Developer API Reference

Programmatic usage of the Indexly inference engine.

The inference engine is modular and side-effect free.

Core entry point:

run_inference_engine(
    df,
    test: str,
    x: list[str] | None = None,
    y: str | None = None,
    group: str | None = None,
    interaction: list[str] | None = None,
    auto_route: bool = True,
    bootstrap: bool = False,
    correction: str | None = None,
    lag: int = 1,
    alpha: float = 0.05,
)

Returns:

InferenceResult

InferenceResult Model

Defined in models.py.

Fields:

Field Description
test_name Name of test
statistic Primary statistic
p_value P-value
effect_size Effect size (optional)
ci_low Lower CI
ci_high Upper CI
paradigm frequentist or bayesian
evidence Bayes factor when Bayesian
additional_table Optional DataFrame
metadata Structured context

Correlation Module

File: correlation.py

Functions:

pearson_corr(df, x, y, alpha=0.05)
spearman_corr(df, x, y)
lag_corr(df, x, y, lag=1)
correlation_matrix(df, columns, correction=None)

Pearson uses Fisher Z CI. correlation_matrix returns Pearson r values and p-values; correction supports bonferroni, holm, and bh.

T-Tests

File: ttest.py

run_ttest(df, y, group, auto_route=True, use_bootstrap=False)
run_paired_ttest(df, x1, x2, use_bootstrap=False)

run_ttest routes to Welch when variances are unequal and to Mann-Whitney when normality fails if auto_route=True. run_paired_ttest returns Cohen’s dz in effect_size and the paired mean-difference CI in ci_low / ci_high.

Bayesian:

run_bayesian_ttest(df, y, group, r=0.707, alpha=0.05)

Returns paradigm="bayesian" and evidence=BF10, where BF10 is evidence for the alternative over the null.

ANOVA

File: anova.py

run_anova(df, y, group, auto_route=True, correction=None)

Posthoc:

run_tukey(df, y, group)

With auto_route=True, ANOVA routes as follows:

  • failed group normality → Kruskal-Wallis
  • acceptable normality but unequal variance → Welch ANOVA
  • acceptable normality and equal variance → classical one-way ANOVA

Classical significant ANOVA runs Tukey HSD automatically for 3+ groups. Tukey already controls family-wise error; correction adds an explicit p_corrected column only when requested.

Regression

File: regression.py

run_ols(df, y, x, interaction_terms=None, auto_route=True, bootstrap_coefficients=False)

When residual diagnostics suggest non-normality or heteroscedasticity and auto_route=True, OLS reports HC3 robust covariance results. Coefficient CIs are recomputed from the final model.

Mixed Effects

File: mixed_effects.py

run_mixed_effects(df, y_col, group_col, x_cols=None, formula=None)

Pass either a formula or CLI-style y_col, x_cols, and group_col. Without a formula, the API builds y_col ~ x1 + x2 and uses random intercepts by group_col.

Multiple Comparison Correction

File: multiple_corrections.py

apply_correction(result, method="bonferroni")

Supported:

  • bonferroni
  • holm
  • bh (Benjamini–Hochberg)

Bootstrap

File: bootstrap.py

Reusable bootstrap utilities for:

  • Coefficients
  • Confidence intervals

Formatting & Export

formatter.py → Console rendering exporter.py → Markdown / PDF reports

Extension Pattern

To add a new test:

  1. Implement function returning InferenceResult
  2. Register in run_inference_engine
  3. Add CLI test choice
  4. Update documentation

The engine is designed for strict modular extension.