assumpcheck

assumpcheck is a small Python package for checking core assumptions for:

• ANOVA
• Linear regression
• Logistic regression

The package is designed to stay simple:

• concise terminal output by default
• plots only when they matter
• mitigation suggestions when something fails
• optional structured output for programmatic use

Installation

pip install assumpcheck

The package currently depends on:

• numpy
• pandas
• scipy
• statsmodels
• matplotlib

Public API

from assumpcheck import (
    check_anova,
    check_linear_regression,
    check_logistic_regression,
)

ANOVA

report = check_anova(y=y, groups=groups)

Linear regression

report = check_linear_regression(model=fitted_ols_model)

Logistic regression

report = check_logistic_regression(model=fitted_logit_model)

Quickstart

import numpy as np
import pandas as pd

from assumpcheck import check_linear_regression

rng = np.random.default_rng(27)
X = pd.DataFrame(
    {
        "x1": rng.normal(size=35),
        "x2": rng.normal(size=35),
    }
)
y = 1.0 + 1.8 * X["x1"] - 0.6 * X["x2"] + rng.normal(scale=0.35, size=35)

report = check_linear_regression(
    X=X,
    y=y,
    design_independent=True,
    plots_on_fail=False,
)

Typical output looks like:

LINEAR REGRESSION ASSUMPTION CHECKS
[PASS] Linearity
[PASS] Independence
[PASS] Normality of residuals
[PASS] Homoscedasticity
[PASS] Multicollinearity
[WARN] Extreme influential points

Summary: 5 pass, 1 warn

Example output

ANOVA ASSUMPTION CHECKS
[INFO] Independence
[PASS] Normality of residuals
[PASS] Equal variance across groups
[FAIL] Extreme outliers

Summary: 2 pass, 1 fail, 1 info

Details:
- Extreme outliers [FAIL]
  Metric: Max |standardized residual| = 3.420; flagged points > 3: 1
  Threshold: Values above 2 deserve review and values above 3 are concerning.
  Interpretation: At least one observation has a standardized residual above the common concern threshold.
  Possible mitigation:
    - Verify data entry for flagged cases.
    - Check whether the observation is legitimate but unusual.
    - Consider a transformation, robust method, or nonparametric alternative if outliers remain influential.

Options

All three public functions support these core options:

• alpha=0.05
• show_all=False
• plots_on_fail=True
• verbose=False
• return_dict=False
• design_independent=None

Additional model-specific option:

• check_linear_regression(..., ordered=False)

Output behavior

Default behavior:

• prints a concise summary
• shows plots for failed or warning-level checks if plots_on_fail=True
• returns an AssumptionReport object

Optional behavior:

• verbose=True prints detail for every check
• show_all=True prints all details and shows all available plots
• return_dict=True returns a serializable dictionary

Independence handling

By default, independence is treated as a design question:

• design_independent=None gives an INFO result
• design_independent=True gives a PASS unless an ordered linear model also shows autocorrelation warnings
• design_independent=False gives a FAIL

Current checks

ANOVA

• Independence
• Normality of residuals
• Equal variance across groups
• Extreme outliers

Linear regression

• Linearity
• Independence
• Normality of residuals
• Homoscedasticity
• Multicollinearity
• Extreme influential points

Logistic regression

• Linearity in the log-odds
• Independence
• Multicollinearity
• Extreme influential points
• Adequate sample / no separation
• Model fit summary via ROC / AUC

Notes on the MVP

• The package prioritizes statsmodels models first.
• Thresholds are intentionally presented as heuristics.
• Logistic ROC / AUC is treated as a fit diagnostic, not a strict assumption.
• Some diagnostics need access to original data or design metadata to be fully informative.
• The current influence heuristics are intentionally conservative, so clean linear or logistic examples may still emit a WARN.

Examples

• Script: examples/basic_usage.py
• Notebook: examples/assumpcheck_examples.ipynb

To rebuild the executed notebook and example plot assets:

python examples/build_workflow_artifacts.py

Tests

Run the test suite with:

python -m pytest -q

Release Process

For the first public release to TestPyPI and PyPI using GitHub Trusted Publishing, see RELEASING.md.