spssmirror 2.0.0

A unified, self-contained statistical analysis library for Python -- an SPSS/R replacement.

SPSSMirror

SPSSMirror is a unified, self-contained statistical analysis library for Python — a modern SPSS/R replacement. Every method wraps scipy/statsmodels/scikit-learn/lifelines/pymc/arch internally and returns a typed, immutable result object with the statistic, p-value, effect size, and confidence interval already computed. You never need to import those libraries yourself to get a complete answer.

from spssmirror import SPSSMirror

mirror = SPSSMirror().load_csv("survey.csv")
result = mirror.regression().linear("score ~ age + C(group)")
print(result.r_squared, result.coefficients)

---

Table of contents

• Why SPSSMirror
• Installation
• What's included
• Quick start
• Design principles
• Testing
• Contributing
• License

---

Why SPSSMirror

Most Python statistics work means juggling scipy.stats, statsmodels, scikit-learn, and reading each library's own conventions for what a "result" looks like. SPSSMirror collapses that into one consistent API:

• One object per analysis. mirror.frequentist().t_test_independent(...) returns a StatTestResult with .statistic, .p_value, .effect_size, and .data_quality — every method across every engine follows the same shape.
• Honest about uncertainty. Regularized regression (Ridge/Lasso) reports std_error/p_value as None instead of fabricating classical inference that regularization invalidates. A mixed model fit with REML reports aic/bic as None rather than silently leaking statsmodels' NaN.
• Refuses to compute nonsense. Running an ANOVA or regression on a column that turns out to be constant raises a clear error instead of a false "p = 0.000016, significant!" result caused by floating-point noise in the underlying model fit.
• Every result tracks its own data quality — n_rows_original, n_nulls_dropped, max_missing_ratio — so you always know what was silently dropped before you trust a number.
• Formula syntax where it belongs. Regression, the ANOVA family, mixed models, and residual diagnostics accept R-like formulas via patsy: "y ~ x1 + C(group) * x2".

Installation

pip install spssmirror

This installs the core engine — descriptive statistics, regression, the full frequentist test suite (parametric and non-parametric), categorical analysis, correlations, psychometrics, effect sizes, power analysis, diagnostics, and mixed models — with a deliberately lean dependency list (pandas, numpy, scipy, statsmodels, pydantic, patsy, rapidfuzz, factor_analyzer).

Four engines depend on heavier, optional libraries and are installed as extras:

pip install spssmirror[bayesian]      # Bayesian t-test/regression (pymc, arviz)
pip install spssmirror[timeseries]    # ARIMA/GARCH forecasting (arch)
pip install spssmirror[survival]      # Kaplan-Meier / Cox PH (lifelines)
pip install spssmirror[multivariate]  # PCA / clustering / discriminant (scikit-learn)
pip install spssmirror[all]           # everything at once

The core install works with zero optional dependencies present — verified by installing the built wheel into a clean virtual environment as part of the test process.

What's included

Engine	Access	Methods
Descriptive	.descriptive()	summary, frequency_table, crosstab
Regression	.regression()	linear, logistic, poisson, glm, robust, ridge, lasso, elastic_net
Frequentist (parametric)	.frequentist()	t_test_one_sample, t_test_independent, t_test_paired, anova_oneway, anova_twoway, ancova, anova_repeated_measures, manova
Frequentist (non-parametric)	.nonparametric()	mann_whitney_u, wilcoxon_signed_rank, kruskal_wallis, friedman_test
Categorical	.categorical()	chi_square_independence, fishers_exact, mcnemar_test
Correlation	.correlations()	pearson, spearman, kendall_tau, point_biserial, partial, correlation_matrix
Psychometrics	.psychometrics()	cronbach_alpha, mcdonald_omega, split_half, kmo, bartlett_sphericity, item_analysis, efa
Effect sizes	.effect_sizes()	cohens_d, hedges_g, glass_delta, eta_squared, omega_squared, cramers_v, odds_ratio
Power analysis	.power()	power_ttest_independent, power_ttest_paired, power_ttest_one_sample, power_anova, power_correlation, power_chisquare, power_curve_ttest, power_curve_anova, power_curve_correlation
Diagnostics	.diagnostics()	normality_tests, homogeneity_of_variance, vif, residual_diagnostics, outliers
Mixed models	.mixed_models()	linear_mixed_model (random intercept/slope, ICC)
Bayesian (extra)	.bayesian()	bayesian_ttest, bayesian_proportion_test, bayesian_linear_regression
Time series (extra)	.timeseries()	arima, auto_arima, exponential_smoothing, garch, acf_pacf, stationarity_test
Survival (extra)	.survival()	kaplan_meier, logrank_test, cox_ph, parametric_survival
Multivariate (extra)	.multivariate()	pca, kmeans_clustering, hierarchical_clustering, linear_discriminant, quadratic_discriminant, canonical_correlation

Every method returns a frozen Pydantic model. Inspect fields directly, or call .model_dump() / .model_dump_json() to export.

Quick start

from spssmirror import SPSSMirror

mirror = SPSSMirror().load_csv("data.csv")

# Reliability
alpha = mirror.psychometrics().cronbach_alpha(["q1", "q2", "q3", "q4"])
print(alpha.statistic)

# Group comparison with effect size
t = mirror.frequentist().t_test_independent("score", "group", "A", "B")
print(t.statistic, t.p_value, t.effect_size)

# Regression — no statsmodels import needed anywhere in your code
reg = mirror.regression().linear("outcome ~ predictor1 + C(category)")
for coef in reg.coefficients:
    print(coef.term, coef.b, coef.p_value)

# Power analysis
power = mirror.power().power_ttest_independent(effect_size=0.5, alpha=0.05, power=0.80)
print(f"Need {power.n:.0f} participants per group")

Loading data:

SPSSMirror().load_csv("data.csv")
SPSSMirror().load_excel("data.xlsx")
SPSSMirror().load_dict({"col1": [...], "col2": [...]})
SPSSMirror().load_dataframe(existing_pandas_df)

Design principles

1. Nothing leaks. Public methods never return a raw scipy/statsmodels/scikit-learn/pymc/lifelines/arch object — only SPSSMirror's own typed models.
2. Honest statistics over convenient statistics. If a number can't be computed validly, the field is None, not a fabricated or silently wrong value.
3. Data quality is never hidden. Every result that drops rows (nulls, non-finite values) reports exactly how many and what fraction.
4. No visualization dependency. Results are plain, inspectable data — pair with whatever plotting library your project already uses.

Testing

git clone https://github.com/<your-username>/spssmirror.git
cd spssmirror
pip install -e ".[all,dev]"
pytest tests/ -v

The test suite checks every engine against engineered ground truth (known true effects and known coefficients, not just "does it run") — see tests/conftest.py for the fixtures.

Contributing

Issues and pull requests are welcome. Please include a test demonstrating the bug or feature — see tests/ for the existing pattern (each test targets one method against either a known analytical result or a clearly engineered scenario).

License

MIT — see LICENSE.