Skip to contents

Changelog

Source: NEWS.md

INLAvaan (development version)

Bug fixes

  • timing() overstated the total: the lavaan-side setup stages were added on top of INLAvaan’s own segments even though they already sit inside init, double-counting the whole lavaan setup (and silently summing the four segment names both sides share). The reported segments are now INLAvaan’s own, disjoint, and sum to a total that matches system.time() on the call; the absolute start_time stamp is no longer exposed as if it were a duration, and the loo/waic segments are now documented.

  • loo() no longer drops units without a second-order term from elpd_loo, p_loo and their standard errors, which summed the model over fewer units and so flattered it. A second-order term is a Gaussian integral that need not converge: the log CPO term exists only where k_max < 1, the lpd term only where the complementary condition on the same curvature holds. Where a log CPO term does not exist, every estimate is now reported at first order over all units, so it remains an approximation of one order rather than a mix of two — substituting first-order terms for the failing units alone drops exactly the curvature that made the integral diverge, giving an error that is systematic, one-directional and concentrated on those units. A missing lpd term gets the opposite remedy, because the two divergences mean opposite things: the case-deletion integral can genuinely be infinite, but the lpd integral is always finite in truth (a density is bounded), so a missing lpd term is an artefact of extrapolating the quadratic rather than a feature of the unit. Its first-order contribution recovers most of the true one — against a sampled reference, roughly 90%, where dropping the unit recovers none — so such a unit now contributes its first-order difference to p_loo, leaving elpd_loo and looic untouched. The log CPO case warns, from loo() and from fitmeasures(); the lpd case is silent at the console, being both the ordinary state of an SEM fit and a smaller error than the second-order lpd’s own bias on the units that keep it, and is instead noted when a result is printed and counted by n_lpd_ok. This changes elpd_loo/looic for fits with any k_max >= 1. The warning names the offending units: since Sigma^-1 + H_u is the deleted posterior precision, k_max >= 1 says that removing the unit leaves the remaining data and the prior unable to identify some combination of parameters — a finding about that unit, worth inspecting rather than worked around.

  • compare() computed se_diff from pointwise contributions that did not match the elpd_diff above them: units without a second-order term were dropped from the paired variance while still counted in the ELPD totals. Both now rest on the same per-unit contributions.

  • compare(loo = TRUE) now scores every model at one common Taylor order, the lowest any of them can supply. Previously a model with all its second-order terms was compared at second order against a model reported at first order, so part of elpd_diff was a change of estimator rather than a difference between the models. The order used is printed with the table.

New features

  • waic() is now deterministic: both WAIC terms are computed in closed form from the fit’s Laplace summary — the same per-unit Taylor quantities behind loo() — instead of being estimated from posterior draws. The penalty is the variance of the unit log-likelihood under the Gaussian posterior, , a polynomial in Gaussian moments with no existence condition of its own; the lpd term is the one loo() already computes, and a unit without a second-order lpd contributes its first-order lpd (noted when printing, counted by n_lpd_ok). This is a change of estimand, not only of computation: the previous estimator took a variance across draws from the skew-normal copula sampler, so p_waic and waic change on every existing fit, most visibly at small N where the two posteriors’ second moments differ most. Consequences: results are exactly reproducible (no seed sensitivity); the nsamp argument is gone (passing it warns); and a second_order argument mirrors loo(), at which order WAIC and LOO coincide exactly. The p_waic > 0.4 reliability rule is removed rather than carried over: it is an empirical threshold calibrated on simulations (Vehtari, Gelman & Gabry, 2017, state plainly that no theory backs it) and it was there to police an estimator’s variability across draws. What replaces it is the estimand’s own existence condition — p_waic is a polynomial in the posterior moments and always finite, so the second-order WAIC exists exactly where its lpd term does, namely where Sigma^-1 - H_u is positive definite (k_min > -1); where it does not, every estimate is reported at first order over all units and warns, and that fallback is exact rather than approximate because first-order WAIC is the first-order LOO score. The log CPO condition k_max < 1 is irrelevant to the WAIC, which reads no case-deletion term. At fit time the WAIC is derived from the same computation as the fit-time LOO at no extra cost (previously it needed nsamp >= 100 and a separate casewise pass over all draws), and is skipped when that LOO is skipped. loo()’s per_unit gains the k_ssq column (, which feeds the penalty) and k_min, the existence diagnostic for the lpd term.

  • loo() now reports two curvature diagnostics per unit, k_max and k_sum, obtained in closed form from the Laplace summary rather than estimated from posterior draws. k_max is the share of the posterior precision the unit carries along its worst direction, and k_max < 1 is exactly the condition for its second-order term to exist; k_sum is the unit’s total leverage. No threshold is applied to either: existence is the only condition acted on.

Minor improvements and fixes

  • INLAvaan now requires lavaan >= 0.7-2. This retires the compatibility layer that supported both lavaan generations at once — the dual-name alias probing for renamed internals, the per-session resolution of renamed argument spellings, and the positional calls that avoided cross-version argument names — and with it the interim warning about lavaan’s slightly inexact two-level FIML gradient for fully-missing-within cases, which lavaan >= 0.7-1.2707 fixed upstream. The unexported lavaan internals INLAvaan relies on are still bound once per session at load time (now under their 0.7 names only), keeping the package correct when lavaan is upgraded in place.

  • The default "nlminb" optimiser now runs with iter.max = 1000 and eval.max = 2000 instead of nlminb()’s stock 150 and 200, which complex models could exhaust — quietly, since hitting the ceiling surfaces only through diagnostics() or the fit-time warning. Values supplied via control still take precedence.

INLAvaan 0.3.1

CRAN release: 2026-07-21

Bug fixes

  • The timing() function did not return the correct total time due to a breaking name change in lavaan.
  • Fixed CRAN errors and notes on certain linux builds relating to .Rd usage and
    convergence checks.

INLAvaan 0.3.0

CRAN release: 2026-07-11

New features

  • loo() computes leave-one-out cross-validation from a single fit without refitting nor sampling, via a Taylor approximation of the case-deletion posterior: per-subject (LOSO) for single-level models, per-cluster (LOCO) for two-level models. Reports first- and second-order estimates and pointwise contributions, with opt-in parallelism (cores) and theta/Sigma overrides for scoring conditioned posterior summaries in user-built model-search workflows.
  • waic() computes the widely applicable information criterion from posterior draws, with pointwise contributions and reliability warnings.
  • Both criteria score fits with exogenous covariates on the likelihood they were fitted with: jointly with the covariates (fixed.x = FALSE) or conditionally on them (fixed.x = TRUE, the lavaan default; exact, no additional approximation), for any covariate placement, including cluster-level and within-level covariates in two-level models. The two flavours are never comparable, as conditional comparisons may differ in their covariate sets, which enables covariate selection.
  • loo() and waic() support multigroup models. Groups are independent, so each unit is scored against its own group’s implied moments, under either mean treatment and either covariate flavour, with cross-group equality constraints (group.equal) flowing through automatically. Units are identified by case number and carry a group column, so results keep their identity across fits that stack groups differently. Multigroup two-level models are not supported yet.
  • loo() and waic() support fits estimated by full-information maximum likelihood (missing = "ml"). Single-level units are scored on the entries they actually have – the observed-data predictive log p(y_i,obs | D_-i) – with casewise kernels evaluated per missing pattern, so a unit with fewer observed entries self-weights in the elpd. Two-level fits are scored per cluster (LOCO), each cluster on its observed-data marginal likelihood via lavaan’s raw-data cluster kernels (no per-cluster sufficient statistics, since LOCO deletes whole clusters). This shares the missing-at-random assumption of the FIML fit itself. Multigroup two-level models remain unsupported under missingness.
  • On two-level models loo() and waic() gain type = "loso", scoring the conditional predictive (leave-one-unit-out: a new observation within an observed cluster) instead of the default marginal predictive (type = "loco", leave-one-cluster-out: a new cluster). These are the two estimands of Merkle, Furr & Rabe-Hesketh (2019); they answer different questions and are easily conflated, so the marginal is the default and the conditional warns. loo() uses the Taylor expansion and waic() the posterior draws, computing the same estimand two ways; both work with and without missing data. (waic() previously had no type.)
  • fitmeasures() gains elpd_loo, se_loo, p_loo, looic and elpd_waic, se_waic, p_waic, waic: included in "all" when stored with the fit, computed on demand when requested by name.
  • compare() gains loo = TRUE. Models sorted by descending ELPD, with p_loo and ELPD differences with paired standard errors (mixed-flavour comparisons are refused). Pairing matches units by id rather than row order, so a pooled fit can be compared against a multigroup fit of the same data, and the measurement-invariance ladder (configural, metric, scalar) is compared on a proper predictive scale.
  • Both criteria can be computed at fit time and stored with the fit. The default test = "standard" does so automatically for supported models with a mean structure. The WAIC reuses the fit’s own posterior draws (when nsamp >= 100), and the LOO runs when its predicted serial cost is within a 10-second budget. test = "loo" forces the LOO regardless of the budget, test = "none" skips everything, and fit <- add_loo(fit) stores it post hoc. Stored results are reused by loo(), waic(), fitmeasures(), and compare().
  • fitted() (and fitted.values()) return the model-implied moments of an INLAvaan fit, evaluated at the posterior means, matching the lavaan and blavaan output structure. type = "ov" gives casewise predicted values.
  • predict() gains a summary argument; summary = TRUE collapses the posterior draws and returns point estimates directly, equivalent to summary(predict(...)) in one call. Default FALSE, so existing code is unaffected.
  • residuals() (and resid()) return the observed-minus-fitted moments of an INLAvaan fit, matching the lavaan and blavaan output structure and supporting all lavaan residual types (raw, cor, cor.bentler, normalized, standardized) plus type = "casewise".
  • anova() on an INLAvaan fit now errors, pointing to compare(). Unlike fitted()/residuals()/predict(), this is a deliberate departure from blavaan (which silently inherits lavaan’s frequentist likelihood-ratio test): there is no direct Bayesian analogue of that test, and compare() already provides the appropriate tools (Bayes factors, DIC/pD, LOO/WAIC).
  • logLik() returns the Laplace-approximated marginal log-likelihood (log evidence) by default, printed with a note that it is not comparable to a classical log-likelihood; type = "plugin" instead returns the classical log-likelihood at the posterior mean, with df/nobs attributes so it supports AIC()/BIC() at the point estimate.
  • deviance() is new for INLAvaan fits (lavaan has no deviance() at all). Follows the BUGS/JAGS/Stan convention: type = "mean" (default) returns the posterior mean deviance with pD/DIC attached as attributes; type = "plugin" returns the deviance at the posterior mean (matching -2 * logLik(type = "plugin")). Both require test != "none".
  • AIC()/BIC() on an INLAvaan fit now error, documented alongside logLik(). Both are large-sample asymptotic approximations to quantities INLAvaan already computes directly – AIC approximates predictive accuracy (loo()/waic()), BIC approximates -2 * log(marginal likelihood) (logLik()) – so reporting them at the posterior mean would be a cruder proxy for numbers already available. The point estimate remains available for reporting-convention purposes via AIC(logLik(object, type = "plugin")) / BIC(...).
  • Fits now self-check their diagnostics: inlavaan() warns once, at the end of the fit, if the optimiser did not converge, the gradient at the reported mode is materially non-zero (Newton step > 0.1 posterior SD), a skew-normal marginal fits poorly (NMAD > 0.1), the VB correction shifted a posterior mean by more than 1 posterior SD, or the Hessian is near-singular – naming the offending parameters. A healthy fit stays silent; suppress via the inlavaan_diagnostics_warning condition class. diagnostics() gains the scale-free mode_shift_max (global) and mode_shift_sigma (per-parameter) measures backing the gradient check. (#18)

Minor improvements and fixes

  • Saturated-means fast path: when the mean structure is saturated (all intercepts free and unconstrained with normal priors, no nonzero latent means), the posterior is exactly block-diagonal between the intercepts and the covariance parameters at the mode. The Hessian intercept block is now computed analytically with an exact zero cross block (finite differences run over the covariance columns only), and the skew-normal marginal scans skip the intercept axes, emitting their exact Gaussian marginals directly. About 25% faster on typical CFA/SEM fits, with results identical to within finite-difference noise.
  • Improved messaging for inlavaan() fit calls.

Bug fixes

  • standardisedsolution() and summary(standardized = TRUE) no longer silently drop their arguments under lavaan >= 0.7-1, which renamed several exported arguments (e.g. cov.std to cov_std, GLIST to glist). INLAvaan now resolves the spelling the installed lavaan expects once per session at load, working across lavaan versions.
  • Two-level FIML loo()/waic() scores are now correct for clusters containing a case fully missing on the within-level variables. lavaan retains such cases but its analytic gradient kernel mishandles the zero-observed pattern; INLAvaan drops these rows before the cluster kernels (exact for the marginal likelihood). Two-level FIML fitting also inherits the upstream gradient issue (fixed in lavaan PR #581), so inlavaan() warns when such cases are present on lavaan versions before the fix.
  • Models fitted with meanstructure = FALSE now use a proper Bayesian likelihood. See “Mean structures” vignette for details, including when model comparisons across the two mean treatments are meaningful.
    • The saturated means are given flat priors and marginalised analytically (closed form), replacing lavaan’s profiled likelihood, which is not a valid Bayesian object.
    • Posterior modes recalibrate by the factor n/(n-1) on the covariance side.
    • loo() and waic() score such fits on the exact exchangeable case-deletion conditionals. The previous zero-mean fallback and its warning are gone, and absolute ELPD values are meaningful and comparable with meanstructure = TRUE fits.
    • Posterior predictive draws include the saturated means and their mean-uncertainty.
    • Requesting meanstructure = FALSE for a two-level model now warns and fits with meanstructure = TRUE (the mean structure is required there).
    • The conditional (fixed.x = TRUE) flavour — the default for SEM with exogenous covariates — is fully supported: the mean marginalisation factorises blockwise, so each unit is scored by the difference of two exchangeable conditionals, with the frozen-covariate term entering as an exact constant.
  • predict() now centres the conditioning data on the model-implied means (or the saturated sample means when the model has no mean structure) when drawing factor scores and predicted observed variables. Previously the kernels conditioned on raw data, offsetting every factor score by a constant that grows with the variable means.
  • sampling() and simulate() draws of observed variables from models without a mean structure now include the saturated (sample) means, so posterior predictive replicates live on the data scale instead of being centred at zero.
  • sampling() and simulate() no longer error for models with a single latent variable, and their saturated-mean recovery is now robust to missing data (replicate columns were previously NA under missing = "pairwise").
  • The PPP’s observed discrepancy now uses the unbiased (divisor n-1) sample covariance, matching the scale of the Wishart-replicated covariances it is compared against; previously the divisor-n form made the PPP very slightly optimistic (an O(1/n) effect, all models).
  • coef() (and the merged parameter table, fitted values, and implied moments) now reports covariance parameters on the covariance scale. Previously these slots carried the posterior-mean correlation, while summary() showed the correct sample-based covariance; the discrepancy is visible whenever the relevant standard deviations are far from 1.

INLAvaan 0.2.5

CRAN release: 2026-06-11

Minor improvements and fixes

  • INLAvaan now works with both the current lavaan 0.6 series and the upcoming lavaan 0.7, which renames many of its internal functions. The lavaan internals INLAvaan relies on are now resolved when the package loads, under whichever naming scheme is available. lavaan (>= 0.6-19) is now declared explicitly, and the package is checked against the oldest supported, current CRAN, and development versions of lavaan on CI.
  • Fixed the trapezoid rule used by compare_mcmc() for density normalisation, overlap, and KL divergence computations.
  • compare_mcmc() and diagnostics() are now robust to NA values in density and diagnostic computations.
  • The dp argument of inlavaan() and friends is now documented in terms of priors_for().

INLAvaan 0.2.4

CRAN release: 2026-04-03

New features

  • bfit_indices() computes per-sample Bayesian fit index vectors (BRMSEA, BCFI, BTLI, BNFI), with summary() and print() methods. Summary statistics are also available via fitmeasures().
  • compare() compares two or more fitted models side by side, reporting marginal log-likelihood, Bayes factors, and DIC, with optional fit measures from fitmeasures().
  • diagnostics() computes global and per-parameter convergence and approximation-quality diagnostics for fitted models.
  • get_inlavaan_internal() is now exported and documented, providing access to the internal list stored in a fitted INLAvaan object.
  • predict() generates predictions for observed data and missing data imputation, respecting multilevel structure if present.
  • sampling() draws from the posterior (or prior) SEM generative model, returning parameter vectors, latent variables, or observed variables.
  • simulate() generates complete replicate datasets from a fitted model, useful for simulation-based calibration and posterior predictive checks.
  • timing() extracts wall-clock timings for individual computation stages of a fitted model.

Minor improvements and fixes

  • Cholesky factorisation of the precision matrix replaces raw solve() for covariance and log-determinant calculations.
  • Copula sampling with NORTA (NORmal To Anything) correlation adjustment is now the default (samp_copula = TRUE), ensuring posterior samples have correct skew-normal marginals and correct Pearson correlations.
  • Pre-computed Owen-scrambled Sobol sequences are used by default, with fallback to {qrng} for larger sequences. QMC sample size now scales with model dimension.
  • Skew-normal fitting now runs in parallel automatically when the number of marginals exceeds 120, using all available cores.
  • Small optimisations to the skew-normal volume correction.
  • acfa(), asem(), and agrowth() gain a vb_correction argument.
  • ggplot2 is now optional; plots fall back to base R graphics when it is not installed.
  • inlavaan() gains an sn_fit_ngrid argument to control the number of grid points per dimension when fitting skew-normal marginals (default 21).
  • inlavaan() now supports sn_fit_sample = TRUE for defined parameters, fitting a skew-normal approximation to their posterior marginals based on drawn samples.
  • plot() method gains improved visualisation options.
  • priors_for() now supports the [prec] scale qualifier for variance parameters (theta, psi), placing the prior on the precision scale with automatic Jacobian adjustment.
  • sampling() and simulate() gain a silent argument to suppress informational messages.
  • summary() now includes 25th and 75th percentile columns.
  • vcov() now returns the covariance matrix of the lavaan-side parameters and supports a type argument for choosing between sample and Laplace covariance.

Bug fixes

  • marginal_correction = "shortcut" no longer produces incorrect volume corrections.
  • qsnorm_fast() no longer incorrectly handles sign symmetries.

INLAvaan 0.2.3

CRAN release: 2026-01-28

  • Improved axis scanning, skewness correction, and VB mean correction routine.
  • Bug fixes for CRAN.
  • Updated README example.

INLAvaan 0.2.2

CRAN release: 2026-01-27

  • Under the hood, use lavaan’s MVN log-likelihood function to compute single- and multi-level log-likelihoods.
  • Added support for multi-level SEM models.
  • Added support for binary data using PML estimator from lavaan. NOTE: Ordinal is possible in theory, but the package still lacks proper prior support for the thresholds.
  • Added support for missing = "ML" to handle FIML for missing data.

INLAvaan 0.2.1

  • Support for lavaan 0.6-21.
  • Implemented variational Bayes mean correction for posterior marginals.
  • Defined parameters are now available, e.g. mediation analysis.
  • Prepare for CRAN release.

INLAvaan 0.2

  • INLAvaan has been rewritten from the ground up specifically for SEM models. The new version does not call R-INLA directly, but instead uses the core approximation ideas to fit SEM models more efficiently.
  • Features are restricted to normal likelihoods only and continuous observations for now.
  • Support for most models that lavaan/blavaan can fit, including CFA, SEM, and growth curve models.
  • Support for multigroup analysis.
  • Added PPP and DIC model fit indices.
  • Added prior specification for all model parameters.
  • Added support for fixed values and parameter constraints.
  • Initial CRAN submission.

INLAvaan 0.1

  • Used rgeneric functionality of R-INLA to implement a basic SEM framework.