0.21.0

Backwards-incompatible changes

1. NIMADS rewritten to be completely immutable #1103
   Classes gone: Annotation, Condition, Note. Annotations are now AnnotationSet/LabelBlock; Study/Analysis/Point/Image survive as read-only accessors.
   Mutation API gone: the setters (studyset.coordinates = df, .images =, .metadata =, .texts =, .annotations_df =), plus set_annotations_df, touch, materialize, is_materialized, is_execution_ready, from_table_cache, NotifyDict. The store is immutable; growth is copy-on-write via with_points, with_images, with_metadata, with_annotation, with_texts, keep_images.
   Studyset.get(dict) removed — the requirements/blocks mechanism (resolve(), coordinate_block(), image_block(), label_block()) replaces it. This is what Estimator.collect_inputs now uses.
   Signature changes: get_metadata, get_images, get_texts, get_annotations all dropped their analyses= argument; get_images gained policy=; get_studies_by_label/get_analyses_by_label/filter_annotations gained annotation=.
   New methods: from_parquet, to_parquet, harmonized(target), select_analyses, select_points, with_context, sample_sizes, analyses, len.
   Studyset.slice now raises ValueError on any id it can't resolve instead of silently returning a subset (#1113). Previously a typo, a foreign id, and "no data" were indistinguishable.
   Estimators now expose studyset and blocks_ alongside inputs_; _required_inputs moved from Estimator up to NiMAREBase.

2. Dataset is now formally deprecated for 1.0.0 (#1107)
   Constructing a Dataset or passing one to any algorithm emits a FutureWarning. Silence with:

warnings.filterwarnings("ignore", message=".*nimare.dataset.Dataset is deprecated")

Also: Dataset.get_studies_by_label documents its real default of label_threshold=0.001 (previously documented as 0.5).

3. CBMR rewritten around a formula interface, write like in R or statsmodels.

CBMREstimator and CBMRInference are removed; the class is now CBMR.
group_categories, moderators, global_moderators, voxelwise_moderators, moderator_effect and penalty are all replaced by a single formula: CBMR("~ s(diagnosis) + sample_size"). s() marks a term as spatially varying.
model=models.PoissonEstimator → distribution="poisson" | "negativebinomial" | "clusterednegativebinomial".
nimare.meta.models no longer exists.
New defaults: n_iter 2000→1000, tol 1e-9→1e-8, new incidence_threshold=0.001 that drops low-incidence voxels from the mask.
Mixed coordinate spaces are now rejected with a message pointing at Studyset.harmonized().
4. IBMA constructor changes — aggressive_mask default flipped True → False. Meta-analyses now run in liberal-mask bags by default, so voxel coverage (and results) change for anyone relying on the old default.
Stouffers(normalize_contrast_weights=...) removed — raises TypeError directing you to groupby.
PermutedOLS gained an explicit constructor (two_sided, use_sample_size, n_jobs, random_state=42).
The PyMARE-backed estimators now share a PyMARERegressionEstimator base.
An IBMA fewer than two analyses now raises instead of returning all-NaN maps (#1112). CBMA is unaffected.
Input images with no usable voxel are dropped (e.g. an all-zero upload) rather than silently emptying the mask or padding inputs["id"] (#1114).
p_to_z on an unsigned p map now warns that the result contributes positive evidence only (#1108).
5. Corrector kwargs are validated
Corrector.init now drops None-valued kwargs and transform validates the rest against the correction method that will receive them, raising TypeError for anything unaccepted (#1109). Correction methods take nlogp instead of p — correct_fdr_indep(p) → correct_fdr_indep(nlogp), which matters for anyone subclassing Corrector.

6. Discrete decoder correction default (#1122)
   BrainMapDecoder / brainmap_decode / NeurosynthDecoder: default changed "fdr_bh" → "bh". "fdr_bh" was never a recognised value, so it silently meant no correction; results will now differ. Unrecognised values raise ValueError instead of falling through to uncorrected.

7. Diagnostics voxel_thresh → target_threshold
   voxel_thresh is a deprecated alias that warns; passing both raises.

8. Dependency bumps
   pymare>=0.0.8 → >=0.0.12, matplotlib>=3.6.0 → >=3.8.0.

New features

IBMA dependence handling (#1090, #1110, #1115) — a groupby parameter on every IBMA estimator identifying images that share participants (defaults to study_id; False opts out). Adds weight_scheme ('rescale'/'individual'/'collapse') and rho, cluster-robust CR2 standard errors with Satterthwaite dof, per-pair null-correlation estimation, and graceful dropping of a group whose images exactly cancel.

small_sample_correction (#1094) — forwards PyMARE's Knapp–Hartung adjustment through the regression estimators.

Log-space p-values (#1099) — new public transforms.z_to_nlogp, nlogp_to_z, t_to_nlogp, chi2_to_nlogp and stats.nlogp_bonferroni, stats.nlogp_fdr. Correction, MKDA chi-square, ALE and the discrete decoders now carry statistics in log space, so p-values no longer bottom out at float32's 1e-45 or at np.spacing(1).

Spatial CBMR estimator + tutorial (#1080), closed-form observed information (#1121), per-experiment exposure() conditioning (#1123), foci stored on CBMRModel (#1126), plus CBMR benchmarks in benchmarks/benchmark_cbmr.py.

ImageTransformer gains d, g, g_var targets (#1092) — standardized effect sizes were previously reachable only inside FixedEffectsHedges.

PermutedOLS emits an uncorrected p map (#1091), so FDRCorrector and Bonferroni work on it.

MetaResult.save(..., with_inputs=False) and a new public results.DroppedInput (#1111) — drops ~29 MB of duplicated input arrays from a 36 MB pickle, raising a clear error if a later operation needs them.

Parquet studyset I/O promoted to the public surface: from_parquet, write_parquet, convert_neurostore_json_to_parquet.

Workflow parameter forwarding fixed — voxel_thresh/cluster_threshold/n_cores are now applied to already-instantiated estimators/correctors/diagnostics that left the parameter at its default, and filtered out for components that don't accept them. Previously they were silently ignored.

Better errors: fetch_* raises naming the query when nothing matches (#1085); Jackknife/FocusCounter fall back to the base z-stat image when handed a cluster-corrected target (#1087).

Fix: off-by-one indexing in the GCLDA log-likelihood, with LAPACK matrix inversion replaced by a closed-form 3×3 inverse (#1098).

What's Changed

• Add CBMR benchmarks by @yifan0330 in #1083
• [FIX] asv env by @jdkent in #1088
• [FIX] table generation with cluster corrected images in Jackknife and focuscounter by @jdkent in #1087
• [ENH] ibma image dependence by @jdkent in #1090
• [ENH] Report an uncorrected p map from PermutedOLS by @jdkent in #1091
• [ENH] Add d, g and g_var targets to ImageTransformer by @jdkent in #1092
• [ENH] Forward PyMARE's small_sample_correction by @jdkent in #1094
• [PERF] reduce test runtime by @jdkent in #1100
• [FIX] Keep assigned Studyset tables through projection rebuilds by @jdkent in #1102
• Correct off-by-one indexing in GCLDA log-likelihood calculation and replace LAPACK matrix inversion by @tsalo in #1098
• [ENH/FIX] switch to using lossless nlogp logp to convert statistical values ove… by @jdkent in #1099
• Document the label_threshold parameter in get_studies_by_label by @lobennett in #1084
• [ENH] columnar studyset by @jdkent in #1103
• Update pymare version to 0.0.12 in setup.cfg by @jdkent in #1106
• Dep/dataset removal 1.0.0 by @jdkent in #1107
• [FIX] warn unsigned p to z by @jdkent in #1108
• Let MetaResult.save leave the input images out of the file by @jdkent in #1111
• Drop an input image that has no usable voxel by @jdkent in #1114
• [FIX] Drop a dependence group that cancels instead of failing the run by @jdkent in #1110
• Raise on an id that Studyset.slice cannot resolve by @jdkent in #1113
• [FIX] Drop None-valued Corrector kwargs and validate the rest against the m… by @jdkent in #1109
• Estimate the null correlation per pair of images by @jdkent in #1115
• Refuse an image-based meta-analysis of fewer than two analyses by @jdkent in #1112
• Raise an informative error when a fetch query matches no files by @lobennett in #1085
• Add spatial CBMR estimator and tutorial by @yifan0330 in #1080
• Compute the CBMR observed information in closed form by @jdkent in #1121
• Fix discrete decoder correction defaults by @const7 in #1122
• Condition CBMR on a per-experiment exposure by @jdkent in #1123
• Store foci on CBMRModel by @yifan0330 in #1126
• Move development conventions into contributing guide by @yifan0330 in #1125
• Update CBMR tutorial formula guidance by @yifan0330 in #1124

New Contributors

• @lobennett made their first contribution in #1084
• @const7 made their first contribution in #1122

Full Changelog: 0.20.0...0.21.0