Coming from Stata#

This chapter has benefitted enormously from Daniel M. Sullivan’s excellent notes.

The biggest difference between Python and Stata is that Python is a fully-fledged programming language, which means it can do lots of things, while Stata is really just for data analysis. What this means in practice is that sometimes the notation to do this or that operation in Python (or any other general purpose programming language) is less concise than in Stata. There is greater competition for each command in Python because it does many more things.

Another difference is that, in Stata, there is one dataset in memory that is represented as matrix where each column is a “variable” with a unique name. In Python, variables can be anything, even functions! But most data analysis in Python is done using dataframes, which are objects that are somewhat similar to a single dataset in Stata. In Python, you can have as many dataframes as you like in action at once. This causes the first major notational differences; in Python, you need to specify which dataframe you want to perform an operation on, in addition to which column (or row, or entry).

Finally, Python and its data analysis packages are free.

Regardless of Python not being a programming language solely dedicated to data analysis, it really does have first class support for data analysis via its pandas package. Support for doing regressions is perhaps less good than Stata, and certainly a bit more verbose—but you can still do pretty much every standard operation you can think of.

Stata <==> Python#

What follows is a giant table of translations between Stata code and Python, leaning heavily on Python’s pandas (panel-data-analysis) package. We’re going to rely on a few packages for econometrics in the below. They are statsmodels as your general purpose and flexible regression library, pyfixest for when you need high dimensional fixed effects, and binsreg for bin scatter.

Many of the examples below assume that, in Python, you have a pandas DataFrame called df. We will use placeholders like varname for Stata variables and df['varname'] for the Python equivalent. Remember that you need to import pandas as pd before running any of the examples that use pd. For the econometrics examples, you will need to import the relevant package.

You can find more on (frequentist) regressions in Regression, Bayesian regressions using formulae appear in Bayesian Inference Made Easier, generalised regression models appear in Generalised regression models, and regression diagnostics and visualisation are in Regression diagnostics and visualisations. For Bayesian regressions, Python is very strong: check out Bayesian Inference Made Easier.

Stata

Python (pandas)

help command

help(command)

cd directory

import os
os.chdir(‘directory’)
Best practice: don’t do this; bring the data to you by opening Visual Studio Code in a project root folder and using relative paths.

use file.dta

df = pd.read_stata('file.dta')

use varlist using dtafile

df = pd.read_stata(‘dtafile’, columns=varlist)

import excel using excelfile

df = pd.read_excel(‘excelfile’)

import delimited using csvfile

df = pd.read_csv(‘csvfile’)

save filename, replace

df.to_stata(‘filename’)
Best practice: don’t save data in .dta files.

outsheet using filename, comma

df.to_csv(‘filename’)

export excel using filename

df.to_excel(‘filename’)
Best practice: don’t save data in Excel files.

keep if condition

df = df[condition]

drop if condition

df = df[~condition]

keep variable

df = df['variable']

keep varstem*

df = df.filter(like='varstem*')

drop variable

df = df.drop('variable', axis=1)

drop varstem*

df = df.drop(df.filter(like='varstem*').columns, axis=1)

describe

df.info()

describe variable

df['variable'].dtype

count

len(df)

count if condition

df[condition].shape[0]

summ variable

df['variable'].describe()

summ variable if condition

df.loc[condition, 'variable'].describe()

gen newvar = expression

df['newvar'] = expression

gen newvar = expression if condition

df.loc[condition, 'newvar'] = expression

replace newvar = expression if condition

df.loc[condition, 'newvar'] = expression

rename var newvar

df = df.rename(columns={var: newvar}) or df.columns=list_new_columns

subinstr(string, " ", "_", .)

df['var'].str.replace(' ', '_')

egen newvar = statistic(var), by(groupvars)

df['newvar'] = df.groupby(groupvars)['var'].transform('statistic')

collapse (sd) var (median) var (max) var (min) <var>, by(groupvars)

df.groupby(groupvars)['var'].agg(['std', 'median', 'min', 'max', 'sum'])

append using filename

df = df1.append(df2)

merge 1:1 vars using filename

df = pd.merge(df1, df2, on=vars) but there are very rich options for merging dataframes (Python is similar to SQL in this respect) and you should check the full documentation.

reshape <wide/long> <stubs>, i(<vars>) j(<var>)

pandas has several reshaping functions, including df.unstack('level') for going to wide, df.stack('column_level') for going to long, pd.melt, and df.pivot. It’s best to check the excellent reshaping documentation to find what best suits your needs.

xi: i.var

pd.get_dummies(df['var'])

reg yvar xvar if condition, r

import pyfixest as pf
fit = pf.feols(“yvar ~ xvar”, data=df[“condition”], vcov=”HC2”)

reg yvar xvar if condition, vce(cluster clustervar)

import pyfixest as pf
fit = pf.feols(“yvar ~ xvar”, data=df[“condition”], vcov={“CRV1”: “clustervar”})

areg yvar xvar, absorb(fe_var)

import pyfixest as pf
fit = pf.feols(“yvar ~ xvar | fe_var”, data=df)

_b[var], _se[var]

results_sw.coef()["var"], results_sw.se()["var"] following creation of results_sw via results_sw = pf.feols(...)

ivreg2 lwage exper expersq (educ=age)

pf.feols(“lwage ~ exper + expersq | educ ~ age”, data=dfiv)

outreg2

results = pf.feols(...) then results.tidy()

binscatter

binsreg from the binsreg package; see Regression diagnostics and visualisations.

twoway scatter var1 var2

df.scatter(var2, var1)

The table below presents further examples of doing regression with both the statsmodels and pyfixest packages.

Note that, in the below, you need only import pf.feols once in each Python session, and the syntax for looking at results is results = pf.feols(...) and then results.summary().

Command

Stata

Python

Fixed Effects (absorbing)

reghdfe y x, absorb(fe)

import pyfixest as pf
fit = pf.feols(“y ~ x | fe”, data=df)

Categorical regression

reghdfe y x i.cat

import pyfixest as pf
fit = pf.feols(“y ~ x + C(cat)”, data=df)
But if cat is of type categorical it can be run with y ~ x + cat

Interacting categoricals

reghdfe y x i.cat#i.cat2

import pyfixest as pf
fit = pf.feols(“yvar ~ xvar + C(cat):C(cat2)”, data=df)
Note that a*b is a short-hand for a + b + a:b, with the last term representing the interaction.

Robust standard errors

reghdfe y x, r

import pyfixest as pf
fit = pf.feols(“y ~ x, data=df, vcov=”HC1”)
Note that a range of heteroskedasticity robust standard errors are available: see Regression for more.

Clustered standard errors

reghdfe y x, cluster(clust)

import pyfixest as pf
fit = pf.feols(“y ~ x”, data=df, vcov={“CRV1”: “clust”})

Two-way clustered standard errors

reghdfe y x, cluster(clust1 clust2)

import pyfixest as pf
fit = pf.feols(“y ~ x”, data=df, vcov={“CRV1”: “clust1 + clust2”})

Instrumental variables

ivreghdfe 2sls y exog (endog = instrument)

import pyfixest as pf
fit = pf.feols(“y ~ exog | endog ~ instrument”, data=df)