smmargins.MarginsResult

class smmargins.MarginsResult(estimate: ndarray, vcov: ndarray, labels: Sequence[str] | None = None, level: float = 0.95, df: int | None = None, stat_name: str = 'margin')

Container for margin estimates with delta-method standard errors.

Holds the point estimate, standard error, confidence interval, p-value, and the full delta-method covariance matrix of a vector of margin statistics (e.g. adjusted predictions or marginal effects).

For a fitted model with parameter vector \(\hat\beta\), estimated covariance \(\widehat V(\hat\beta)\), and a (possibly vector-valued) statistic \(g(\beta)\), the delta method gives

\[\widehat{\mathrm{Var}}[g(\hat\beta)] \approx G \, \widehat V \, G^\top,\]

where \(G = \partial g / \partial \beta|_{\hat\beta}\).

Parameters:

• estimate (ndarray) – Point estimates.

• vcov (ndarray) – Full delta-method covariance of the estimates (useful for joint tests).

• labels (sequence of str, optional) – Row labels. Defaults to ["m1", "m2", ...].

• level (float) – Confidence level for intervals (default 0.95).

• df (int or None) – Residual degrees of freedom; if set, uses t-distribution for p-values and CIs. Otherwise uses N(0,1).

• stat_name (str) – Column name for the statistic in summary().

estimate

Type:

ndarray

se

Type:

ndarray

labels

Type:

list[str]

vcov

Full delta-method covariance of the estimates (useful for joint tests).

Type:

ndarray

level

Type:

float

df

Type:

int or None

ci_lower

Type:

ndarray

ci_upper

Type:

ndarray

pvalue

Type:

ndarray

zstat

Type:

ndarray

_test_name

"z" or "t" depending on whether the normal or t-distribution is used.

Type:

str

Examples

Most users obtain a MarginsResult by calling Margins.predict or Margins.dydx rather than constructing one directly. Once you have a result, the typical workflow is to inspect .summary() or pull fields off it:

res = M.dydx(["x1", "x2"])
res.summary()                # tidy table of estimates / SE / CI
res.estimate                 # ndarray of point estimates
res.vcov                     # joint delta-method covariance

Forming linear contrasts of the estimates uses the joint covariance that is already on the result, so no additional differentiation is required. For example, the difference between the AMEs of x1 and x2:

res.contrast([1.0, -1.0], labels=["x1 - x2"])

Constructing one directly (mostly useful in tests):

>>> import numpy as np
>>> from smmargins import MarginsResult
>>> est = np.array([1.0, 2.0])
>>> vcov = np.array([[0.25, 0.10],
...                  [0.10, 0.16]])
>>> res = MarginsResult(est, vcov, labels=["m1", "m2"])
>>> res.se.round(2).tolist()
[0.5, 0.4]
>>> res.contrast([1.0, -1.0], labels=["m1 - m2"]).estimate.tolist()
[-1.0]

See also

Margins.predict, Margins.dydx, Margins.did

__init__(estimate: ndarray, vcov: ndarray, labels: Sequence[str] | None = None, level: float = 0.95, df: int | None = None, stat_name: str = 'margin')

Methods

__init__(estimate, vcov[, labels, level, ...])
contrast(c[, labels, name])	Form linear contrasts of the estimates, with delta-method SEs.
summary()	Return a summary DataFrame with estimates, SEs, and confidence intervals.