A short discussion of how one might evaluate a new method for discovering patterns.
This is the text and code for Bowers, Jake. forthcoming. "Method Games," Sociological Methodology.
The results of the machine method game are in socmeth.Rout
.
To rerun this game, or to modify it for your own use, you will need to install the R statistical computing environment.
If you are using OS X or some other version of unix, you can rerun the
game in socmeth.R
to recreate socmeth.Rout
by typing the
following at the command line:
R CMD BATCH socmeth.R
Be prepared to wait a while. The code is designed to return NA and to keep
running if the problem becomes intractable for a given method. This means one
should not worry about errors --- the proportion of NAs will be reported below
the report of the proportion of successes. In this version, we see, for
example, that 5/800=.00625 of the versions of the easy data problem (5
variables, all involved in the truth) were intractable by the SIS()
function.
If you are using Windows, you will need to edit the code to avoid the
parallel
library and to substitute lapply()
or replicate()
because, as far as I know, the shared memory multicore parallel capabilities
of R have not yet been ported to Windows. You can also setup your own local or remote cluster
using distributed memory parallelism using the snow
package and the parLapply()
function.