Ways of obtaining empirical data in software engineering
For as long as I can remember I have been a collector of empirical data. Writing a book that involves analysis of empirical lots of data has added some focus to my previous scatter gun approach. I have been using three methods to obtain data relating to a recently read paper+one other approach:
- Download from researchers website,
- Emailing the author requesting a copy of the data,
- Reverse engineering numbers from the original paper (using tools like WebPlotDigitizer).
- Roll my sleeves up and do the experiment, write the extraction tool or convince a company to make its data available.
A sea change in attitudes to making data available seems to be underway. Until recently it was rare to find a researcher who provided a link for downloading data; in the last 12 months there has been a noticeable increase in the number of researchers making data, associated with a paper, available for download. I hope this increase continues and making data freely available becomes the accepted norm.
I regularly (once or twice a week) email the authors of a paper asking if I can have a copy of their data, typical responses include:
- Yes, here it is,
- Yes, but you cannot share it with anybody else (i.e., everybody has to get it from the original author). I have said “Thanks, but no thanks” in these cases since I make all the data I use freely available for download,
- I no longer have a copy of the data (changed jobs, lost in a computer crash, etc). In one case an established repository at a university lost funding and has gone dark.
- Data is confidential,
- Plan to write more papers based on the data, will release it when done (obtaining good data can be very time consuming and I can appreciate researchers wanting to maximize their return on investment),
- No response.
When analysing data the most common ‘mistake’ I encounter is researchers failing to get the most out of the data they have. An example of this is two researchers who made some structural changes to the way a Java library worked and then ran a thorough before/after benchmark to investigate the impact; their statistical analysis consisted of reducing the extensive data down to mean+variance and comparing these across before/after (I built a regression model that makes a much stronger case for their claims).
Of course the usual incorrect use of statistical techniques does occur, but I have not spotted anything major. However, one study found: Willingness to Share Research Data Is Related to the Strength of the Evidence and the Quality of Reporting of Statistical Results, based on 49 papers published in two major psychology journals. Since I am concentrating on papers where the data is available I am probably painting an overly rosy picture of not getting things wrong.
As always, if anybody knows of ways of obtaining data that I have not mentioned (e.g., a twitter account to follow) do please let me know.