The NLMIXED Procedure

(View the complete code for this example.)

A popular application of nonlinear mixed models is in the field of pharmacokinetics, which studies how a drug disperses through a living individual. This example considers the theophylline data from Pinheiro and Bates (1995). Serum concentrations of the drug theophylline are measured in 12 subjects over a 25-hour period after oral administration. The data are as follows.

data theoph;
   input subject time conc dose wt;
   datalines;
 1  0.00  0.74 4.02 79.6
 1  0.25  2.84 4.02 79.6
 1  0.57  6.57 4.02 79.6
 1  1.12 10.50 4.02 79.6
 1  2.02  9.66 4.02 79.6
 1  3.82  8.58 4.02 79.6
 1  5.10  8.36 4.02 79.6
 1  7.03  7.47 4.02 79.6

   ... more lines ...   

12 24.15  1.17 5.30 60.5
;

Pinheiro and Bates (1995) consider the following first-order compartment model for these data:

where is the observed concentration of the ith subject at time t, D is the dose of theophylline, is the elimination rate constant for subject i, is the absorption rate constant for subject i, is the clearance for subject i, and are normal errors. To allow for random variability between subjects, they assume

where the s denote fixed-effects parameters and the s denote random-effects parameters with an unknown covariance matrix.

The PROC NLMIXED statements to fit this model are as follows:

proc nlmixed data=theoph;
   parms beta1=-3.22 beta2=0.47 beta3=-2.45
         s2b1 =0.03  cb12 =0    s2b2 =0.4 s2=0.5;
   cl   = exp(beta1 + b1);
   ka   = exp(beta2 + b2);
   ke   = exp(beta3);
   pred = dose*ke*ka*(exp(-ke*time)-exp(-ka*time))/cl/(ka-ke);
   model conc ~ normal(pred,s2);
   random b1 b2 ~ normal([0,0],[s2b1,cb12,s2b2]) subject=subject;
run;

The PARMS statement specifies starting values for the three s and four variance-covariance parameters. The clearance and rate constants are defined using SAS programming statements, and the conditional model for the data is defined to be normal with mean pred and variance s2. The two random effects are b1 and b2, and their joint distribution is defined in the RANDOM statement. Brackets are used in defining their mean vector (two zeros) and the lower triangle of their variance-covariance matrix (a general matrix). The SUBJECT= variable is subject.

The results from this analysis are as follows.

The "Specifications" table lists the setup of the model (Figure 18). The "Dimensions" table indicates that there are 132 observations, 12 subjects, and 7 parameters. PROC NLMIXED selects 5 quadrature points for each random effect, producing a total grid of 25 points over which quadrature is performed (Figure 19).

The "Parameters" table lists the 7 parameters, their starting values, and the initial evaluation of the negative log likelihood using adaptive Gaussian quadrature (Figure 20). The "Iteration History" table indicates that 10 steps are required for the dual quasi-Newton algorithm to achieve convergence.

The "Fit Statistics" table lists the final optimized values of the log-likelihood function and information criteria in the "smaller is better" form (Figure 21).

The "Parameter Estimates" table contains the maximum likelihood estimates of the parameters (Figure 22). Both s2b1 and s2b2 are marginally significant, indicating between-subject variability in the clearances and absorption rate constants, respectively. There does not appear to be a significant covariance between them, as seen by the estimate of cb12.

The estimates of , , and are close to the adaptive quadrature estimates listed in Table 3 of Pinheiro and Bates (1995). However, Pinheiro and Bates use a Cholesky-root parameterization for the random-effects variance matrix and a logarithmic parameterization for the residual variance. The PROC NLMIXED statements using their parameterization are as follows, and results are similar.

proc nlmixed data=theoph;
   parms ll1=-1.5 l2=0 ll3=-0.1 beta1=-3 beta2=0.5 beta3=-2.5 ls2=-0.7;
   s2   = exp(ls2);
   l1   = exp(ll1);
   l3   = exp(ll3);
   s2b1 = l1*l1*s2;
   cb12 = l2*l1*s2;
   s2b2 = (l2*l2 + l3*l3)*s2;
   cl   = exp(beta1 + b1);
   ka   = exp(beta2 + b2);
   ke   = exp(beta3);
   pred = dose*ke*ka*(exp(-ke*time)-exp(-ka*time))/cl/(ka-ke);
   model conc   ~ normal(pred,s2);
   random b1 b2 ~ normal([0,0],[s2b1,cb12,s2b2]) subject=subject;
run;