Missing Data (Quantitative Applications in the Social Sciences)

Rating:
Summary: Fabulous primer on handling missing data
Review: As usual, Paul Allison has produced an accessible and practical treatment of conceptual and methodological issues that commonly confound social scientists. His discussion of the meaning, effects, and remedies for missing data is thorough and clear. In particular, the section on multiple imputation is extremely well-done.

This is a reference work that will improve the scholarship of even the most rigorous researcher, and yet can serve as a wonderful introductory text on the subject of missing data for students at many levels.

Rating:
Summary: Fabulous primer on handling missing data
Review: As usual, Paul Allison has produced an accessible and practical treatment of conceptual and methodological issues that commonly confound social scientists. His discussion of the meaning, effects, and remedies for missing data is thorough and clear. In particular, the section on multiple imputation is extremely well-done.

This is a reference work that will improve the scholarship of even the most rigorous researcher, and yet can serve as a wonderful introductory text on the subject of missing data for students at many levels.

Rating:
Summary: Dealing with an ugly problem
Review: Beginning stats students never see the real world of dirty data. They imagine that everyone responds fully to their surveys, and that every experiment yields legible results. Oh, for such a simple world.

Allison deals with the harsh reality of incomplete data sets. The book starts with a brief description of techniques that drop incomplete data from analysis. The large majority of the book, however, discusses ways to fill in the blanks.

The author rightly points out that "imputation", or creating values to replace what's missing, is not to be taken lightly. He gives techniques, each suited to the statistical character of some set of problems, and each matched to some technique for analysis. The mathematical goal is to create proxy values that won't upset the outcome of analysis.

That is quite a bit different from finding values that represent reality. Even though imputation is supposed to be mathematically innocuous, faking experimental data seems almost immoral to me. My data sets are about as dirty as any around. Also, they have the opposite of usual form: instead of a few dozen measurements on large numbers of samples, they have thousands of measurements on relatively few individuals. I have not convinced myself that Allison's manipulations are valid in this case. I would have been grateful for more discussion of techniques for stepping around the dropouts, and for statistically deciding whether I can ignore them.

Still, this book has worthwhile content. It's brief, clear, and informative about a very important topic. I will refer back to it, but maybe not the way the author intended.