Plausible-value / weight-aware multiverse (specification-curve) analysis

Runs an lsastrat estimator across a grid of analytic specifications – the "forking paths" of a multiverse analysis – and collects one tracked quantity from each, so the robustness of a finding to defensible analytic choices can be read off a specification curve. Every specification inherits correct plausible-value pooling and replicate-weight inference from the underlying estimator.

Usage

lsa_multiverse(
  data,
  estimator,
  grid,
  fixed = list(),
  design = NULL,
  term = NULL,
  .progress = FALSE
)

Arguments

data: A data frame of student-level records.
estimator: The estimator to run: a function (e.g. social_gradient) or its name.
grid: A named list of forking paths. Each name is an argument of estimator; each element is a named list of the alternative values for that argument (the inner names label the choices). All combinations are crossed.
fixed: A named list of arguments held constant across specifications.
design: Optional lsa_design() whose fields are added to fixed.
term: Which output term (row of the estimator's coefficients) to track; defaults to the first term.
.progress: Logical; message each specification as it runs.

Value

An object of class "lsa_multiverse": a list with specs (one row per specification: a column per forking path plus estimate, ci_lo, ci_hi) and a summary. Plot it with plot.lsa_multiverse().

Examples

data(pisa_mini)
des <- lsa_design(weight = "W_FSTUWT", repweights = paste0("W_FSTURWT", 1:64))
mv <- lsa_multiverse(
  pisa_mini, estimator = social_gradient,
  grid = list(
    achievement = list(math = paste0("PV", 1:10, "MATH"),
                       read = paste0("PV", 1:10, "READ")),
    curvature = list(linear = FALSE, quadratic = TRUE)
  ),
  fixed = list(escs = "ESCS"), design = des, term = "slope")
mv
#> Multiverse: `social_gradient`, term = slope  (4 specifications)
#>   median estimate: 39.1   range: [36.8, 41.4]
#>   100% of specifications same-signed; 100% exclude 0 (95% CI)