Python, R, Stata and EViews
in a single notebook

Why this exists

The four-program problem

An applied econometrics paper rarely lives in one program. The unit-root test is in EViews because that is where the output is readable. The panel estimator is in Stata. The plots are in R. The data cleaning is in Python. So the working day looks like this:

Four windows, four copies of the same data slowly drifting apart. A missing value that meant .a in Stata arrives as an empty cell in R. A quarterly index becomes a string. And when a referee asks why your robust standard error differs from theirs, there is no way to answer without redoing the chain by hand.

What EconEnv does instead

From zero

Installation

This assumes a computer with nothing installed. If you already have Python and Jupyter, jump to step 3 — that is the only required one.

1. Python 3.9 or newer

   Download from python.org. On the first installer screen, tick “Add python.exe to PATH” — the single most commonly missed step.

   python --version

   Or, if you prefer conda:

   conda create -n econ python=3.12
   conda activate econ

2. Jupyter

   pip install jupyterlab

   jupyter lab

3. EconEnv

   This is the only step that is required. It installs pandas, numpy, IPython and statsmodels, and nothing commercial.

   pip install econenv

   From inside a notebook cell, use this form instead — it installs into the kernel actually running, rather than whichever Python happens to be first on your PATH:

   %pip install econenv

   Optional extras

   pip install "econenv[eviews]"   # comtypes, for EViews on Windows
   pip install "econenv[arrow]"    # faster data transfer to R
   pip install "econenv[all]"

4. R — optional, free

   Install from CRAN and accept the defaults. Nothing else is needed: EconEnv searches the registry and the usual install locations, so you do not have to set R_HOME or touch PATH.

   You do not need rpy2. EconEnv's default R backend is a persistent Rterm child process — no compiler, no wheels to hunt for. rpy2 publishes no Windows wheels, so pip install rpy2 usually leaves a broken build.

5. Stata — optional, commercial

   You need Stata 17 or newer, because that is when StataCorp began shipping PyStata. Install and licence it normally; EconEnv finds it, validates the executable and detects the edition (BE, SE or MP).

   %econ config stata.home "C:/Program Files/StataNow19"
   %econ config stata.edition mp

6. EViews — optional, commercial, Windows only

   Install and licence EViews, then open it once by hand and close it. That registers its COM automation server; until you have done it, automation cannot connect.

   pip install "econenv[eviews]"

   Several EViews versions installed? The generic identifier binds to whichever registered last — not the newest. Pin one with %econ config eviews.progid EViews.Manager.14. Note the format: EViews.Manager.14, not EViews14.Manager.

Verify

Your first cell

Load the extension once per notebook. It reports which magics it registered and, importantly, who owns each one.

%load_ext econenv

EconEnv 1.0.9 loaded — one notebook, multiple econometric engines.
  %econ                  econenv
  %Rec / %%Rec           econenv
  %R / %%R               econenv (rpy2 not installed)
  %eviews / %%eviews     econenv
  %stata / %%stata       pystata (official)
  %econ status · %econ doctor · %econ help

What is available on this machine

%econ status

             state version edition        backend  \
engine
Python  configured  3.11.0                cpython
EViews  configured      13               comtypes
MATLAB  configured  R2024a          matlab.engine
R       configured   4.5.2             subprocess
Stata      running    19.5      mp        pystata

                                                 location
engine
Python  C:\Users\HP\AppData\Local\Programs\Python\Pyth...
EViews                         C:\Program Files\EViews 13
...

And if something will not start

%econ doctor checks every layer — Python, Jupyter, R, Stata, EViews, COM, temp directories, permissions — and every warning or error carries a suggested fix. That is enforced by a test.

%econ doctor

Reference

Every magic command

Line magics start with % and act on one line. Cell magics start with %% and take the whole cell.

%econ — managing everything

%%R -i df -o results
fit <- lm(y ~ x1 + x2, data = df)
results <- coef(fit)

%econ push stata df

%%stata
regress y x1 x2, robust
estat ic

%%eviews -i df
equation eq1.ls y c x1 x2
eq1.output
line y

import econenv

econenv.push("r", "df", df)
econenv.pull("stata")
econenv.move("stata", "eviews", "mydata")
econenv.compare_ols(df, "y ~ x1 + x2")
econenv.doctor()
econenv.snapshot()

Real data, real output

A complete worked analysis

Everything below is genuine output from the example notebook, executed against R 4.5.2, StataNow 19.5 MP and EViews 13. The data is real US quarterly macroeconomic data, 1959Q1–2009Q3 (203 observations), shipped with statsmodels — so the notebook needs no download and no private file.

1 · Python builds the data

import numpy as np, pandas as pd, statsmodels.api as sm, econenv

raw = sm.datasets.macrodata.load_pandas().data
idx = pd.PeriodIndex(year=raw.year.astype(int),
                     quarter=raw.quarter.astype(int), freq='Q').to_timestamp()

macro = pd.DataFrame({
    'lrgdp':  np.log(raw.realgdp.values),
    'lrcons': np.log(raw.realcons.values),
    'realint': raw.realint.values,
}, index=idx)

         lrgdp   lrcons    lrinv    unemp     infl    tbill  realint
count  203.000  203.000  203.000  203.000  203.000  203.000  203.000
mean     8.781    8.362    6.750    5.885    3.961    5.312    1.337
std      0.466    0.501    0.596    1.459    3.253    2.803    2.669
min      7.905    7.443    5.560    3.400   -8.790    0.120   -6.790
25%      8.398    7.963    6.252    4.900    2.270    3.515   -0.085
50%      8.789    8.366    6.798    5.700    3.240    5.010    1.340
75%      9.173    8.764    7.270    6.800    4.975    6.665    2.630
max      9.504    9.145    7.725   10.700   14.620   15.330   10.950

2 · R estimates and diagnoses

%%R -i macro
fit_r <- lm(lrcons ~ lrgdp + realint, data = macro)
summary(fit_r)

Call:
lm(formula = lrcons ~ lrgdp + realint, data = macro)

Residuals:
      Min        1Q    Median        3Q       Max
-0.030434 -0.011066 -0.000498  0.009559  0.038223

Coefficients:
              Estimate Std. Error t value Pr(>|t|)
(Intercept) -1.0735725  0.0189306 -56.711  < 2e-16 ***
lrgdp        1.0746806  0.0021514 499.527  < 2e-16 ***
realint     -0.0010893  0.0003754  -2.902  0.00412 **
---
Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

Residual standard error: 0.01424 on 200 degrees of freedom
Multiple R-squared:  0.9992,	Adjusted R-squared:  0.9992
F-statistic: 1.248e+05 on 2 and 200 DF,  p-value: < 2.2e-16

R's four-panel diagnostic view, rendered straight into the notebook:

3 · Stata supplies HAC standard errors

Those residuals are strongly autocorrelated, so ordinary standard errors are too small. newey is Stata's reference implementation of Newey–West.

%econ push stata macro

%%stata
tsset t
newey lrcons lrgdp realint, lag(4)

. gen t = _n

. tsset t

Time variable: t, 1 to 203
        Delta: 1 unit

. newey lrcons lrgdp realint, lag(4)

4 · EViews does the time-series work

%%eviews -i macro
lrgdp.uroot(adf)

Null Hypothesis: LRGDP has a unit root
Exogenous: Constant
Lag Length: 1 (Automatic - based on SIC, maxlag=14)

                                                                                         t-Statistic  Prob.*

Augmented Dickey-Fuller test statistic                                                   -1.820451    0.3698
Test critical values:                                1% level                            -3.462901
                                                     5% level                            -2.875752
                                                     10% level                           -2.574423

*MacKinnon (1996) one-sided p-values.
...

Log real GDP fails to reject a unit root — exactly as it should on trending macro data. Next, whether the two series move together in the long run:

%%eviews
group gci lrcons lrgdp
gci.coint(e)

                   Johansen Cointegration Test
=================================================================
Date: 09/05/26   Time: 07:07
Sample: 1959Q1 2009Q3
Included observations: 203
Lags interval (in first differences): 1 to 4
Endogenous variables: LRCONS LRGDP
Deterministic assumptions: Case 5 (Johansen-Hendry-Juselius): Both the
        cointegrating relationship and short-run dynamics include a constan
        trend.
=================================================================

=================================================================
Unrestricted Cointegration Rank Test (Trace)
=================================================================
Hypothesized                  Trace        0.05        Prob.**
...

%%eviews
equation eq_ev.ls lrcons c lrgdp realint
eq_ev.output

Dependent Variable: LRCONS
Method: Least Squares
Date: 09/05/26   Time: 07:07
Sample: 1959Q1 2009Q3
Included observations: 203

Variable                      Coefficient  Std. Error             t-Statistic  Prob.

C                             -1.073573    0.018931               -56.71091    0.0000
LRGDP                         1.074681     0.002151               499.5275     0.0000
REALINT                       -0.001089    0.000375               -2.902182    0.0041

R-squared                     0.999199     Mean dependent var                  8.361722
Adjusted R-squared            0.999191     S.D. dependent var                  0.500639
S.E. of regression            0.014236     Akaike info criterion               -5.651383
Sum squared resid             0.040534     Schwarz criterion                   -5.602419
Log likelihood                576.6154     Hannan-Quinn criter.                -5.631574
F-statistic                   124805.2     Durbin-Watson stat                  0.278552
...

5 · EViews plots, captured into the notebook

The part that is hard to do any other way

The same model in all four engines

One specification, four programs, one table — and an honest account of where they differ.

econenv.compare_ols(macro, "lrcons ~ lrgdp + realint")

C:\Users\HP\Documents\econenv\src\econenv\engines\stata_engine.py:356: UserWarning: EconEnv push -> stata: [warning] date: datetime converted to Stata %tc (ms since 1960-01-01); apply `format ... %tc` in Stata to display it as a date
  push_frame(self, "__econenv_ols", data.loc[:, columns], clear=True)

For researchers who click

EViews without the menus

In EViews you normally click. In a notebook there is nothing to click, and that — not any missing feature — is the real obstacle. So the command reference lives inside the notebook, where the question gets asked.

%econ eviews                      # the list of tasks
%econ eviews graph                # everything about plotting
%econ eviews find cointegration   # search all of it

✓ g.coint(e)
    Johansen system cointegration test — trace and maximum-eigenvalue
    statistics for how many cointegrating relations exist.
    GUI: Group > View > Cointegration Test > Johansen
    e.g. g1.coint(e)

136 commands, 94 verified against EViews 13. Each entry gives the menu path you already know, the command, what it does, and a line to copy.

The one rule that makes EViews guessable

EViews objects work as object_name.what_you_want. Anything in an object's View menu is object.thatview; anything in Proc is object.thatproc. That single rule covers most of the language.

Three ways to draw the same plot

All three work, and the GUI hides the distinction:

line x            ' command form  — quickest
x.line            ' view form     — matches the View menu
graph gr1.line x  ' object form   — keeps the graph so you can edit it

The complete command reference →

No installation at all

Running on Google Colab

One line, no local setup — but only for two of the four engines, and the reason is worth stating plainly rather than leaving you to discover it.

%pip install econenv

All four engines, still in Colab

There is a way round every limitation above, and it is Google's own feature: connect Colab to a local runtime. Colab already runs in a browser on your PC — point it at a Jupyter server on that same PC and the interface stays Colab while the kernel, and every engine, is your Windows machine.

Nothing is exposed to the internet: your browser talks to localhost, Google's servers never reach your machine, and EViews is driven by local COM exactly as in a local notebook.

python -m venv colab-runtime
colab-runtime/Scripts/pip install "notebook==6.4.12" jupyter_http_over_ws econenv

colab-runtime/Scripts/jupyter serverextension enable --py jupyter_http_over_ws

colab-runtime/Scripts/jupyter notebook --no-browser --port=8888 --NotebookApp.port_retries=0 --NotebookApp.allow_origin="https://colab.research.google.com"

Copy the http://localhost:8888/?token=… line, then in Colab click the Connect arrow and choose Connect to a local runtime.

Open the Colab notebook

When something goes wrong

Troubleshooting

ModuleNotFoundError: No module named 'econenv'

import sys; print(sys.executable)
!{sys.executable} -m pip install econenv

%econ doctor

%econ config eviews.progid EViews.Manager.14

ImportError: cannot import name
'SexpVectorCCompatibleAbstract'

pip uninstall -y rpy2
conda install -c conda-forge rpy2