A Python implementation of the linear interaction-effect estimator from the R package interflex. Estimates heterogeneous treatment or marginal effects of a treatment variable across values of a moderating variable, with support for multiple GLM methods, variance estimation paths, and fixed effects.
- Five GLM methods: linear (OLS), logit, probit, Poisson, negative binomial
- Three variance paths: delta method (analytical), simulation (MVN draws), nonparametric bootstrap
- Four vcov types: homoscedastic, robust (HC1), cluster-robust, panel-corrected (PCSE)
- Fixed effects: Frisch-Waugh-Lovell iterative demeaning for one-way and two-way FE
- Instrumental variables: 2SLS with optional FE (IV-FWL)
- Treatment types: discrete (binary or multi-arm) and continuous
- Uniform confidence intervals: bootstrap bisection and delta-method MVN
- ATE/AME: average treatment effects (discrete) and average marginal effects (continuous)
- Plotting: multi-panel matplotlib figures with CI ribbons and distribution strips
# From source (development mode)
git clone https://github.com/statsclaw/example-R2PY.git
cd example-R2PY
pip install -e ".[dev]"Note: The current pyproject.toml has a build-backend issue. If pip install -e . fails, use:
export PYTHONPATH="$PWD:$PYTHONPATH"- Python >= 3.10
- numpy >= 1.24
- scipy >= 1.10
- statsmodels >= 0.14
- pandas >= 2.0
- matplotlib >= 3.7
- seaborn >= 0.12
Both the R and Python versions of interflex produce equivalent marginal effect estimates. The plots below use the same DGP (Y = 1 + 2D + 0.5X + 1.5DX + ε, n=500, same seed):
R interflex |
Python interflex |
|---|---|
 |
 |
Both recover the true marginal effect ∂E[Y|D,X]/∂D = 2 + 1.5X with matching point estimates and confidence intervals.
import interflex
# The module itself is callable — no need for interflex.interflex()
result = interflex(data=df, Y="Y", D="D", X="X", estimator="linear")import numpy as np
import pandas as pd
import interflex
# Generate sample data
rng = np.random.default_rng(42)
n = 500
X = rng.uniform(-2, 2, n)
D = rng.choice(["control", "treated"], n)
Z1 = rng.normal(0, 1, n)
D_num = (D == "treated").astype(float)
Y = 1 + 0.5 * X + 2 * D_num + 1.5 * D_num * X + 0.3 * Z1 + rng.normal(0, 0.5, n)
data = pd.DataFrame({"Y": Y, "D": D, "X": X, "Z1": Z1})
# Estimate interaction effects
result = interflex(
data=data, Y="Y", D="D", X="X", estimator="linear",
Z=["Z1"],
method="linear",
vartype="delta",
vcov_type="robust",
figure=True,
)
# Access results
print(result.est_lin) # Treatment effects at evaluation points
print(result.avg_estimate) # Average treatment effect (ATE)D_cont = rng.normal(0, 1, n)
Y_cont = 1 + 0.5 * X + 0.8 * D_cont + 0.6 * D_cont * X + rng.normal(0, 0.5, n)
data_cont = pd.DataFrame({"Y": Y_cont, "D": D_cont, "X": X})
result = interflex(
"linear", data_cont, "Y", "D", "X",
treat_type="continuous",
method="linear",
vartype="delta",
figure=True,
)
print(result.est_lin) # Marginal effects at evaluation points
print(result.avg_estimate) # Average marginal effect (AME)interflex(
estimator="linear", # Estimator type (only "linear" supported)
data=None, # pandas DataFrame
Y="", # Outcome column name
D="", # Treatment column name
X="", # Moderator column name
treat_type=None, # "discrete" or "continuous" (auto-detected if None)
base=None, # Base treatment group for discrete treatment
Z=None, # List of covariate column names
IV=None, # List of instrument column names
FE=None, # List of fixed effect column names
full_moderate=False, # Include Z*X interaction terms
weights=None, # Weight column name
na_rm=False, # Remove rows with NA values
method="linear", # "linear", "logit", "probit", "poisson", "nbinom"
vartype="delta", # "delta", "simu", "bootstrap"
vcov_type="robust", # "homoscedastic", "robust", "cluster", "pcse"
cl=None, # Cluster column name (required for "cluster" and "pcse")
time=None, # Time column name (required for "pcse")
neval=50, # Number of evaluation points
nboots=200, # Number of bootstrap replications
nsimu=1000, # Number of simulation draws
figure=True, # Generate matplotlib plot
# ... additional plotting parameters
)Returns: InterflexResult dataclass.
Key attributes:
| Attribute | Type | Description |
|---|---|---|
est_lin |
dict[str, ndarray] |
TE/ME table per treatment arm: columns (X, TE, sd, lower, upper, [uniform_lower, uniform_upper]) |
pred_lin |
dict[str, ndarray] |
Predicted values at evaluation points |
link_lin |
dict[str, ndarray] |
Link (linear predictor) values |
diff_estimate |
dict[str, DataFrame] |
Differences of TE/ME across moderator values |
vcov_matrix |
dict[str, ndarray] |
TE/ME covariance matrix per arm |
avg_estimate |
dict[str, DataFrame] |
ATE (discrete) or AME (continuous) |
figure |
Figure or None |
matplotlib figure if figure=True |
use_fe |
bool |
Whether fixed effects were used |
Methods:
result.predict(type="response")-- re-plot predictions
| Method | Link Function | Use Case |
|---|---|---|
"linear" |
Identity | Continuous outcome |
"logit" |
Logistic | Binary outcome |
"probit" |
Normal CDF | Binary outcome |
"poisson" |
Log | Count outcome |
"nbinom" |
Log | Overdispersed count |
| VarType | Description |
|---|---|
"delta" |
Analytical delta method (fastest, default) |
"simu" |
Simulation from MVN coefficient distribution |
"bootstrap" |
Nonparametric bootstrap (block bootstrap for clustered data) |
| VcovType | Description | Requirements |
|---|---|---|
"homoscedastic" |
Classical OLS variance | -- |
"robust" |
HC1 heteroscedasticity-consistent | -- |
"cluster" |
Cluster-robust sandwich | cl specified |
"pcse" |
Panel-corrected (Beck & Katz) | cl and time specified |
# Install dev dependencies
pip install -e ".[dev]"
# Run tests
pytest tests/ -v --tb=short
# Run with coverage
pytest tests/ -v --cov=interflex --cov-report=term-missing- Hainmueller, J., Mummolo, J., & Xu, Y. (2019). How Much Should We Trust Estimates from Multiplicative Interaction Models? Simple Tools to Improve Empirical Practice. Political Analysis, 27(2), 163-192.
- Original R package: github.com/xuyiqing/interflex