Threshold determination in dose-response studies: Benchmark analysis with non-parametric smoothing splines

The primary aim of this research was to develop and test a method for detecting clinically meaningful thresholds where the criteria were specified a priori, the functional form of the relationship was not imposed and confidence intervals were defined. This was accomplished by extending benchmark analysis to include non-parametric smoothing splines as the estimate of the dose-response relationship. A secondary aim of this research was to estimate the point of departure for the allowable fetal exposure levels of alcohol and Polychlorinated Biphenyls (PCB) using the new method. The study was conducted in two parts. In the first, a simulation was conducted to compare a linear and a non-linear method of benchmark analysis across a variety of conditions. In the second part, the two methods were compared using data from two empirical studies.; Based on the simulation study the following conclusions were made. The first was that the linear estimation method could break down and become unstable, producing very large BMD estimates when the population effect size is small. This problem is compounded when the sample size is small. The second was that transformations of positively skewed exposure variables, to reduce the impact of outliers, results in BMD estimates that are less accurate and too large. The third was that when heterogeneity is present and controlled, the BMD estimates become more stable. Finally, as expected, the performance of the non-linear estimation method was superior when the dose-response relationship was non-linear. Its estimates had lower bias, better accuracy, efficiency, and coverage probability.; Using the non-linear benchmark method, the point of departure for the allowable fetal exposure level of alcohol was estimated to be about 17 drinks per week. This was based on a P₀ of 0.16 and BMR of 0.05, which corresponded to the smallest increase in risk. This is preferred given the serious impact of maternal drinking on the fetus. The non-linear method was not applied to the PCB data because the dose-response relationships were all linear. Directions for future research include the application of this method to dichotomous outcomes and the inclusion of multiple predictors, covariates and moderators.