Skip to contents

Changelog

Source: NEWS.md

lsastrat 0.3.0

General estimation, a shared design object, and tidyverse integration.

  • lsa_model(): fit any model (lm, glm, multilevel via lme4::lmer, …) to plausible-value outcomes, with replicate-weight variance and Rubin pooling. Lifts the package beyond the built-in estimators and beyond the OLS-only support of comparable tools.
  • All estimators now accept a design = argument – an [lsa_design()] object bundling weight, repweights, rep_method and fay – so those arguments need not be repeated on every call.
  • lsa_trend(): compare a quantity across assessment cycles and report the change from baseline with the linking (equating) error folded into the standard error.
  • tidy() and glance() methods (via the generics package) for broom/tidyverse integration.
  • Performance: social_gradient(), ieop() and ieop_decompose() now reuse a precomputed model matrix across plausible values and replicate weights (lm.wfit), giving identical results several times faster on large data (the Shapley decomposition is ~3x faster).
  • Parallel processing: set options(lsastrat.cores = n) to run the replicate-weight refits across n cores (Unix/macOS; sequential fallback on Windows). Results are identical regardless of the core count.

lsastrat 0.2.0

Comparative analysis, decompositions, a multiverse engine, and ggplot2 support.

Comparative analysis

  • lsa_by(): run any estimator across the levels of a grouping variable (e.g. countries) and stack the pooled coefficients into one tidy, comparable table.
  • equity_scatter(): the OECD-style quality-vs-equity scatter (mean achievement against gradient strength) across groups.

New estimators

  • variance_decomposition(): between- vs within-school variance decomposition and the intraclass correlation (between-school share).
  • ieop_decompose(): Shapley decomposition attributing the inequality of educational opportunity to each circumstance (contributions sum to the total).
  • oaxaca_gap(): twofold and threefold Oaxaca-Blinder decomposition of an achievement gap into explained (composition) and unexplained (structure).

Multiverse

  • lsa_multiverse() and lsa_design(): a plausible-value / replicate-weight aware specification-curve engine – run an estimator across a grid of analytic forking paths and summarise the robustness of a finding, with a specification-curve plot.

Visualisation

  • plot() methods for lsastrat_by, equity_scatter and lsa_multiverse.
  • autoplot() methods (ggplot2, Suggests) for estimates, the gradient, by-group forests, the equity scatter and the specification curve.

lsastrat 0.1.0

First public release. lsastrat computes educational stratification and inequality quantities from large-scale assessment data (PISA, TIMSS, PIRLS) as first-class outputs, with plausible-value pooling and replicate-weight variance.

Estimators

  • social_gradient(): the socio-economic achievement gradient — slope, strength (R-squared x 100) and curvature, following the OECD definition. Cross-validated against intsvy::pisa.reg.pv().
  • segregation_index(): school segregation indices (Dissimilarity D, Gorard’s S, isolation, Hutchens square-root, Theil’s H, mutual information M) for social or achievement-based (plausible-value pooled) grouping, with a choice of replicate-weight or two-stage clustered-bootstrap variance and finite-sample bias correction (Elbers, 2023). D and H cross-validated against the segregation package.
  • ieop(): the Ferreira-Gignoux decomposition of inequality of educational opportunity (relative IOp = R-squared, with a degrees-of-freedom adjusted alternative and an overfitting warning, plus absolute IOp and total variance).

Variance engine

  • pool_pv() and rep_factor(): the shared Rubin’s-rules pooling and replicate-weight machinery, exported for reuse. Inference uses a Rubin t reference, not the normal, which correctly widens intervals for the small number of plausible values typical in LSA.

Infrastructure

  • Shared S3 class lsastrat_estimate with coef(), confint(), as.data.frame(), print()/summary(), plus plot() methods for the gradient and segregation profile.
  • Bundled simulated data set pisa_mini for examples, tests and the vignette.
  • Note: only the segregation indices receive finite-sample bias correction; the gradient slope/strength and IOp are not subject to the same small-cell upward bias.