Update cookies preferences

E-book: Flexible Imputation of Missing Data, Second Edition 2nd edition [Taylor & Francis e-book]

(TNO Quality of Life, Leiden, The Netherlands)
Other books in subject:
  • Taylor & Francis e-book
  • Price: 133,87 €*
  • * this price gives unlimited concurrent access for unlimited time
  • Regular price: 191,24 €
  • Save 30%
Flexible Imputation of Missing Data, Second Edition 2nd editionLarger Image
Other books in subject:
Taylor & Francis e-booksMore info about Taylor & Francis e-books
Companion website: https://www.taylorfrancis.com/books/9780429492259

This is the second edition of a popular book on multiple imputation, focused on explaining the application of methods through detailed worked examples using the MICE package as developed by the author. This new edition incorporates the recent developments in this fast-moving field.



Missing data pose challenges to real-life data analysis. Simple ad-hoc fixes, like deletion or mean imputation, only work under highly restrictive conditions, which are often not met in practice. Multiple imputation replaces each missing value by multiple plausible values. The variability between these replacements reflects our ignorance of the true (but missing) value. Each of the completed data set is then analyzed by standard methods, and the results are pooled to obtain unbiased estimates with correct confidence intervals. Multiple imputation is a general approach that also inspires novel solutions to old problems by reformulating the task at hand as a missing-data problem.





This is the second edition of a popular book on multiple imputation, focused on explaining the application of methods through detailed worked examples using the MICE package as developed by the author. This new edition incorporates the recent developments in this fast-moving field.



This class-tested book avoids mathematical and technical details as much as possible: formulas are accompanied by verbal statements that explain the formula in accessible terms. The book sharpens the reader’s intuition on how to think about missing data, and provides all the tools needed to execute a well-grounded quantitative analysis in the presence of missing data.

Foreword xv
Donald B. Rubin
Preface to second edition xvii
Preface to first edition xxi
About the author xxiii
List of symbols xxv
List of algorithms xxvii
I Basics 1(160)
1 Introduction
3(26)
1.1 The problem of missing data
3(5)
1.1.1 Current practice
3(3)
1.1.2 Changing perspective on missing data
6(2)
1.2 Concepts of MCAR, MAR and MNAR
8(1)
1.3 Ad-hoc solutions
9(10)
1.3.1 Listwise deletion
9(2)
1.3.2 Pairwise deletion
11(1)
1.3.3 Mean imputation
12(1)
1.3.4 Regression imputation
13(1)
1.3.5 Stochastic regression imputation
14(2)
1.3.6 LOCF and BOCF
16(1)
1.3.7 Indicator method
17(1)
1.3.8 Summary
18(1)
1.4 Multiple imputation in a nutshell
19(4)
1.4.1 Procedure
19(1)
1.4.2 Reasons to use multiple imputation
20(1)
1.4.3 Example of multiple imputation
21(2)
1.5 Goal of the book
23(1)
1.6 What the book does not cover
23(3)
1.6.1 Prevention
24(1)
1.6.2 Weighting procedures
24(1)
1.6.3 Likelihood-based approaches
25(1)
1.7 Structure of the book
26(1)
1.8 Exercises
26(3)
2 Multiple imputation
29(34)
2.1 Historic overview
29(4)
2.1.1 Imputation
29(1)
2.1.2 Multiple imputation
30(2)
2.1.3 The expanding literature on multiple imputation
32(1)
2.2 Concepts in incomplete data
33(8)
2.2.1 Incomplete-data perspective
33(1)
2.2.2 Causes of missing data
33(2)
2.2.3 Notation
35(1)
2.2.4 MCAR, MAR, and MNAR again
36(2)
2.2.5 Ignorable and nonignorable
38(1)
2.2.6 Implications of ignorability
39(2)
2.3 Why and when multiple imputation works
41(8)
2.3.1 Goal of multiple imputation
41(1)
2.3.2 Three sources of variation
41(3)
2.3.3 Proper imputation
44(2)
2.3.4 Scope of the imputation model
46(1)
2.3.5 Variance ratios
46(1)
2.3.6 Degrees of freedom
47(1)
2.3.7 Numerical example
48(1)
2.4 Statistical intervals and tests
49(2)
2.4.1 Scalar or multi-parameter inference?
49(1)
2.4.2 Scalar inference
50(1)
2.4.3 Numerical example
50(1)
2.5 How to evaluate imputation methods
51(4)
2.5.1 Simulation designs and performance measures
51(1)
2.5.2 Evaluation criteria
52(1)
2.5.3 Example
53(2)
2.6 Imputation is not prediction
55(2)
2.7 When not to use multiple imputation
57(1)
2.8 How many imputations?
58(3)
2.9 Exercises
61(2)
3 Univariate missing data
63(42)
3.1 How to generate multiple imputations
63(4)
3.1.1 Predict method
65(1)
3.1.2 Predict + noise method
65(1)
3.1.3 Predict + noise + parameter uncertainty
65(1)
3.1.4 A second predictor
66(1)
3.1.5 Drawing from the observed data
66(1)
3.1.6 Conclusion
66(1)
3.2 Imputation under the normal linear normal
67(7)
3.2.1 Overview
67(1)
3.2.2 Algorithms
67(2)
3.2.3 Performance
69(1)
3.2.4 Generating MAR missing data
70(2)
3.2.5 MAR missing data generation in multivariate data
72(1)
3.2.6 Conclusion
73(1)
3.3 Imputation under non-normal distributions
74(3)
3.3.1 Overview
74(1)
3.3.2 Imputation from the t-distribution
75(2)
3.4 Predictive mean matching
77(7)
3.4.1 Overview
77(2)
3.4.2 Computational details
79(2)
3.4.3 Number of donors
81(1)
3.4.4 Pitfalls
82(2)
3.4.5 Conclusion
84(1)
3.5 Classification and regression trees
84(3)
3.5.1 Overview
84(3)
3.6 Categorical data
87(4)
3.6.1 Generalized linear model
87(2)
3.6.2 Perfect prediction
89(1)
3.6.3 Evaluation
90(1)
3.7 Other data types
91(5)
3.7.1 Count data
91(1)
3.7.2 Semi-continuous data
92(1)
3.7.3 Censored, truncated and rounded data
93(3)
3.8 Nonignorable missing data
96(6)
3.8.1 Overview
96(1)
3.8.2 Selection model
97(1)
3.8.3 Pattern-mixture model
98(1)
3.8.4 Converting selection and pattern-mixture models
99(1)
3.8.5 Sensitivity analysis
100(1)
3.8.6 Role of sensitivity analysis
101(1)
3.8.7 Recent developments
102(1)
3.9 Exercises
102(3)
4 Multivariate missing data
105(34)
4.1 Missing data pattern
105(6)
4.1.1 Overview
105(2)
4.1.2 Summary statistics
107(2)
4.1.3 Influx and outflux
109(2)
4.2 Issues in multivariate imputation
111(1)
4.3 Monotone data imputation
112(3)
4.3.1 Overview
112(1)
4.3.2 Algorithm
113(2)
4.4 Joint modeling
115(4)
4.4.1 Overview
115(1)
4.4.2 Continuous data
115(2)
4.4.3 Categorical data
117(2)
4.5 Fully conditional specification
119(11)
4.5.1 Overview
119(1)
4.5.2 The MICE algorithm
120(2)
4.5.3 Compatibility
122(2)
4.5.4 Congeniality or compatibility?
124(1)
4.5.5 Model-based and data-based imputation
125(1)
4.5.6 Number of iterations
126(1)
4.5.7 Example of slow convergence
126(3)
4.5.8 Performance
129(1)
4.6 FCS and JM
130(5)
4.6.1 Relations between FCS and JM
130(1)
4.6.2 Comparisons
130(1)
4.6.3 Illustration
131(4)
4.7 MICE extensions
135(2)
4.7.1 Skipping imputations and overimputation
135(1)
4.7.2 Blocks of variables, hybrid imputation
135(1)
4.7.3 Blocks of units, monotone blocks
136(1)
4.7.4 Tile imputation
136(1)
4.8 Conclusion
137(1)
4.9 Exercises
137(2)
5 Analysis of imputed data
139(22)
5.1 Workflow
139(6)
5.1.1 Recommended workflows
140(2)
5.1.2 Not recommended workflow: Averaging the data
142(2)
5.1.3 Not recommended workflow: Stack imputed data
144(1)
5.1.4 Repeated analyses
144(1)
5.2 Parameter pooling
145(2)
5.2.1 Scalar inference of normal quantities
145(1)
5.2.2 Scalar inference of non-normal quantities
146(1)
5.3 Multi-parameter inference
147(6)
5.3.1 D1 Multivariate Wald test
147(2)
5.3.2 D2 Combining test statistics
149(1)
5.3.3 D3 Likelihood ratio test
150(2)
5.3.4 D1, D2 or D3?
152(1)
5.4 Stepwise model selection
153(4)
5.4.1 Variable selection techniques
153(1)
5.4.2 Computation
154(1)
5.4.3 Model optimism
155(2)
5.5 Parallel computation
157(1)
5.6 Conclusion
158(1)
5.7 Exercises
158(3)
II Advanced techniques 161(96)
6 Imputation in practice
163(34)
6.1 Overview of modeling choices
163(2)
6.2 Ignorable or nonignorable?
165(1)
6.3 Model form and predictors
166(4)
6.3.1 Model form
166(1)
6.3.2 Predictors
167(3)
6.4 Derived variables
170(14)
6.4.1 Ratio of two variables
170(5)
6.4.2 Interaction terms
175(1)
6.4.3 Quadratic relations
176(1)
6.4.4 Compositional data
177(4)
6.4.5 Sum scores
181(1)
6.4.6 Conditional imputation
182(2)
6.5 Algorithmic options
184(5)
6.5.1 Visit sequence
184(3)
6.5.2 Convergence
187(2)
6.6 Diagnostics
189(5)
6.6.1 Model fit versus distributional discrepancy
190(1)
6.6.2 Diagnostic graphs
190(4)
6.7 Conclusion
194(1)
6.8 Exercises
195(2)
7 Multilevel multiple imputation
197(44)
7.1 Introduction
197(1)
7.2 Notation for multilevel models
197(3)
7.3 Missing values in multilevel data
200(4)
7.3.1 Practical issues in multilevel imputation
201(1)
7.3.2 Ad-hoc solutions for multilevel data
202(1)
7.3.3 Likelihood solutions
203(1)
7.4 Multilevel imputation by joint modeling
204(1)
7.5 Multilevel imputation by fully conditional specification
205(2)
7.5.1 Add cluster means of predictors
206(1)
7.5.2 Model cluster heterogeneity
207(1)
7.6 Continuous outcome
207(7)
7.6.1 General principle
208(1)
7.6.2 Methods
209(1)
7.6.3 Example
209(5)
7.7 Discrete outcome
214(4)
7.7.1 Methods
214(1)
7.7.2 Example
215(3)
7.8 Imputation of level-2 variable
218(1)
7.9 Comparative work
219(1)
7.10 Guidelines and advice
220(20)
7.10.1 Intercept-only model, missing outcomes
222(1)
7.10.2 Random intercepts, missing level-1 predictor
222(2)
7.10.3 Random intercepts, contextual model
224(2)
7.10.4 Random intercepts, missing level-2 predictor
226(2)
7.10.5 Random intercepts, interactions
228(4)
7.10.6 Random slopes, missing outcomes and predictors
232(2)
7.10.7 Random slopes, interactions
234(4)
7.10.8 Recipes
238(2)
7.11 Future research
240(1)
8 Individual causal effects
241(16)
8.1 Need for individual causal effects
241(2)
8.2 Problem of causal inference
243(2)
8.3 Framework
245(1)
8.4 Generating imputations by FCS
246(8)
8.4.1 Naive FCS
246(1)
8.4.2 FCS with a prior for p
247(6)
8.4.3 Extensions
253(1)
8.5 Bibliographic notes
254(3)
III Case studies 257(80)
9 Measurement issues
259(36)
9.1 Too many columns
259(12)
9.1.1 Scientific question
260(1)
9.1.2 Leiden 85+ Cohort
260(1)
9.1.3 Data exploration
261(2)
9.1.4 Outflux
263(2)
9.1.5 Finding problems: loggedEvents
265(2)
9.1.6 Quick predictor selection: quickpred
267(1)
9.1.7 Generating the imputations
268(2)
9.1.8 A further improvement: Survival as predictor variable
270(1)
9.1.9 Some guidance
270(1)
9.2 Sensitivity analysis
271(6)
9.2.1 Causes and consequences of missing data
272(2)
9.2.2 Scenarios
274(1)
9.2.3 Generating imputations under the δ-adjustment
274(1)
9.2.4 Complete-data model
275(2)
9.2.5 Conclusion
277(1)
9.3 Correct prevalence estimates from self-reported data
277(6)
9.3.1 Description of the problem
277(1)
9.3.2 Don't count on predictions
278(2)
9.3.3 The main idea
280(1)
9.3.4 Data
281(1)
9.3.5 Application
281(2)
9.3.6 Conclusion
283(1)
9.4 Enhancing comparability
283(11)
9.4.1 Description of the problem
283(1)
9.4.2 Full dependence: Simple equating
284(2)
9.4.3 Independence: Imputation without a bridge study
286(2)
9.4.4 Fully dependent or independent?
288(1)
9.4.5 Imputation using a bridge study
289(3)
9.4.6 Interpretation
292(1)
9.4.7 Conclusion
293(1)
9.5 Exercises
294(1)
10 Selection issues
295(16)
10.1 Correcting for selective drop-out
295(7)
10.1.1 POPS study: 19 years follow-up
295(1)
10.1.2 Characterization of the drop-out
296(1)
10.1.3 Imputation model
296(3)
10.1.4 A solution "that does not look good"
299(2)
10.1.5 Results
301(1)
10.1.6 Conclusion
302(1)
10.2 Correcting for nonresponse
302(7)
10.2.1 Fifth Dutch Growth Study
303(1)
10.2.2 Nonresponse
303(1)
10.2.3 Comparison to known population totals
304(1)
10.2.4 Augmenting the sample
304(2)
10.2.5 Imputation model
306(1)
10.2.6 Influence of nonresponse on final height
307(1)
10.2.7 Discussion
308(1)
10.3 Exercises
309(2)
11 Longitudinal data
311(26)
11.1 Long and wide format
311(2)
11.2 SE Fireworks Disaster Study
313(7)
11.2.1 Intention to treat
314(1)
11.2.2 Imputation model
315(2)
11.2.3 Inspecting imputations
317(1)
11.2.4 Complete-data model
318(1)
11.2.5 Results from the complete-data model
319(1)
11.3 Time raster imputation
320(12)
11.3.1 Change score
321(1)
11.3.2 Scientific question: Critical periods
322(2)
11.3.3 Broken stick model
324(2)
11.3.4 Terneuzen Birth Cohort
326(2)
11.3.5 Shrinkage and the change score
328(1)
11.3.6 Imputation
328(2)
11.3.7 Complete-data model
330(2)
11.4 Conclusion
332(2)
11.5 Exercises
334(3)
IV Extensions 337(14)
12 Conclusion
339(12)
12.1 Some dangers, some do's and some don'ts
339(3)
12.1.1 Some dangers
339(1)
12.1.2 Some do's
340(1)
12.1.3 Some don'ts
341(1)
12.2 Reporting
342(3)
12.2.1 Reporting guidelines
343(1)
12.2.2 Template
344(1)
12.3 Other applications
345(2)
12.3.1 Synthetic datasets for data protection
345(1)
12.3.2 Analysis of coarsened data
345(1)
12.3.3 File matching of multiple datasets
346(1)
12.3.4 Planned missing data for efficient designs
346(1)
12.3.5 Adjusting for verification bias
347(1)
12.4 Future developments
347(2)
12.4.1 Derived variables
347(1)
12.4.2 Algorithms for blocks and batches
347(1)
12.4.3 Nested imputation
348(1)
12.4.4 Better trials with dynamic treatment regimes
348(1)
12.4.5 Distribution-free pooling rules
348(1)
12.4.6 Improved diagnostic techniques
349(1)
12.4.7 Building block in modular statistics
349(1)
12.5 Exercises
349(2)
References 351(42)
Author index 393(12)
Subject index 405
Stef van Buuren is a statistical consultant at the Netherlands Organisation for Applied Scientific Research TNO in Leiden with a broad knowledge of quantitative issues in public health. Since 2015, Van Buuren holds is the world's first Professor of Missing Data at the department of Methodology & Statistics, FSS, University of Utrecht. He is the originator of various new statistical tools.
Permanent link: https://www.kriso.lv/db/9780429492259_pe.html
Keywords: