Skip to content

linearmodels/manual-two-stage

Rule manual-two-stage
Pack linearmodels
Severity warning
Fix none, edit by hand
Library versions every version

Fitting a second regression on the fitted values of a first one is two stage least squares by hand, and its standard errors come from the wrong residuals. Use IV2SLS.

first = OLS(data.educ, data[["const", "exper", "fatheduc", "motheduc"]]).fit()
second = OLS(data.lwage, pd.concat([data[["const", "exper"]], first.fitted_values], axis=1)).fit()
result = IV2SLS(data.lwage, data[["const", "exper"]], data.educ, data[["fatheduc", "motheduc"]]).fit()

The rule reports a model whose exog holds .fitted_values of another fit. The model is OLS or a panel class, or IV2SLS, IVLIML or IVGMM with None for endog and for instruments, which is least squares. Such a model is two stage least squares by hand. The coefficient is right, but the standard errors come from residuals that use the fitted values in place of the regressor.

xpd has no automatic fix, so make the change by hand.

Measurement By hand IV2SLS
coefficient of educ on the mroz data 0.0664 0.0664
unadjusted standard error on the mroz data 0.0327 0.0311
coverage of the 95% interval, over 2,000 draws of 500 rows with one instrument 99.6% 95.7%

In the draws, the standard error by hand was 1.37 times the other on average. Which way the error goes depends on the data. The rule sees only the form where the fitted values are written in the call, and it misses a second stage that reads them from a column stored earlier. Neither set of code has the form the rule reads.

xpd prints this message with the issue. A name that starts with $ stands for code from your file.

This regression takes the fitted values of another one as a regressor, which is two stage least squares by hand. The coefficient is the one IV2SLS gives, but the standard errors come from the wrong residuals. In a simulation the 95% interval of the second stage covered the true value in 99.6% of 2,000 draws, against 95.7% for IV2SLS. Use IV2SLS(dependent, exog, endog, instruments).

Browse the other rules of the linearmodels pack.