- •brief contents
- •contents
- •preface
- •acknowledgments
- •about this book
- •What’s new in the second edition
- •Who should read this book
- •Roadmap
- •Advice for data miners
- •Code examples
- •Code conventions
- •Author Online
- •About the author
- •about the cover illustration
- •1 Introduction to R
- •1.2 Obtaining and installing R
- •1.3 Working with R
- •1.3.1 Getting started
- •1.3.2 Getting help
- •1.3.3 The workspace
- •1.3.4 Input and output
- •1.4 Packages
- •1.4.1 What are packages?
- •1.4.2 Installing a package
- •1.4.3 Loading a package
- •1.4.4 Learning about a package
- •1.5 Batch processing
- •1.6 Using output as input: reusing results
- •1.7 Working with large datasets
- •1.8 Working through an example
- •1.9 Summary
- •2 Creating a dataset
- •2.1 Understanding datasets
- •2.2 Data structures
- •2.2.1 Vectors
- •2.2.2 Matrices
- •2.2.3 Arrays
- •2.2.4 Data frames
- •2.2.5 Factors
- •2.2.6 Lists
- •2.3 Data input
- •2.3.1 Entering data from the keyboard
- •2.3.2 Importing data from a delimited text file
- •2.3.3 Importing data from Excel
- •2.3.4 Importing data from XML
- •2.3.5 Importing data from the web
- •2.3.6 Importing data from SPSS
- •2.3.7 Importing data from SAS
- •2.3.8 Importing data from Stata
- •2.3.9 Importing data from NetCDF
- •2.3.10 Importing data from HDF5
- •2.3.11 Accessing database management systems (DBMSs)
- •2.3.12 Importing data via Stat/Transfer
- •2.4 Annotating datasets
- •2.4.1 Variable labels
- •2.4.2 Value labels
- •2.5 Useful functions for working with data objects
- •2.6 Summary
- •3 Getting started with graphs
- •3.1 Working with graphs
- •3.2 A simple example
- •3.3 Graphical parameters
- •3.3.1 Symbols and lines
- •3.3.2 Colors
- •3.3.3 Text characteristics
- •3.3.4 Graph and margin dimensions
- •3.4 Adding text, customized axes, and legends
- •3.4.1 Titles
- •3.4.2 Axes
- •3.4.3 Reference lines
- •3.4.4 Legend
- •3.4.5 Text annotations
- •3.4.6 Math annotations
- •3.5 Combining graphs
- •3.5.1 Creating a figure arrangement with fine control
- •3.6 Summary
- •4 Basic data management
- •4.1 A working example
- •4.2 Creating new variables
- •4.3 Recoding variables
- •4.4 Renaming variables
- •4.5 Missing values
- •4.5.1 Recoding values to missing
- •4.5.2 Excluding missing values from analyses
- •4.6 Date values
- •4.6.1 Converting dates to character variables
- •4.6.2 Going further
- •4.7 Type conversions
- •4.8 Sorting data
- •4.9 Merging datasets
- •4.9.1 Adding columns to a data frame
- •4.9.2 Adding rows to a data frame
- •4.10 Subsetting datasets
- •4.10.1 Selecting (keeping) variables
- •4.10.2 Excluding (dropping) variables
- •4.10.3 Selecting observations
- •4.10.4 The subset() function
- •4.10.5 Random samples
- •4.11 Using SQL statements to manipulate data frames
- •4.12 Summary
- •5 Advanced data management
- •5.2 Numerical and character functions
- •5.2.1 Mathematical functions
- •5.2.2 Statistical functions
- •5.2.3 Probability functions
- •5.2.4 Character functions
- •5.2.5 Other useful functions
- •5.2.6 Applying functions to matrices and data frames
- •5.3 A solution for the data-management challenge
- •5.4 Control flow
- •5.4.1 Repetition and looping
- •5.4.2 Conditional execution
- •5.5 User-written functions
- •5.6 Aggregation and reshaping
- •5.6.1 Transpose
- •5.6.2 Aggregating data
- •5.6.3 The reshape2 package
- •5.7 Summary
- •6 Basic graphs
- •6.1 Bar plots
- •6.1.1 Simple bar plots
- •6.1.2 Stacked and grouped bar plots
- •6.1.3 Mean bar plots
- •6.1.4 Tweaking bar plots
- •6.1.5 Spinograms
- •6.2 Pie charts
- •6.3 Histograms
- •6.4 Kernel density plots
- •6.5 Box plots
- •6.5.1 Using parallel box plots to compare groups
- •6.5.2 Violin plots
- •6.6 Dot plots
- •6.7 Summary
- •7 Basic statistics
- •7.1 Descriptive statistics
- •7.1.1 A menagerie of methods
- •7.1.2 Even more methods
- •7.1.3 Descriptive statistics by group
- •7.1.4 Additional methods by group
- •7.1.5 Visualizing results
- •7.2 Frequency and contingency tables
- •7.2.1 Generating frequency tables
- •7.2.2 Tests of independence
- •7.2.3 Measures of association
- •7.2.4 Visualizing results
- •7.3 Correlations
- •7.3.1 Types of correlations
- •7.3.2 Testing correlations for significance
- •7.3.3 Visualizing correlations
- •7.4 T-tests
- •7.4.3 When there are more than two groups
- •7.5 Nonparametric tests of group differences
- •7.5.1 Comparing two groups
- •7.5.2 Comparing more than two groups
- •7.6 Visualizing group differences
- •7.7 Summary
- •8 Regression
- •8.1 The many faces of regression
- •8.1.1 Scenarios for using OLS regression
- •8.1.2 What you need to know
- •8.2 OLS regression
- •8.2.1 Fitting regression models with lm()
- •8.2.2 Simple linear regression
- •8.2.3 Polynomial regression
- •8.2.4 Multiple linear regression
- •8.2.5 Multiple linear regression with interactions
- •8.3 Regression diagnostics
- •8.3.1 A typical approach
- •8.3.2 An enhanced approach
- •8.3.3 Global validation of linear model assumption
- •8.3.4 Multicollinearity
- •8.4 Unusual observations
- •8.4.1 Outliers
- •8.4.3 Influential observations
- •8.5 Corrective measures
- •8.5.1 Deleting observations
- •8.5.2 Transforming variables
- •8.5.3 Adding or deleting variables
- •8.5.4 Trying a different approach
- •8.6 Selecting the “best” regression model
- •8.6.1 Comparing models
- •8.6.2 Variable selection
- •8.7 Taking the analysis further
- •8.7.1 Cross-validation
- •8.7.2 Relative importance
- •8.8 Summary
- •9 Analysis of variance
- •9.1 A crash course on terminology
- •9.2 Fitting ANOVA models
- •9.2.1 The aov() function
- •9.2.2 The order of formula terms
- •9.3.1 Multiple comparisons
- •9.3.2 Assessing test assumptions
- •9.4 One-way ANCOVA
- •9.4.1 Assessing test assumptions
- •9.4.2 Visualizing the results
- •9.6 Repeated measures ANOVA
- •9.7 Multivariate analysis of variance (MANOVA)
- •9.7.1 Assessing test assumptions
- •9.7.2 Robust MANOVA
- •9.8 ANOVA as regression
- •9.9 Summary
- •10 Power analysis
- •10.1 A quick review of hypothesis testing
- •10.2 Implementing power analysis with the pwr package
- •10.2.1 t-tests
- •10.2.2 ANOVA
- •10.2.3 Correlations
- •10.2.4 Linear models
- •10.2.5 Tests of proportions
- •10.2.7 Choosing an appropriate effect size in novel situations
- •10.3 Creating power analysis plots
- •10.4 Other packages
- •10.5 Summary
- •11 Intermediate graphs
- •11.1 Scatter plots
- •11.1.3 3D scatter plots
- •11.1.4 Spinning 3D scatter plots
- •11.1.5 Bubble plots
- •11.2 Line charts
- •11.3 Corrgrams
- •11.4 Mosaic plots
- •11.5 Summary
- •12 Resampling statistics and bootstrapping
- •12.1 Permutation tests
- •12.2 Permutation tests with the coin package
- •12.2.2 Independence in contingency tables
- •12.2.3 Independence between numeric variables
- •12.2.5 Going further
- •12.3 Permutation tests with the lmPerm package
- •12.3.1 Simple and polynomial regression
- •12.3.2 Multiple regression
- •12.4 Additional comments on permutation tests
- •12.5 Bootstrapping
- •12.6 Bootstrapping with the boot package
- •12.6.1 Bootstrapping a single statistic
- •12.6.2 Bootstrapping several statistics
- •12.7 Summary
- •13 Generalized linear models
- •13.1 Generalized linear models and the glm() function
- •13.1.1 The glm() function
- •13.1.2 Supporting functions
- •13.1.3 Model fit and regression diagnostics
- •13.2 Logistic regression
- •13.2.1 Interpreting the model parameters
- •13.2.2 Assessing the impact of predictors on the probability of an outcome
- •13.2.3 Overdispersion
- •13.2.4 Extensions
- •13.3 Poisson regression
- •13.3.1 Interpreting the model parameters
- •13.3.2 Overdispersion
- •13.3.3 Extensions
- •13.4 Summary
- •14 Principal components and factor analysis
- •14.1 Principal components and factor analysis in R
- •14.2 Principal components
- •14.2.1 Selecting the number of components to extract
- •14.2.2 Extracting principal components
- •14.2.3 Rotating principal components
- •14.2.4 Obtaining principal components scores
- •14.3 Exploratory factor analysis
- •14.3.1 Deciding how many common factors to extract
- •14.3.2 Extracting common factors
- •14.3.3 Rotating factors
- •14.3.4 Factor scores
- •14.4 Other latent variable models
- •14.5 Summary
- •15 Time series
- •15.1 Creating a time-series object in R
- •15.2 Smoothing and seasonal decomposition
- •15.2.1 Smoothing with simple moving averages
- •15.2.2 Seasonal decomposition
- •15.3 Exponential forecasting models
- •15.3.1 Simple exponential smoothing
- •15.3.3 The ets() function and automated forecasting
- •15.4 ARIMA forecasting models
- •15.4.1 Prerequisite concepts
- •15.4.2 ARMA and ARIMA models
- •15.4.3 Automated ARIMA forecasting
- •15.5 Going further
- •15.6 Summary
- •16 Cluster analysis
- •16.1 Common steps in cluster analysis
- •16.2 Calculating distances
- •16.3 Hierarchical cluster analysis
- •16.4 Partitioning cluster analysis
- •16.4.2 Partitioning around medoids
- •16.5 Avoiding nonexistent clusters
- •16.6 Summary
- •17 Classification
- •17.1 Preparing the data
- •17.2 Logistic regression
- •17.3 Decision trees
- •17.3.1 Classical decision trees
- •17.3.2 Conditional inference trees
- •17.4 Random forests
- •17.5 Support vector machines
- •17.5.1 Tuning an SVM
- •17.6 Choosing a best predictive solution
- •17.7 Using the rattle package for data mining
- •17.8 Summary
- •18 Advanced methods for missing data
- •18.1 Steps in dealing with missing data
- •18.2 Identifying missing values
- •18.3 Exploring missing-values patterns
- •18.3.1 Tabulating missing values
- •18.3.2 Exploring missing data visually
- •18.3.3 Using correlations to explore missing values
- •18.4 Understanding the sources and impact of missing data
- •18.5 Rational approaches for dealing with incomplete data
- •18.6 Complete-case analysis (listwise deletion)
- •18.7 Multiple imputation
- •18.8 Other approaches to missing data
- •18.8.1 Pairwise deletion
- •18.8.2 Simple (nonstochastic) imputation
- •18.9 Summary
- •19 Advanced graphics with ggplot2
- •19.1 The four graphics systems in R
- •19.2 An introduction to the ggplot2 package
- •19.3 Specifying the plot type with geoms
- •19.4 Grouping
- •19.5 Faceting
- •19.6 Adding smoothed lines
- •19.7 Modifying the appearance of ggplot2 graphs
- •19.7.1 Axes
- •19.7.2 Legends
- •19.7.3 Scales
- •19.7.4 Themes
- •19.7.5 Multiple graphs per page
- •19.8 Saving graphs
- •19.9 Summary
- •20 Advanced programming
- •20.1 A review of the language
- •20.1.1 Data types
- •20.1.2 Control structures
- •20.1.3 Creating functions
- •20.2 Working with environments
- •20.3 Object-oriented programming
- •20.3.1 Generic functions
- •20.3.2 Limitations of the S3 model
- •20.4 Writing efficient code
- •20.5 Debugging
- •20.5.1 Common sources of errors
- •20.5.2 Debugging tools
- •20.5.3 Session options that support debugging
- •20.6 Going further
- •20.7 Summary
- •21 Creating a package
- •21.1 Nonparametric analysis and the npar package
- •21.1.1 Comparing groups with the npar package
- •21.2 Developing the package
- •21.2.1 Computing the statistics
- •21.2.2 Printing the results
- •21.2.3 Summarizing the results
- •21.2.4 Plotting the results
- •21.2.5 Adding sample data to the package
- •21.3 Creating the package documentation
- •21.4 Building the package
- •21.5 Going further
- •21.6 Summary
- •22 Creating dynamic reports
- •22.1 A template approach to reports
- •22.2 Creating dynamic reports with R and Markdown
- •22.3 Creating dynamic reports with R and LaTeX
- •22.4 Creating dynamic reports with R and Open Document
- •22.5 Creating dynamic reports with R and Microsoft Word
- •22.6 Summary
- •afterword Into the rabbit hole
- •appendix A Graphical user interfaces
- •appendix B Customizing the startup environment
- •appendix C Exporting data from R
- •Delimited text file
- •Excel spreadsheet
- •Statistical applications
- •appendix D Matrix algebra in R
- •appendix E Packages used in this book
- •appendix F Working with large datasets
- •F.1 Efficient programming
- •F.2 Storing data outside of RAM
- •F.3 Analytic packages for out-of-memory data
- •F.4 Comprehensive solutions for working with enormous datasets
- •appendix G Updating an R installation
- •G.1 Automated installation (Windows only)
- •G.2 Manual installation (Windows and Mac OS X)
- •G.3 Updating an R installation (Linux)
- •references
- •index
- •Symbols
- •Numerics
- •23.1 The lattice package
- •23.2 Conditioning variables
- •23.3 Panel functions
- •23.4 Grouping variables
- •23.5 Graphic parameters
- •23.6 Customizing plot strips
- •23.7 Page arrangement
- •23.8 Going further
Advanced methods for missing data
This chapter covers
■Identifying missing data
■Visualizing missing data patterns
■Complete-case analysis
■Multiple imputation of missing data
In previous chapters, we focused on analyzing complete datasets (that is, datasets without missing values). Although doing so helps simplify the presentation of statistical and graphical methods, in the real world, missing data are ubiquitous.
In some ways, the impact of missing data is a subject that most of us want to avoid. Statistics books may not mention it or may limit discussion to a few paragraphs. Statistical packages offer automatic handling of missing data using methods that may not be optimal. Even though most data analyses (at least in social sciences) involve missing data, this topic is rarely mentioned in the methods and results sections of journal articles. Given how often missing values occur, and the degree to which their presence can invalidate study results, it’s fair to say that the subject has received insufficient attention outside of specialized books and courses.
414
Steps in dealing with missing data |
415 |
Data can be missing for many reasons. Survey participants may forget to answer one or more questions, refuse to answer sensitive questions, or grow fatigued and fail to complete a long questionnaire. Study participants may miss appointments or drop out of a study prematurely. Recording equipment may fail, internet connections may be lost, or data may be miscoded. Analysts may even plan for some data to be missing. For example, to increase study efficiency or reduce costs, you may choose not to collect all data from all participants. Finally, data may be lost for reasons that you’re never able to ascertain.
Unfortunately, most statistical methods assume that you’re working with complete matrices, vectors, and data frames. In most cases, you have to eliminate missing data before you address the substantive questions that led you to collect the data. You can eliminate missing data by (1) removing cases with missing data or (2) replacing missing data with reasonable substitute values. In either case, the end result is a dataset without missing values.
In this chapter, we’ll look at both traditional and modern approaches for dealing with missing data. We’ll primarily use the VIM and mice packages. The command install.packages(c("VIM", "mice")) will download and install both.
To motivate the discussion, we’ll look at the mammal sleep dataset (sleep) provided in the VIM package (not to be confused with the sleep dataset describing the impact of drugs on sleep provided in the base installation). The data come from a study by Allison and Chichetti (1976) that examined the relationship between sleep and ecological and constitutional variables for 62 mammal species. The authors were interested in why animals’ sleep requirements vary from species to species. The sleep variables served as the dependent variables, whereas the ecological and constitutional variables served as the independent or predictor variables.
Sleep variables included length of dreaming sleep (Dream), nondreaming sleep (NonD), and their sum (Sleep). The constitutional variables included body weight in kilograms (BodyWgt), brain weight in grams (BrainWgt), life span in years (Span), and gestation time in days (Gest). The ecological variables included degree to which species were preyed upon (Pred), degree of their exposure while sleeping (Exp), and overall danger (Danger) faced. The ecological variables were measured on 5-point rating scales that ranged from 1 (low) to 5 (high).
In their original article, Allison and Chichetti limited their analyses to the species that had complete data. We’ll go further, analyzing all 62 cases using a multiple imputation approach.
18.1 Steps in dealing with missing data
If you’re new to the study of missing data, you’ll find a bewildering array of approaches, critiques, and methodologies. The classic text in this area is Little and Rubin (2002). Excellent, accessible reviews can be found in Allison (2001); Schafer
416 |
CHAPTER 18 Advanced methods for missing data |
and Graham (2002); and Schlomer, Bauman, and Card (2010). A comprehensive approach usually includes the following steps:
1Identify the missing data.
2Examine the causes of the missing data.
3Delete the cases containing missing data, or replace (impute) the missing values with reasonable alternative data values.
Unfortunately, identifying missing data is usually the only unambiguous step. Learning why data are missing depends on your understanding of the processes that generated the data. Deciding how to treat missing values will depend on your estimation of which procedures will produce the most reliable and accurate results.
A classification system for missing data
Statisticians typically classify missing data into one of three types. These types are usually described in probabilistic terms, but the underlying ideas are straightforward. We’ll use the measurement of dreaming in the sleep study (where 12 animals have missing values) to illustrate each type in turn:
■Missing completely at random—If the presence of missing data on a variable is unrelated to any other observed or unobserved variable, then the data are missing completely at random (MCAR). If there’s no systematic reason why dream sleep is missing for these 12 animals, the data are said to be MCAR. Note that if every variable with missing data is MCAR, you can consider the complete cases to be a simple random sample from the larger dataset.
■Missing at random—If the presence of missing data on a variable is related to other observed variables but not to its own unobserved value, the data are missing at random (MAR). For example, if animals with smaller body weights are more likely to have missing values for dream sleep (perhaps because it’s harder to observe smaller animals), and the “missingness” is unrelated to an animal’s time spent dreaming, the data are considered MAR. In this case, the presence or absence of dream sleep data is random, once you control for body weight.
■Not missing at random—If the missing data for a variable are neither MCAR nor MAR, the data are not missing at random (NMAR). For example, if animals that spend less time dreaming are also more likely to have a missing dream value (perhaps because it’s harder to measure shorter events), the data are considered NMAR.
Most approaches to missing data assume that the data are either MCAR or MAR. In this case, you can ignore the mechanism producing the missing data and (after replacing or deleting the missing data) model the relationships of interest directly.
Data that are NMAR can be difficult to analyze properly. When data are NMAR, you have to model the mechanisms that produced the missing values, as well as the relationships of interest. (Current approaches to analyzing NMAR data include the use of selection models and pattern mixtures. The analysis of NMAR data can be complex and is beyond the scope of this book.)
|
|
|
|
|
Identifying missing values |
417 |
||||||||||||
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
Identify Missing Values |
|
|
|
|
|
|
|
||||
|
|
|
|
|
|
|
|
|
is.na() |
|
|
|
|
|
|
|
||
|
|
|
|
|
|
|
!complete.cases() |
|
|
|
|
|
|
|
||||
|
|
|
|
|
|
|
VIM package |
|
|
|
|
|
|
|
||||
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
||
|
Delete Missing Values |
|
|
|
Maximum Likelihood |
|
|
|
Impute Missing Values |
|
|
|||||||
|
|
|
|
|
|
|
|
Estimation |
|
|
|
|
|
|
|
|||
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|||||
|
|
|
|
|
|
|
mvmle package |
|
|
|
|
|
|
|
||||
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|||||
|
|
|
|
|
|
|
|
|
|
|||||||||
Casewise (Listwise) |
|
Available Case |
|
|
Single (simple) |
|
|
Multiple Imputation |
|
|||||||||
|
omit.na() |
|
|
(Pairwise) |
|
|
Imputation |
|
|
mi package |
|
|||||||
|
|
|
Option available for |
|
|
Hmisc Package |
|
|
mice package |
|
||||||||
|
|
|
|
|
|
|
amelia package |
|
||||||||||
|
|
|
|
some functions |
|
|
|
|
|
|
|
|
|
|||||
|
|
|
|
|
|
|
|
|
|
|
||||||||
|
|
|
|
|
|
|
|
|
|
|
|
mitools package |
|
|||||
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
||
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|||||
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
Figure 18.1 Methods for handling incomplete data, along with the R packages that support them
There are many methods for dealing with missing data—and no guarantee that they’ll produce the same results. Figure 18.1 describes an array of methods used for handling incomplete data and the R packages that support them.
A complete review of missing-data methodologies would require a book in itself. In this chapter, we’ll review methods for exploring missing-values patterns and focus on the three most popular methods for dealing with incomplete data (a rational approach, listwise deletion, and multiple imputation). We’ll end the chapter with a brief discussion of other methods, including those that are useful in special circumstances.
18.2 Identifying missing values
To begin, let’s review the material introduced in section 4.5, and expand on it. R represents missing values using the symbol NA (not available) and impossible values using the symbol NaN (not a number). In addition, the symbols Inf and -Inf represent positive infinity and negative infinity, respectively. The functions is.na(), is.nan(), and is.infinite() can be used to identify missing, impossible, and infinite values, respectively. Each returns either TRUE or FALSE. Examples are given in table 18.1.
Table 18.1 Examples of return values for the is.na(), is.nan(), and is.infinite() functions
|
|
|
|
x |
|
is.na(x) |
|
is.nan(x) |
is.infinite(x) |
|
|
|
|
|
|
|
|
||
x <- NA |
|
|
TRUE |
|
FALSE |
|
FALSE |
||
x |
<- |
0 |
/ |
0 |
TRUE |
|
TRUE |
|
FALSE |
x |
<- |
1 |
/ |
0 |
FALSE |
|
FALSE |
|
TRUE |
|
|
|
|
|
|
|
|
|
|