← Back to all sparks
S

svrep

INFRA · APIS
Velocity0.0

Tools for Creating, Updating, and Analyzing Survey Replicate Weights

Steadily absorbing every replicate-weight method the survey literature has to offer

survey-statisticsreplicate-weightsvariance-estimationbootstrapapi-consolidation
Current state
svrep is a broad implementation of replication-based variance estimation for complex surveys, covering the generalized bootstrap, Fay's generalized replication, successive difference replication and the random-groups jackknife. The last two years added methods quickly: the Antal-Tille doubled half bootstrap, the Beaumont-Emond and kernel-based BOSB variance estimators, and optional torch-backed computation of replicate weights. Version 0.9.0 tightened the package boundary, promoting 'survey' from Imports to Depends and deleting the long-deprecated rescale_reps().
Where it's heading
The recent arc is consolidation rather than expansion. 0.9.0 removed a deprecated function, added accessors and refactored for memory, while 0.9.1 exists only to answer a Journal of Statistical Software review. That submission explains the shape of the changelog: the API is being frozen and documented rather than extended. Method coverage is broad enough that additions are now gap-fills for particular sampling designs.
Prediction
Expect the next releases to track the JSS review through to publication with documentation and output-structure tidying, and any new estimator to arrive only when an issue or reviewer asks for a specific design.

Recent moves

  1. 10mo ago

    Documentation and output tidying for JSS review

    A patch cut purely to answer the Journal of Statistical Software editorial team: a documentation note about which bootstrap-sizing function has a plot method, and a simplified output object from estimate_boot_reps_for_target_cv(). No statistical behaviour changes.

    View source ↗
  2. 11mo ago

    'survey' moves to Depends; deprecated rescale_reps() removed

    The clearest boundary-setting release in the package's history: 'survey' is attached whenever svrep loads, the deprecated rescale_reps() is finally gone, and get_rep_scale_coefs() and get_rep_type() give users a supported way to read design metadata. Fits the consolidation phase — the surface is being tidied, not widened.

    View source ↗
  3. 1y ago

    Antal-Tille doubled half bootstrap added

    Adds the doubled half bootstrap for single-stage designs, whose practical draw is that it produces integer replication factors. Another entry in the steady accumulation of published replication methods rather than a change of direction.

    View source ↗
  4. 1y ago

    Successive difference replication, kernel variance estimator, optional torch backend

    The densest release in the window: successive difference replication, the Breidt-Opsomer-Sanchez-Borrego kernel estimator for finely stratified samples, Beaumont-Emond, and an optional torch backend for building replicate weights. The torch option is the first sign that weight-matrix size, rather than method coverage, is the binding constraint for some users.

    View source ↗
  5. 2y ago

    Helpers to shuffle, subsample and pad replicate columns

    Adds shuffle_replicates(), subsample_replicates() and add_inactive_replicates() — plumbing for users assembling replicate sets independently across strata. Practical rather than statistical, and typical of how this package fills the gaps its users actually hit.

    View source ↗
  6. 2y ago

    Version bump for CRAN check failures

    A CRAN-driven bump with no user-facing change, alongside removal of unmarked UTF-8 strings and LaTeX formatting that broke manual builds on an older macOS release. Housekeeping between substantive versions.

    View source ↗