The more I read on data formats, the more I think it is getting difficult to arrive at a common data format. Each data analyst does his or her own thing, with his/her preferred statistical package
What's more striking is that some data files come with their own header, like
sam files or Weka
arff files, while other data files remain divided into several pieces and do not necessarily include much information other than the data itself. This is often the case when there's only one type of measurement (e.g., read sequence or activity level) but we need to convey different kind of information (e.g., subject's level information and target measurement), like in
bed+bam+fam files or
hdr+img (NIFTI) files. This is mostly a way to avoid using a database given that dedicated software will handle inner joins for us. Moreover, there may exist plain text version (e.g.,
ped files) and binary version (
bed fields, respectively) of the same data file, for obvious reasons of disk space saving. I'm fine with all that.
However, I've always been a bit bothered by the fact that it is quite difficult to handle a dictionary file, also called a codebook, with a raw data file in R or Python. It looks to me that this is a fundamental piece of information when you're working on a new dataset. What I noticed over the years I was working in medical statistics is that data managers tend to record their codebook in Excel, while the raw data are stored in a statistical database (SAS or Stata) — even if Stata allows to define a dictionary that contains the type and label of each variable. There's nothing close in R, except perhaps the csvy package which tries to append metadata as a YAML header in CSV files. There is an R package that handles CSVY file format quite nicely. To the best of my knowledge, such files cannot be used by Python unless you skip the header lines like you would do with comments in a CSV file.
Note that I am not talking about handling categorical variables. There are many solutions available in R (again, not as nice as in Stata, even if I'm quite impressed by the work done by Daniel Lüdecke). Personally, I stand by the Hmisc package (it also handles variable labels, by the way), since it integrates quote nicely with the rms package which I have been using for years. I keep hoping some day we will see full integration of
Hmisc into RStudio data browser or the tibble package. I must confess I don't know enough about the state of the art in Python since I do not use Python for managing statistical data, but I doubt there are that many solutions available in that ecosystem.
In short, the beauty of data analysis can be summed up in the diversity of data formats. The more we learn about their value and possible limitations, the more we learn new techniques for investigating raw data.