R.E. Port
German Cancer Research Center, Heidelberg
When drug treatment affects a variable that is also measurable without treatment, e.g. blood pressure, heart rate, or the concentration of an endogenous compound, modeling response requires to estimate the pretreatment level of this variable. Individual pretreatment levels may vary widely in a population and cannot necessarily be assumed to be normally distributed, especially in pathological conditions. Often, pretreatment measurements are available in each individual to support estimating pretreatment levels individually. Subtracting, for each individual, a pretreatment measurement from all posttreatment measurements in order to produce “observations” of the drug’s “net effect” is unsatisfactory as it puts undue weight on the pretreatment measurements of which usually only one or a few are available for each individual. One can incorporate the mean population pretreatment level as a structural parameter in a population model and assign an interindividual variance to it. Fixing this variance to an arbitrarily large value will alleviate the assumption of a normal (or log-normal) interindividual distribution if this is the default in the modeling program used, as in NONMEM. If, in practice, this approach fails, the estimates of individual pretreatment levels may be “loosely tied” to individual pretreatment measurements in the following way: If just one pretreatment measurement is available for each individual this measured value is defined to be the expected individual pretreatment level. True individual pretreatment levels will randomly deviate from this expectation. The interindividual distribution of these deviations can be assumed to be normal, and it is plausible to assume that its variance is equal to the residual variance of the posttreatment measurements. If more than one pretreatment measurements are available per individual, their average may be used as the expected individual pretreatment level, and true individual pretreatment levels may be assigned a variance which is a proper fraction of overall residual variance. This approach is easily implemented in NONMEM.
Its application will be illustrated using data on the response of blood hemoglobin concentration to erythropoietin in anemic children.
Reference: PAGE 5 (1996) Abstr 549 [www.page-meeting.org/?abstract=549]
Poster: oral presentation