ddof in regression, and use-t

[josef-pkt]

xtgls invest mvalue kstock, corr(ar1) FGLS for timeseries-cross-section

estimates scale = ssr / nobs instead of ssr / (nobs - k_vars)

and reports z-values, i.e. normal distribution based

xtregar with re reports z-values and chi2, with fe it reports t-values and F

In PanelAR1, I have the same params (using the same rho as stata) but bse and cov_params differ because the scale is normalized in a different way.
I'm returning a OLS regression, the same would be with GLS.

In the documentation for another estimator (don't remember which), Stata has a user option to use nobs or (nobs - k_vars) as denominator in the scale calculation, referring to Greene that both cases are used and no clear advantage to either one.

-> need a possibility to change ddof for scale, e.g. ssr / (nobs - k_vars - ddof) with possibility to set ddof = - k_vars
-> need possibility to switch between t and normal distribution (and f and t) use_t as in summary.

(I thought I opened already an issue for ddof with pre-estimation removal of groups means for fixed effects models, but didn't find it.)

[josef-pkt]

for cluster robust standard errors, I needed to introduce df_resid_inference
In another case in Stata I have seen something like df_resid_var the df that are used in calculating the scale.
R also has in several places (???) different df, e.g. specifically for an F-test.

There are several related issues open. Until now we only have a case specific pattern, except for our old standard definitions.
For example sandwich robust covariances still need streamlining of the different options for df and denominator corrections.

[josef-pkt]

BTW: use_t is now generically implemented.

[josef-pkt]

Note: this is different from ddof for absorbed variables, e.g. demeaned exog and endog. See #2568 and #2774

The issue here mainly affects the cov_params which is now handled through cov_type arguments. AFAIR, that includes now a generic scaling factor.