Changelog

All notable changes to this project are documented here.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

Econometric correctness audit (independently verified against R's plm::pgmm/ plm::mtest on the Arellano-Bond (1991) EmplUK replication, plus targeted simulated DGPs). These change numeric output — not a patch release.

DifferenceGMM one-step Sargan/Hansen J-statistic was computed as N · q instead of q / σ_v̂² (the weight matrix used to obtain β1 was arbitrarily N-scaled, which is harmless for β1 and the one-step robust sandwich SEs but not for the J-statistic). EmplUK one-step Sargan: was ≈70.9, now ≈63.8-64.9 (paper: 65.8).
SystemGMM/AndersonHsiao one-step Sargan J-statistic likewise omitted the /σ̂² scaling of the classical Sargan statistic e'Z(Z'Z)⁻¹Z'e / σ̂² (EmplUK System GMM one-step: was ≈1.2 for 38 df, now ≈74).
DifferenceGMM one-step non-robust (homoskedastic) standard errors were too small by a factor of 2/N for the same reason. EmplUK: was 0.0169, now 0.1414 (independently verified formula match).
ar_test (Arellano-Bond AR(m) test) used an incorrect variance formula (summed squared per-observation products with no instrument/weight term) instead of Arellano & Bond (1991) eq. (8)'s per-individual cross moments plus the correct delta-method estimation-uncertainty term. Now matches R's plm::mtest to 3 decimals on EmplUK (one-step and two-step).
SystemGMM had no level-equation intercept, biasing estimates whenever the fixed effect and/or regressors have nonzero mean (the realistic case for levels data, e.g. logs) — confirmed via simulation: ρ̂=0.796 for true ρ=0.5 became ρ̂=0.497 after adding a "_cons" regressor/instrument to the level block.
time_effects=true used a hardcoded T-2 dummy count, which is only correct for lag-1-only formulas; a formula with a deeper lag (e.g. lag(y, 2), as in the AB(1991) replication) produced an unidentified (all-zero) dummy column and a spurious collinearity error. Now derives the dummy start from the deepest lag actually used.
**First-differencing subtracted adjacent rows, not calendar-adjacent *periods*** — an internal gap in an unbalanced panel (e.g. t=2 then t=4, missing t=3) was differenced as y(4) - y(2) instead of being skipped.
diff_data.n_groups counted groups with no retained observations (too-short groups, or groups fully dropped by lag/gap requirements), inflating the instrument-count/group ratio reported by diagnostics.

With these fixes the package reproduces Arellano & Bond (1991) Table 4 to the published precision on EmplUK (examples/example.jl, now with the paper's year dummies): two-step n(-1)=0.6287, n(-2)=-0.0652, w=-0.5258, k=0.2784, ys=0.5919, Hansen J=31.38 (paper 31.4); one-step n(-1)=0.6862 (0.1446), Sargan 65.82 (paper 65.8); AR(1)/AR(2) equal to R's plm::mtest to three decimals.

Added

AndersonHsiao now supports multiple lags of the dependent variable: each lag(y, k) regressor gets its own level instrument y_{t-k-1} (FOD: y_{t-k}), so y ~ lag(y) + lag(y, 2) + … is exactly identified instead of failing as under-identified. get_diff_data exposes the lag orders as y_lag_orders.
Versioned git hooks (.githooks/, enable via git config core.hooksPath .githooks): pre-commit formats staged Julia files and blocks accidentally staged gitignored artifacts (Manifest.toml, docs/build/, *.pdf, changelog.md); commit-msg rejects empty/wip/fixup messages on main; pre-push runs the full test suite.
Dedicated Aqua quality-checks CI job (previously only ran inside the main test job, not separately visible in CI status).
.github/dependabot.yml for GitHub Actions version updates (CompatHelper already covered Julia package deps).
PR template and issue templates (bug report / feature request).
CITATION.cff for machine-readable citation.
Makefile with short aliases (make test, make format, make docs, make example, make hooks) for the long julia --project=... commands.
CONTRIBUTING.md.
benchmarks/ — ad-hoc Chairmarks.jl timing script for the GMM solver on larger panels (N=500/5,000/50,000), not wired into CI.
test/test_replication.jl with a bundled copy of the EmplUK data (test/data/): asserts the Arellano-Bond (1991) Table 4 (a1)/(a2) coefficients, SEs, Sargan/Hansen and AR(1)/AR(2) statistics against the published values and against plm, without network access.

Changed

README/docs: the package is registered in General (Pkg.add("DynamicPanelModels")), not "unregistered"; added a Validation section.
CI: GitHub Actions bumped (actions/checkout@v7, julia-actions/setup-julia@v3, julia-actions/cache@v3, codecov/codecov-action@v6).

Each fix has a dedicated regression test (see test/) and the reasoning is recorded in the relevant docstrings (_one_step_variance, ar_test, get_diff_data, initial_weight_matrix).

[0.3.3] — 2026-08-12

Changed

jarque_bera_test now computes skewness/kurtosis via StatsBase.skewness/ StatsBase.kurtosis instead of hand-rolled moment formulas (identical result; StatsBase.kurtosis is excess kurtosis, so the - 3.0 term moved into the library call).
The three residual-standardization plot recipes (:residuals, :histogram, :qq) now use StatsBase.zscore instead of a repeated (x .- mean(x)) ./ std(x) expression.
StatsAPI.nobs(model; count_groups=true) removed — it duplicated the existing ngroups(model) accessor. Use ngroups(model) for the number of groups; nobs(model) now only returns total observations.

[0.3.2] — 2026-07-15

Added

Regression tests for three previously-uncovered defensive branches, found via line-coverage inspection rather than code reading alone: build_instruments's abstract-model fallback (must raise an informative error, not a generic MethodError), ar_test's degenerate total_variance <= 0 guard (all-zero residuals), and wald_test's near-singular R*V*R' regularization branch.

Fixed

estimate now rejects collinear regressors (rank(X) < n_regressors) with a clear error naming the affected columns, instead of failing deep inside the GMM solver with an opaque LinearAlgebra.SingularException. Found via adversarial testing (e.g. a regressor and an exact rescaling of it), not previously covered by any test.
examples/example.jl's diagnostics comment mischaracterized the AR(1)/AR(2) results: AR(1) is expected to be significant after differencing (not a misspecification signal, per the Getting Started guide); only AR(2) significance would indicate misspecification. Reworded for precision after running the example against live data and checking the actual output.

Performance

calculate_clustered_weight_matrix and _ab_h_weight_matrix's per-cluster accumulation loops used non-mutating A += ..., reallocating the full n_instruments × n_instruments matrix on every individual/cluster instead of accumulating in place; changed to A .+= ... (identical result, no behavior change). Benchmarked on simulated panels: roughly 20-28% less memory allocated by fit at N=20,000-50,000 (e.g. two-step DifferenceGMM at N=50,000, T=10: ~4.08 GB → ~2.97 GB), confirming the README's large-N performance claim under an actual adversarial/stress-test pass rather than just reading the sparse-matrix code.

[0.3.1] — 2026-07-14

Added

lag documented in docs/src/api.md (it was exported but missing from the API reference).
Behavioral regression tests: collapse/max_lags are now verified to actually change fit's returned coefficients/vcov (not just the intermediate instrument-matrix column count), and SystemGMM(windmeijer=false) is now verified to skip the correction the same way the existing DifferenceGMM case does.

Changed

DifferenceGMM/SystemGMM are now generated from a shared @dynpanel_spec macro to remove duplicated boilerplate between the two model spec structs; fields and behavior are unchanged.

Fixed

DifferenceGMM's one-step weighting matrix used plain (Z'Z)^-1 instead of the Arellano-Bond (1991) (Z'HZ)^-1 (H reflecting the MA(1) structure of differenced homoskedastic errors). Coefficients were only mildly affected, but the one-step Sargan/Hansen J-statistic was biased downward by roughly a factor of the number of groups. Validated against the EmplUK dataset from the original paper (Table 4): one-step Sargan chi2(25) now ≈70.9 vs. the paper's 65.8, versus ≈0.5 before the fix.
The new DifferenceGMM weighting-matrix code re-densified the sparse instrument matrix once per individual instead of once upfront, and called Matrix() twice on the same slice; fixed to match the existing densify-once pattern already used elsewhere in the solver. The cluster-boundary-walking loop this needed was also duplicated three times across the file; extracted into a shared, memoized _cluster_ranges helper (results threaded via a ranges keyword so callers don't recompute them).
calculate_clustered_weight_matrix/calculate_windmeijer_correction now guard that an explicitly passed ranges keyword argument is actually consumed, instead of silently ignoring it.
Added a genuine (non-tautological) Windmeijer correction regression test against a hand-computed expected value, replacing a check that only verified internal wiring.
_lag_vector's (id, time) => row index map was rebuilt from scratch for every regressor column in ar_test (once per column, all with Any-typed Dict keys causing boxing on every insert/lookup); it is now built once per ar_test call, concretely typed on eltype(id)/eltype(time), and reused across all columns.
_solve_gmm densified the instrument matrix Z unconditionally, even on the one-step, non-robust path where the densified copy is never used (calculate_clustered_weight_matrix is only called when steps == 2 or robust); it is now only densified when actually needed.

[0.3.0] — 2026-07-08

Added

Difference GMM (Arellano-Bond, 1991) and System GMM (Blundell-Bond, 1998) estimators, plus Anderson-Hsiao (1981) as a baseline IV estimator; one-step and two-step estimation with cluster-robust and Windmeijer (2005) finite-sample corrected standard errors.
Forward orthogonal deviations (transform=:fod, Arellano-Bover) for Difference GMM, as an alternative to first differencing (better for unbalanced panels).
exog keyword: strictly exogenous regressors are used as their own ("IV-style") GMM instruments (Roodman, 2009). A lag of a strictly exogenous covariate may itself be exogenous; the dependent variable and its lags may not.
time_effects keyword: adds T-2 exogenous period dummies to absorb common time shocks.
xtabond2-style instrument lag limits (min_lag, max_lag) and automatic dropping of linearly dependent instrument columns (drop_collinear, on by default).
Richer formula syntax: whitelisted transforms (log, log10, log2, exp, sqrt, abs) on either side and nested in lags (log(y) ~ lag(log(y)) + log(x)); explicit and ranged lags (lag(x, 2), lag(x, 0:1)).
StatsModels @formula support (formula = @formula(y ~ lag(y) + x)) in addition to the string form, plus the lag formula marker.
Tables.jl input: fit/get_diff_data accept any Tables-compatible source, not only DataFrame.
Diagnostics: Sargan/Hansen J-test, Arellano-Bond AR(1)/AR(2) tests (with cached results, so ar_test(model, order) works after diagnose), Wald test, Jarque-Bera test, pseudo-R², instrument-proliferation check, and diff_hansen_test (a Durbin-Wu-Hausman-style difference-in-Hansen test for a nested instrument subset).
StatsAPI RegressionModel interface (coef, vcov, stderror, confint, …), Stata-like result printing, and diagnostic plot recipes.
Aqua.jl quality tests; ~99% line coverage.
Package infrastructure: GitHub Actions CI (+ CompatHelper, TagBot), Documenter.jl docs, .JuliaFormatter.toml, CHANGELOG.md, CONTRIBUTING.md, and integration tests against a simulated panel with known parameters.

[Unreleased]: https://github.com/MichalS16/DynamicPanelModels.jl/compare/v0.4.0...HEAD [0.4.0]: https://github.com/MichalS16/DynamicPanelModels.jl/compare/v0.3.3...v0.4.0 [0.3.3]: https://github.com/MichalS16/DynamicPanelModels.jl/compare/v0.3.2...v0.3.3 [0.3.2]: https://github.com/MichalS16/DynamicPanelModels.jl/compare/v0.3.1...v0.3.2 [0.3.1]: https://github.com/MichalS16/DynamicPanelModels.jl/compare/v0.3.0...v0.3.1 [0.3.0]: https://github.com/MichalS16/DynamicPanelModels.jl/releases/tag/v0.3.0