Statsmodels api

[s3alfisc]

I have written a first 'API' for statsmodels as a wildboottest() function, that works similarly to fwildclusterboot::boottest(). As the bootstrap algorithm (as implemented) does not produce a bootstrapped vcov matrix, I believe that it might be best not to introduce the bootstrap algo via get_robustcov_results() but to write a custom method / function to work with OLS estimation in statsmodels. wildboottest() is a first draft - I am open to any changes you suggest @amichuda.

We can also introduce wild bootstrapped vcov matrices via get_robustcov_results(), but I would then suggest to do so at a later point in time. As mentioned the other day, computing the full bootstrapped vcov slows down the bootstrap quite a bit.

Note that there are still a few open questions regarding the pre-processing of the design matrix and cluster variable - e.g., does

model = sm.OLS(Y, X)
model.exog

drop all missing values, collinear variables, etc?

Example:

from wildboottest.wildboottest import wildboottest
import statsmodels.api as sm
import numpy as np

N = 1000
k = 10
G= 12
X = np.random.normal(0, 1, N * k).reshape((N,k))
beta = np.random.normal(0,1,k)
beta[0] = 0.005
u = np.random.normal(0,1,N)
Y = 1 + X @ beta + u
cluster = np.random.choice(list(range(0,G)), N)
B = 99999


model = sm.OLS(Y, X)
model.exog
results = model.fit(cov_type = 'cluster', cov_kwds = {
   'groups': cluster
})
results.summary()
# >>> results.summary()
# <class 'statsmodels.iolib.summary.Summary'>
# """
#                                  OLS Regression Results                                
# =======================================================================================
# Dep. Variable:                      y   R-squared (uncentered):                   0.799
# Model:                            OLS   Adj. R-squared (uncentered):              0.797
# Method:                 Least Squares   F-statistic:                              790.3
# Date:                Sat, 15 Oct 2022   Prob (F-statistic):                    3.16e-14
# Time:                        12:03:43   Log-Likelihood:                         -1784.6
# No. Observations:                1000   AIC:                                      3589.
# Df Residuals:                     990   BIC:                                      3638.
# Df Model:                          10                                                  
# Covariance Type:              cluster                                                  
# ==============================================================================
#                  coef    std err          z      P>|z|      [0.025      0.975]
# ------------------------------------------------------------------------------
# x1             0.0128      0.064      0.200      0.841      -0.113       0.138

wildboottest(model, "X1", cluster, B)
# 0.8408408408408409

[amichuda]

Alex, this looks amazing! Let me take a look at the pull request when I'm near my computer, but this looks great.

I think you're right in terms of statsmodels and the vcov matrix. However, I think what I would propose then is to add it to statsmodels in such a way that the resulting summary table would just show pvalues without a standard error. Perhaps either just NaN or we just cut that column out whenever a user chooses it. How does that sound?

[s3alfisc]

Yes, I will add an option to bootstrap the "full" vcov matrix :) See #24