NEWS

---

spaci 0.2.0

• Spatial-dependence-valid inference: vcov_hac() (Conley spatial HAC variance for recoverU()/recoverUplus() fits, with a data-driven bandwidth), boot_spatial() (spatial block bootstrap that re-runs the whole pipeline on resampled blocks), and rand_test() (conditional randomization test for the sharp null of no direct effect). The default i.i.d. standard error is anticonservative under spatial dependence — these give wider, better-calibrated intervals. recoverU/recoverU+ fits now carry their influence values.
• The "mle" Matérn recovery engine now fixes the smoothness at nu = 0.5 (exponential) by default, controlled by the new matern_nu argument of recoverU() / recoverUplus(). The free four-parameter fit is poorly identified on weak residual fields and could send the smoothness estimate to a numerical boundary; pass matern_nu = NULL for the previous behaviour. The "geoR" engine is unchanged.
• New bias_bound(): an estimated post-matching bias bound for an idaps() fit (covariate, exposure and variogram-based confounding terms), with a print() method.
• Matching estimators (idaps(), daps(), naive_ps()) gain a match_method argument: "greedy" (default, as before) or "optimal", a deterministic 1:1 assignment via clue::solve_LSAP() that needs no seed.
• boot_spatial() and rand_test() restore the caller's random-number state when seed is supplied, as the other estimators already do.

spaci 0.1.1 (2026-09-21)

• First CRAN release.
• simulate_spatial_causal() now defaults to n = 150 (previously 250).
• Examples run on smaller simulated data sets to keep check times short.

spaci 0.1.0

• First release. Implements the idaps() (distance-adjusted propensity score with interference) and recoverUplus() (doubly robust with recovered spatial confounder and neighbourhood exposure) estimators, together with the naive_ps(), daps() and recoverU() comparators.
• simulate_spatial_causal() generates data with joint spatial confounding and interference; spatial_ate() runs all estimators and collects the results.
• The recovered-confounder step uses a self-contained Matérn maximum-likelihood fit by default (matern_method = "mle"), with an optional geoR engine for faithful reproduction of the report.