statspai.mht

mht

Multiple Hypothesis Testing (MHT) module for StatsPAI.

Provides corrections for simultaneous inference across many outcomes or subgroups --- the single most common gap when Python users replicate empirical economics workflows that rely on Stata's rwolf or wyoung.

Estimators and utilities:

• Romano-Wolf stepdown (Romano & Wolf 2005, 2016) --- bootstrap FWER control that exploits dependence across test statistics.
• Westfall-Young maxT (Westfall & Young 1993) --- single-step resampling-based FWER control.
• Bonferroni, Holm (1979), Benjamini-Hochberg (1995) --- classical non-resampling adjustments included for comparison.
• adjust_pvalues() --- convenience dispatcher across all methods.

RomanoWolfResult dataclass

Bases: ResultProtocolMixin

Container for Romano-Wolf multiple hypothesis testing results.

Attributes:

Name	Type	Description
table	DataFrame	One row per outcome with columns: outcome, coef, se, t, p_value, p_rw, p_bonf, p_holm, p_bh.
n_outcomes	int
n_boot	int
n_obs	int

Examples:

>>> import statspai as sp
>>> df = sp.cps_wage()
>>> results = sp.romano_wolf(
...     data=df,
...     y=["log_wage", "tenure", "experience"],
...     x="union",
...     controls=["education"],
...     n_boot=200,
...     seed=42,
... )
>>> type(results).__name__
'RomanoWolfResult'
>>> results.n_outcomes
3
>>> bool("p_rw" in results.table.columns)  # stepdown-adjusted p-values
True

summary

summary() -> str

Pretty-print the results table.

plot

plot(figsize: tuple[float, float] = (8, 5)) -> tuple[Any, Any]

Dot-plot comparing unadjusted and adjusted p-values.

Returns:

Type	Description
fig, ax : matplotlib Figure and Axes

adjust_pvalues

adjust_pvalues(pvalues: PValueInput, method: str = 'holm') -> ndarray

Adjust p-values for multiple comparisons.

Parameters:

Name	Type	Description	Default
pvalues	array - like	Unadjusted p-values.	required
method	str	Adjustment method. One of: 'bonferroni' -- Bonferroni correction. 'holm' -- Holm (1979) step-down. 'bh' or 'fdr' -- Benjamini-Hochberg FDR. For Romano-Wolf or Westfall-Young adjustments (which require the original data and bootstrap), use :func:romano_wolf directly.	``'holm'``

Returns:

Type	Description
ndarray	Adjusted p-values.

Examples:

>>> import statspai as sp
>>> sp.adjust_pvalues([0.01, 0.04, 0.03, 0.20], method='holm')
array([0.04, 0.12, 0.09, 0.20])

bonferroni

bonferroni(pvalues: PValueInput) -> ndarray

Bonferroni correction: p_adj = min(p * S, 1).

Parameters:

Name	Type	Description	Default
pvalues	array - like	Unadjusted p-values.	required

Returns:

Type	Description
ndarray	Bonferroni-adjusted p-values.

Examples:

>>> import statspai as sp
>>> sp.bonferroni([0.001, 0.01, 0.03, 0.2]).round(3).tolist()
[0.004, 0.04, 0.12, 0.8]

holm

holm(pvalues: PValueInput) -> ndarray

Holm (1979) step-down correction.

Sort p-values ascending; for rank i (1-based): p_adj(i) = max(p_adj(i-1), min(p(i) * (S - i + 1), 1)).

Parameters:

Name	Type	Description	Default
pvalues	array - like	Unadjusted p-values.	required

Returns:

Type	Description
ndarray	Holm-adjusted p-values (in original order).

Examples:

>>> import statspai as sp
>>> sp.holm([0.001, 0.01, 0.03, 0.2]).round(3).tolist()
[0.004, 0.03, 0.06, 0.2]

benjamini_hochberg

benjamini_hochberg(pvalues: PValueInput) -> ndarray

Benjamini-Hochberg (1995) FDR correction.

Sort p-values ascending; for rank i (1-based): p_adj(i) = min(p(i) * S / i, 1), with reverse-cumulative-min to enforce monotonicity.

Parameters:

Name	Type	Description	Default
pvalues	array - like	Unadjusted p-values.	required

Returns:

Type	Description
ndarray	BH-adjusted p-values (in original order).

Examples:

>>> import statspai as sp
>>> sp.benjamini_hochberg([0.001, 0.01, 0.03, 0.2]).round(3).tolist()
[0.004, 0.02, 0.04, 0.2]