 |
 |
The question was regarding the use of Bayesian Inference, especially prior distributions, to try to overcome the shortcomings of the data available to the statistician (or, more commonly, the engineer). The question is repeated at the end of this email. It was generally agreed that Bayesian methods could help in these cases by allowing subjective and objective data to be used to construct prior distributions. There is also a body of work concerned with elicitation of belief from experts, to help with the construction of subjective priors and the use of expert systems/networks. "Conventional" techniques for handling missing data (i.e., imputation) may also be useful. I was referred to the work of Dr. Jayanta Kumar Ghosh. of the Indian Statistical Institute (but I have not, as yet, been able to obtain any of his papers!) For my specific interest on cost estimation, the work of Chulani was especially interesting. The suggested literature (with their own comments): A good book on using Bayesian methods is the recently published book by Spiegelhalter, Abrams and Myles: Bayesian Approaches to Health-Care Evaluation, Wiley 2004. The book by Parmigiani (2002) in the same series, Modelling in Medical Decision Making, A Bayesian Approach can also be recommended. Chulani,S., 2003. Bayesian analysis of software cost estimating model: COCOMO II ISBA Bulletin, 10, no. 4., 13. (International Society for Bayesian Analysis). Cowell, R.G., Dawid, A.P., Lauritzen, S.L., Spiegelhalter, D.J., 1999. Probabilistic Networks and Expert Systems. New York: Springer. Farrow,M., 2003. Practical building of subjective covariance structures for large complicated systems. The Statistician, 52, 553-573. Gelman, A., Carlin, J.B., Stern, H.S. and Rubin, D.B., 1995. Bayesian Data Analysis, London: Chapman and Hall. Korb, K.B. and Nicholson, A.E., 2003. Bayesian Artificial Intelligence. London: Chapman and Hall. O'Hagan,A., 1998. Eliciting expert beliefs in substantial practical applications, The Statistician, 47, 21-35. I would like to thank Malcolm Farrow, Philip Good, Stephen Senn, Indrajit Sengupta, David Stephens and Martina Mittlboeck for their interest and suggestions. "In applications where prior statistical data is either sparse, incomplete or not directly comparable, using Bayesian techniques may offer advantages over traditional techniques. This is a problem in the aerospace industry, for example, where cost data from previous projects is limited and usually not directly comparable to the current application; using Bayesian priors may allow more effective and more rigorous use of this data. Is anyone aware of any work that has been done in using Bayesian ideas to compensate for incomplete and limited information? A Williamson