Measures of Interobserver Agreement

(2001), reviewed for Technometrics by Ziegel (2003). The present authors indicate that replacing nondetects with half the detection limit may be good enough as long as the nondetects are not too numerous. Regression and maximum likelihood procedures are presented. The authors favor the simplicity of regression as long as the percent of nondetects is not too high, and there is only a single detection limit. Other topics in this chapter include estimation of upper bounds for means, estimation when some of the responses are true zeros, and estimation when all of the data are below the detection limit. For the latter topic, methods are given for estimating the mean and standard deviation under certain assumptions. The situation is applied to the problem of demonstrating compliance in groundwater monitoring. Fiducial intervals are utilized. In Chapter 6, “The Promise of the Bootstrap,” the authors’ enthusiasm is obvious. They present their case with several effective illustrations of situations where it would be impossible to designate an appropriate reference distribution for tests and confidence intervals. This particular chapter should work well for nonstatisticians, because the authors make their case very effectively from their experiences and their compelling illustrations. Chapters 7 and 8 deal with spatial data and temporal data. The statistical description of methods for spatial data is certainly much easier to understand than the corresponding material from the classic statistical treatise by Cressie (1993). On the other hand, nonstatisticians likely still will not understand many of the author’s statements. The authors develop arguments for the use of the nonparametric geostatistical method known as probability kriging. A total of 27 different semivariograms and cross-semivariograms are needed to represent the anisotropy in the chapter’s principal illustration, 9 in each of 3 directions. They utilize both SAS and the software GSLIB from Deutsch and Journel (1998) (see Ziegel 1998). The example is both thorough and complete and results in the estimation of the volume of the contaminated soil. Concluding material and two more examples emphasize understanding the variography for a site. The final chapter presents the analysis of temporal data via autoregressive integrated moving average (ARIMA) models. This is another somewhat advanced discussion for nonstatisticians. The authors do not do a good job explaining how patterns in the autocorrelation and partial autocorrelation functions dictate the selection of certain model forms. Nonstatisticians can use these without an understanding of Yule–Walker equations. The modeling process is not helped by the selection of a nonseasonal model that includes an autoregressive parameter of order 9 and a moving average parameter of order 15 for dealing with hourly PM10 concentrations. The parade of interesting examples continues with a second set of PM10 values that are adjusted for both wind speed and wind direction before they are fitted by an ARIMA transfer function model. However, the authors show a fitted model graphically without providing any indication about how one identifies or estimates the necessary model. Eventually, crosscorrelation methodology is discussed. Intervention variables are also used with little presentation of the necessary technical background. The authors must have decided that the book was long enough. This is not the book about statistics for the environmental scientist that I would have written. However, it has great examples, and it covers a lot of ground. It is a good book for statisticians and probably now the best available book even for environmental scientists. After all, they are going to need to get some help from their favorite statistician.