NONMEM Users Guide Part I - Users Basic Guide - Chapter F
F. Nonlinear Regression with One-Level Nested Random Effects
F.1 An Example
F.2 Implementation of the Example
F.2.1 Inputs
F.2.2 Selected Printout

NONMEM Users Guide Part I - Users Basic Guide - Chapter F

F. Nonlinear Regression with One-Level Nested Random Effects

F.1 An Example

In this chapter we return to theophylline plasma concentration vs time data, but where such data from 12 subjects, rather than from a single subject, are available. This is done to illustrate a regression with one-level nested random effects where the regression function is nonlinear. The fact that the regression function is nonlinear really does not introduce any new considerations regarding the inputs required by NONMEM. However, often with a nonlinear regression function and one-level nested random effects, a modeling approximation is necessary, and this is described here. Also, this example does bring together a number of concepts discussed in the earlier examples.

Each subject is given a single oral dose, the same dose for each subject. Each subject has a different weight. Often dose is expressed as the amount of drug administered per unit weight of the subject (weight-adjusted dose), and in this example the dose data item is the weight-adjusted dose. However, a weight data item is also included in the data record because it will be assumed that interindividual differences in plasma concentrations may be due to interindividual weight differences beyond those expressed through weight-adjusted dose. Also, 11 plasma concentrations are observed per individual at different times, and these times vary between individuals. (The clearance and rate constant measurements used in the examples of chapter E are obtained from this concentration vs time data and from similar data using an additional five doses per subject.)

A model for the jth observation from the ith individual might be given by

Image grohtml-206533.png

where

Image grohtml-206534.png

Image grohtml-206535.png

Image grohtml-206536.png

where Image grohtml-20653-7.png , Image grohtml-20653-8.png , and Image grohtml-20653-9.png are the (non-weight-adjusted) dose, time, and weight variables. Here Image grohtml-20653-10.png and Image grohtml-20653-11.png are not subscripted with a j, indicating that the values of the dose and weight variables do not vary within the individual. This model is similar to one used in chapter C for data from a single individual, but there are some notable differences. First, the ith individual is regarded as having his own set of pharmacokinetic parameters, these parameters are denoted by Image grohtml-20653-12.png , Image grohtml-20653-13.png , and Image grohtml-20653-14.png . Second, two of the pharmacokinetic parameters are rate constant of absorption, Image grohtml-20653-15.png , and rate constant of elimination, Image grohtml-20653-16.png , as previously, but the third basic parameter is clearance, Image grohtml-20653-17.png , rather than volume of distribution. Third, these parameters are affected by random interindividual affects, and thus random interindividual variability is expressed in the model. Fourth, residual error is an intraindividual effect. Note that an individual’s clearance is linearly related to his weight as in chapter E. The variance-covariance of the random interindividual effects, Image grohtml-20653-18.png , is regarded as a full matrix in this example.

As stated in section A.1, with the current version of NONMEM random effects must enter the model (for the observations) linearly. This requirement is not met in the above model; the random interindividual effects enter nonlinearly. One device that has been found useful under these circumstances is to approximate the above model, A, with another, B, obtained by expanding A with a first-term Taylor Series in the random effects about their mean values (0). In the case at hand B is given by

Image grohtml-206537.png

where

Image grohtml-206538.png

Image grohtml-206539.png

Image grohtml-2065310.png

Image grohtml-2065311.png

Image grohtml-2065312.png

Written this way, the model is also displayed as the NONMEM linear model schematic. Use of this first-order approximation to the original model, along with use of the ELS objective function, has been called the First-Order Method for analyzing nonlinear mixed effects modeled data. This method has been shown to be statistically efficacious in particular situations (Sheiner and Beal, 1980, 1981, and 1983, and Beal 1984a). The first-order approximation itself may be called the First-Order Model. One practical problem with this method is that it can require some nontrivial effort to obtain the partial derivatives defining the g’s. Moreover, there is little to be gained by examining these derivatives. Indeed, rather than try to display explicit formulae for the g’s in this example in this text, we refer the reader to the PRED routine of Fig. 73 where code is given for these formulae. Certain tools are available to help the user obtain the first-order model. PREDPP is a package which can be used with NONMEM and with pharmacokinetic data and which automatically obtains the derivatives Image grohtml-20653-25.png , when, as in the example, the effect of the Image grohtml-20653-26.png is through Image grohtml-20653-27.png . PREDPP is actually a very elaborate PRED subroutine. It then remains for the user to supply code for the derivatives Image grohtml-20653-28.png ; these are relatively simple to obtain. Also, NM-TRAN, a computer program which facilitates the problem of constructing inputs to NONMEM, can be used to automatically obtain the derivatives Image grohtml-20653-29.png . (Both PREDPP and NM-TRAN are distributed with NONMEM.)

Let I denote the number of individuals. Also, for fixed i, let Image grohtml-20653-30.png denote the column vector of values of the Image grohtml-20653-31.png , let Image grohtml-20653-32.png denote the column vector of values of the Image grohtml-20653-33.png , let Image grohtml-20653-34.png denote the column vector of values of the Image grohtml-20653-35.png , let Image grohtml-20653-36.png denote the column vector of values of the Image grohtml-20653-37.png , and let Image grohtml-20653-38.png denote the column vector of values of the Image grohtml-20653-39.png . Then the ELS objective function is given by

Image grohtml-2065313.png

where

Image grohtml-2065314.png

Image grohtml-2065315.png

The last term in the expression for Image grohtml-20653-43.png is just a fancy way of writing the diagonal matrix whose elements are all Image grohtml-20653-44.png . The matrix Image grohtml-20653-45.png is the variance-covariance matrix of Image grohtml-20653-46.png . The vector Image grohtml-20653-47.png is the vector of weighted residuals from the observations Image grohtml-20653-48.png . As with previous examples, it has the form residual (vector) divided by standard deviation (matrix), and it is "squared" in the expression for the objective function. The weighted residuals are defined to be the weighted residuals from all obervations Image grohtml-20653-49.png .

F.2 Implementation of the Example

F.2.1 Inputs

A code for PRED which implements the example is given in Fig. 74. It is similar to that in Fig. 1. However, the values returned in G are now very different, and a value is also returned in H. The same rules for determining what is returned in G and H, and that are given in chapter E, apply here too. For clarity, code to compute the partial derivatives that are returned in G is indented from the other code. Note that in the expression for F the weight-adjusted dose (DOSE) appears, rather than the non-weight-adjusted dose, but that also THETA(3) occurs in the denominator (E=THETA(3)*C) of that same expression, so that weight itself need not enter this expression. On the other hand, since Image grohtml-20653-50.png adds to mean clearance, weight does enter the expression for G(3).

A control stream for this example is given in Fig. 75. The data set is embedded in it. Note that for readability and for the purpose of conveniently keying the data, the weight-adjusted dose and weight data items are blank for all data records of an individual record except the first data record. The PRED routine stores these data items in its local storage whenever the first data record of an individual record is passed to it (review the argument NEWIND described in section C.3.5.2).

The initial STRUCTURE record for the problem specification has 1’s in fields 7 and 8, indicating that Image grohtml-20653-51.png is a full matrix, but that Image grohtml-20653-52.png is constrained to be diagonal. (Again, since Image grohtml-20653-53.png is a scalar, it can be regarded as an unconstrained Image grohtml-20653-54.png matrix, but for the sake of a more perspicuous problem summary, it is taken to be diagonal.)

F.2.2 Selected Printout

The final estimate, standard errors, and correlation matrix are shown in Figs. 76-78. It may interest the reader to see how remarkably well the final estimates in Figs. 66 and 76 agree for those parameters that occur in both the model in section E.4 and the model in section F.1. The final estimates of these parameters from both figures, their standard errors, and the ratios of standard error to estimate are given in Table F.2.2.i. Recall that the estimates in Fig. 76 are obtained using one-sixth the amount of data used to obtain the estimates in Fig. 66, since in the present example only the concentration data from one dose per individual are used, while in the previous example this same data, plus similar data from five additional doses per individual, are used.

Table F.2.2.i Estimate Comparison

Image grohtml-2065316.png

The first page of the requested table is shown in Fig. 79. Scatterplots of residual vs time and of weighted residual vs time, both separated by ID, are requested. The four scatterplots corresponding to individuals 4 and 5 are shown as examples in Figs. 80-83.

TOP

TABLE OF CONTENTS

NEXT