clusterIV

Estimation and fixed effects

• cjive(), cjar(), cjscore(), iv_infer(), and iv_compare() now accept high-dimensional fixed effects through fixed_effects. A base-R, matrix-free joint projection absorbs multiple factor dimensions without constructing a dummy matrix and verifies weighted orthogonality before estimation.
• Formula methods accept either y ~ x | z | fe or the controls-inside-formula layout y ~ exog | fe | x ~ z. Formula sections deliberately support a restricted additive grammar of bare variable names.
• Formula methods now support subset and na.action. A single complete-case filter is applied to every model component. The fitted-object interfaces (cjive(), cjar(), cjscore(), iv_infer()) store and report the number of omitted rows; iv_compare() returns a plain data frame and does not.
• inference = "t" uses a t(G - 1) reference convention for CJIVE Wald inference. It does not change the coefficient or standard error and is not presented as an exact finite-sample law.
• The dense Frisch–Waugh–Lovell path remains the default throughout the package. method = "leaveout_mean" remains an explicit special case for a pure grouping-instrument design and is never selected automatically.

Diagnostics and output

• Dense-path fits from cjive(), cjar(), cjscore(), and iv_infer() report the clustered Montiel Olea–Pflueger effective first-stage statistic (F_eff) and its simplified-TSLS critical value. CJAR and CJS also report their polynomial tail diagnostics (F_CJ and F_CJS) alongside the existing maximum within-cluster leverage diagnostic. iv_compare() intentionally reports estimates only, without these diagnostics.
• New summary() methods for CJAR and CJS report the test, confidence-set topology, first-stage diagnostics, and design advisories together. summary.cjive() includes the same strength block.
• nobs() methods are available for fitted objects. Supported tidy() and glance() methods are registered through generics when that suggested package is installed; disjoint and unbounded sets remain available in a list column rather than being forced into a single interval.
• iv_compare() is now an S3 generic with vector and formula methods. Its historical "JIVE" row label is retained for compatibility; the observation-level calculation after FWL is IJIVE.

Documentation

• Two precomputed vignettes work through the package on real, named datasets: vignette("queens-workflow") (Dube and Harish 2020; the full workflow, the published specification set, and the LaTeX export path) and vignette("miami-bail") (Frandsen, Leslie and McIntyre 2025; the judge-leniency design at scale). The data cannot be redistributed, so the vignettes ship with their outputs baked in and state where to obtain the files, verified by hash. knitr and rmarkdown enter Suggests for vignette building only.
• The README gained a Validation section summarising the external Stata cross-checks and the frozen dense oracles, and five FAQ entries (which function to call first, CJAR versus CJS, interpreting F_CJ/F_CJS, robust sets versus Wald intervals, and LaTeX export).

Performance

• The cross-fit variance kernel dispatches each leave-two-cluster-out solve on the stacked size of the two left-out clusters, mirroring the CJIVE leave-out kernel: many-instrument, small-cluster pairs are solved on the small side through a Woodbury identity, and the leave-out Gram matrices are cached lazily instead of always held. Together this reduces the per-pair time and the memory footprint by orders of magnitude in many-instrument, small-cluster designs; coefficients still pass the dense eq. (7) oracle at the 1e-10 gate, the exact spectral guard and its fail-closed classification are unchanged, and the large-side path is unchanged bit for bit.
• cjar() and cjscore() with variance = "crossfit" now compute only the variance polynomial they consume. iv_infer() still shares one pass when both jackknife tests are requested.
• Whitening and the first-stage hat-diagonal accumulation allocate far less memory, which matters at high-dimensional-fixed-effects scale.

Input and numerical safety

• Input validation is shared across the public functions. Outcomes, endogenous regressors, and weights must be numeric; weights must be finite and positive; absorbed variables, rank-deficient instrument designs, and numerically zero IV denominators produce explicit errors.
• Factor and character grouping instruments use reference coding whenever an intercept or fixed effects are partialled out, and full-level coding only when neither is present. Data-frame factor controls follow the same rule, avoiding a redundant dummy when fixed effects already span the intercept. Dense paths assess support from the encoded, residualized instrument matrix; the stricter outside-cluster support rule is applied only to the explicit leaveout_mean shortcut where that rule is mathematically required.
• Dense controls that are numerically absorbed by the joint fixed-effect span are removed before the second FWL step, rather than allowing pivoted QR to promote projection noise into a regressor. Surviving controls are normalized before a dimension-aware rank decision, making the result invariant to their units. The reported k_controls is now the non-redundant reference-coded nuisance-column upper bound rather than intercept plus every FE level.
• Numerically scaled inference preserves coefficient, standard-error, and test-statistic invariance under common rescaling and common rescaling of precision weights. Very large finite null values in CJAR/CJS and their plot methods are evaluated without forming overflowing powers of the null.
• Cross-fit variance estimation now rejects a leave-one- or leave-two-cluster- out instrument Gram whose spectral reciprocal condition number is at or below sqrt(.Machine$double.eps), instead of returning a result that can depend materially on the instrument basis.
• CJAR inversion now handles two-cluster designs and variance-boundary tangencies consistently with pointwise test acceptance. JIVE leverage equal to one and unsupported plain-test covariance ranks stop or return a typed unavailable result rather than propagating non-finite output.

Important changes

• confint() now defaults to the confidence level stored on the fitted object instead of silently reverting to 95%.
• Fast fixed effects make high-control designs practical, but the reported plug-in CJIVE sandwich can over-reject when the number of instruments and controls is large relative to the sample. The help pages now state this limitation and cite Kolesar, Min, Wang and Zhang (2026).
• The package title and description now cover both clustered IV estimation and weak-instrument-robust inference. The CRAN housekeeping corrections noted after version 0.1.0 are included.