EpiPyLab

EpiPyLab is a computational epidemiology framework for Python.

The package provides epidemiology-focused tools for mortality, morbidity, descriptive and analytic epidemiology, measures of association, precision and uncertainty, validity and diagnostic performance, stratified analysis, regression, survival and competing-risks analysis, complex survey analysis, surveillance, pharmacoepidemiology, sample-size planning, data preparation, machine learning, visualization, interpretation, teaching, publication-oriented workflows, and causal inference.

Package status

Version 0.2.4 is an alpha release. It adds the epipylab.causal subpackage to the existing EpiPyLab distribution.

Installation

The distribution name and Python import package are both epipylab.

pip install epipylab

import epipylab
print(epipylab.__version__)

There is no separate epipylab-causal distribution. Causal methods are imported from epipylab.causal.

Causal inference (epipylab.causal)

epipylab.causal is a time-fixed binary-treatment causal-inference subpackage. Numerical estimation does not establish causal identification. Observational causal interpretation depends on identification assumptions that the analyst states; the software does not empirically prove them.

The public API is:

• estimands: average_treatment_effect, average_treatment_effect_treated, average_treatment_effect_controls
• identification records: observational_assumptions, randomized_assumptions
• result helpers: causal_result, format_causal_result, validate_binary_treatment, validate_binary_outcome
• algebraic contrasts of supplied means: causal_difference, causal_ratio, causal_odds_ratio, causal_effect_from_means
• parametric g-formula / standardization: g_formula
• propensity-score estimation: propensity_score
• IPTW: iptw
• 1:1 greedy nearest-neighbor propensity-score matching: propensity_score_matching
• standardized mean differences: standardized_mean_differences
• causal diagnostics: causal_diagnostics
• AIPW: aipw
• TMLE: tmle
• bootstrap inference: bootstrap_inference

Estimand constructors name ATE, ATT, or ATC targets; they do not estimate effects from data. Algebraic contrast functions require already targeted mu1 and mu0. propensity_score estimates an assignment probability; it is not itself an effect estimator. standardized_mean_differences and causal_diagnostics summarize observed data and do not return a treatment effect. bootstrap_inference wraps a supplied estimator; it is not a causal method of its own.

g_formula and iptw support ATE, ATT, and ATC. Matching estimates ATT only. AIPW and TMLE estimate ATE only.

import pandas as pd
from epipylab.causal import (
    observational_assumptions,
    g_formula,
)

data = pd.DataFrame(
    {
        "A": [0, 0, 0, 0, 1, 1, 1, 1],
        "Y": [1.0, 2.0, 1.5, 2.5, 3.0, 4.0, 3.5, 5.0],
        "X": [0.0, 1.0, 0.0, 1.0, 0.0, 1.0, 0.0, 1.0],
    }
)

assumptions = observational_assumptions(
    treatment="A",
    outcome="Y",
    covariates=["X"],
)

result = g_formula(
    data=data,
    treatment="A",
    outcome="Y",
    outcome_type="continuous",
    covariates=["X"],
    estimand="ATE",
    scale="difference",
    identification_assumptions=assumptions,
)

print(result["estimate"])
print(result["causal_language"])

result["causal_language"] is True only because an identification record was attached. That flag does not mean the assumptions were verified.

Limitations

• Observational causal interpretation depends on stated identification assumptions and is not proven by a fitted estimate, measured-covariate balance, overlap summaries, or a bootstrap interval.
• Matching is ATT-only. If treated units remain unmatched, the contrast is among the matched treated subset and their matched controls; it is not automatically the ATT for the entire original treated population and is not relabeled as ATE.
• AIPW and TMLE are ATE-only.
• This implementation does not cross-fit, does not use Super Learner, and is not CV-TMLE.
• Propensity scores and weights are not automatically clipped or trimmed.
• AIPW and TMLE do not provide an analytic standard error or confidence interval in this release.
• Bootstrap inference is available separately through bootstrap_inference. It does not upgrade causal_language and does not add cross-fitting, Super Learner, CV-TMLE, or analytic influence-curve inference.
• g_formula is a time-fixed parametric g-formula / standardization estimator, not a longitudinal g-formula.

Author

Shams Rahman
Email: shams16@gmail.com

License

EpiPyLab is released under the MIT License.

Copyright (c) 2026 Shams Rahman