Newey-West standard errors for OLS in Python?

Question

I want to have a coefficient and Newey-West standard error associated with it.

I am looking for Python library (ideally, but any working solutions is fine) that can do what the following R code is doing:

library(sandwich)
library(lmtest)

a <- matrix(c(1,3,5,7,4,5,6,4,7,8,9))
b <- matrix(c(3,5,6,2,4,6,7,8,7,8,9))

temp.lm = lm(a ~ b)

temp.summ <- summary(temp.lm)
temp.summ$coefficients <- unclass(coeftest(temp.lm, vcov. = NeweyWest))

print (temp.summ$coefficients)

Result:

             Estimate Std. Error   t value  Pr(>|t|)
(Intercept) 2.0576208  2.5230532 0.8155281 0.4358205
b           0.5594796  0.4071834 1.3740235 0.2026817

I get the coefficients and associated with them standard errors.

I see statsmodels.stats.sandwich_covariance.cov_hac module, but I don't see how to make it work with OLS.

Falstaffian

Answer 1

Edited (10/31/2015) to reflect preferred coding style for statsmodels as fall 2015.

In statsmodels version 0.6.1 you can do the following:

import pandas as pd
import numpy as np
import statsmodels.formula.api as smf

df = pd.DataFrame({'a':[1,3,5,7,4,5,6,4,7,8,9],
                   'b':[3,5,6,2,4,6,7,8,7,8,9]})

reg = smf.ols('a ~ 1 + b',data=df).fit(cov_type='HAC',cov_kwds={'maxlags':1})
print(reg.summary())

                                OLS Regression Results
==============================================================================
Dep. Variable:                      a   R-squared:                       0.281
Model:                            OLS   Adj. R-squared:                  0.201
Method:                 Least Squares   F-statistic:                     1.949
Date:                Sat, 31 Oct 2015   Prob (F-statistic):              0.196
Time:                        03:15:46   Log-Likelihood:                -22.603
No. Observations:                  11   AIC:                             49.21
Df Residuals:                       9   BIC:                             50.00
Df Model:                           1
Covariance Type:                  HAC
==============================================================================
                 coef    std err          z      P>|z|      [95.0% Conf. Int.]
------------------------------------------------------------------------------
Intercept      2.0576      2.661      0.773      0.439        -3.157     7.272
b              0.5595      0.401      1.396      0.163        -0.226     1.345
==============================================================================
Omnibus:                        0.361   Durbin-Watson:                   1.468
Prob(Omnibus):                  0.835   Jarque-Bera (JB):                0.331
Skew:                           0.321   Prob(JB):                        0.847
Kurtosis:                       2.442   Cond. No.                         19.1
==============================================================================

Warnings:
[1] Standard Errors are heteroscedasticity and autocorrelation robust (HAC) using 1 lags and without small sample correction

Or you can use the get_robustcov_results method after fitting the model:

reg = smf.ols('a ~ 1 + b',data=df).fit()
new = reg.get_robustcov_results(cov_type='HAC',maxlags=1)
print(new.summary())


                                OLS Regression Results
==============================================================================
Dep. Variable:                      a   R-squared:                       0.281
Model:                            OLS   Adj. R-squared:                  0.201
Method:                 Least Squares   F-statistic:                     1.949
Date:                Sat, 31 Oct 2015   Prob (F-statistic):              0.196
Time:                        03:15:46   Log-Likelihood:                -22.603
No. Observations:                  11   AIC:                             49.21
Df Residuals:                       9   BIC:                             50.00
Df Model:                           1
Covariance Type:                  HAC
==============================================================================
                 coef    std err          z      P>|z|      [95.0% Conf. Int.]
------------------------------------------------------------------------------
Intercept      2.0576      2.661      0.773      0.439        -3.157     7.272
b              0.5595      0.401      1.396      0.163        -0.226     1.345
==============================================================================
Omnibus:                        0.361   Durbin-Watson:                   1.468
Prob(Omnibus):                  0.835   Jarque-Bera (JB):                0.331
Skew:                           0.321   Prob(JB):                        0.847
Kurtosis:                       2.442   Cond. No.                         19.1
==============================================================================

Warnings:
[1] Standard Errors are heteroscedasticity and autocorrelation robust (HAC) using 1 lags and without small sample correction

The defaults for statsmodels are slightly different than the defaults for the equivalent method in R. The R method can be made equivalent to the statsmodels default (what I did above) by changing the vcov, call to the following:

temp.summ$coefficients <- unclass(coeftest(temp.lm, 
               vcov. = NeweyWest(temp.lm,lag=1,prewhite=FALSE)))
print(temp.summ$coefficients)

             Estimate Std. Error   t value  Pr(>|t|)
(Intercept) 2.0576208  2.6605060 0.7733945 0.4591196
b           0.5594796  0.4007965 1.3959193 0.1962142

You can also still do Newey-West in pandas (0.17), although I believe the plan is to deprecate OLS in pandas:

print(pd.stats.ols.OLS(df.a,df.b,nw_lags=1))

-------------------------Summary of Regression Analysis-------------------------

Formula: Y ~ <x> + <intercept>

Number of Observations:         11
Number of Degrees of Freedom:   2

R-squared:         0.2807
Adj R-squared:     0.2007

Rmse:              2.0880

F-stat (1, 9):     1.5943, p-value:     0.2384

Degrees of Freedom: model 1, resid 9

-----------------------Summary of Estimated Coefficients------------------------
      Variable       Coef    Std Err     t-stat    p-value    CI 2.5%   CI 97.5%
 --------------------------------------------------------------------------------
             x     0.5595     0.4431       1.26     0.2384    -0.3090     1.4280
     intercept     2.0576     2.9413       0.70     0.5019    -3.7073     7.8226
*** The calculations are Newey-West adjusted with lags     1

---------------------------------End of Summary---------------------------------

Interdental

to get the same result as pandas, we need to use the small sample option: use_correction=True, use_t=True. I don't know where the difference to R comes from, especially the difference in the standard error of the constant. Most robust covariance matrices were verified against Stata, but statsmodels doesn't always have the same default values for the options. – Crustacean

looked at R. It looks like R becomes equivalent to the statsmodels defaults with maxlags=1 if the vcov option is changed to the following: vcov. = NeweyWest(temp.lm,lag=1,prewhite=FALSE) – Interdental

Sorry, I´ve a question to codeline print pd.stats.ols.OLS(df.a,df.b,nw_lags=1). Does this line works with multiple independent variables? Like a as the dependent variable and b and c (instead of only b) as independent variables? If yes, how should the line look like? Can I just add df.c behind df.b? Thanks a lot1!! – Septi