Determining Uncertainty
More Information
Errors in inputs to a predictor can lead to bias in the final predictor, which can be corrected if the magnitude of the error is understood. View the animation of this graph.
We must assume three models to analyze uncertainty: 1) A dynamical model that predicts the consequences for specific parameters, 2) a hierarchical model that describes population variability between individuals and 3) a measurement model that describes how observations, including errors, were made. We can be uncertain about any of the three models as well as the parameters that describe those models.
Uncertainty and Variability Sources in CRA. View the full size image.
The "Uncertainty" project develops and uses advanced statistical tools to analyze biological models (primarily pharmacokinetics - absorption, distribution, metabolism, and elimination of chemicals in the body). This allows decisions to be made based on quantitative measures of uncertainty about how a chemical or chemicals effect people.
The Uncertainty project:
- Uses multiple statistical methods to determine plausible ranges of parameter values and make comparisons between multiple models with different equations on the same data.
- maximum likelihood estimation, which produces a single set of most likely parameter values
- Bayesian methods to determine distributions of parameter values that are consistent with the data given assumptions about how data was obtained
- Discriminates among multiple, potentially complicated models for the same data.
- Distinguishes uncertainty (how much or little is known about the model), from true variability, for example differences in body weights between the individuals being modeled.
- Allows for varying levels of uncertainty that depend on available data quantity and quality. For example, there may be uncertainty about the specific value of the parameters within an equation or even about the equation itself.
- In collaboration with other ORD laboratories, provides support for regulatory decisions.
- Supports statistically-based experimental design, to minimize the number and cost of experiments and maximize the information gained.
- Uses a wide range of computers, from standard laptop computers with one or two CPUs to high performance computing clusters with hundreds of CPUs, depending on the problem.