LoginSign Up
← Back to all sparks
M

multimput

ANALYTICS
Velocity0.0

Using Multiple Imputation to Address Missing Data

github.com/inbo/multimput ↗

Ecological imputation tooling whose recent work is mostly about not falling over.

multiple-imputationinlaecological-monitoringr-packagebayesian
◆Current state
multimput handles multiple imputation for ecological monitoring counts, wrapping INLA and glm-style models with aggregation helpers for the follow-up analysis. The mature capability arrived with hurdle models and broader zero-inflated distribution support; recent releases have focused on degenerate inputs — empty data, identical imputations, models that never finish. It shares INBO's checklist packaging machinery with its sibling packages.
◆Where it's heading
The direction is defensive hardening rather than new statistics. Each recent release names a specific way the pipeline failed in practice — an empty join, all-identical imputed values, a runaway model — and closes it. That is characteristic of a package used in production monitoring workflows where the input data cannot be assumed well-behaved.
◆Prediction
Expect further edge-case handling in model_impute() and aggregate_impute(); nothing here suggests new model families are queued.

◆Recent moves

  1. 1y ago

    Fixes single-covariate selection in hurdle_impute()

    A single bugfix in hurdle_impute() when selecting single covariates. Narrow follow-up to the covariate filtering added in the previous release.

    View source ↗
  2. 1y ago

    Timeout, covariate filtering and degenerate-input handling

    model_impute() short-circuits to a single model when all imputed values are identical, handles empty data, accepts a user-supplied covariate filter and gains a timeout. Four separate ways the imputation pipeline used to stall or waste work.

    View source ↗
  3. 2y ago

    Accepts model functions by name; handles empty joins

    aggregate_impute() copes with joins that produce an empty dataset, and model_fun can now be passed as a namespaced string such as INLA::inla. The string form removes a hard dependency edge for a package that cannot assume INLA is installed.

    View source ↗
  4. 3y ago

    Hurdle models and wider zero-inflated distribution support

    Adds hurdle_impute() for separate presence and count models, an extra argument to carry observations excluded from the model into aggregation, and INLA support for binomial and both zero-inflated Poisson and negative binomial types. The most substantive statistical release in the window.

    View source ↗
  5. 4y ago

    Vignette builds without INLA installed

    Makes the vignette run without INLA so the package builds on the INBO r-universe. Build infrastructure, invisible to users.

    View source ↗
  6. 4y ago

    Adopts INBO checklist packaging infrastructure

    Adopts the shared INBO checklist infrastructure and folds in an older change letting aggregate_impute() work on already-aggregated objects. Packaging alignment across the institute's R packages.

    View source ↗