good job description.

The first link to DuckDuckGo on the page is a link to Google's main search page. Typo or some joke I'm not getting?

yep - and I agree about the dubiousness (see my comment above).

Thanks - a few responses:

1. For that/which:

http://www.kentlaw.edu/academics/lrw/grinker/LwtaThat_Versus...

I should note that some of these "rules" are bogus (a point made by Williams) and are regularly broken by great writers. Williams argues that writers should feel free to break them, but that they should just be aware of them and the possibility that they will be judged by others not aware of these rule's dubious status.

2. I would agree with you it would be impractical if someone were going to mechanically just apply all the "hits" and treat them as changes to be made---but that's not what it is for. The idea is that the writer would be reviewing the hits and making judgement calls in each case. For example, in the serial comma case, a smart writer would see that "oh, this is fine" and continue. So I think of the high false positive rate as a plus, esp. for the kind of writing this is most useful for (i.e., stuff you're going to publish & need to proofread/edit the heck out of).

I definitely agree that more linguistic knowledge would be awesome (and my Mathematica version had some of this).

yeah - sorry. I'll add a GPL notice to the repository tomorrow.

On the data science models => production code problem, could PMML<http://www.dmg.org/pmml-v3-0.html>; be (part) of the answer? Anyone have any experience using it?

Thanks for the tip!

Actually, that's my fault. We (oDesk) recently introduced a job post text parser that examines the text of the job description and auto-suggests skills. The parser uses the levenshtein distance to find approximate matches, and in the cases you flagged, the job descriptions used the word "closure" which the parser converts to "clojure." I probably need to add "clojure" to the parser's stop word list...

For a little background on this debate: http://tuvalu.santafe.edu/~jdf/papers/farmer0606.pdf

This is exactly what prompted us making this thing---some of my co-workers at oDesk were talking about how dissatisfied we are with the whole 1-10/NPS score way of getting feedback.

As far as the slider---my motivation was that humans are sensitive to the emotional content of even very tiny changes in a human face (which is the logic behind the Chernoff face http://en.wikipedia.org/wiki/Chernoff_face) and that the smiley-faced slider would get people to (a) use the "same" scale for conveying their feelings and (b) convey their feelings more precisely. It's on our agenda to do some tests with it---shoot me an email if you want to try it on your site & see how it does. It probably won't make it into our product too soon :(

This is why there's a huge literature in economics on trying to measure the causal impact of more schooling. They've used everything from state-level changes in compulsory schooling laws, proximity to college, changes in grants policies, variation around cut-offs that determine eligibility etc. In all of these studies, the picture is quite clear: more schooling has a positive, causal effect on average earnings, with the estimates around ~6-10%.

That regression (which is called a mincerian wage equation & economists have been running it for about 40 years) will give you the average percentage increase in earnings (if you do it in logs) from another year of education. That average measure can be consistent with lots of different dynamics---which is Mankiw's point. Perhaps another year of education increases earnings by 10% across the board (deterministic view) or perhaps is increase by 10% the chance that your income will increase 100% but 90% of the time does nothing.

Note: All of this is just correlational---there are serious problems w/ treating these regressions as establishing causality, but that's another issue.

FWIW, this is apparent in the data as well---employers are price sensitive, but not nearly as price sensitive as people might think. They good employers aren't stupid and realize you get what you pay for. In fact, the number of one predictor (by far) of who gets hired on oDesk is not experience, hours, tests, county etc.---it's how customized their cover letter is (I took the Levenshtein distance between an applicant's cover letter & their previous cover letter to create a "customization score" that I used as a predictor in a model of hiring).

At least for development work, where each job is only receiving ~10-15 applicants, a well written cover letter that talks intelligently about the employer's problem can cut through the noise of low-quality applicants & lead to a hire.

I think he's essentially correct on the macro trends that will move the world more to this kind of model.

On the specific point on what's holding back a more distributed/networked model of production, I think one obstacle is that labor is not easily "commoditized" in a way that facilitates trade (see http://crowdresearch.org/chi2011-workshop/papers/horton.pdf for more on this).

But isn't their ability to make money for the company exactly what determines demand for engineers (or any worker for that matter)?

I think working on the funding agencies (i.e., getting them to adopt sensible policies) is the right strategy. For the individual researcher, the incentives for code & data sharing are pretty limited right now, at least until the culture changes.

On the technical side of code/data sharing--I think one obstacle (in least in the fields I'm familiar with) is that may researchers put together papers in a way that makes reproducibility needlessly hard. If you do all your stats in something like Stata or SPSS, then paste tables/figures into an MS Word document (which is passed around among colleagues), finding your own errors is hard enough---never mind some third party trying to re-produce your results. If instead, you use tools like Sweave & script the data analysis & paper assembly process (ideally with version control), reproduction/sharing becomes much simpler.

It seems like there is little tactical payoff in doing this, as large, foreign militaries would already have detailed, accurate maps covering the whole world, China included. I bet this is an internal-resistance oriented policy.

This article is more about how multiple sets of parameters can fit the same data equally well. This is why economists draw a distinction between calibration and estimation. If a parameter is "identified" in some estimation procedure, they mean they have an experiment or quasi-experiment that gives them a credible CI for the true parameter.

Strongly second that---USAA is fantastic.

I'd like the same feature but I don't think it's possible right now. FWIW, I asked whether this was possible on Quora awhile back and got a negative response: http://www.quora.com/Is-there-a-way-to-make-a-Google+-post-p...