The current default for randomization inference is the "randomization-c" method, despite evidence that the "randomization-t" procedure is preferable. The main reason for the current default is that the "randomization-c" method is currently much faster. Before making the "randomization-t" the default, I'd like to significantly improve its performance.
Why is the implementation of the "randomization-t" slow, and "randomization-c" fast at the moment?
The reason is that the "randomization-t" procedure currently loops over calls to either feols() or fepois, whereas the "randomization-c" is based on a more optimized numba implementation without all the overhead of the two functions. This is the case because in order to compute the RI t-statistics, we have to compute variance-covariance matrices. For that, we currently have no standalone interface that is numba compatible - the variance covariance matrices are not defined in a standalone, callable function, but are attached to a method of the Feols class.
To do
Add an option to compute vcov's and t-stats in _run_ri.
Change the ritest default of type from "randomization-c" to "randomization-t".
Blocked by
Moving the content of Feols.vcov to a stand-alone, callable, numba compatible function.
Context
The current default for randomization inference is the "randomization-c" method, despite evidence that the "randomization-t" procedure is preferable. The main reason for the current default is that the "randomization-c" method is currently much faster. Before making the "randomization-t" the default, I'd like to significantly improve its performance.
Why is the implementation of the "randomization-t" slow, and "randomization-c" fast at the moment?
The reason is that the "randomization-t" procedure currently loops over calls to either
feols()orfepois, whereas the "randomization-c" is based on a more optimized numba implementation without all the overhead of the two functions. This is the case because in order to compute the RI t-statistics, we have to compute variance-covariance matrices. For that, we currently have no standalone interface that is numba compatible - the variance covariance matrices are not defined in a standalone, callable function, but are attached to a method of the Feols class.To do
Blocked by