A Monte Carlo exchange algorithm for finding near-optimal designs under model contamination
A Monte Carlo exchange algorithm for finding near-optimal designs under model contamination
A Monte Carlo exchange algorithm is presented for finding efficient designs under bias-based criteria. The true model is assumed to differ from the model to be fitted by an additive contamination term, which is considered to be a realisation of a random variable. This results in the bias for any given design also being a random variable. Any available prior knowledge can be used to inform the distribution of the contamination. Criteria for design selection are described whose objective functions are based upon properties of the bias distribution.
These criteria are implemented in an exchange algorithm, where efficient designs are found by exchanging points between the design and a candidate list of possible points. The objective functions are approximated via Monte Carlo simulation within the algorithm. Although for some criteria a single approximation can be used for all exchanges of design and candidate points, for others a simulation is required for each swap. Therefore the approach is highly computationally intensive.
Examples are used to demonstrate the designs produced using the algorithm and criteria. These designs tend to have more distinct support points than traditional optimal designs and hence are of more practical value. The adequacy of the simulation is considered for each example, and the sensitivity of the designs to the prior information is investigated.
Woods, D.C.
ae21f7e2-29d9-4f55-98a2-639c5e44c79c
16 September 2003
Woods, D.C.
ae21f7e2-29d9-4f55-98a2-639c5e44c79c
Woods, D.C.
(2003)
A Monte Carlo exchange algorithm for finding near-optimal designs under model contamination.
International Biometric Society British Region Conference, Reading, UK.
16 - 17 Sep 2003.
Record type:
Conference or Workshop Item
(Poster)
Abstract
A Monte Carlo exchange algorithm is presented for finding efficient designs under bias-based criteria. The true model is assumed to differ from the model to be fitted by an additive contamination term, which is considered to be a realisation of a random variable. This results in the bias for any given design also being a random variable. Any available prior knowledge can be used to inform the distribution of the contamination. Criteria for design selection are described whose objective functions are based upon properties of the bias distribution.
These criteria are implemented in an exchange algorithm, where efficient designs are found by exchanging points between the design and a candidate list of possible points. The objective functions are approximated via Monte Carlo simulation within the algorithm. Although for some criteria a single approximation can be used for all exchanges of design and candidate points, for others a simulation is required for each swap. Therefore the approach is highly computationally intensive.
Examples are used to demonstrate the designs produced using the algorithm and criteria. These designs tend to have more distinct support points than traditional optimal designs and hence are of more practical value. The adequacy of the simulation is considered for each example, and the sensitivity of the designs to the prior information is investigated.
Text
biometrics.pdf
- Other
More information
Published date: 16 September 2003
Venue - Dates:
International Biometric Society British Region Conference, Reading, UK, 2003-09-16 - 2003-09-17
Organisations:
Statistics
Identifiers
Local EPrints ID: 15831
URI: http://eprints.soton.ac.uk/id/eprint/15831
PURE UUID: f88d1f17-5ad1-4703-ad14-306e0270b7b9
Catalogue record
Date deposited: 06 Jun 2005
Last modified: 16 Mar 2024 03:14
Export record
Download statistics
Downloads from ePrints over the past year. Other digital versions may also be available to download e.g. from the publisher's website.
View more statistics
