Explicit conversion of supplied tests and fitted models — as

Convert recognized finalized test objects or summarize an existing fitted model once, without refitting.

Usage

# S3 method for class 'htest'
as_inference(x, method_id = NULL, n = NA_real_, population,
  groups = NULL, event = NULL, n_excluded = NA_real_, ...)
# S3 method for class 'lm'
as_inference(x, conf.level = .95, population, ...)
# S3 method for class 'glm'
as_inference(x, conf.level = .95, population,
  interval = "wald", ...)

Arguments

x: A recognized htest, lm or glm object.
method_id: Optional identity, which must agree with the stored test. For an imported signed-rank test, a compatible explicit identity declares the otherwise unrecoverable one-sample or paired design.
n: Known analysis sample counts, or explicit NA when unavailable. Never reconstructed from degrees of freedom.
population: Required description of the supplied analysis population.
groups: Two ordered labels for independent two-sample location, odds-ratio or variance-ratio contrasts; optionally, labels of the two measurements for a declared paired signed-rank contrast.
event: Explicit event definition for binary/proportion and odds-ratio tests.
n_excluded: Known exclusion count, otherwise NA.
conf.level: Requested confidence level for summaries of supplied models.
interval: Fitted-glm conversion explicitly supports "wald". The named logistic procedure provides profile-likelihood slope intervals separately.
...: Reserved; unknown arguments are rejected.

Details

Wilcoxon intervals retain the native-reported confidence level. Older R versions can record a nominal or rounded level whose exact attainment differs; newer versions can return a fully unbounded conservative interval. The common interval-method label and metadata identify this convention. No level, endpoint, or captured native diagnostic is replaced.

Recognized htest objects include standard t, Wilcoxon, binomial, proportion, Fisher, correlation, variance and common omnibus tests. Stored estimates, intervals and p-values are preserved with method-specific estimands. Two group means become their difference; no odds ratio is invented for a general contingency table. Proportion tests across more than two groups retain an omnibus row without a fabricated scalar probability. Unknown htest methods are rejected.

A signed-rank htest has the same method label whether it came from a one-sample analysis or paired differences. Its default identity is therefore imported_signed_rank, with unknown design. Supplying method_id = "signed_rank_wilcoxon_one_sample" or method_id = "signed_rank_wilcoxon" declares the one-sample or paired design, respectively; metadata records that this design was declared by the caller. The converter never infers design from the printed data expression. Stored estimates, confidence intervals, null values and p-values are unchanged. For a paired result, supplied n counts analyzed pairs, not twice the pair count.

Linear models use classical model-based marginal t intervals. GLM conversion is restricted to binomial-logit, Poisson-log and Gaussian-identity models; it uses model-based covariance and explicitly labeled Wald or t intervals. It does not invoke generic profile dispatch or switch covariance backends. Binomial slopes may also be shown as odds ratios and Poisson slopes as rate ratios; intercepts are not ratios. Exponentiation overflow is a numerical failure, distinct from a legitimate unbounded sampling interval. Aliased coefficients remain unavailable.

Value

A common finalized inference_result. The supplied source object is unchanged.

Examples

fit <- stats::t.test(c(1, 3, 4, 8), c(2, 4, 5, 6, 9))
as_inference(fit, n = c(4, 5), population = "Synthetic independent groups",
  groups = c("x", "y"))
#> <inference> t_means_two_independent_welch
#>                           row_id                   estimand estimate p.reported
#>  t_means_two_independent_welch:1 population mean difference     -1.2  0.5449251
#>  status
#>      ok