STATISTICAL COMPUTING SECTION, ROYAL STATISTICAL SOCIETY.

Half day meeting on:

Sensitivity Analysis of Computer Models

Wednesday, 28th May.  14.00 - 17.00
Royal Statistical Society, Errol Street, London

This topic concerns the quantification and analysis of
uncertainty induced in the outputs of computer models, as a
consequence of uncertainty in the inputs, including
uncertainty about the assumptions in the model and
internal parameters, as well as uncertainty in the input
values to be used for a specific model run.

The meeting will be held at the Royal Statistical Society
and will be immediately preceded by the AGM of the
Statistical Computing Section. There is no charge, and no
prior registration is needed.

See http://www.rss.org.uk/about/direction.html for details
on how to find the RSS.

14.00 AGM

14.05 Sensitivity analysis as a tool for model assessment
      Marian Scott (University of Glasgow)

Abstract: Sensitivity analysis (SA) is a general
methodology used to evaluate the sensitivity of model
output to changes in model input, i.e. the rate of change
of the response function relative to the input parameters.
There are a number of different methods for carrying out a
sensitivity analysis, ranging from simple one-at-a-time
methods to global, multivariate methods.  There are also
strong links to classical design of experiments. SA is
closely linked to Uncertainty Analysis (UA), another
computational method, where the objective is to evaluate
the uncertainty on the model response as a result of
uncertainties on the model input parameters (parametric
uncertainty) and on the model form itself (structural
uncertainty).  In this talk, SA and UA tools, their use and
the challenges presented in their application to some
complex models will be discussed.

14.45 Modelling, making inferences and making decisions:
      the roles of sensitivity analysis
      Simon French (Manchester Business School)

Abstract:n Sensitivity analysis, robustness studies and
uncertainty analyses are key stages in the modelling,
inference and evaluation used in operational research,
decision analytic and risk management studies.  However,
sensitivity methods - or others so similar technically
that they are difficult to distinguish from sensitivity methods
- are used in many different circumstances for many
different purposes; and the manner of their use in one
context may be inappropriate in another.  In this talk, I
categorise and explore the use of sensitivity analysis and
its parallels, and in doing so hope to provide a guide and
typology to a large growing literature.

15.30 Tea/coffee break

16.00 Sensitivity Analysis using Value of Information
      Jeremy Oakley (University of Sheffield)

Abstract: When using a computer model to guide a decision, the decision-
maker will need to understand the uncertainty present in the model.
One source of uncertainty will typically be the values of various input
parameters in the model that should be used for the situation at
hand. Sensitivity analysis is then concerned with identifying which
of these parameters is influential in some sense in driving the model
output uncertainty. The value of information approach to sensitivity
analysis directly relates the influence of each uncertain input
parameter to the decision problem in question; it quantifies the value
to the decision maker of learning the true value of any uncertain
input parameter before making their final decision. Unfortunately,
severe computational problems can arise when calculating these
measures of sensitivity for complex models. In this talk the value of
information framework for sensitivity analysis is reviewed and
efficient computational tools for complex models are presented. The
talk is motivated by an application in medical decision making using
health economic models.

16.45 General Discussion

17.00 Close of meeting

----------------------
Suzanne Evans
[log in to unmask]