Changelog

Source: NEWS.md

sondage 0.8

Initial CRAN release.

Sampling dispatchers

Five dispatchers returning S3 design objects with class c("<prob_class>", "<replacement>", "sondage_sample"):

• equal_prob_wor(N, n, method=) - Simple random sampling, systematic, Bernoulli (random size).
• equal_prob_wr(N, n, method=) - Simple random sampling with replacement.
• unequal_prob_wor(pik, method=) - Conditional Poisson / maximum entropy (CPS), Brewer, systematic PPS, Poisson (random size), Sequential Poisson (SPS), Pareto.
• unequal_prob_wr(hits, method=) - Chromy’s minimum replacement, multinomial PPS.
• balanced_wor(pik, aux, strata, method=) - Cube method for balanced sampling (Deville & Tille, 2004), with optional stratification (Chauvet & Tille, 2006; Chauvet, 2009).

Design queries

Five generics for variance estimation quantities:

• inclusion_prob() - Compute from size measures or extract from a WOR design.
• expected_hits() - Compute from size measures or extract from a WR design.
• joint_inclusion_prob() - N x N matrix of joint inclusion probabilities (WOR). Exact for CPS, systematic, Poisson, SRS, Bernoulli; high-entropy approximation (Brewer & Donadio, 2003) for Brewer, SPS, Pareto, cube. sampled_only = TRUE returns only the n x n submatrix for sampled units.
• joint_expected_hits() - N x N matrix of pairwise expectations E(n_i n_j) (WR). Exact analytic for multinomial and SRS and simulation-based for Chromy (nsim parameter). sampled_only = TRUE returns only the submatrix for selected units.
• sampling_cov() - Sampling covariance matrix. weighted = TRUE returns Sen-Yates-Grundy check quantities. sampled_only = TRUE computes only the submatrix for sampled units.

Sampling methods (13 built-in)

Method	Dispatcher	Fixed size	PRN	Joint probs
srs	equal_prob_wor	yes	no	Exact
systematic	equal_prob_wor	yes	no	Exact (some = 0)
bernoulli	equal_prob_wor	no	yes	Exact (independent)
srs	equal_prob_wr	yes	no	Exact
cps	unequal_prob_wor	yes	no	Exact
brewer	unequal_prob_wor	yes	no	HE approximation
systematic	unequal_prob_wor	yes	no	Exact (some = 0)
poisson	unequal_prob_wor	no	yes	Exact (independent)
sps	unequal_prob_wor	yes	yes	HE approximation
pareto	unequal_prob_wor	yes	yes	HE approximation
chromy	unequal_prob_wr	yes	no	Simulated
multinomial	unequal_prob_wr	yes	no	Exact
cube	balanced_wor	yes	no	HE approximation

Custom method registration

• register_method() lets users plug custom unequal-probability sampling algorithms into the existing dispatchers and generics. Registered methods work with unequal_prob_wor(), unequal_prob_wr(), joint_inclusion_prob(), sampling_cov(), batch mode (nrep), and all downstream tooling.
• method_spec() queries method metadata (type, fixed size, PRN support) for any method, built-in or registered. Covers all five dispatchers including balanced sampling.
• Helper functions registered_methods(), is_registered_method(), and unregister_method() manage the registry.
• he_jip() and hajek_jip() are exported joint inclusion probability approximations that can be passed directly as joint_fn to register_method(). he_jip() uses the Brewer & Donadio (2003) high-entropy formula (C implementation); hajek_jip() uses the Hajek (1964) conditional Poisson approximation.
• New vignette “Extending sondage with Custom Methods” with worked examples (Sampford’s method with he_jip(), wrapping sampling::UPtille).

Features

• All functions return design objects usable with s$sample, s$pik, s$hits, etc.
• Batch sampling via nrep argument for Monte Carlo simulations. Fixed-size designs return a matrix; random-size designs return a list.
• Permanent random numbers (prn) for sample coordination (Bernoulli, Poisson, SPS, Pareto).
• C implementations for all sampling algorithms.
• CPS batch optimisation: the O(N^2) design matrix is computed once, then each replicate is a cheap sequential draw.
• Cube batch optimisation: C-level batch entry points reuse the workspace allocation across replicates.
• Cube auxiliary conditioning: condition_aux = TRUE pre-conditions aux by weighted centering/scaling and QR-pivot rank pruning.
• Stratified cube: within-stratum size constraints are preserved exactly when per-stratum sum(pik) is close to an integer.
• High-entropy joint inclusion probabilities handle certainty units correctly and clamp values to valid bounds.
• N-size guard: joint_inclusion_prob() and joint_expected_hits() refuse N > 10,000 to prevent accidental allocation of multi-GB dense matrices. sampled_only = TRUE bypasses this limit.
• All long-running C loops call R_CheckUserInterrupt() so Ctrl-C works.