Using SAS for non-linear mixed modelling – An overview
The NLMIXED procedure in SAS fits nonlinear mixed models, that is, models in which both fixed and random effects may enter nonlinearly. The conditional distribution of the data (given the random effects) can be specified with either a standard form (normal, binomial, Poisson) or a general distribution. The mixed models are fitted by maximizing an approximation to the likelihood integrated over the random effects. Different integral approximations are available, the principal ones being adaptive Gaussian quadrature and a first-order Taylor series approximation. A variety of alternative optimisation techniques are available to carry out the maximization; the default is a dual quasi-Newton algorithm.
This presentation will give a brief overview of the integral approximation methods, the function optimisation methods, finite difference approximations of derivatives (for gradient and hessian determinations), and the specification of the model, initial estimates, and associated random effects. A selection of the 73 options shall be considered, including the specification of the convergence criteria, and requesting additional output. Obtaining all modelling results can be achieved using the Output Delivery System (ODS). To improve performance, some comments on coding will be made. The advantages and limitations of this procedure will also be discussed. Time permitting, an example of using NLMIXED to perform adaptive trial design with dynamic patient allocation will be presented.