Journal of Biopharmaceutical Statistics: Book Review

In “Statistical Rules of Thumb,” Gerald van Belle covers a wide range of topics, everything, it seems, except a definition of the phrase “rule of thumb.” Wikipedia offers a useful definition (Wikipedia, 2009). “A rule of thumb is a principle with broad application that is not intended to be strictly accurate or reliable for every situation. It is an easily learned and easily applied procedure for approximately calculating or recalling some value, or for making some determination.” The key words in the definition that you need to keep in mind are “not intended to be strictly accurate or reliable for every situation.” No one ever got thrown in jail for ignoring a rule of thumb. So although this reviewer is sorely tempted to say “yes but” to about half of the rules in this book, he will resist the temptation. Dr. van Belle, himself, often acknowledges these “yes buts” in his book. This requires a careful balance. He avoids a rigid listing of statistical dogmatisms but doesn’t timidly offer all possible perspectives as if they carried equal weight. Dr. van Belle wrote these rules of thumb partly from his own experience but more from discussions with other statisticians and a review of the classic references in statistics. The bibliography is outstanding. The three quartiles of publication years in the bibliography are 1984, 1997, and 2001, but they reach as far back as 1925. Rules of thumb represent established practice developed over decades of experience, so the emphasis on older publications should be expected. I had seen this book when it was just a single chapter on the Web prior to publication of the first edition. Some of the strongest material of the book is the rules of thumb involving sample size determination. I have used many of the sample size rules in this chapter, and I credit van Belle’s work in this area with providing me the ability to rapidly recognize when a client is proposing a grossly inadequate sample size. Several sample size rules in Chapter 2 are worth quoting. In a two-group comparison, 16 divided by the square of the effect size produces a reasonable sample size per group. Thus, a good power for a half standard deviation shift requires about 16/0 52 = 64 patients in each group. A rule for calculating power for a relative change in means given only a coefficient of variation (relative standard deviation) is also very useful. Finally, if your client is expecting to prove that an outcome is rare, it helps to set sample size targets based on the rule of 3 (if you observe 0 events out of n trials, 3/n is an approximate 95% upper confidence bound). To establish that an outcome occurs less than 1% of the time, sample at least 300 patients and hope that none of them experience the event.