The Mahalanobis–Taguchi Strategy

generally nonparametric statistics in the book, because forcing data to Ž t some mathematically convenient cdf destroys potentially actionable information and forces me to defend the choice of parametric cdf. The book contains a lot of modeling. I have known the academic pressure of “publish or perish,” and I know some journals function as repositories of knowledge. Authors publish models and statistical analyses of these models that were analyzed because they could be. I wonder if those authors were inspired by real problems or because they could solve the problems. One does not read those articles for interest or entertainment. One might read them in the book, because they have been collected, summarized, and compared. Recurrent events in the new Chapter 9 cannot be modeled with dead-forever failure time distributions. Unfortunately, not too many reliability engineers recognize the distinction between recurrent events, such as product repairs, and dead-forever failures; fewer still have time-between events data. Finally, there is not much point to modeling recurrent events if one does not know enough about them. A student in India keeps sending me helicopter failure counts— dozens, between scheduled maintenances, without classiŽ cation of cause or failure mode. Modeling and extrapolation hardly seems worthwhile. Reliability engineers have a lot to learn from biomedical recurrent events analyses and from Nelson (2003). The second edition of The Statistical Analysis of Failure Time Data is a great book—improved, modernized, and comprehensive. Keep this book-level reference handy. Buy the book if you do biomedical survival analysis or reliability statistics. The second edition is a better textbook than the Ž rst. It has more exercises, although they are largely academic exercises (stuff that did not make it into the text itself). It is amazing how hard innocent graduate students will work to learn what they may never use. Discipline. Tradition. Just a minute, please, I am not Ž nished. What if I do not have a random sample, iid, possibly censored, truncated, or grouped? What if members of the cohort are not tracked by name, serial number, or other identiŽ cation? What if privacy considerations prevent identiŽ cation of members of a cohort by name or serial number, for example, AIDS, counts of third world disease incidences, or ships (births, installations, sales) and returns (deaths, complaints, failures, repairs, spares sales) counts by calendar interval. Look for other work. Compute monthly average failure rates, failures/installed base? This book does not deal with such real life concerns. Also, cost considerations weigh against tracking products by serial number. About the same time the hard disk drive industry began to recognize that failure rates were not constant (http://www.idema.org), computer companies quit tracking parts to failure by serial number. Others, like auto equipment manufacturers, collect warranty repair records by vehicle identiŽ cation number, which requires 1,000 times the storage as ships and returns counts and incurs at least that many times more errors. People who use the automotive computer system say that it is easy to put data in, but it sure is hard to get information out. Generally accepted accounting principles require ships and returns counts. They are sufŽ cient to make nonparametric estimates of survival functions (http://web.utk.edu/»asaqp/newsletters/1299newsletter.pdf, pp. 13–16). With a little more work, ships and returns counts could be classiŽ ed by concomitant variables, allowing the same analyses as done by Kalb eisch and Prentice. For transient processes such as AIDS, hantavirus, other epidemics, and transplants, clinical trials would not have to kill controls, because the control survivor function could be estimated from population data. Ronald Fisher taught the controltreatment paradigm for agricultural tests, not tests of life-saving drugs. We need statistics for samples other than iid.