Design and Analysis of Clinical Experiments

Rating:
Summary: A must for biostatisticians
Review: Dear Sirs I do not know if I can use this part as a tool for contacting the autour or not; however please excuse me if I have not used it properly: I am a student of epidemiology, and presently we are studying the following book: The Design and Analysis of Clinical Experiments; Joseph L. Fleiss; John Wiley & Sons; 1986. But on page 67 it seems that there is some misunderstandings: When we do a log transformation, in fact we are changing (or shifting) the previous distribution to a normal distribution. Without any doubts in this transformation the means of our previous distributions do not transfer to the means of the new distributions! In fact these are the medians of the previous distributions which are transferred to the place of the means of new distributions (of course if we presume that the new distributions are almost perfectly normal) and by finding the confidence interval of the difference of the means of new distributions (lambda1 - lambda2) we are finding the CI of ratio Median1/Median2. In this way it seems reasonable that the formula 3.25 be changed to Median1/ Median2. It would be very kind of you if you help me in this problem!

With best regards Dr. Shahrokh Izadi

Rating:
Summary: A question from the autour
Review: Dear Sirs I do not know if I can use this part as a tool for contacting the autour or not; however please excuse me if I have not used it properly: I am a student of epidemiology, and presently we are studying the following book: The Design and Analysis of Clinical Experiments; Joseph L. Fleiss; John Wiley & Sons; 1986. But on page 67 it seems that there is some misunderstandings: When we do a log transformation, in fact we are changing (or shifting) the previous distribution to a normal distribution. Without any doubts in this transformation the means of our previous distributions do not transfer to the means of the new distributions! In fact these are the medians of the previous distributions which are transferred to the place of the means of new distributions (of course if we presume that the new distributions are almost perfectly normal) and by finding the confidence interval of the difference of the means of new distributions (lambda1 - lambda2) we are finding the CI of ratio Median1/Median2. In this way it seems reasonable that the formula 3.25 be changed to Median1/ Median2. It would be very kind of you if you help me in this problem!

With best regards Dr. Shahrokh Izadi

Rating:
Summary: now a classic and still a great reference
Review: This book was first published in 1986. As Fleiss states in his preface, the intention is to fill a gap in the standard texts on experimental designs by emphasizing and illustrating those that are useful in clinical studies. This book was clearly marketed for the rapidly growing and highly regulated pharmaceutical industry. In addition to the classic experimental designs, Fleiss covers cross-over designs and repeated measure designs that are important in clinical trials. He writes clearly and deals with the important issues in clinical trials including potential biases, blinding, randomized controls, multiple comparisons and repeated measures. The book starts off with a chapter that emphasizes the effect of measurement error and also provides some simple experiments on reliability of measurements.

There is a wealth of methods included, many different designs and various parametric and nonparametric analysis techniques. It is aimed at biostatisticians in the pharmaceutical industry and the medical research field. The book is very much suited for an advanced undergraduate or graduate level course for students majoring in statistics or biostatistics. The level of mathematics is high but not excessively used. Mathematical results on sample size determination are deferred to an appendix.

The Wiley editors choose only successful books to be included in their Classics Library series. The intent of the Classics series is to take popular books by distinguished authors and create a paperback edition that may be more affordable than the hardcovered edition still in print. It is not a revision of the book. This book entered the Classics series in 1999.

It is a great reference source and I plan to consult it a great deal in the future. The only drawback to it that I see is that it is not up to date. The last 15 years has seen many advances in group sequential methods, Bayesian designs and longitudinal data analysis that this text misses. So Fleiss' book is not one stop shopping for a clinical biostatistician but it does offer a lot and presents it eloquently.

With regard to reviewer Izadi's comments on Amazon, I think the appropriate way to ask this question is really to write the author. Since it is here in print for the readers, I will attempt a reply. I think there is a misinterpretation of the terminology. When Fleiss refers to mean logarithms he is not referring to the population means on the log scale but rather the logarithm of the population mean on the original scale. With the latter interpretation equation 3.25 makes perfect sense. It is the former interpretation of the parameters that the reviewers point addresses. The importance in the example is to demonstrate the lack of robustness of t or F tests to non-normal (e.g. lognormal) data and to show that tests and confidence intervals can still be accomplished using the normal theory after the transformation. The key point is that it is the ratio of the parameters that is transformed into differences of the log of the parameters.

Rating:
Summary: now a classic and still a great reference
Review: This book was first published in 1986. As Fleiss states in his preface, the intention is to fill a gap in the standard texts on experimental designs by emphasizing and illustrating those that are useful in clinical studies. This book was clearly marketed for the rapidly growing and highly regulated pharmaceutical industry. In addition to the classic experimental designs, Fleiss covers cross-over designs and repeated measure designs that are important in clinical trials. He writes clearly and deals with the important issues in clinical trials including potential biases, blinding, randomized controls, multiple comparisons and repeated measures. The book starts off with a chapter that emphasizes the effect of measurement error and also provides some simple experiments on reliability of measurements.

There is a wealth of methods included, many different designs and various parametric and nonparametric analysis techniques. It is aimed at biostatisticians in the pharmaceutical industry and the medical research field. The book is very much suited for an advanced undergraduate or graduate level course for students majoring in statistics or biostatistics. The level of mathematics is high but not excessively used. Mathematical results on sample size determination are deferred to an appendix.

The Wiley editors choose only successful books to be included in their Classics Library series. The intent of the Classics series is to take popular books by distinguished authors and create a paperback edition that may be more affordable than the hardcovered edition still in print. It is not a revision of the book. This book entered the Classics series in 1999.

It is a great reference source and I plan to consult it a great deal in the future. The only drawback to it that I see is that it is not up to date. The last 15 years has seen many advances in group sequential methods, Bayesian designs and longitudinal data analysis that this text misses. So Fleiss' book is not one stop shopping for a clinical biostatistician but it does offer a lot and presents it eloquently.

With regard to reviewer Izadi's comments on Amazon, I think the appropriate way to ask this question is really to write the author. Since it is here in print for the readers, I will attempt a reply. I think there is a misinterpretation of the terminology. When Fleiss refers to mean logarithms he is not referring to the population means on the log scale but rather the logarithm of the population mean on the original scale. With the latter interpretation equation 3.25 makes perfect sense. It is the former interpretation of the parameters that the reviewers point addresses. The importance in the example is to demonstrate the lack of robustness of t or F tests to non-normal (e.g. lognormal) data and to show that tests and confidence intervals can still be accomplished using the normal theory after the transformation. The key point is that it is the ratio of the parameters that is transformed into differences of the log of the parameters.

Rating:
Summary: A must for biostatisticians
Review: This is the standard text on this subject. Recommend you have at least a Master's degree in statistics to take full advantage of this book. This book is too technical for non-statisticians, although they may get some useful infomation from the non-statistical discussions.