Generalized covariance models estimation

Selecting a mixed effects model means identifying a structural or mean model, the components of variance, and the covariance matrix for the residuals. The basic rationale for model selection will be parsimony in parameters, i.e., to obtain the most efficient estimation of fixed effects, one selects the covariance model that has the most parsimonious structure that fits the data (Wol-finger, 1996). Estimation of the fixed effects is dependent on the covariance matrix and statistical significance may change if a different covariance structure is used. The general strategy to be used follows the ideas presented in... 

The Poisson regression model is an example of the generalized linear model. The maximum likelihood estimates of the coefficients of the predictors can be found by iteratively reweighted least squares. This also finds the covariance matrix of the normal distribution that matches the curvature of the likelihood... 

In this chapter, we will examine the variance-covariance matrix to see how the location of experiments in factor space (i.e., the experimental design) affects the individual variances and covariances of the parameter estimates. Throughout this section we will be dealing with the specific two-parameter first-order model y, = Pq + + li only the resulting principles are entirely general, however, and can be... 

When the analysis is based on a continuous outcome there is commonly the choice of whether to use the raw outcome variable or the change from baseline as the primary endpoint. Whichever of these endpoints is chosen, the baseline value should be included as a covariate in the primary analysis. The use of change from baseline without adjusting for baseline does not generally constitute an appropriate covariate adjustment. Note that when the baseline is included as a covariate in the model, the estimated treatment effects are identical for both change from baseline and the raw outcome analysis. ... 

The methods of Chapter 6 are not appropriate for multiresponse investigations unless the responses have known relative precisions and independent, unbiased normal distributions of error. These restrictions come from the error model in Eq. (6.1-2). Single-response models were treated under these assumptions by Gauss (1809, 1823) and less completely by Legendre (1805), co-discoverer of the method of least squares. Aitken (1935) generalized weighted least squares to multiple responses with a specified error covariance matrix his method was extended to nonlinear parameter estimation by Bard and Lapidus (1968) and Bard (1974). However, least squares is not suitable for multiresponse problems unless information is given about the error covariance matrix we may consider such applications at another time. 

In practice, at first one may wish to try to obtain a separate estimate of the variance on each occasion, i.e., obtain estimates of coj, w2, etc. If the variance terms are approximately equal (in general if the ratio of the largest to smallest variance component is less than four, the variances are treated as equivalent) then one can assume that Wj = w2 =. .. (a , or that there is a common variance between occasions, and reestimate the model. If, however, there is a trend in the variances over time then one may wish to treat as a function of time. Alternatively, one may wish to examine whether IOV can be explained by any covariates in the data set. For most data sets, such complex IOV models cannot be supported by the data and these complications will not be explored any further. 

Suppose Y = f(x, 0, t ) + g(z, e) where nr] — (0, il), (0, ), x is the set of subject-specific covariates x, z, O is the variance-covariance matrix for the random effects in the model (t ), and X is the residual variance matrix. NONMEM (version 5 and higher) offers two general approaches towards parameter estimation with nonlinear mixed effects models first-order approximation (FO) and first-order conditional estimation (FOCE), with FOCE being more accurate and computationally difficult than FO. First-order (FO) approximation, which was the first algorithm derived to estimate parameters in a nonlinear mixed effects models, was originally developed by Sheiner and Beal (1980 1981 1983). FO-approximation expands the nonlinear mixed effects model as a first-order Taylor series approximation about t) = 0 and then estimates the model parameters based on the linear approximation to the nonlinear model. Consider the model... 

In summary, the Type I error rate from using the LRT to test for the inclusion of a covariate in a model was inflated when the data were heteroscedastic and an inappropriate estimation method was used. Type I error rates with FOCE-I were in general near nominal values under most conditions studied and suggest that in most cases FOCE-I should be the estimation method of choice. In contrast, Type I error rates with FO-approximation and FOCE were very dependent on and sensitive to many factors, including number of samples per subject, number of subjects, and how the residual error was defined. The combination of high residual variability with sparse sampling was a particularly disastrous combination using... 

---