RLM cov_params, t_test, f_test don't use bcov_scaled #1164

Closed
josef-pkt opened this Issue Oct 30, 2013 · 4 comments

Projects

None yet

2 participants

@josef-pkt
Member

RLMResults does not have bse = sqrt(diag(cov_params))

generic results based on cov_params (e.g. t_test) are not the same as those by the specific methods.
generic tvalues uses bse which is correctly based on bcov_scaled

cov_params should return the default covariance which is bcov_scaled

found while testing that t_test(eye(..)) returns the same as tvalues, pvalues across models (after adjusting for normal/t distribution.

@josef-pkt
Member

In my robustcov branch I can just set it in __init__ with

self.cov_params_default = self.bcov_scaled

see #1158

@josef-pkt josef-pkt added a commit to josef-pkt/statsmodels that referenced this issue Oct 31, 2013
@josef-pkt josef-pkt REF: use robustcov in RLM as default, see #1164 6117549
@josef-pkt
Member

Note: the above commit 6117549 broke bcov_unscaled, (and RLM tests were failing)
fixed in 87684b7 by replacing with normalized_cov_params

Currently there is no cov option unscaled, the only available options are H1, H2, H3

Should there be a "non-robust" option?

@PierreBdR PierreBdR pushed a commit to PierreBdR/statsmodels that referenced this issue Sep 2, 2014
@josef-pkt josef-pkt REF: use robustcov in RLM as default, see #1164 c99566d
@josef-pkt
Member

Looks like this is fixed in master with the two commits.

missing unit tests for this?

@jseabold
Member

Closing. Create a separate issue to add tests, if desired.

@jseabold jseabold closed this Sep 26, 2014
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment