IJE Advance Access originally published online on May 21, 2007
International Journal of Epidemiology 2007 36(4):934-935; doi:10.1093/ije/dym100
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
Book Reviews |
Applied Multilevel Analysis: A Practical Guide. Jos WR Twisk.
E-mail: Liza.Bowen{at}lshtm.ac.uk
Applied Multilevel Analysis: A Practical Guide. Jos WR Twisk. Cambridge University Press, UK 2006, pp. 196 , £24.99 ISBN 100521614989 (PB), ISBN 100521849756 (HB)
The popularity of multilevel analysis (alias hierarchical analysis) has rapidly increased over the past 10 years. The number of articles archived in PubMed using this technique rose >7-fold from 1995 to 2004. Multilevel techniques arose with the recognition that one of the fundamental assumptions underlying many of the statistics used in epidemiology and medical research—that of independence of observations—is often violated. For example, in studies sampling patients from a selection of doctors, patients with the same doctor are likely to be more similar to each other than to the average, leading to correlated (i.e. dependent) observations. This dependence may also be found at other levels, such as with doctors clustered within hospitals. While there is an increasing literature surrounding the use of multilevel modelling, it is almost all authored by statisticians and in language that is difficult for many epidemiologists to understand. However, epidemiologists are still expected to use these techniques, hence Twisk's aim to develop explanations of multilevel analyses for epidemiologists, by an epidemiologist. The result is a highly accessible account of why multilevel analyses are necessary, when it is appropriate to use them, and how such analyses can be carried out.
The book emphasizes that all multilevel techniques are simply extensions of standard regression techniques, and while assuming some prior knowledge of such techniques, it guides the reader comprehensively through each stage in conducting and interpreting regression analyses. An investigation into the effects of age on blood pressure forms the basis of all examples given to illustrate the use of different multilevel regression techniques; blood pressure is classified as a linear variable for linear regression, binary variable (high/low) for logistic regression, categorical variable (high/medium/low) for multinomial regression and with a count of risk factors to explain Poisson multilevel analysis. Focusing the reader on this single research question eases the transition between the different techniques covered.
For the majority of the book, output from MLwiN (multilevel analysis for Windows) is used to illustrate the use of techniques, explaining the use and interpretation of each line of output, and comparing ‘naïve’ to multilevel output so the reader is regularly reminded of why these techniques are being employed and what their effect is on results. Points of confusion are well anticipated, and the reader is directed back to earlier sections of the book to revise explanations. References to more detailed mathematical coverage are also given throughout, for those who desire a fuller explanation of any given point.
Final chapters include the application to analysis of longitudinal studies, an account of how to create prediction or association models, and calculating sample sizes for multilevel studies. The last chapter goes through the use of each multilevel regression technique in several statistical packages—SPSS, STATA, SAS and R—offering insights into the strengths and limitations of each, and which analyses they are most appropriate and user-friendly for.
While concentration is required to grasp some of the finer details of the methods described, Twisk succeeds in his aim of breaking down the statistical language barrier, and in a mere 168 (small) pages manages to convey complex mathematical ideas in a readily digestible manner.
CiteULike 
Connotea 
Del.icio.us What's this?
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||