Semiparametric theory and missing data. (Q2494657): Difference between revisions

Semiparametric models (SMs) are statistical models with finite-dimensional parameters of interest and infinite-dimensional nuisance parameters. The book under review deals with estimation for SMs with missing, coarsened, and censored data. Much of the theory is taken from a series of papers by Robins and Rotnitzky (together with colleagues) (1986--2001). The main topics of the monograph are missing at random and coarsening at random. Imputation methods and inverse probability weighted methods due to \textit{J. M. Robins} et al. [ J. Am. Stat. Assoc. 89, No. 427, 846--866 (1994; Zbl 0815.62043)] are studied in detail. The first five chapters consider SMs when missing data are absent. The semiparametric theory is developed from a geometric perspective, based on the geometry of Hilbert spaces. Missing-data mechanisms are discussed in Ch. 6 and extended to the broader class of coarsened data in Ch.~7. In Chs.~7--11 various efficient and robust estimators of parameters with missing and coarsened data are constructed, based on geometric ideas. The proposed methods can be computationally expensive, therefore in Ch. 12 some approximate methods are developed, which are easier to implement. Chap. 13 applies the methods to estimation of causal treatment effects in a point exposure study. The first chapter investigates multiple-imputation estimators for finite-dimensional parametric models. The book is very clearly and informally written. The exposition is instructive and rigorous enough. There are many important examples, oriented to biomedical applications. The monograph will be useful for graduate and post-graduate students in statistics and biostatistics, as well as researchers in statistics and survival analysis. Contents: Ch.~1, Introduction to semiparametric models; Ch.~2, Hilbert spaces for random vectors; Ch.~3, The geometry of influence functions; Ch.~4, Semiparametric models; Ch.~5, Other examples of semiparametric models; Ch.~6, Models and methods for missing data; Ch.~7, Missing and coarsening at random semiprametric models; Ch.~8, The nuisance tangent space and its orthogonal complement; Ch.~9, Augmented inverse probability weighted complete-case estimators; Ch. 10, Improving efficiency and double robustness with coarsened data; Ch.~11, Locally efficient estimators for coarsened-data semiparametric models; Ch.~12, Approximate methods for gaining efficiency; Ch.~13, Double-robust estimators of the average causal treatment effect; Ch.~14, Multiple imputation: a frequentist perspective. Note of the reviewer: A mistake in explanation is made on pp. 30, 55, and 109: for a vector function, an expansion with single intermediate point need not exist. In (7.2) the density should contain the absolute value of the Jacobian determinant.