Advocating the Devil's advocate


Ruxton, G.D. & Colegrave, N. ( 2003 ) Experimental design for the life sciences . Oxford University Press , Oxford, UK . xviii + 114 pp . , figs, line diagrams, index. Paperback: price £16.99 , ISBN 0-19-925232-7.

Books such as this are extremely useful. It is the just the right length and just the right style for reading over a day or two and exactly the book that every fresh student should read as soon as embarking on an empirically-based research programme. Nor will it waste the time of more established researchers. I would generally expect that those who read this book would make substantially fewer time- and effort-wasting blunders than those who do not. We all might like to think that experimental design and interpretational errors are ‘things that happen to other people’ but, as the authors point out, we all have great potential to make mistakes, and to substantiate their point they gallantly admit to a few red-faced experiences of their own. Benefiting from someone else's hindsight rather than having to acquire one's own has got to be worth £16.99 of any researcher's money.

Books on experimental design are usually books about statistics. This one is not, although I am going to keep it next to my collection of statistics books because it complements them perfectly. Ruxton and Colegrave are at pains to make the point that biologists, and not statisticians, are the ones who should design life-science experiments. To do this they must, however, be statistically informed, otherwise they will be unable to analyse the data deriving from their work; so this is a book partially about being ‘statistically informed’ rather than about statistics per se. Linking in with intended subsequent statistical analysis, the following major design issues are covered clearly and in sufficient detail to grasp the essential points: randomization, replication, pseudoreplication, statistical power, controls, factorial experiments, cross-over designs and split-plot designs. If you are ‘mathematically challenged’, you will be pleased that all this is dealt with without equations: indeed, there are none in the book, just many useful concepts and tips.

As well as complementing texts on statistics sensu stricto, this book contains useful advice on good working practice. For example, there is a whole chapter on ‘taking measurements’, and ‘ethical considerations’ are a recurring theme. Perhaps one of the most useful tricks of the trade promoted by Ruxton and Colegrave, although not original to them, is for research scientists to imagine that their work was being scrutinized constantly by a highly intelligent, informed and critical peer, personified by them as ‘the Devil's advocate’. I have been using this technique for years and I am sure it has helped, but I am not going to reveal here exactly who my own Devil's advocate has been.

Ruxton and Colegrave point out that, in spite of the need to keep the Devil's advocate at bay, there is no one ‘universal’ way of doing things and no perfect study. Good design, they say, is ‘all about maximizing the amount of information that we can get, given the resources that we have available’. Similarly, there is probably no perfect book, and authors brave enough to write a ‘how-to’ book are going to be more vulnerable than most to nit-picking criticism, so here is mine: during a discussion of pseudoreplication (p. 36) there is an example involving testing whether the sex ratio of turtle hatchlings is skewed towards females. One possibility would be to take 10 turtle hatchlings from a nest and assess the sex of one. If the sex of that one turtle is likely to give information on the sex of the other nine (because turtle nests tend to contain only one sex of offspring due to temperature-dependent sex determination) then the 10 turtles are not independent replicates and ‘we would be better collecting only single individuals from each nest’. The point here is that the nest is the sampling unit, and not the hatchling, so one should acquire one rather than 10 data per nest: quite so. However, and ignoring the ethics of disturbing more turtles (I am an entomologist, and am ashamed to say that ethics seldom come into design considerations in my field), the best way of extracting information from the handful of 10 hatchlings would actually be to sex them all and analyse the sexual composition data as a proportional response variable in a logistic analysis which would, appropriately, give more weight to samples of 10 hatchlings than samples of one and yet retain the nest as the sampling unit. To be fair, Ruxton and Colegrave do make a point akin to this on p. 42.

Generally, the book is well structured, although the chapter entitled ‘Final thoughts’ seemed to me to be somewhat of a rag-bag of useful comments that were well worth including, but that nonetheless could have been integrated into other areas of the book. Each chapter starts with a bulleted list of key points and ends with another bullet-point summary: short of including free instant coffee sachets, books do not come much more user-friendly than this. There is also a useful bibliography that includes informative mini-reviews of the referenced publications.

In summary, this is a highly useful book that delivers its messages in a very accessible way. It even verges on the fun, which is quite an achievement for a book in this area. For instance, there are some experimental design flow charts at the start of the book that contain the almost penultimate instruction ‘sleep on it’: who would wish to disobey that? Ruxton and Colegrave's account of ‘things they wish they had known earlier’ will be a ‘memoir’ that benefits many. I am certainly going to be using this book when teaching my (virtually maths-free) course on statistics and experimental design.