Article type
Year
Abstract
Background: Assessing the effect of a continuous covariate in a single study requires the determination of a dose-response relationship in a model adjusting for other covariates. Even with individual patient data (IPD), summarizing the results of several dose-response functions in a meta-analysis is not straightforward.
Objectives: To describe a new procedure which determines in single studies an adjusted dose-response function for a continuous variable and which also summarizes such functions across studies.
Methods: In single studies we determine functions using fractional polynomials, with or without adjustment for confounders. Three different methods to select the FP function are proposed. To average functions from several studies, a fixed or random effect approach can be used.
Results: Using data from the US SEER database in which each individual registry is treated as a single study, we derive adjusted overall estimates of the functional forms for the association between time to breast cancer death and a continuous covariate. Whereas the individual functions for nodes, a factor with a large effect, are similar across studies, those for age, a covariate with a weak effect, show considerable variability. Because of their different weights, whether to use a fixed or a random effects model affects the average function.
Conclusions: For an 'ideal’ situation with IPD and only minor variations between studies with respect to measurement techniques and confounders, our approach allows one to model dose-response relationships in single studies and to summarize them in an average function. Modifications are available if the data situation is less than ideal.
Objectives: To describe a new procedure which determines in single studies an adjusted dose-response function for a continuous variable and which also summarizes such functions across studies.
Methods: In single studies we determine functions using fractional polynomials, with or without adjustment for confounders. Three different methods to select the FP function are proposed. To average functions from several studies, a fixed or random effect approach can be used.
Results: Using data from the US SEER database in which each individual registry is treated as a single study, we derive adjusted overall estimates of the functional forms for the association between time to breast cancer death and a continuous covariate. Whereas the individual functions for nodes, a factor with a large effect, are similar across studies, those for age, a covariate with a weak effect, show considerable variability. Because of their different weights, whether to use a fixed or a random effects model affects the average function.
Conclusions: For an 'ideal’ situation with IPD and only minor variations between studies with respect to measurement techniques and confounders, our approach allows one to model dose-response relationships in single studies and to summarize them in an average function. Modifications are available if the data situation is less than ideal.