antedep 0.3.1

Documentation

• The examples of ci_inad(), test_homogeneity_inad(), and run_homogeneity_tests_inad() use the first four occasions of the bolus data and Poisson innovations, so that they run in seconds; test_homogeneity_inad() now shows how fits are reused across tests.
• The reference in test_homogeneity_inad() now cites the INAD manuscript as under review.
• No change in computations: every function returns the same results as in 0.3.0.

antedep 0.3.0

New features

• fit_inad() gains a constrain argument for time-invariance constraints ("alpha", "alpha1", "alpha2", "theta", "kappa", "innovation", "both", "all"). Constrained models are refitted by maximizing the observed-data log-likelihood jointly over every remaining free parameter, including negative binomial innovation sizes, and are finished with Newton steps using the exact score and observed information. n_params, AIC, and BIC count only the free parameters.
• ci_inad() now computes the observed information exactly by Louis’ identity for all free parameters jointly: models of order 0, 1, and 2, with or without block effects, with or without time-invariance constraints, and including the negative binomial innovation size. method = "hessian" gives the numerical observed-data Hessian (fixed absolute steps) as a check. Intervals for positive parameters are formed on the log scale by default (log_scale = TRUE); the full covariance matrix is returned as vcov. Block effects receive Wald intervals in tau; for unconstrained fits, profile-likelihood intervals are also returned in tau_profile.
• em_inad() is now an exact EM algorithm for all nine thinning-innovation combinations. The E-step evaluates the full conditional distribution of the latent thinned counts, so the expected complete-data log-likelihood is exact even for negative binomial innovations, whose log-likelihood is not linear in the latent counts; negative binomial innovation sizes are now estimated. With block effects the innovation parameters are updated by conditional maximization (ECM). The returned fit includes n_params, aic, and bic.
• test_stationarity_inad() and run_stationarity_tests_inad() support constraints on negative binomial innovation sizes ("kappa", "innovation", "all").
• Fit objects now inherit from a shared ad_fit parent class while retaining their family-specific classes (gau_fit, cat_fit, and inad_fit).
• Added fit_ad() as a unified fitting entry point for Gaussian, categorical, and INAD workflows.
• Added anova() methods for likelihood-ratio comparison of nested antedependence fits. For INAD fits of different orders no chi-square p-value is given, because the comparison is a boundary problem.
• New kkt_inad() evaluates the exact score of a fitted INAD model and checks the first-order (Karush-Kuhn-Tucker) conditions, reporting estimates on the boundary of the parameter space and on the negative binomial size cap. Interior stationarity requires a positive definite observed information, a Newton decrement and single-parameter quantities s_j^2 / I_jj no larger than the tolerance. A boundary estimate whose one-sided score points into the parameter space passes only if a line search on the observed-data log-likelihood along that parameter finds no increase above the tolerance.
• fit_inad() gains newton_polish = TRUE: complete-data fits are finished by Newton steps with the exact score and observed information, holding boundary estimates whose score points outward fixed. A coefficient exactly on a bound where its thinned count is degenerate, and whose one-sided score points inward, is first moved off the bound by a line search on the observed-data log-likelihood.

Changes in behavior

• test_order_inad() no longer reports an ordinary chi-square p-value, which is not valid when the added thinning coefficients are on the boundary under the null hypothesis. By default (reference = "none") it reports the likelihood ratio statistic and BIC values only; reference = "chibar" gives a working binomial-weight chi-bar-square p-value, with a warning for models with block effects, for which these weights have not been established. use_chibar is deprecated.

Bug fixes

• Direct maximization in fit_inad() (models with and without block effects, constrained fits, and missing-data fits) restricted every thinning coefficient to be below 1. Only binomial thinning requires this; Poisson and negative binomial thinning coefficients may now exceed 1, as in em_inad(), so that both algorithms maximize over the same parameter space. simulate_inad() likewise accepts negative binomial thinning coefficients of 1 or more.
• For models without block effects, fit_inad() eliminated the innovation mean through the sample moment equation and required the implied innovation mean to be positive, so maxima with the innovation mean on its boundary (zero) were missed. Each time point’s parameters are now optimized jointly from several starting values (including the nested lower-order fit), so that an order-2 fit is never worse than the order-1 fit. For order-2 Bell models, the reported Bell parameter at the second occasion was the innovation mean.
• The missing-data likelihood (na_action = "marginalize") restricted each thinned count to be no larger than the count being thinned, which is wrong for Poisson and negative binomial thinning; fits to incomplete data with these operators were therefore incorrect.
• A thinning coefficient whose predecessor counts are all zero is not identifiable; it is now set to zero by both fit_inad() and em_inad().
• The final Newton steps could stall when a coefficient was just above zero with its score pointing out of the parameter space (as after EM, which approaches boundary values slowly), when a parameter on its bound had a Newton step pointing outward, or far from the maximum, where the observed information need not be positive definite. Such coefficients are now moved onto or held at the bound, and damped (Levenberg-Marquardt) steps are used when the Newton step fails. Constrained fits record in settings$path whether they were completed by these Newton steps or needed the derivative-free search.
• At a thinning coefficient of exactly 0 (or binomial 1) the thinned count is degenerate, and the conditional mean of the complete-data score is 0/0. The score routines replaced the coefficient by 1e-10, which gives minus the predecessor count instead of the one-sided derivative, and was also wrong for coefficients between 0 and 1e-10. As a result kkt_inad() could accept a fit with a zero coefficient that a small increase would improve, and the Newton steps held such coefficients at zero. The exact one-sided derivative is now used. kkt_inad() also accepted a negative Newton decrement when the observed information was indefinite; it now requires positive definiteness.
• The negative binomial thinning pmf was evaluated with prob = 1/(1 + alpha), from which dnbinom() recovers alpha/(1 + alpha) by subtraction. For small alpha this loses about -log10(alpha) digits in the probabilities of positive thinned counts. At alpha = 1e-12, where EM leaves a coefficient at its lower limit, the score was off by about 0.4%; log-likelihoods were affected only by rounding. The pmf now uses the equivalent mean parameterization.
• Constrained fits in test_stationarity_inad() held negative binomial innovation sizes fixed at their unconstrained estimates instead of re-estimating them, giving incorrect likelihood ratio statistics and BIC values for negative binomial innovations.
• .chi_bar_pvalue() (used by test_order_inad(use_chibar = TRUE)) counted the point mass at zero as contributing 1 to the upper tail, inflating every chi-bar-square p-value by 2^-r.
• Parameter counts in the tables of test_order_inad(), bic_order_inad(), and test_stationarity_inad() now include one negative binomial innovation size per time point and the coefficient alpha_{2,1} of order-2 models, as in fit_inad(). For order 2, the "alpha1" constraint now includes alpha_{2,1} instead of fixing it at zero.
• em_inad() restricted the latent thinned counts to be no larger than the count being thinned, which is wrong for Poisson and negative binomial thinning, and capped their coefficients below 1.
• The Bell parameter is now obtained from the Bell mean to machine precision (previously uniroot() default tolerance), so Bell log-likelihoods with block effects are smooth in the parameters.
• Louis-based intervals in ci_inad() for order-2 models used only the lag-1 coefficient, and Bell intervals with block effects omitted the chain rule from the innovation mean to the Bell parameter.

Improvements

• Constrained fits start from pooled estimates and are completed by Newton steps; the derivative-free algorithm is used only when the first-order conditions are not met, which makes constrained fits much faster.
• Block-effect tau standard errors and confidence intervals are aligned in downstream tables (Wald in tau; profile likelihood in tau_profile).
• coef() now returns a named numeric vector by default for standard R compatibility; use type = "list" for the previous structured inspection format.
• confint() now returns a standard two-column matrix for all fit families; the structured ci_gau(), ci_cat(), and ci_inad() helpers remain available.
• Added broader vcov() support using confidence-interval standard errors where available, improving compatibility with tools such as lmtest::coeftest().
• The family-specific aic_*() and bic_*() helpers are retained as thin compatibility wrappers for existing code, but examples and new documentation prefer the standard AIC() and BIC() generics.

antedep 0.2.0

New features

• New S3 methods deviance(), confint(), and vcov() for gau_fit, cat_fit, and inad_fit objects, improving compatibility with standard R model-fitting workflows.
• New summary.partial_corr() method prints a lag-by-lag table of intervenor-adjusted partial correlations (mean absolute value, range, and number of significant pairs at each lag).

Improvements

• Matrix inversion in ci_inad(), fit_gau_em(), and cat_test_stats.R now uses chol2inv(chol(.)) instead of solve() for symmetric positive-definite matrices, improving numerical stability and efficiency.

antedep 0.1.0

New features

• fit_inad() gains a nb_inno_size_ub argument (default 50) that caps the upper bound of the negative-binomial innovation size parameter during optimization, improving numerical stability for near-Poisson data.
• test_order_gau() accepts order_null and order_alt as convenience aliases for p and the absolute alternative order; both are also returned in the result object.

Bug fixes

• ci_inad(): fixed a sign error in the observed Fisher information for the negative-binomial innovation size parameter; the Hessian term (r + u) / (r + λ)² was added instead of subtracted, producing confidence intervals that were too wide.
• ci_inad(): the numerical second derivative for nb_inno_size CIs now retries with progressively smaller step sizes (×0.1, ×0.01) before falling back to NA, avoiding spurious failures when the default step lands in a non-finite region.
• test_homogeneity_inad(): degrees of freedom for LRT tests involving innovation = "nbinom" are now computed from the actual number of NB size parameters in the fitted models rather than assuming a fixed count of 1. This corrects LRT statistics and p-values whenever nb_inno_size is fitted as a time-varying vector.
• ci_inad() tau profile CI: nb_inno_size (negative-binomial innovation dispersion) is now held fixed at its full-model MLE during profile refits, consistent with the constrained-fit paradigm used throughout the package. Previously it was re-optimised as a nuisance parameter, which could widen the interval to the point of crossing zero even when the LRT clearly rejects the null (Variant 1 vs Variant 2 fix).
• ci_inad() tau profile CI: the bracket search in .ci_tau_profile_inad no longer imposes an artificial upper cap (max(|tau_mle| + 1, 1)) on the search range. The maximum bracket iterations are increased from 20 to 50 and the initial step size is set to max(0.1, |tau_mle| * 0.2), preventing the search from stalling for large or near-zero MLEs.

Initial release

• Initial CRAN submission candidate for Gaussian AD, categorical AD, and INAD workflows.
• Added/expanded examples for key user-facing modeling functions (fit_*, em_*, simulate_*, logL_*).
• Refreshed missing-data notes and harmonized documentation links and metadata.
• logL_gau() default missing-data behavior is now na_action = "fail" (previously marginalization-first in earlier drafts). For missing inputs, pass na_action = "marginalize" or na_action = "complete" explicitly.
• Added packaged datasets labor_force_cat (categorical labor-force sequences) and race_100km (continuous 100km race split times).

mirror server hosted at Truenetwork, Russian Federation.