It seems like there needs to be a real break between the collectors of data and the analyzers. At the moment heaps of data is collected by the same teams that analyze it, but just as an API opens up a codebase to new ideas, so would a standard practice of placing the data into a library before starting analysis.
Fortunately there's already a growing number of places (typically government in my experience) making an effort just to keep data libraries up and open to further research.
I agree that this is a problem, but I fear it's not as easy to solve as with traditional libraries that store books. For data to be useful you need a lot of very formal high quality metadata, and that metadata has to be collected by the original researchers as part of their process. So I think there has to be an incentive or mandate that makes researchers collect and document data in a more systematic way. Librarians cannot do it for them.
"Are titles being chosen simply to maximize click rates?"
I was under the impression that the editors usually choose four or five different titles, and then after the first few thousand impressions the software chooses the one that does the best to become the permanent title.
Apropos of this, a cool hack is that the more days in a row you click on an item in Amazon the cheaper it gets. So if you want to save an extra 10% on something, just click it once a day for a week or so until the price goes down. (After a certain point though it starts going back up again, so you have to time it right.)
Straying a bit from the original topic: For amazon price tracking I've used this and it's pretty good: http://camelcamelcamel.com/ (They apparently cover some other retailers too.)
I suppose this is as good a place as any for this rant:
If you go out seeking a specific type of data, it is almost impossible to find it. Multiple times in the past, I've wanted to sanity-check against public results, and gone looking for experimental data. I find plenty of papers, some of which even purport to measure the exact thing I'm interested in, but every single fucking one of them just has a plot. A god damn plot. Give me the data, I can make a god damn plot. I have MATLAB, I have gnuplot, it's no problem. But give me a plot, and I've got to fire up an image editor and count pixels - and of course it's a log-log plot, so I have to convert "pixels over" and "pixels up" to whatever units I actually need.
I'm tired of this shit. I can't read half the papers I see because they're behind a paywall, and when I could read them, none of them included data. If you're an experimental scientist, your prose is probably worthless. I am wading through your entirely passive-voice description of your experimental setup (separate rant...) solely so that I can find some god damn data. Give me a drawing of your experimental setup, a csv of your data, and whatever MATLAB/LabView/Excel kludge you used for any derived quantities, and you can avoid something you obviously don't like (writing - or else you wouldn't do it so badly) and I can avoid something I don't like (reading your writing).
Now that the internet exists, there's no reason not to communicate actual data sets - in the past, you'd have to deliver a phone book to your subscribers if you did that, so this all sort of makes sense in an anachronistic way, but it'd be so much more useful for the scientific community if the funding required "post your data set, drawings of your setup with specs for any instruments, and whatever MATLAB/LabView/Excel you used for any derived quantities" instead of "publish a short paper, it better have plots".
(This doesn't really apply to multi-TB data sets like the author describes - obviously you're going to have to contact the PI and ask him/her to ship you a hard drive)
Everything from incentives to the attitude seems to make academic research groups act like sealed little islands of "brilliance".
Researchers should be strongly "incentivized" to make their data and source code available and immediately usable (ie, not in Matlab). And the only way to do that is to gradually devalue data-less, source-code-less research in the way non-peer-review research is now devalued. (and yes, it will be messy, embarrassing, fragile code sometimes but we'll have to live with that).
The amazing thing is that small, for-profit, private research organizations can still operate with this "what we produce is a pdf with a picture of our research on it" attitude.
I disagree with this one. Matlab is nice, and I'm going to use it if it's available (and if I'm doing linear algebra - if the problem is more graph/tree-ish, I might reach for SciPy).
The problem with Matlab is that while it is fast for prototyping, it's difficult to write maintainable code in it (which is why I switched to python). Furthermore, you run into the issue of sharing code with people at institutions that don't have Matlab and Octave may or may not be compatible. Sometimes, an institution may have Matlab, but even if they do, they might not have access to the same toolkits you use (for example, I use a savitsky-golay filter as part of a peak finding algorithm--because it's in Matlab and we have a DSP toolkit, I just used that one. Because my collaborators didn't have it, I ended up rewriting the code in python--and this was for a group at a national lab...Now, imagine people at small universities and colleges...)
I like being able to see function signatures in a header file.
In Python or Matlab, I have to read the implementation to see what types it wants. To me, having collections of type signatures in a separate file is a huge boon when I have to come back to code I haven't seen for a while - it's a 10,000 foot view of what the code does that helps jog my memory - and it's guaranteed to be accurate, unlike documentation.
To the extent that static typing supposedly prevents bugs, I agree that it is overrated. My neutron transport solver I wrote in Haskell had plenty of god damn bugs after I got it to compile, thank you very much (perhaps a very deep Haskell-fu is required to attain the mythical "neutron transport solver that works right the first time" - in any case I realized that compiler-aided bug prevention is not a low-hanging fruit in Haskell, so I lost interest).
That was a long-winded way to say that I feel lost and alone without function signatures, and that this preference is possibly only quasi-rational.
The university and/or the journal. Supposedly, part of their existence and social contract is to maintain knowledge archives for public benefit (even if its not free).
University members, even students, nearly always get free hosting space. The problem is not the cost of hosting, or even the time it takes to put it up. The problem is the competition between research groups. If you publish your data, the competition can use it to beat you at the next discovery in the often narrow field.
> Have you had any luck emailing the author of the paper?
This works sometimes. People want to be cited after all. Sometimes you just never get a response. Also not viable for lab courses in graduate school, where you need to "compare your results with the literature", or for "calibrate my instruments" type situations (rather than "cite you" type situations) - there's some data that would be useful for things that aren't really worth bothering a PI over.
If I were a funding agency, I would earmark funding specifically for hosting data, because that's the biggest value of a lot of experimental science. A lot of times the data analysis is pedestrian, but actually taking measurements is difficult and expensive.
not much motivation: most research is done on a "works for me" and "good enough, got the 5 sigma" level. Reproducible industrial ISO protocols to which you would need to add user support is certainly not granted for in grant applications. Not to mention data formats, software versions and the number of flies in the lab.
It's an incentive question: Researchers are often evaluated based on the papers they produce (impact and numbers)--not based on how maintainable our code base is. If I can keep a cobbled together code base going and still produce code, then it's in my best interest to do science rather than write recyclable code.
The other problem that happens is that some people don't release their source code (they may release binaries) because either they're embarrassed by it, or because they want to maintain competitive advantage....
oh, come on. If you have done real science, you should know, that asking for scientific data is like asking for the source code of Windows. You have valid points (it has been published, you have licensed access, you can do cool stuff with it, verify and iron out the bugs etc.): but this is not the way it works. Reality does not match the sharing knowledge mantra - not to mention fudge factors and data context (experiment design, knowledge of measuring device, detector etc.). LEP data is not only useless because it is old and accessible via the dinosaurish FORTRAN, but because the knowledge of the metadata (i.e. shower profile of an endcap) almost gone and certainly not available to outsiders.
Reproducible experiments themselves are becoming the exception and not the norm. In every field. (not to mention specialisation).
Now that the internet exists, there's no reason not to communicate actual data sets
That is not always true. For many sets involving humans, data may have been collected under the express promise tat anything published will not be used for anything else and/or cannot be de-anonymized. Verifying that that is the case can take quite an effort, especially when considering that third parties can combine raw data from multiple papers building on the same data.
if they made it a bit more accessible, scientists could potentially share some of their data, and new research could associate their related findings with the original data.
22 comments
[ 2.9 ms ] story [ 40.7 ms ] threadFortunately there's already a growing number of places (typically government in my experience) making an effort just to keep data libraries up and open to further research.
This title to me meant "Scientists are complaining that greater availability of data is making it harder to do research."
Instead it turns out the post is about scientists losing data due to not having places/formats for storing it.
I have trouble understanding how someone thought this was a good title for this post. Are titles being chosen simply to maximize click rates ?
I was under the impression that the editors usually choose four or five different titles, and then after the first few thousand impressions the software chooses the one that does the best to become the permanent title.
Apropos of this, a cool hack is that the more days in a row you click on an item in Amazon the cheaper it gets. So if you want to save an extra 10% on something, just click it once a day for a week or so until the price goes down. (After a certain point though it starts going back up again, so you have to time it right.)
If you go out seeking a specific type of data, it is almost impossible to find it. Multiple times in the past, I've wanted to sanity-check against public results, and gone looking for experimental data. I find plenty of papers, some of which even purport to measure the exact thing I'm interested in, but every single fucking one of them just has a plot. A god damn plot. Give me the data, I can make a god damn plot. I have MATLAB, I have gnuplot, it's no problem. But give me a plot, and I've got to fire up an image editor and count pixels - and of course it's a log-log plot, so I have to convert "pixels over" and "pixels up" to whatever units I actually need.
I'm tired of this shit. I can't read half the papers I see because they're behind a paywall, and when I could read them, none of them included data. If you're an experimental scientist, your prose is probably worthless. I am wading through your entirely passive-voice description of your experimental setup (separate rant...) solely so that I can find some god damn data. Give me a drawing of your experimental setup, a csv of your data, and whatever MATLAB/LabView/Excel kludge you used for any derived quantities, and you can avoid something you obviously don't like (writing - or else you wouldn't do it so badly) and I can avoid something I don't like (reading your writing).
Now that the internet exists, there's no reason not to communicate actual data sets - in the past, you'd have to deliver a phone book to your subscribers if you did that, so this all sort of makes sense in an anachronistic way, but it'd be so much more useful for the scientific community if the funding required "post your data set, drawings of your setup with specs for any instruments, and whatever MATLAB/LabView/Excel you used for any derived quantities" instead of "publish a short paper, it better have plots".
(This doesn't really apply to multi-TB data sets like the author describes - obviously you're going to have to contact the PI and ask him/her to ship you a hard drive)
Have you had any luck emailing the author of the paper?
Everything from incentives to the attitude seems to make academic research groups act like sealed little islands of "brilliance".
Researchers should be strongly "incentivized" to make their data and source code available and immediately usable (ie, not in Matlab). And the only way to do that is to gradually devalue data-less, source-code-less research in the way non-peer-review research is now devalued. (and yes, it will be messy, embarrassing, fragile code sometimes but we'll have to live with that).
The amazing thing is that small, for-profit, private research organizations can still operate with this "what we produce is a pdf with a picture of our research on it" attitude.
Btw, one organization can and I think is help is PLOS: http://www.plos.org/
I disagree with this one. Matlab is nice, and I'm going to use it if it's available (and if I'm doing linear algebra - if the problem is more graph/tree-ish, I might reach for SciPy).
If I plan on maintaining something, I want static typing, so the whole SciPy/Matlab/Octave/R debate is a little moot for me there.
In Python or Matlab, I have to read the implementation to see what types it wants. To me, having collections of type signatures in a separate file is a huge boon when I have to come back to code I haven't seen for a while - it's a 10,000 foot view of what the code does that helps jog my memory - and it's guaranteed to be accurate, unlike documentation.
To the extent that static typing supposedly prevents bugs, I agree that it is overrated. My neutron transport solver I wrote in Haskell had plenty of god damn bugs after I got it to compile, thank you very much (perhaps a very deep Haskell-fu is required to attain the mythical "neutron transport solver that works right the first time" - in any case I realized that compiler-aided bug prevention is not a low-hanging fruit in Haskell, so I lost interest).
That was a long-winded way to say that I feel lost and alone without function signatures, and that this preference is possibly only quasi-rational.
The university and/or the journal. Supposedly, part of their existence and social contract is to maintain knowledge archives for public benefit (even if its not free).
This works sometimes. People want to be cited after all. Sometimes you just never get a response. Also not viable for lab courses in graduate school, where you need to "compare your results with the literature", or for "calibrate my instruments" type situations (rather than "cite you" type situations) - there's some data that would be useful for things that aren't really worth bothering a PI over.
If I were a funding agency, I would earmark funding specifically for hosting data, because that's the biggest value of a lot of experimental science. A lot of times the data analysis is pedestrian, but actually taking measurements is difficult and expensive.
The other problem that happens is that some people don't release their source code (they may release binaries) because either they're embarrassed by it, or because they want to maintain competitive advantage....
Reproducible experiments themselves are becoming the exception and not the norm. In every field. (not to mention specialisation).
That is not always true. For many sets involving humans, data may have been collected under the express promise tat anything published will not be used for anything else and/or cannot be de-anonymized. Verifying that that is the case can take quite an effort, especially when considering that third parties can combine raw data from multiple papers building on the same data.
if they made it a bit more accessible, scientists could potentially share some of their data, and new research could associate their related findings with the original data.