Preface to second edition
Welcome to the second edition of Flexible Imputation of Missing Data, a book that can help you to solve missing data problems using
mice. I am tremendously grateful for the success of the first edition. The
mice community has been steadily growing over the last years, and the
mice package is now downloaded from CRAN at a rate of about 750 downloads per day. My hope is that this book will sharpen your intuition on how to think about missing data, and provide you the tools to apply your ideas in practice.
Since the first edition was published in 2012, multiple imputation of missing data has become one of the great academic industries. Many analysts now employ multiple imputation on a regular basis as a generic solution to the omnipresent missing-data problem, and a substantial group of practitioners are doing the calculations in
mice. This book aspires to combine a state-of-the-art overview of the field with a set of how-to instructions for practical data analysis.
Some sections of the first edition are still perfectly fine, but many others appear outdated. And of course, some of the newer developments are missing from the first edition. The second edition brings the text up-to-date. So what’s new?
Multiple imputation of multilevel data has been a hot spot of statistical research. Multilevel data can arise from nested data collection designs, but also emerge when data are combined from multiple sources. Imputers and analysts now have a bewildering array of options. The three pages in the first edition have expanded into a full-blown new chapter on multilevel imputation. This chapter translates the current insights among the leading developers into practical advice for end users.
Another hot spot in statistics and data science is the creation of personalized estimates. Causal inference by multiple imputation of the potential outcomes is an innovative approach that attempts to answer “what if” questions on the level of the individual. This edition contains a short new chapter on individual causal effects that demonstrates how multiple imputation is applied to obtain well-grounded personalized estimates.
Data science has continued to grow at a phenomenal pace. The
Rlanguage is now the dominant software for developing new statistical techniques.
RStudiohas successfully introduced the open
tidyverseecosystem for data acquisition, organization, analysis and visualization. This edition targets this growing audience of data scientists by including new sections on parallel computation and MICE workflows using pipes within the
New algorithms for creating imputations have appeared, in particular methodology based on predictive mean matching, for imputing binary and ordered variables, for interactions using classification and regression trees, and many types of machine learning methods. Chapters 3 and 4 incorporate these developments.
Important theoretical advances have been made on the convergence, compatibility, misspecification and stability of the simulation algorithms underlying MICE. Chapter 4 in this edition highlights these developments.
In parallel to the book, I worked on a significant update of the software:
mice 3.0. The main MICE algorithm now iterates over blocks of variables instead of individual columns, so we may now easily combine univariate and multivariate imputation methods. In addition, it is now possible to specify exactly which cells in the data should be imputed. There are new functions for multivariate tests, the support for native formula’s has improved, and, thanks to the
broom package, parameter pooling is now available for a much wider selection of complete-data models. The calculations use better numerical algorithms for low-level imputation functions. I have tried hard to remain code-compatible with previous versions of
mice. Existing code should run properly, but do not expect exact replication of the results. All code used in this book was tested with
The previous edition had two colors, and some of the plots did not work as well as I had intended. I am very glad that this edition is in full color, so that the differences between the blue and red points stand out clearly and provide a unifying look to the book. There is also syntax coloring of the
R code, which makes it very easy to distinguish the various language elements.
All data are incomplete, and so are all books. I had the luxury that I could devote my time during the period December 2017-March 2018 to this revision. A block of four months may seem like a formidable amount of time, but in retrospect it passed very quickly. While some topics I had planned have remained in the conceptual stage, overall I think that this edition covers the relevant developments in the field.
New statistical techniques will only be applied if there is high-quality and user-friendly software available. I would like to thank the following people for their contribution to the
mice package over the years: Karin Groothuis-Oudshoorn, Gerko Vink, Lisa Doove, Shahab Jolani, Roel de Jong, Rianne Schouten, Florian Meinfelder, Philipp Gaffert, Alexander Robitzsch and Bernie Gray.
There is a growing ecosystem of related
R packages that extend the functionality of
mice in some way. Currently, these include
ImputeRobust. There is also a
Python version in the works, which could result in an enormous expansion of the user base. I thank the authors of these packages for the time and effort they have put into creating these programs: Alexander Robitzsch, Simon Grund, Thorsten Henke, Oliver Lüdtke, Vincent Audigier, Matthieu Resche-Rigon, Kristian Kleinke, Jost Reinecke, Anoop Shah, Tobias Schumacher, Philipp Gaffert, Daniel Salfran, Martin Spiess, Sergey Feldman and Rianne Schouten.
I wish to thank Rob Calver, Statistics Editor at Chapman & Hall/CRC for his encouragement during both the first and second edition. Lara Spieker, Suzanne Lassandro and Shashi Kumar have been very helpful in meeting the ambitious production schedule. I thank Daan Kloet of TNO for his support for the idea of a mini-sabbatical, and his assistence in realizing the idea within TNO. I also wish to thank Peter van der Heijden of the University of Utrecht for his support over the years. Several people read and commented on parts of the manuscript. I thank Gerko Vink, Shahab Jolani, Iris Eekhout, Simon Grund, Tom Snijders and Joop Hox for their insightful and useful feedback. This has helped me a lot to understand the details much clearer, allowing me to improve my fumbled writings.
Last but not least, I thank my wife Eveline for her patience in living with an individual who can be so preoccupied with something else.