I am trying to rebuild a paper (Dreher et al. 2020. Aid, China, and Growth: Evidence from a New Global Development Finance Dataset) with Python. The paper's calculations were performed with Stata. I managed to rebuild everything until now, but I am stuck.
The data of this project is PanelData and I have to include year specific effects and country specific effects,
time_specific_effects = True, entity_effects = True / other_effects = data4.code
(because the variable data4.code is here the same as the entity).
The plan is to use 2SLS:
Regression IV Strategy The Stata code is given by the author:
*xtivreg2 growth_pc (l2.OFn_all= l3.IV_reserves_OFn_all_1_ln l3.IV_factor1_OFn_all_1_ln) l.population_ln time* if code!="CHN", fe first savefprefix(first) cluster(code) endog(l2.OFn_all)*
I rebuilt all the lagged variables in Python using shift() and it worked:
data4["l3IV_reserves_OFn_all_1_ln"] = data4["IV_reserves_OFn_all_1_ln"].shift(3)
data4["l3IV_factor1_OFn_all_1_ln"] = data4["IV_factor1_OFn_all_1_ln"].shift(3)
So the setup is the same as it is for the author:
As far as I know, there is no library in Python that can perform 2SLS with fixed effects. So I thought that I will just use linear model PanelOLS (which is suited for panel data with fixed effects) to perform the First Stage and Second Stage separately:
dependendFS = data4.l2OFn_all
exog2 = sm.tools.add_constant(data4[["l1population_ln", "l3IV_reserves_OFn_all_1_ln","l3IV_factor1_OFn_all_1_ln"]])
mod = lm.panel.PanelOLS(dependendFS, exog2, time_effects = True, entity_effects=True, drop_absorbed=True)
mod_new21c = mod.fit(cov_type='clustered', clusters = data4.code)
# Safe the fitted values
fitted_c = mod_new21c.fitted_values
data4["fitted_values_c"] = fitted_c
dependentSS = data4.growth_pc
exog = sm.tools.add_constant(data4[["fitted_values_c", "l1population_ln"]])
mod = lm.panel.PanelOLS(dependentSS, exog, time_effects=True, entity_effects= True)
mod_new211c = mod.fit(cov_type='clustered', clusters = data4.code)
I tried several combinations of the fixed effects and for the covariance, but it did not so far deliver the results I need. Here is my output for the Second Stage: Results after Second Stage and this is what they should look like: dependent variable is growth p.c, SE in brackets
Where is my mistake? Do I have to adjust my data or the output of the First Stage since I am separately performing 2SLS? Is there a mistake or a better method of estimating 2SLS in Python?