MEMWAS

Mixed-Effects Models with Autocorrelation Structures

Fits longitudinal generalized mixed-effects models through the 'MEMWAS' interface and a registered 'C++' numerical backend. Supported serial covariance structures include first-order autoregressive (AR(1)), exponential or Ornstein-Uhlenbeck, higher-order autoregressive (AR(p)), first-order autoregressive moving-average (ARMA(1,1)), compound symmetry, Toeplitz, and unstructured covariance. Serial processes can be unified or attached independently to numeric predictor loadings. Candidate temporal structures can be ranked on a common sample by primary-cluster grouped cross-validation, the Akaike information criterion, the Bayesian information criterion, or log-likelihood. Clustered, crossed, and nested random intercepts and slopes are assembled jointly with diagonal or term-specific unstructured covariance. Available approximation methods include Laplace, saddlepoint likelihood with latent Laplace integration, adaptive Gaussian quadrature, full-covariance Gaussian variational inference, and penalized quasi-likelihood. Subject-grouped tuning requires every validation fold to succeed and supports fold-local nonlinear screening, bootstrap inference, prediction inference, and effective degrees of freedom for penalized information criteria. The mixed-effects framework follows Laird and Ware (1982) <doi:10.2307/2529876>; generalized-model approximations follow Breslow and Clayton (1993) <doi:10.1080/01621459.1993.10594284>; and serial covariance formulations follow Pinheiro and Bates (2000) <doi:10.1007/b98882>. The run-time fitting interface imports no third-party 'R' packages.

Versions across snapshots

Dependencies (latest)