Softverse

Softverse counts how often each R, Python and Stata package is loaded by the code behind published social science papers. It covers 3,223 packages across 8,349 replication deposits at economics and political science journals that check an author’s code before publication.

The counts exist to support credit, and they are built to be expensive to fake. What a software author can point to today is a download count, which records how often a file was fetched, or a mention in someone’s methods section. A build server installing a package on every commit runs the download count up on its own, and cutting more frequent releases runs it up further, so a number like that tells a hiring committee about release practice as much as about use. Adding one to a count here takes a paper published at a journal that checks its authors’ code, with the package loaded in that code.

Three things the counts show

The most loaded package is estout, a Stata command that formats regression output into a table, in 2,440 deposits. three of the four most loaded packages also present results rather than estimate them, so the tools research leans on hardest are the ones that prepare its exhibits, and those are the tools least likely to be cited.

Stata is the most used of the three languages and the least studied. It appears in 6,212 deposits against 3,518 for R. Work on research software covers R and Python because CRAN and PyPI will tell you which project a library name belongs to, and Stata has no equivalent registry. This project reconstructed one from the Statistical Software Components archive and released it.

Some heavily used software belongs to no registry at all. grc1leg is called 1,224 times in this corpus, StataCorp distributes it from their own website, and no archive lists it, so a credit system that works by registry lookup cannot reach whoever wrote it.

Where to go

Look up a package to see its count, the denominator it is a share of, and which of the two repositories is doing the crediting.

Read the paper for the sampling frame, the extraction method, and the validation, which includes measuring the R extractor against R’s own parser.

Take the data, published as CSV under CC0, along with the Stata command-to-package index at 10.5281/zenodo.21926099.

What the counts do not show

These counts say a package was loaded by code in the deposit. They do not say the code ran. Authors often leave older scripts in a deposit, and a script can load a package inside a branch that never executes. Whatever an author kept out of the deposit cannot be counted at all. The paper measures how wide that gap is.