ADA1 Notes F14

Lecture notes for
Advanced Data Analysis 1 (ADA1)

Stat 427/527
University of New Mexico
Erik B. Erhardt Edward J. Bedrick Ronald M. Schrader
Fall 2014
Contents
0 Introduction to R, Rstudio, and ggplot 1
0.1 Syllabus . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
0.2 Rstudio . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
0.3 R building blocks . . . . . . . . . . . . . . . . . . . . . . . . . 6
0.4 Plotting with ggplot2 . . . . . . . . . . . . . . . . . . . . . . 13
0.4.1 Improving plots . . . . . . . . . . . . . . . . . . . . . 20
0.5 Course Overview . . . . . . . . . . . . . . . . . . . . . . . . . 26
1 Summarizing and Displaying Data 27

1.1 Random variables . . . . . . . . . . . . . . . . . . . . . . . . 27
1.2 Numerical summaries . . . . . . . . . . . . . . . . . . . . . . 28
1.3 Graphical summaries for one quantitative sample . . . . . . . . 33
1.3.1 Dotplots . . . . . . . . . . . . . . . . . . . . . . . . . 34
1.3.2 Histogram . . . . . . . . . . . . . . . . . . . . . . . . 35
1.3.3 Stem-and-leaf plot . . . . . . . . . . . . . . . . . . . . 36
1.3.4 Boxplot or box-and-whiskers plot . . . . . . . . . . . . 38
1.4 Interpretation of Graphical Displays for Numerical Data . . . . 43
1.5 Interpretations for examples . . . . . . . . . . . . . . . . . . . 57
2 Estimation in One-Sample Problems 59

2.1 Inference for a population mean . . . . . . . . . . . . . . . . . 59
2.1.1 Standard error, LLN, and CLT . . . . . . . . . . . . . 60
2.1.2 t-distribution . . . . . . . . . . . . . . . . . . . . . . . 66
2.2 CI for µ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
2.2.1 Assumptions for procedures . . . . . . . . . . . . . . . 70
2.2.2 The effect of α on a two-sided CI . . . . . . . . . . . . 73
CONTENTS iii
2.3 Hypothesis Testing for µ . . . . . . . . . . . . . . . . . . . . . 73
2.3.1 P-values . . . . . . . . . . . . . . . . . . . . . . . . . 75
2.3.2 Assumptions for procedures . . . . . . . . . . . . . . . 77
2.3.3 The mechanics of setting up hypothesis tests . . . . . . 84
2.3.4 The effect of α on the rejection region of a two-sided test 86
2.4 Two-sided tests, CI and p-values . . . . . . . . . . . . . . . . 88
2.5 Statistical versus practical significance . . . . . . . . . . . . . 89
2.6 Design issues and power . . . . . . . . . . . . . . . . . . . . . 90
2.7 One-sided tests on µ . . . . . . . . . . . . . . . . . . . . . . . 91
2.7.1 One-sided CIs . . . . . . . . . . . . . . . . . . . . . . 95
3 Two-Sample Inferences 98
3.1 Comparing Two Sets of Measurements . . . . . . . . . . . . . 98
3.1.1 Plotting head breadth data: . . . . . . . . . . . . . . . 100
3.1.2 Salient Features to Notice . . . . . . . . . . . . . . . . 106
3.2 Two-Sample Methods: Paired Versus Independent Samples . . 106
3.3 Two Independent Samples: CI and Test Using Pooled Variance 108
3.4 Satterthwaite’s Method, unequal variances . . . . . . . . . . . 109
3.4.1 R Implementation . . . . . . . . . . . . . . . . . . . . 109
3.5 One-Sided Tests . . . . . . . . . . . . . . . . . . . . . . . . . 118
3.6 Paired Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . 119
3.6.1 R Analysis . . . . . . . . . . . . . . . . . . . . . . . . 120
3.7 Should You Compare Means? . . . . . . . . . . . . . . . . . . 128
4 Checking Assumptions 130

4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
4.2 Testing Normality . . . . . . . . . . . . . . . . . . . . . . . . 131
4.2.1 Normality tests on non-normal data . . . . . . . . . . . 134
4.3 Formal Tests of Normality . . . . . . . . . . . . . . . . . . . . 138
4.4 Testing Equal Population Variances . . . . . . . . . . . . . . . 146
4.5 Small sample sizes, a comment . . . . . . . . . . . . . . . . . 148
5 One-Way Analysis of Variance 150

5.1 ANOVA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150
iv CONTENTS
5.2 Multiple Comparison Methods: Fisher’s Method . . . . . . . . 158
5.2.1 FSD Multiple Comparisons in R . . . . . . . . . . . . . 161
5.2.2 Bonferroni Comparisons . . . . . . . . . . . . . . . . . 163
5.3 Further Discussion of Multiple Comparisons . . . . . . . . . . 168
5.4 Checking Assumptions in ANOVA Problems . . . . . . . . . . 170
5.5 Example from the Child Health and Development Study (CHDS) 173
6 Nonparametric Methods 181

6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 181
6.2 The Sign Test and CI for a Population Median . . . . . . . . . 182
6.3 Wilcoxon Signed-Rank Procedures . . . . . . . . . . . . . . . 188
6.3.1 Nonparametric Analyses of Paired Data . . . . . . . . . 192
6.3.2 Comments on One-Sample Nonparametric Methods . . 194
6.4 (Wilcoxon-)Mann-Whitney Two-Sample Procedure . . . . . . 195
6.5 Alternatives for ANOVA and Planned Comparisons . . . . . . 204
6.5.1 Kruskal-Wallis ANOVA . . . . . . . . . . . . . . . . . 205
6.5.2 Transforming Data . . . . . . . . . . . . . . . . . . . . 205
6.5.3 Planned Comparisons . . . . . . . . . . . . . . . . . . 218
6.5.4 Two final ANOVA comments . . . . . . . . . . . . . . 221
6.6 Permutation tests . . . . . . . . . . . . . . . . . . . . . . . . 221
6.6.1 Automated in R . . . . . . . . . . . . . . . . . . . . . 225
6.7 Density estimation . . . . . . . . . . . . . . . . . . . . . . . . 228
7 Categorical Data Analysis 232

7.1 Categorical data . . . . . . . . . . . . . . . . . . . . . . . . . 232
7.2 Single Proportion Problems . . . . . . . . . . . . . . . . . . . 235
7.2.1 A CI for p . . . . . . . . . . . . . . . . . . . . . . . . 236
7.2.2 Hypothesis Tests on Proportions . . . . . . . . . . . . 238
7.2.3 The p-value for a two-sided test . . . . . . . . . . . . . 239
7.2.4 Appropriateness of Test . . . . . . . . . . . . . . . . . 241
7.2.6 One-Sided Tests and One-Sided Confidence Bounds . . 242
7.2.7 Small Sample Procedures . . . . . . . . . . . . . . . . 245
7.3 Analyzing Raw Data . . . . . . . . . . . . . . . . . . . . . . . 247
CONTENTS v
7.4 Goodness-of-Fit Tests (Multinomial) . . . . . . . . . . . . . . 249
7.4.1 Adequacy of the Goodness-of-Fit Test . . . . . . . . . . 252
7.4.3 Multiple Comparisons in a Goodness-of-Fit Problem . . 255
7.5 Comparing Two Proportions: Independent Samples . . . . . . 257
7.5.1 Large Sample CI and Tests for p1 − p2 . . . . . . . . . 258
7.6 Effect Measures in Two-by-Two Tables . . . . . . . . . . . . . 263
7.7 Analysis of Paired Samples: Dependent Proportions . . . . . . 265
7.8 Testing for Homogeneity of Proportions . . . . . . . . . . . . . 268
7.8.1 Adequacy of the Chi-Square Approximation . . . . . . 273
7.9 Testing for Homogeneity in Cross-Sectional and Stratified Studies 274
7.9.1 Testing for Independence in a Two-Way Contingency Table275
7.9.2 Further Analyses in Two-Way Tables . . . . . . . . . . 277
8 Correlation and Regression 281

8.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 281
8.2 Testing that ρ = 0 . . . . . . . . . . . . . . . . . . . . . . . . 284
8.2.1 The Spearman Correlation Coefficient . . . . . . . . . . 284
8.3 Simple Linear Regression . . . . . . . . . . . . . . . . . . . . 289
8.3.1 Linear Equation . . . . . . . . . . . . . . . . . . . . . 289
8.3.2 Least Squares . . . . . . . . . . . . . . . . . . . . . . . 290
8.4 ANOVA Table for Regression . . . . . . . . . . . . . . . . . . 293
8.4.1 Brief discussion of the output for blood loss problem . . 298
8.5 The regression model . . . . . . . . . . . . . . . . . . . . . . 298
8.5.1 Back to the Data . . . . . . . . . . . . . . . . . . . . . 301
8.6 CI and tests for β1 . . . . . . . . . . . . . . . . . . . . . . . . 301
8.6.1 Testing β1 = 0 . . . . . . . . . . . . . . . . . . . . . . 302
8.7 A CI for the population regression line . . . . . . . . . . . . . 303
8.7.1 CI for predictions . . . . . . . . . . . . . . . . . . . . . 303
8.7.2 A further look at the blood loss data . . . . . . . . . . 305
8.8 Model Checking and Regression Diagnostics . . . . . . . . . . 306
8.8.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . 306
8.8.2 Residual Analysis . . . . . . . . . . . . . . . . . . . . 307
8.8.3 Checking Normality . . . . . . . . . . . . . . . . . . . 312
vi CONTENTS
8.8.4 Checking Independence . . . . . . . . . . . . . . . . . 313
8.8.5 Outliers . . . . . . . . . . . . . . . . . . . . . . . . . . 313
8.8.6 Influential Observations . . . . . . . . . . . . . . . . . 314
8.8.7 Summary of diagnostic measures . . . . . . . . . . . . 317
8.9 Regression analysis suggestion . . . . . . . . . . . . . . . . . . 318
8.9.1 Residual and Diagnostic Analysis of the Blood Loss Data 319
8.9.2 Gesell data . . . . . . . . . . . . . . . . . . . . . . . . 326
8.10 Weighted Least Squares . . . . . . . . . . . . . . . . . . . . . 330
9 Introduction to the Bootstrap 335

9.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 335
9.2 Bootstrap . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 337
9.2.1 Ideal versus Bootstrap world, sampling distributions . . 337
9.2.2 The accuracy of the sample mean . . . . . . . . . . . . 340
9.2.3 Comparing bootstrap sampling distribution from popu-
lation and sample . . . . . . . . . . . . . . . . . . . . 346
10 Power and Sample size 351

10.1 Power Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . 351
10.2 Effect size . . . . . . . . . . . . . . . . . . . . . . . . . . . . 357
10.3 Sample size . . . . . . . . . . . . . . . . . . . . . . . . . . . . 358
10.4 Power calculation via simulation . . . . . . . . . . . . . . . . . 363
A Custom R functions 368

A.1 Ch 2. Estimation in One-Sample Problems . . . . . . . . . . . 368
A.2 Ch 3. Two-Sample Inferences . . . . . . . . . . . . . . . . . . 369
Chapter 0
Introduction to R,
Rstudio, and ggplot
Learning objectives
identify a function or operation and describe its use
apply functions and operations to achieve a specific result
predict answers of calculations written in R
use R’s functions to get help and numerically summarize data
apply ggplot() to organize and reveal patterns visually
explain what each plotting option does
Achieving these goals contributes to mastery in these course learning outcomes:
1. organize knowledge
6. summarize data visually, numerically, and descriptively
8. use statistical software
2 Ch 0: Introduction to R, Rstudio, and ggplot
0.1 Syllabus
Tools for this course
Computer: Windows/Mac/Linux
Software: R, text editor (Rstudio)
Brain: scepticism, curiosity, organization
planning, execution, clarity
Syllabus Website: http://statacumen.com/teaching/ada1
Let’s see what’s there:

Step 0
Tentative timetable
Textbook
i>Clickers
Learning outcomes – HW00
Rubrics
Grading: Homework, team projects, midterm exam, clickers
To do by next class:
HW 00 due Thursday (or at end of class today).
Set up R and Rstudio.
Obtain and register an iClicker.
Read Hadley Wickham’s R style guide stat405.had.co.nz/r-style.
html
Read the rubric http://statacumen.com/teach/rubrics.pdf for goal
setting and definitions of quality work.
Obtain (optional) text book for an R programming reference.
If you need a disability requiring accommodation, please see me and reg-
ister with the UNM Accessibility Resource Center.
Learning outcomes, syllabus

Which of the learning outcomes on the syllabus are important for you to
achieve?
0.2: Rstudio 3
Circle these on your HW00 sheet.
Put a check mark by the one that is most important to you.
ClickerQ s — Ch 0, Learning outcomes, General
Statistics can be challenging because we operate at the higher levels of

Bloom’s Taxonomy1.
6 * Create/synthesize
5 * Evaluate
4 * Analyze
3 Apply
2 Understand
1 Remember
0.2 Rstudio
Setup Install R and Rstudio on your computer, as outlined in Step 0 on the

course webpage.
Quick tour Note that I changed my background to black for stealth coding
at night. . .
1
en.wikipedia.org/wiki/Bloom’s_Taxonomy
1. Upper-left: Program editor window, where you’ll write code.

2. Lower-left: Console, where you’ll submit commands to be executed.
3. Upper-right: Workspace lists variables in memory, History lists the com-
mands you’ve previously submitted in the console.
4. Lower-right: Files in current working directory, Plots you’ve produced,
Packages installed and whether they’re loaded, Help on commands.
Learning the keyboard shortcuts will make your life more wonderful. (Un-
der Help menu)
0.2: Rstudio 5
In particular:
Ctrl-Enter (submit current line or selection in editor window to console)

Ctrl-1 (move cursor to editor window)
Ctrl-2 (move cursor to console)
Ctrl-6 (move cursor to plots)
Esc (interrupt currently executing command)
Workflow is how you get things done. Here’s my recommendation:
1. Work out commands in the editor, then copy/paste into console (or use
Ctrl-Enter to submit current line or selection).
2. Your editor will keep a history of what you’ve done.
3. Make comments to help your future-self recall why you did what you did.
0.3 R building blocks
0.3: R building blocks 7
R as calculator In the following sections, look at the commands and output

and understand how the command was interpretted to give the output.
#### Calculator
# Arithmetic
2 * 10
## [1] 20
1 + 2
## [1] 3
# Order of operations is preserved
1 + 5 * 10
## [1] 51
(1 + 5) * 10
## [1] 60
# Exponents use the ^ symbol
2^5
## [1] 32
9^(1/2)
## [1] 3
Vectors A vector is a set of numbers like a column of a spreadsheet. Below

these sets of numbers are not technically “vectors”, since they are not in a
row/column structure, though they are ordered and can be referred to by index.
#### Vectors
# Create a vector with the c (short for combine) function
c(1, 4, 6, 7)
## [1] 1 4 6 7
c(1:5, 10)
## [1] 1 2 3 4 5 10
# or use a function
# (seq is short for sequence)
seq(1, 10, by = 2)
## [1] 1 3 5 7 9
seq(0, 50, length = 11)
## [1] 0 5 10 15 20 25 30 35 40 45 50
seq(1, 50, length = 11)
## [1] 1.0 5.9 10.8 15.7 20.6 25.5 30.4 35.3 40.2 45.1 50.0
1:10 # short for seq(1, 10, by = 1), or just
## [1] 1 2 3 4 5 6 7 8 9 10
seq(1, 10)
## [1] 1 2 3 4 5 6 7 8 9 10
5:1
## [1] 5 4 3 2 1
# non-integer sequences
# (Note: the [1] at the beginning of lines indicates
# the index of the first value in that row)
seq(0, 100*pi, by = pi)
## [1] 0.000 3.142 6.283 9.425 12.566 15.708 18.850 21.991
## [9] 25.133 28.274 31.416 34.558 37.699 40.841 43.982 47.124
## [17] 50.265 53.407 56.549 59.690 62.832 65.973 69.115 72.257
## [25] 75.398 78.540 81.681 84.823 87.965 91.106 94.248 97.389
## [33] 100.531 103.673 106.814 109.956 113.097 116.239 119.381 122.522
## [41] 125.664 128.805 131.947 135.088 138.230 141.372 144.513 147.655
## [49] 150.796 153.938 157.080 160.221 163.363 166.504 169.646 172.788
## [57] 175.929 179.071 182.212 185.354 188.496 191.637 194.779 197.920
## [65] 201.062 204.204 207.345 210.487 213.628 216.770 219.911 223.053
## [73] 226.195 229.336 232.478 235.619 238.761 241.903 245.044 248.186
## [81] 251.327 254.469 257.611 260.752 263.894 267.035 270.177 273.319
## [89] 276.460 279.602 282.743 285.885 289.027 292.168 295.310 298.451
## [97] 301.593 304.734 307.876 311.018 314.159
Assign variables Assigning objects to varibles.

#### Assign variables
# Assign a vector to a variable with <-
a <- 1:5
a
## [1] 1 2 3 4 5
b <- seq(15, 3, length = 5)
b
## [1] 15 12 9 6 3
c <- a * b
c
## [1] 15 24 27 24 15
Basic functions There are lots of functions available in the base package.
Type ?base and click on Index at the bottom of the help page for a complete
list of functions. Other functions to look at are in the ?stats and ?datasets
packages.
#### Basic functions
# Lots of familiar functions work
a
## [1] 1 2 3 4 5
sum(a)
## [1] 15
prod(a)
## [1] 120
mean(a)
## [1] 3
sd(a)
## [1] 1.581
var(a)
## [1] 2.5
min(a)
## [1] 1
median(a)
## [1] 3
max(a)
## [1] 5
range(a)
## [1] 1 5
Extracting subsets It will be an extremely important skill to understand

the structure of your variables and pull out the information you need from
them.
#### Extracting subsets

# Specify the indices you want in the square brackets []
a <- seq(0, 100, by = 10)
# blank = include all
a
## [1] 0 10 20 30 40 50 60 70 80 90 100
a[]
## [1] 0 10 20 30 40 50 60 70 80 90 100
# integer +=include, 0=include none, -=exclude
a[5]
## [1] 40
a[c(2, 4, 6, 8)]
## [1] 10 30 50 70
a[0]
## numeric(0)
a[-c(2, 4, 6, 8)]
## [1] 0 20 40 60 80 90 100
a[c(1, 1, 1, 6, 6, 9)] # subsets can be bigger than the original set
## [1] 0 0 0 50 50 80
a[c(1,2)] <- c(333, 555) # update a subset
a
## [1] 333 555 20 30 40 50 60 70 80 90 100
Boolean: True/False Subsets can be selected based on which elements

meet specific conditions.
#### Boolean
a
## [1] 333 555 20 30 40 50 60 70 80 90 100
(a > 50) # TRUE/FALSE for each element
## [1] TRUE TRUE FALSE FALSE FALSE FALSE TRUE TRUE TRUE TRUE TRUE
which(a > 50) # which indicies are TRUE
## [1] 1 2 7 8 9 10 11
a[(a > 50)]
## [1] 333 555 60 70 80 90 100
!(a > 50) # ! negates (flips) TRUE/FALSE values
## [1] FALSE FALSE TRUE TRUE TRUE TRUE FALSE FALSE FALSE FALSE FALSE
a[!(a > 50)]
## [1] 20 30 40 50
Comparison functions All your favorite comparisons are available.
#### Comparison
# Here they are: < > <= >= != == %in%
a
## [1] 333 555 20 30 40 50 60 70 80 90 100
# equal to
a[(a == 50)]
## [1] 50
# equal to
a[(a == 55)]
## numeric(0)
# not equal to
a[(a != 50)]
## [1] 333 555 20 30 40 60 70 80 90 100
# greater than
a[(a > 50)]
## [1] 333 555 60 70 80 90 100
# less than
a[(a < 50)]
## [1] 20 30 40
# less than or equal to
a[(a <= 50)]
## [1] 20 30 40 50
# which values on left are in the vector on right
(c(10, 14, 40, 60, 99) %in% a)
## [1] FALSE FALSE TRUE TRUE FALSE
Boolean operators compare TRUE/FALSE values and return TRUE/-

FALSE values.
#### Boolean
# & and, | or, ! not
a
## [1] 333 555 20 30 40 50 60 70 80 90 100
a[(a >= 50) & (a <= 90)]
## [1] 50 60 70 80 90
a[(a < 50) | (a > 100)]
## [1] 333 555 20 30 40
a[(a < 50) | !(a > 100)]
## [1] 20 30 40 50 60 70 80 90 100
a[(a >= 50) & !(a <= 90)]
## [1] 333 555 100
Missing values The value NA (not available) means the value is missing.
Any calculation involving NA will return an NA by default.
#### Missing values
NA + 8
## [1] NA
3 * NA
## [1] NA
mean(c(1, 2, NA))
## [1] NA
# Many functions have an na.rm argument (NA remove)
mean(c(NA, 1, 2), na.rm = TRUE)
## [1] 1.5
sum(c(NA, 1, 2))
## [1] NA
sum(c(NA, 1, 2), na.rm = TRUE)
## [1] 3
# Or you can remove them yourself
a <- c(NA, 1:5, NA)
a
## [1] NA 1 2 3 4 5 NA
is.na(a) # which values are missing?
## [1] TRUE FALSE FALSE FALSE FALSE FALSE TRUE
!is.na(a) # which values are NOT missing?
## [1] FALSE TRUE TRUE TRUE TRUE TRUE FALSE
a[!is.na(a)] # return those which are NOT missing
## [1] 1 2 3 4 5
a # note, this did not change the variable a
## [1] NA 1 2 3 4 5 NA
# To save the results of removing the NAs,
# assign to another variable or reassign to the original variable
# Warning: if you write over variable a then the original version is gone forever!
a <- a[!is.na(a)]
a
## [1] 1 2 3 4 5
Your turn!
ClickerQ s — Ch 0, R building blocks, Subset

0.4: Plotting with ggplot2 13
ClickerQ s — Ch 0, R building blocks, T/F selection 1
ClickerQ s — Ch 0, R building blocks, T/F selection 2
How’d you do?

Outstanding Understanding the operations and how to put them together,
without skipping steps.
Good Understanding most of the small steps, missed a couple details.
Hang in there Understanding some of the concepts but all the symbols
make my eyes spin.
Reading and writing a new language takes work.
You’ll get better as you practice, practice, practice2.
Having a buddy to work with will help.
R command review This is a summary of the commands we’ve seen so

far.
#### Review
<-
+ - * / ^
c()
seq() # by=, length=
sum(), prod(), mean(), sd(), var(),
min(), median(), max(), range()
a[]
(a > 1), ==, !=, >, <, >=, <=, %in%
&, |, !
NA, mean(a, na.rm = TRUE), !is.na()
0.4 Plotting with ggplot2

There are three primary strategies for plotting in R. The first is base graphics,
using functions such as plot(), points(), and par(). A second is use of the
package lattice, which creates very nice graphics quickly, though can be more
difficult to create high-dimensional data displays. A third, and the one I will use
throughout this course, is using package ggplot2, which is an implementation
2
What’s your Carnegie Hall that you’re working towards?
of the “Grammar of Graphics”. While creating a very simple plot requires an
extra step or two, creating very complex displays are relatively straightforward
as you become comfortable with the syntax.
This section is intended as an introduction to ggplot() and some of its
capabilities (we will cover these plotting functions and others in detail in later
chapters). As a basic introduction, it requires a data.frame object as input (a
table where each row represents an observation or data point, and each column
can have a different data type), and then you define plot layers that stack on
top of each other, and each layer has visual/text elements that are mapped to
aesthetics (colors, size, opacity). In this way, a simple set of commands can be
combined to produce extremely informative displays.
In the example that follows, we consider a dataset mpg consisting of fuel
economy data from 1999 and 2008 for 38 popular models of car.
#### Installing
# only needed once after installing or upgrading R
install.packages("ggplot2")
#### Library
# each time you start R
# load package ggplot2 for its functions and datasets
library(ggplot2)
# ggplot2 includes a dataset "mpg"
# ? gives help on a function or dataset

?mpg
Let’s get a look at the dataset we’re using.
#### mpg dataset
# head() lists the first several rows of a data.frame
head(mpg)
## manufacturer model displ year cyl trans drv cty hwy fl class
## 1 audi a4 1.8 1999 4 auto(l5) f 18 29 p compact
## 2 audi a4 1.8 1999 4 manual(m5) f 21 29 p compact
## 4 audi a4 2.0 2008 4 auto(av) f 21 30 p compact
## 5 audi a4 2.8 1999 6 auto(l5) f 16 26 p compact
# str() gives the structure of the object
str(mpg)
## 'data.frame': 234 obs. of 11 variables:
## $ manufacturer: Factor w/ 15 levels "audi","chevrolet",..: 1 1 1 1 1 1 1 1 1 1 ...
## $ model : Factor w/ 38 levels "4runner 4wd",..: 2 2 2 2 2 2 2 3 3 3 ...
## $ displ : num 1.8 1.8 2 2 2.8 2.8 3.1 1.8 1.8 2 ...
## $ year : int 1999 1999 2008 2008 1999 1999 2008 1999 1999 2008 ...
## $ cyl : int 4 4 4 4 6 6 6 4 4 4 ...
## $ trans : Factor w/ 10 levels "auto(av)","auto(l3)",..: 4 9 10 1 4 9 1 9 4 10 ...
## $ drv : Factor w/ 3 levels "4","f","r": 2 2 2 2 2 2 2 1 1 1 ...
## $ cty : int 18 21 20 21 16 18 18 18 16 20 ...
## $ hwy : int 29 29 31 30 26 26 27 26 25 28 ...
## $ fl : Factor w/ 5 levels "c","d","e","p",..: 4 4 4 4 4 4 4 4 4 4 ...
## $ class : Factor w/ 7 levels "2seater","compact",..: 2 2 2 2 2 2 2 2 2 2 ...
# summary() gives frequeny tables for categorical variables
# and mean and five-number summaries for continuous variables
summary(mpg)
## manufacturer model displ
## dodge :37 caravan 2wd : 11 Min. :1.60
## toyota :34 ram 1500 pickup 4wd: 10 1st Qu.:2.40
## volkswagen:27 civic : 9 Median :3.30
## ford :25 dakota pickup 4wd : 9 Mean :3.47
## chevrolet :19 jetta : 9 3rd Qu.:4.60
## audi :18 mustang : 9 Max. :7.00
## (Other) :74 (Other) :177
## year cyl trans drv cty
## Min. :1999 Min. :4.00 auto(l4) :83 4:103 Min. : 9.0
## 1st Qu.:1999 1st Qu.:4.00 manual(m5):58 f:106 1st Qu.:14.0
## Median :2004 Median :6.00 auto(l5) :39 r: 25 Median :17.0
## Mean :2004 Mean :5.89 manual(m6):19 Mean :16.9
## 3rd Qu.:2008 3rd Qu.:8.00 auto(s6) :16 3rd Qu.:19.0
## Max. :2008 Max. :8.00 auto(l6) : 6 Max. :35.0
## (Other) :13
## hwy fl class
## Min. :12.0 c: 1 2seater : 5
## 1st Qu.:18.0 d: 5 compact :47
## Median :24.0 e: 8 midsize :41
## Mean :23.4 p: 52 minivan :11
## 3rd Qu.:27.0 r:168 pickup :33
## Max. :44.0 subcompact:35
## suv :62
#### ggplot_mpg_displ_hwy
# specify the dataset and variables
p <- ggplot(mpg, aes(x = displ, y = hwy))
p <- p + geom_point() # add a plot layer with points
print(p)
40
●
●
●
●
● ●
● ●
● ● ●
30 ● ● ●
● ● ● ● ● ● ●
hwy
● ● ● ● ●
● ● ● ● ● ● ● ●
● ● ● ● ● ● ● ● ● ● ● ● ● ●
● ● ● ● ● ● ● ● ●
● ● ● ● ● ● ● ●
● ● ● ● ● ● ●
● ● ● ● ●
● ●
20 ● ● ● ●
● ● ● ● ● ●
● ● ● ● ● ● ●
● ● ● ● ● ● ● ● ● ● ● ● ●
● ● ● ●
● ● ● ● ● ● ● ●
● ●
2 3 4 5 6 7
displ
Geoms, aesthetics, and facets are three concepts we’ll see in this section.
Geom: is the “type” of plot

Aesthetics: shape, colour, size, alpha
Faceting: “small multiples” displaying different subsets
Help is available. Try searching for examples, too.
had.co.nz/ggplot2
had.co.nz/ggplot2/geom_point.html
When certain aesthetics are defined, an appropriate legend is chosen and

displayed automatically.
#### ggplot_mpg_displ_hwy_colour_class
p <- p + geom_point(aes(colour = class))
print(p)
40
●
●
●
●
class
● ●
● ● ● 2seater
● ● ●
● compact
30 ● ● ●
● ● ● ● ● ● ● ● midsize
hwy
● ● ● ● ●
● ● ● ● ● ● ● ●
● minivan
● ● ● ● ● ● ● ● ● ● ● ● ● ● ● pickup
● ● ● ● ● ● ● ● ●
● ● ● ● ● ● ● ● ● subcompact
● ● ● ● ● ● ●
● suv
● ● ● ● ●
● ●
20 ● ● ● ●
● ● ● ● ● ●
● ● ● ● ● ● ●
● ● ● ● ● ● ● ● ● ● ● ● ●
● ● ● ●
● ● ● ● ● ● ● ●
● ●
2 3 4 5 6 7
displ
I encourage you to experiment with aesthetics!
1. Assign variables to aesthetics colour, size, and shape.

2. What’s the difference between discrete or continuous variables?
3. What happens when you combine multiple aesthetics?
The behavior of the aesthetics is predictable and customizable.
Aesthetic Discrete Continuous

colour Rainbow of colors Gradient from red to blue
size Discrete size steps Linear mapping between radius and value
shape Different shape for each Shouldn’t work
Let’s see a couple examples.

#### ggplot_mpg_displ_hwy_colour_class_size_cyl_shape_drv
p <- p + geom_point(aes(colour = class, size = cyl, shape = drv))
print(p)
drv
● 4
40 f
r
cyl
●
4
● 5
● 6
30
● 7
hwy
●8
●
● ●
● ● ●
● ●
● ● class
●
●
●
●
●
● ● 2seater
● compact
20 ●
● ● midsize
● ● ● ●● ● ● minivan
● ● ●● ●●
●● ●● ● ●● ●●● ● ● ● pickup
●● ●
● ●● ●● ●● ● subcompact
● ● ● suv
●
2 3 4 5 6 7
displ
#### ggplot_mpg_displ_hwy_colour_class_size_cyl_shape_drv_alpha
p <- p + geom_point(aes(colour = class, size = cyl, shape = drv)
, alpha = 1/4) # alpha is the opacity
print(p)
drv
4
40 f
r
cyl
4
5
6
30
7
hwy
class
2seater
compact
20 midsize
minivan
pickup
subcompact
suv
2 3 4 5 6 7
displ
Faceting A small multiple3 (sometimes called faceting, trellis chart, lattice
chart, grid chart, or panel chart) is a series or grid of small similar graphics or
charts, allowing them to be easily compared.
Typically, small multiples will display different subsets of the data.

Useful strategy for exploring conditional relationships, especially for large
data.
Experiment with faceting of different types. What relationships would you

like to see?
#### ggplot_mpg_displ_hwy_facet
# start by creating a basic scatterplot
p <- p + geom_point()
## two methods
# facet_grid(rows ~ cols) for 2D grid, "." for no split.
# facet_wrap(~ var) for 1D ribbon wrapped into 2D.
# examples of subsetting the scatterplot in facets

p1 <- p + facet_grid(. ~ cyl) # columns are cyl categories
p2 <- p + facet_grid(drv ~ .) # rows are drv categories
p3 <- p + facet_grid(drv ~ cyl) # both
p4 <- p + facet_wrap(~ class) # wrap plots by class category
# plot all four in one arrangement

library(gridExtra)
grid.arrange(p1, p2, p3, p4, ncol = 2)
3
According to Edward Tufte (Envisioning Information, p. 67): “At the heart of quantitative reasoning
is a single question: Compared to what? Small multiple designs, multivariate and data bountiful, answer
directly by visually enforcing comparisons of changes, of the differences among objects, of the scope of
alternatives. For a wide range of problems in data presentation, small multiples are the best design
solution.”
4 5 6 8
40
●
30 ●
4
● ● ● ● ●
●
● ●
● ●
● ●
40 ● ● ● ●
20 ● ● ●
●
●
● ●● ●
●● ●●
● ●
● ● ●
●●● ● ●
● ●
●●
● ●
● ● ● ● ● ●● ● ● ●
● ●
●
●
●
●
●● 40
● ● ●
●
●●● ●
●
●
● ● ●
30
hwy
●● ● ● ●
●●
30 ● ●
● ●
● ● ● ● ●
hwy
●●●●● ● ● ● ● ● ● ●
f
●
● ● ●
●● ● ● ●
● ● ●
● ●
● ● ●●● ● ●●
● ●
● ● ●● ● ●
●●●
● ●●●● ● ●
● ●
●
●●●●
● ●●●●●●●● ● ● 20
● ● ●●● ●● ● ● ●
●
● ●
●●● ● ●
● ● ●● ●● ●
● ●● ● ●
● ●
20 ● ● ●
● 40
●●● ●
● ●
●● ●● ●●●
●
● ●●● ●
●●●●●●● ● 30
r
●
● ●● ● ● ● ●
● ●● ●●
●●● ● ● ● ●
●
● ●
● ● ●
20 ●●
●
● ● ● ●
● ●●
2 3 4 5 6 7 2 3 4 5 6 7 2 3 4 5 6 7 2 3 4 5 6 7 2 3 4 5 6 7
displ displ
4 5 6 8 2seater compact midsize

●
40 40 ●
●
● ●
● ● ●
●●
30 30 ●●
●●
●●●●
●●●● ●●●
●●
●●
●●
●●
●● ●●●●●
● ●
● ● ● ●
●● ● ● ●●●
4
● ● ● ● ●●
● ●●● ● ●
●
● ●●
● ●● ● ● ●
● ●
●● ● ● 20
20 ● ●●●
●●
● ●●●
● ●●
●● ●
●● ●
● ●●●
●●●
● ● ● ●
● ●● ●●●●
●
●
●
minivan pickup subcompact
● ●
40 40 ●
●
●
●
● ●
●
●●● ●
● ●●
●
hwy
hwy
●
30 ●●
●●
●●●
● ● ●
● ●● 30 ● ● ●
●●
●● ●
●● ●
f
●●●
● ● ●●●● ●
●●
●● ● ● ●●
● ●●
● ●
●
●●●●●●● ● ● ● ● ●● ● ●
● ● ●
● ● ●
●
20 ● 20 ●
● ●● ●
● ●●●● ●
●
●●
●
● ● ●
●●●
●● ●
●
suv
40
40
30
30
r
●●
●● ● ●
● ●
● ● ● ●
●
● ●
●
● ●
20 ●
●
●●● 20 ● ●●●● ●● ● ●●●●
● ●
● ●● ●●●●●● ●●● ●● ●
● ● ●● ● ●● ●●●
●
2 3 4 5 6 7 2 3 4 5 6 7 2 3 4 5 6 7 2 3 4 5 6 7 2 3 4 5 6 7
displ displ
0.4.1 Improving plots
How can this plot be improved?

#### ggplot_mpg_cty_hwy
p <- ggplot(mpg, aes(x = cty, y = hwy))
print(p)
● ●
40
●
● ●
●
●
● ●
● ● ●
● ● ● ●
30 ● ● ●
● ● ● ● ● ●
hwy
● ● ●
● ● ● ● ●
● ● ● ● ● ●
● ● ● ● ● ●
● ● ● ●
● ● ●
● ● ●
●
20 ● ● ●
● ● ●
● ● ● ●
● ● ● ● ●
● ● ●
●
●
10 15 20 25 30 35
cty
Problem: points lie on top of each other, so it’s impossible to tell how many
observations each point represents.
A solution: Jitter the points to reveal the individual points and reduce the
opacity to 1/2 to indicate when points overlap.
#### ggplot_mpg_cty_hwy_jitter
p <- ggplot(mpg, aes(x = cty, y = hwy))
p <- p + geom_point(position = "jitter", alpha = 1/2)
print(p)
40
30
hwy
20
10 20 30
cty
How can this plot be improved?

#### ggplot_mpg_class_hwy
p <- ggplot(mpg, aes(x = class, y = hwy))
print(p)
● ●
40
●
●
●
●
● ●
● ●
● ●
30 ● ●
● ● ●
hwy
● ● ●
● ● ● ●
● ● ● ● ●
● ● ● ● ●
● ● ● ● ● ●
● ● ● ● ● ●
● ● ● ●
● ●
20 ● ● ●
● ●
● ●
● ● ●
● ●
● ●
●
● ●
2seater compact midsize minivan pickup subcompact suv

class
Problem: The classes are in alphabetical order, which is somewhat arbitrary.
A solution: Reorder the class variable by the mean hwy for a meaningful
ordering. Get help with ?reorder to understand how this works.
#### ggplot_mpg_reorder_class_hwy
p <- ggplot(mpg, aes(x = reorder(class, hwy), y = hwy))
print(p)
● ●
40
●
●
●
●
● ●
● ●
● ●
30 ● ●
● ● ●
hwy
● ● ●
● ● ● ●
● ● ● ● ●
● ● ● ● ●
● ● ● ● ● ●
● ● ● ● ● ●
● ● ● ●
● ●
20 ● ● ●
● ●
● ●
● ● ●
● ●
● ●
●
● ●
pickup suv minivan 2seater midsize subcompact compact

reorder(class, hwy)
. . . add jitter
#### ggplot_mpg_reorder_class_hwy_jitter
p <- p + geom_point(position = "jitter")
print(p)
● ●
40
●
●
●
●
●
● ●
●
● ● ●
●
●● ● ●●
●
●
30 ●
●
●
●
●● ● ● ● ● ● ● ● ●
●● ●●●●● ●
hwy
●
●
● ●
● ● ●
● ●
● ● ●●
● ● ●●
● ● ● ● ●
●
● ● ●● ●● ● ●● ● ●●
● ●● ●●● ●
● ●
● ●●● ● ● ●● ●
● ● ● ● ● ● ●
● ● ● ●
● ●
●
● ● ● ● ●
●● ● ● ● ●
● ●
●
● ● ●
● ● ●
● ● ●●
● ●●
●
●
● ●
20 ● ● ●● ● ● ●
● ●
● ●
● ●
● ●●
●
● ● ● ● ●
● ● ●●●
● ● ● ●
● ● ● ●● ●
●● ● ●
●
●●
● ●● ●● ●● ●
● ●●
●
●●● ● ● ●
● ●●● ● ●●
● ●● ●
●
●● ● ● ●
●
●
● ● ● ●
●

reorder(class, hwy)
. . . or replace with boxplots

#### ggplot_mpg_reorder_class_hwy_boxplot
p <- p + geom_boxplot()
print(p)
● ●
40
30
hwy
●
●
●
●
●
●
20
● ●

reorder(class, hwy)
. . . or jitter those points with reduced-opacity boxplots on top

#### ggplot_mpg_reorder_class_hwy_jitter_boxplot
p <- p + geom_boxplot(alpha = 0.5)
print(p)
● ● ●●
● ●
40
●
●
● ●
● ● ●
●
● ●
● ●●
●
●● ●
●
● ● ●
● ● ●●
30 ●
● ● ● ● ●● ●●● ● ●
● ● ● ●● ● ● ● ● ●
hwy
●● ●
● ● ● ●
● ● ● ● ● ● ●
● ●
● ● ● ●
●
●
● ● ●●
● ●
● ● ●● ●● ● ● ● ●
● ● ● ●● ●●
●● ● ●
●
●● ● ●
● ●●
● ● ● ●● ● ● ● ● ● ● ●●
● ● ● ● ● ● ● ● ● ●
● ●
● ●
● ●
● ● ● ● ●
● ● ● ● ● ●
● ●
● ●
● ●●
20 ● ● ● ●
● ●●● ● ●●
● ● ● ● ● ● ●●
●
● ● ●● ●
●
●
●
● ● ● ●● ● ● ● ●●
● ●●
● ●● ● ●
● ●● ● ●
● ●
●
●● ●● ● ●
●
● ●
● ● ●
● ●
●● ● ● ● ● ●●
●
●
● ● ● ● ● ●
●
10
reorder(class, hwy)
. . . even better with the boxplots with jittered points on top

#### ggplot_mpg_reorder_class_hwy_boxplot_jitter
print(p)
●
● ● ●
● ●
40
● ●
●
●
● ●●
●
● ●
● ●
● ●
● ● ●
●
●● ●
● ●
30 ●
●
● ● ●●● ● ● ● ●●●●
● ● ● ● ●● ● ●
hwy
● ●
● ●● ● ●
● ●
●● ● ●
● ● ● ● ● ● ●●
● ● ●
● ●● ●
● ● ● ●
●● ●
●
● ● ●● ● ● ● ● ● ●
● ● ● ● ●●●● ● ● ●●
● ●
● ● ● ●
● ●
● ● ●● ● ● ●
● ● ●
● ● ● ●● ●
● ●
● ● ●
● ● ● ● ●
● ●
●
● ● ● ● ●
● ● ●
● ●
● ● ●
20 ● ● ● ● ●● ● ●
●
●
●● ●
● ● ● ● ●● ● ●
●● ● ●
●● ● ● ●
● ●● ● ● ● ●●●● ●
●●● ●●
● ● ● ●● ● ● ●● ●●● ● ●
● ● ●
● ●
● ● ●● ●
● ● ● ●●
● ● ● ●
●
● ●
●
● ● ● ● ●●

reorder(class, hwy)
. . . and can easily reorder by median() instead of mean() (mean is the default)
#### ggplot_mpg_reorder_class_hwy_boxplot_jitter_median
p <- ggplot(mpg, aes(x = reorder(class, hwy, FUN = median), y = hwy))
print(p)
● ● ●
●
● ●
40
●
●
●
●
● ●
●
●
● ●
●● ●
●
● ● ●
●● ● ●
● ●
30 ●●
● ● ●● ●● ● ● ● ● ●
●
● ● ● ● ●
● ● ● ●
hwy
● ● ● ●
●● ●
● ● ● ●
● ● ● ●● ●● ● ● ●
● ● ●● ●● ● ● ● ●● ● ●●
●
● ● ●● ●
● ● ● ● ● ● ● ● ●● ●● ●
● ●●●● ● ●
●●● ● ● ●●
●●
● ● ● ● ● ● ●
● ● ● ● ● ●
●
●
● ● ● ● ● ● ●
● ●
● ● ● ●
● ●
● ●
● ● ●● ●
20 ● ● ● ● ● ●
● ● ●
● ● ● ●● ●
● ●●
●
● ● ●
● ●● ● ●●
● ●● ● ● ●●
●●● ● ●● ● ● ● ●● ●●
● ● ● ●●
● ● ●●●
● ●
● ● ● ●
● ●
●
● ● ● ● ● ●●
● ●
●
● ●
● ●
● ● ● ●●
pickup suv minivan 2seater subcompact compact midsize

reorder(class, hwy, FUN = median)
R command review This is a summary of the commands we saw in this

plotting section.
#### Review
library(ggplot2)
?help
head()
str()
summary()
ggplot(df)
geom_point()
aes()
colour, size, shape, alpha
position = "jitter"
geom_boxplot()
facet_grid()
facet_wrap()
reorder()
median()
One-minute paper:
Muddy Any “muddy” points — anything that doesn’t make sense yet?
Thumbs up Anything you really enjoyed or feel excited about?
0.5 Course Overview
See ADA2 Chapter 1 notes for a brief overview of all we’ll cover in this semester
of ADA1.
Chapter 1
Summarizing and
Displaying Data
Learning objectives
After completing this topic, you should be able to:
use R’s functions to get help and numerically summarize data.
apply R’s base graphics and ggplot to visually summarize data in several
ways.
explain what each plotting option does.
describe the characteristics of a data distribution.
1. organize knowledge.
6. summarize data visually, numerically, and descriptively.
8. use statistical software.
1.1 Random variables

A random variable is a variable whose value is subject to variations due
to chance. Random variables fall into two broad categories: qualitative and
quantitative.
Qualitative data includes categorical outcomes:
Nominal Outcome is one of several categories.
Ex: Blood group, hair color.
28 Ch 1: Summarizing and Displaying Data
Ordinal Outcome is one of several ordered categories.
Ex: Likert data such as (strongly agree, agree, neutral, disagree, strongly
disagree).
Quantitative data includes numeric outcomes:
Discrete Outcome is one of a fixed set of numerical values.
Ex: Number of children.
Continuous Outcome is any numerical value.
Ex: Birthweight.
It may not always be perfectly clear which type data belong to, and may
sometimes be classified based on the question being asked of the data. Distinc-
tion between nominal and ordinal variables can be subjective. For example,
for vertebral fracture types: (Wedge, Concavity, Biconcavity, Crush), one could
argue that a crush is worse than a biconcavity which is worse than a concav-
ity . . . , but this is not self-evident. Distinction between ordinal and discrete
variables can be subjective. For example, cancer staging (I, II, III, IV) sounds
discrete, but better treated as ordinal because the “distance” between stages
may be hard to define and unlikely to be equal. Continuous variables generally
measured to a fixed level of precision, which makes them discrete. This “dis-
creteness” of continuous variables is not a problem, providing there are enough
levels.
ClickerQ s — Random variables
1.2 Numerical summaries

Suppose we have a collection of n individuals, and we measure each individ-
ual’s response on one quantitative characteristic, say height, weight, or systolic
blood pressure. For notational simplicity, the collected measurements are de-
noted by Y1, Y2, . . . , Yn, where n is the sample size. The order in which the
measurements are assigned to the place-holders (Y1, Y2, . . . , Yn) is irrelevant.
Among the numerical summary measures we’re interested in are the sample
mean Ȳ and the sample standard deviation s. The sample mean is a
measure of central location, or a measure of a “typical value” for the data
1.2: Numerical summaries 29
set. The standard deviation is a measure of spread in the data set. These
summary statistics should be familiar to you. Let us consider a simple example
to refresh your memory on how to compute them.
Suppose we have a sample of n = 8 children with weights (in pounds): 5,
9, 12, 30, 14, 18, 32, 40. Then
P
i Yi Y1 + Y2 + · · · + Yn
Ȳ = =
n n
5 + 9 + 12 + 30 + 14 + 18 + 32 + 40 160
= = = 20.
8 8
#### Numerical summaries
#### mean
y <- c(5, 9, 12, 30, 14, 18, 32, 40)
mean(y)
## [1] 20
The sample standard deviation is the square root of the sample variance
2
(Y1 − Ȳ )2 + (Y2 − Ȳ )2 + · · · + (Yk − Ȳ )2
P
2 i (Yi − Ȳ )
s = =
n−1 n−1
(5 − 20) + (9 − 20)2 + · · · + (40 − 20)2
2
= = 156.3,
√ 7
s = s2 = 12.5.
#### variance
var(y)
## [1] 156.3
sd(y)
## [1] 12.5
Summary statistics have well-defined units of measurement, for example,

Ȳ = 20 lb, s2 = 156.3 lb2, and s = 12.5 lb. The standard deviation is often
used instead of s2 as a measure of spread because s is measured in the same
units as the data.
Remark If the divisor for s2 was n instead of n − 1, then the variance would
be the average squared deviation observations are from the center of the data
as measured by the mean.
The following graphs should help you to see some physical meaning of the
sample mean and variance. If the data values were placed on a “massless”
ruler, the balance point would be the mean (20). The variance is basically the
“average” (remember n − 1 instead of n) of the total areas of all the squares
obtained when squares are formed by joining each value to the mean. In both
cases think about the implication of unusual values (outliers). What happens
to the balance point if the 40 were a 400 instead of a 40? What happens to the
squares?
The sample median M is an alternative measure of central location.

The measure of spread reported along with M is the interquartile range,
IQR = Q3 − Q1, where Q1 and Q3 are the first and third quartiles of the data
set, respectively. To calculate the median and interquartile range, order the
data from lowest to highest values, all repeated values included. The ordered
weights are
5 9 12 14 18 30 32 40.
#### sorting
sort(y)
## [1] 5 9 12 14 18 30 32 40
The median M is the value located at the half-way point of the ordered
string. There is an even number of observations, so M is defined to be half-way
between the two middle values, 14 and 18. That is, M = 0.5(14 + 18) = 16
lb. To get the quartiles, break the data into the lower half: 5 9 12 14, and the
upper half: 18 30 32 and 40. Then
Q1 = first quartile = median of lower half of data = 0.5(9+12)=10.5 lb,
and
Q3 = third quartile = median of upper half of data = 0.5(30+32) = 31 lb.
1.2: Numerical summaries 31
The interquartile range is
IQR = Q3 − Q1 = 31 − 10.5 = 20.5 lb.
#### quartiles
median(y)
## [1] 16
fivenum(y)
## [1] 5.0 10.5 16.0 31.0 40.0
# The quantile() function can be useful, but doesn't calculate Q1 and Q3
# as defined above, regardless of the 9 types of calculations for them!
# summary() is a combination of mean() and quantile(y, c(0, 0.25, 0.5, 0.75, 1))
summary(y)
## Min. 1st Qu. Median Mean 3rd Qu. Max.
## 5.0 11.2 16.0 20.0 30.5 40.0
# IQR
fivenum(y)[c(2,4)]
## [1] 10.5 31.0
fivenum(y)[4] - fivenum(y)[2]
## [1] 20.5
diff(fivenum(y)[c(2,4)])
## [1] 20.5
The quartiles, with M being the second quartile, break the data set roughly
into fourths. The first quartile is also called the 25th percentile, whereas the
median and third quartiles are the 50th and 75th percentiles, respectively. The
IQR is the range for the middle half of the data.
Suppose we omit the largest observation from the weight data:

5 9 12 14 18 30 32.
How do M and IQR change? With an odd number of observations, there is a
unique middle observation in the ordered string which is M . Here M = 14 lb.
It is unclear which half the median should fall into, so M is placed into both
the lower and upper halves of the data. The lower half is 5 9 12 14, and the
upper half is 14 18 30 32. With this convention, Q1 = 0.5(9 + 12) = 10.5 and
Q3 = 0.5(18 + 30) = 24, giving IQR = 24 − 10.5 = 13.5 (lb).
#### remove largest
# remove the largest observation by removing the last of the sorted values
y2 <- sort(y)[-length(y)]
y2
## [1] 5 9 12 14 18 30 32
median(y2)
## [1] 14
fivenum(y2)
## [1] 5.0 10.5 14.0 24.0 32.0
diff(fivenum(y2)[c(2,4)])
## [1] 13.5
If you look at the data set with all eight observations, there actually are many
numbers that split the data set in half, so the median is not uniquely defined1,
although “everybody” agrees to use the average of the two middle values. With
quartiles there is the same ambiguity but no such universal agreement on what
to do about it, however, so R will give slightly different values for Q1 and
Q3 when using summary() and some other commands than we just calculated,
and other packages will report even different values. This has no practical
implication (all the values are “correct”) but it can appear confusing.
Example The data given below are the head breadths in mm for a sample
of 18 modern Englishmen, with numerical summaries generated by R.
#### Englishmen
hb <- c(141, 148, 132, 138, 154, 142, 150, 146, 155
, 158, 150, 140, 147, 148, 144, 150, 149, 145)
# see sorted values

sort(hb)
## [1] 132 138 140 141 142 144 145 146 147 148 148 149 150 150 150 154
## [17] 155 158
# number of observations is the length of the vector (when no missing values)
length(hb)
## [1] 18
# default quartiles
summary(hb)
1
The technical definition of the median for an even set of values includes the entire range between the
two center values. Thus, selecting any single value in this center range is convenient and the center of this
center range is one sensible choice for the median, M .
1.3: Graphical summaries for one quantitative sample 33

## 132 142 148 146 150 158
# standard quartiles
fivenum(hb)
## [1] 132.0 142.0 147.5 150.0 158.0
# range() gives the min and max values
range(hb)
## [1] 132 158
# the range of the data is the (max - min), calculated using diff()
diff(range(hb))
## [1] 26
mean(hb)
## [1] 146.5
# standard deviation
sd(hb)
## [1] 6.382
# standard error of the mean
se <- sd(hb)/sqrt(length(hb))
√
Note that se is the standard error of the sample mean, SEȲ = s/ n, and
is a measure of the precision of the sample mean Ȳ .
ClickerQ s — Numerical summaries
1.3 Graphical summaries for one quantita-

tive sample
There are four graphical summaries of primary interest: the dotplot, the
histogram, the stem-and-leaf display, and the boxplot. There are many
more possible, but these will often be useful. The plots can be customized.
Make liberal use of the help for learning how to customize them. Plots can also
be generated along with many statistical analyses, a point that we will return
to repeatedly.
1.3.1 Dotplots
The dotplot breaks the range of data into many small-equal width intervals,
and counts the number of observations in each interval. The interval count is
superimposed on the number line at the interval midpoint as a series of dots,
usually one for each observation. In the head breadth data, the intervals are
centered at integer values, so the display gives the number of observations at
each distinct observed head breadth.
A dotplot of the head breadth data is given below. Of the examples below,
the R base graphics stripchart() with method="stack" resembles the traditional
dotplot.
#### stripchart-ggplot
# stripchart (dotplot) using R base graphics
# 3 rows, 1 column
par(mfrow=c(3,1))
stripchart(hb, main="Modern Englishman", xlab="head breadth (mm)")
stripchart(hb, method="stack", cex=2
, main="larger points (cex=2), method is stack")
stripchart(hb, method="jitter", cex=2, frame.plot=FALSE
, main="no frame, method is jitter")
# dotplot using ggplot

library(ggplot2)
# first put hb vector into a data.frame
hb_df <- data.frame(hb)
p1 <- ggplot(hb_df, aes(x = hb))
p1 <- p1 + geom_dotplot(binwidth = 2)
p1 <- p1 + labs(title = "Modern Englishman head breadth")
p1 <- p1 + xlab("head breadth (mm)")

p2 <- p2 + geom_dotplot(binwidth = 2, stackdir = "center")
p2 <- p2 + labs(title = "Modern Englishman head breadth, stackdir=center")

p3 <- p3 + geom_dotplot(binwidth = 2, stackdir = "centerwhole")
p3 <- p3 + labs(title = "Modern Englishman head breadth, stackdir=centerwhole")
library(gridExtra)
grid.arrange(p1, p2, p3, ncol=1)
Modern Englishman Modern Englishman head breadth
1.00
0.75
count
0.50
0.25
0.00
135 140 145 150 155
130 140 150 160
head breadth (mm) head breadth (mm)
larger points (cex=2), method is stack Modern Englishman head breadth, stackdir=center
0.50
0.25
count
0.00
−0.25
−0.50
135 140 145 150 155
130 140 150 160
head breadth (mm)
no frame, method is jitter Modern Englishman head breadth, stackdir=centerwhole

0.50
0.25
count
0.00
−0.25
−0.50
135 140 145 150 155
130 140 150 160
head breadth (mm)
1.3.2 Histogram
The histogram and stem-and-leaf displays are similar, breaking the range
of data into a smaller number of equal-width intervals. This produces graphical
information about the observed distribution by highlighting where data values
cluster. The histogram can use arbitrary intervals, whereas the intervals for the
stem-and-leaf display use the base 10 number system. There is more arbitrari-
ness to histograms than to stem-and-leaf displays, so histograms can sometimes
be regarded a bit suspiciously.
#### hist
# histogram using R base graphics
# par() gives graphical options
# mfrow = "multifigure by row" with 1 row and 3 columns
par(mfrow=c(1,3))
# main is the title, xlab is x-axis label (ylab also available)
hist(hb, main="Modern Englishman", xlab="head breadth (mm)")
# breaks are how many bins-1 to use
hist(hb, breaks = 15, main="Histogram, 15 breaks")

# freq=FALSE changes the vertical axis to density,
# so the total area of the bars is now equal to 1
hist(hb, breaks = 8, freq = FALSE, main="Histogram, density")
# histogram using ggplot

library(ggplot2)
p <- ggplot(hb_df, aes(x = hb))
# always specify a binwidth for the histogram (default is range/30)
# try several binwidths
p <- p + geom_histogram(binwidth = 5)
p <- p + labs(title = "Modern Englishman head breadth")
print(p)
Modern Englishman Histogram, 15 breaks Histogram, density Modern Englishman head breadth
6
8
0.08
6
0.06
4
Frequency
Frequency
Density
count
4
0.04
2
0.02
2
0.00
0
0
130 140 150 160 135 145 155 130 140 150 160
130 140 150 160
head breadth (mm) hb hb hb
R allows you to modify the graphical display. For example, with the his-
togram you might wish to use different midpoints or interval widths. I will let
you explore the possibilities.
1.3.3 Stem-and-leaf plot

A stem-and-leaf plot is a character display histogram defining intervals for a
grouped frequency distribution using the base 10 number system. Intervals are
generated by selecting an appropriate number of lead digits for the data values
to be the stem. The remaining digits comprise the leaf. It is useful for small
samples.
Character plots use typewriter characters to make graphs, and can be con-
venient for some simple displays, but require use of fixed fonts (like Courier)
when copied to a word processing program or they get distorted.
The display almost looks upside down, since larger numbers are on the
bottom rather than the top. It is done this way so that if rotated 90 degrees
counter-clockwise it is a histogram.
The default stem-and-leaf display for the head breadth data is given below.
The two columns give the stems and leaves. The data have three digits. The
first two comprise the stem. The last digit is the leaf. Thus, a head breadth
of 154 has a stem of 15 and leaf of 4. The possible stems are 13, 14, and 15,
whereas the possible leaves are the integers from 0 to 9. In the first plot, each
stem occurs once, while in the second each stem occurs twice. In the second
instance, the first (top) occurrence of a stem value only holds leaves 0 through
4. The second occurrence holds leaves 5 through 9. The display is generated
by placing the leaf value for each observation on the appropriate stem line.
For example, the top 14 stem holds data values between 140 and 144.99. The
stems on this line in the display tell us that four observations fall in this range:
140, 141, 142 and 144. Note that this stem-and-leaf display is an elaborate
histogram with intervals of width 5. An advantage of the stem-and-leaf display
over the histogram is that the original data values can essentially be recovered
from the display.
#### stem-and-leaf
# stem-and-leaf plot
stem(hb)
##
## The decimal point is 1 digit(s) to the right of the |
##
## 13 | 28
## 14 | 0124567889
## 15 | 000458
# scale=2 makes plot roughly twice as wide
stem(hb, scale=2)
##
##
## 13 | 2
## 13 | 8
## 14 | 0124
## 14 | 567889
## 15 | 0004
## 15 | 58
# scale=5 makes plot roughly five times as wide
stem(hb, scale=5)
##
## The decimal point is at the |
##
## 132 | 0
## 134 |
## 136 |
## 138 | 0
## 140 | 00
## 142 | 0
## 144 | 00
## 146 | 00
## 148 | 000
## 150 | 000
## 152 |
## 154 | 00
## 156 |
## 158 | 0
The data values are always truncated so that a leaf has one digit. The leaf
unit (location of the decimal point) tells us the degree of round-off. This will
become clearer in the next example.
Of the three displays, which is the most informative? I think the middle
option is best to see the clustering and shape of distributions of numbers.
ClickerQ s — Stem-and-leaf plot
1.3.4 Boxplot or box-and-whiskers plot
The boxplot breaks up the range of data values into regions about the center
of the data, measured by the median. The boxplot highlights outliers and
provides a visual means to assess “normality”. The following help entry
outlines the construction of the boxplot, given the placement of data values on
the axis.
The endpoints of the box are placed at the locations of the first and third
quartiles. The location of the median is identified by the line in the box. The
whiskers extend to the data points closest to but not on or outside the outlier
fences, which are 1.5IQR from the quartiles. Outliers are any values on or
outside the outlier fences.
The boxplot for the head breadth data is given below. There are a lot
of options that allow you to clutter the boxplot with additional information.
Just use the default settings. We want to see the relative location of data (the
median line), have an idea of the spread of data (IQR, the length of the box),
see the shape of the data (relative distances of components from each other –
to be covered later), and identify outliers (if present). The default boxplot has
all these components.
Note that the boxplots below are horizontal to better fit on the page. The
horizontal=TRUE and coord_flip() commands do this.
#### boxplot
fivenum(hb)
## [1] 132.0 142.0 147.5 150.0 158.0
# boxplot using R base graphics
par(mfrow=c(1,1))
boxplot(hb, horizontal=TRUE
, main="Modern Englishman", xlab="head breadth (mm)")
# boxplot using ggplot

library(ggplot2)
p <- ggplot(hb_df, aes(x = "hb", y = hb))
p <- p + coord_flip()
p <- p + labs(title = "Modern Englishman head breadth")
print(p)
Modern Englishman Modern Englishman head breadth
"hb"
hb
135 140 145 150 155
head breadth (mm) 140 150

hb
ClickerQ s — Boxplots
Improvements to the boxplot
As a quick aside, a violin plot is a combination of a boxplot and a kernel density

plot. They can be created using the vioplot() function from vioplot package.
#### vioplot
# vioplot using R base graphics
# 3 rows, 1 column
par(mfrow=c(3,1))
# histogram
hist(hb, freq = FALSE
, main="Histogram with kernel density plot, Modern Englishman")
# Histogram overlaid with kernel density curve
points(density(hb), type = "l")
# rug of points under histogram
rug(hb)
# violin plot
library(vioplot)
vioplot(hb, horizontal=TRUE, col="gray")
title("Violin plot, Modern Englishman")
# boxplot
boxplot(hb, horizontal=TRUE
, main="Boxplot, Modern Englishman", xlab="head breadth (mm)")
Histogram with kernel density plot, Modern Englishman
0.08
Density
0.04
0.00
130 135 140 145 150 155 160
hb
Violin plot, Modern Englishman
1
135 140 145 150 155
Boxplot, Modern Englishman
135 140 145 150 155
head breadth (mm)
Example: income The data below are incomes in $1000 units for a sample
of 12 retired couples. Numerical and graphical summaries are given. There are
two stem-and-leaf displays provided. The first is the default display.
#### Income examples
income <- c(7, 1110, 7, 5, 8, 12, 0, 5, 2, 2, 46, 7)
# sort in decreasing order
income <- sort(income, decreasing = TRUE)
income
## [1] 1110 46 12 8 7 7 7 5 5 2 2 0
summary(income)
## 0.0 4.2 7.0 101.0 9.0 1110.0
stem(income)
##
##
## 0 | 00000000000
## 0 |
## 1 | 1
Because the two large outliers, I trimmed them to get a sense of the shape
of the distribution where most of the observations are.
#### remove largest
# remove two largest values (the first two)
income2 <- income[-c(1,2)]
income2
## [1] 12 8 7 7 7 5 5 2 2 0
summary(income2)
## 0.00 2.75 6.00 5.50 7.00 12.00
stem(income2)
##
##
## 0 | 022
## 0 | 557778
## 1 | 2
# scale=2 makes plot roughly twice as wide
stem(income2, scale=2)
##
##
## 0 | 0
## 2 | 00
## 4 | 00
## 6 | 000
## 8 | 0
## 10 |
## 12 | 0
Boxplots with full data, then incrementally removing the two largest outliers.
#### income-boxplot
# 1 row, 3 columns
par(mfrow=c(1,3))
boxplot(income, main="Income")
boxplot(income[-1], main="(remove largest)")
boxplot(income2, main="(remove 2 largest)")
Income (remove largest) (remove 2 largest)

12
● ●
1000
40
10
800
8
30
600
6
20
400
4
10
200
●
0
0
1.4: Interpretation of Graphical Displays for Numerical Data 43
1.4 Interpretation of Graphical Displays for
Numerical Data
In many studies, the data are viewed as a subset or sample from a larger
collection of observations or individuals under study, called the population.
A primary goal of many statistical analyses is to generalize the information in
the sample to infer something about the population. For this generalization
to be possible, the sample must reflect the basic patterns of the population.
There are several ways to collect data to ensure that the sample reflects the
basic properties of the population, but the simplest approach, by far, is to
take a random or “representative” sample from the population. A random
sample has the property that every possible sample of a given size has the
same chance of being the sample (eventually) selected (though we often do this
only once). Random sampling eliminates any systematic biases associated with
the selected observations, so the information in the sample should accurately
reflect features of the population. The process of sampling introduces random
variation or random errors associated with summaries. Statistical tools are used
to calibrate the size of the errors.
Whether we are looking at a histogram (or stem-and-leaf, or dotplot) from
a sample, or are conceptualizing the histogram generated by the population
data, we can imagine approximating the “envelope” around the display with a
smooth curve. The smooth curve that approximates the population histogram
is called the population frequency curve or population probability
density function or population distribution2. Statistical methods for
inference about a population usually make assumptions about the shape of the
population frequency curve. A common assumption is that the population has
a normal frequency curve. In practice, the observed data are used to assess
the reasonableness of this assumption. In particular, a sample display should
resemble a population display, provided the collected data are a random or rep-
resentative sample from the population. Several common shapes for frequency
distributions are given below, along with the statistical terms used to describe
2
“Distribution function” often refers to the “cumulative distribution function”, which is a different (but
one-to-one related) function than what I mean here.
them.
Unimodal, symmetric, bell-shaped, and no outliers The first dis-

play is unimodal (one peak), symmetric, and bell-shaped with no out-
liers. This is the prototypical normal curve. The boxplot (laid on its side for
this display) shows strong evidence of symmetry: the median is about halfway
between the first and third quartiles, and the tail lengths are roughly equal. The
boxplot is calibrated in such a way that 7 of every 1000 observations are outliers
(more than 1.5(Q3 − Q1) from the quartiles) in samples from a population with
a normal frequency curve. Only 2 out of every 1 million observations are ex-
treme outliers (more than 3(Q3 − Q1) from the quartiles). We do not have any
outliers here out of 250 observations, but we certainly could have some without
indicating nonnormality. If a sample of 30 observations contains 4 outliers, two
of which are extreme, would it be reasonable to assume the population from
which the data were collected has a normal frequency curve? Probably not.
#### Unimodal, symmetric, bell-shaped, and no outliers (Normal distribution)
## base graphics
# sample from normal distribution
x1 <- rnorm(250, mean = 100, sd = 15)
par(mfrow=c(3,1))
hist(x1, freq = FALSE, breaks = 20)
points(density(x1), type = "l")
rug(x1)
# violin plot
library(vioplot)
vioplot(x1, horizontal=TRUE, col="gray")
# boxplot
boxplot(x1, horizontal=TRUE)
## ggplot
x1_df <- data.frame(x1)
p1 <- ggplot(x1_df, aes(x = x1))
# Histogram with density instead of count on y-axis
p1 <- p1 + geom_histogram(aes(y=..density..)
, binwidth=5
, colour="black", fill="white")
# Overlay with transparent density plot
p1 <- p1 + geom_density(alpha=0.1, fill="#FF6666")

p1 <- p1 + geom_point(aes(y = -0.001)
, position = position_jitter(height = 0.0005)
, alpha = 1/5)
# violin plot
p2 <- ggplot(x1_df, aes(x = "x1", y = x1))
p2 <- p2 + geom_violin(fill = "gray50")
p2 <- p2 + geom_boxplot(width = 0.2, alpha = 3/4)
p2 <- p2 + coord_flip()
# boxplot
p3 <- p3 + geom_boxplot()
library(gridExtra)
Histogram of x1 0.03
0.02
density
0.020
Density
0.01
0.000
0.00
60 80 100 120 140
50 75 100 125 150
x1 x1
"x1"
●
x1 ● ●
1
60 80 100 120 140

75 100 125
x1
"x1"
● ●
x1 ● ●
60 80 100 120 140

75 100 125
x1
#### Central statistical moments

# moments package for 3rd and 4th moments: skewness() and kurtosis()
library(moments)
summary(x1)
## 59.4 89.0 99.3 99.3 109.0 143.0
sd(x1)
## [1] 15.1
skewness(x1)
## [1] 0.1424
kurtosis(x1)
## [1] -0.14
stem(x1)
##
##
## 5 | 9
## 6 | 4
## 6 | 5889
## 7 | 3333344
## 7 | 578888899
## 8 | 01111122222223344444
## 8 | 55555666667777888889999999
## 9 | 000111111122222233333344
## 9 | 5555555556666666677777888888899999999
## 10 | 00000111222222233333344444
## 10 | 555555555666666667777777788888999999999
## 11 | 0000011111122233444
## 11 | 566677788999
## 12 | 00001123444
## 12 | 5679
## 13 | 00022234
## 13 | 6
## 14 | 3
Unimodal, symmetric, heavy-tailed The boxplot is better at highlight-

ing outliers than are other displays. The histogram and stem-and-leaf displays
below appear to have the same basic shape as a normal curve (unimodal, sym-
metric). However, the boxplot shows that we have a dozen outliers in a sample
of 250 observations. We would only expect about two outliers in 250 observa-
tions when sampling from a population with a normal frequency curve. The
frequency curve is best described as unimodal, symmetric, and heavy-tailed.
#### Unimodal, symmetric, heavy-tailed
x2.temp <- rnorm(250, mean = 0, sd = 1)
x2 <- sign(x2.temp)*x2.temp^2 * 15 + 100
par(mfrow=c(3,1))
rug(x2)
# violin plot
library(vioplot)
# boxplot

, binwidth=5
, alpha = 1/5)
# violin plot
# boxplot
library(gridExtra)
Histogram of x2
0.04
density
0.015
Density
0.02
0.000
0.00
−150 −100 −50 0 50 100 150
−100 0 100
x2 x2
"x2"
●
x2 ● ● ●●● ●●●
●●●
●●
●● ●
●
●●●● ●
●
●●●●
●
●●
●●●
●●●
●●●●
●
●
1
−150 −100 −50 0 50 100 150

−100 0 100
x2
"x2"
● ● ● ●● ●●●
●●●
●●●● ● ●
●●
●● ●
● ●●
●●● ●
●●
●●
●●●● ●●●●●●●
●
x2 ● ● ●●● ●●●
●●●
●●
●● ●
●
●●●● ●
●
●●●●
●
●●
●●●
●●●
●●●●
●
●
−150 −100 −50 0 50 100 150

−100 0 100
x2
summary(x2)
## -139.0 95.3 100.0 99.5 108.0 155.0

sd(x2)
## [1] 25.79
skewness(x2)
## [1] -3.345
kurtosis(x2)
## [1] 28.61
stem(x2)
##
##
## -12 | 9
## -10 |
## -8 |
## -6 |
## -4 |
## -2 |
## -0 |
## 0 | 3
## 2 | 69
## 4 | 168002444778
## 6 | 078992559
## 8 | 00223333455577889999000112233334445555666666666677777778888888888899
## 10 | 00000000000000000000000000000000000011111111111122222222233333344444+32
## 12 | 001244667789902566678
## 14 | 001267901345
Symmetric, (uniform,) short-tailed Not all symmetric distributions

are mound-shaped, as the display below suggests. The boxplot shows symme-
try, but the tails of the distribution are shorter (lighter) than in the normal
distribution. Note that the distance between quartiles is roughly constant here.
#### Symmetric, (uniform,) short-tailed
# sample from uniform distribution
x3 <- runif(250, min = 50, max = 150)
par(mfrow=c(3,1))
rug(x3)
# violin plot
library(vioplot)
# boxplot

, binwidth=5
, alpha = 1/5)
# violin plot
# boxplot
library(gridExtra)
Histogram of x3 0.015
0.010
density
0.010
Density
0.005
0.000
0.000
60 80 100 120 140
40 60 80 100 120 140 160
x3 x3
"x3"
●
x3
1
60 80 100 120 140

50 75 100 125 150
x3
"x3"
x3
60 80 100 120 140

50 75 100 125 150
x3
summary(x3)
## 50.4 75.1 97.4 99.5 125.0 150.0
sd(x3)
## [1] 28.87
skewness(x3)
## [1] 0.0481
kurtosis(x3)
## [1] -1.225
stem(x3)
##
##
## 5 | 011122223344
## 5 | 5666677778
## 6 | 000111123334
## 6 | 55667788999
## 7 | 0000112233344444
## 7 | 5556677788999
## 8 | 0000111222223333444
## 8 | 577888899
## 9 | 00001222333444
## 9 | 5566777777888888999
## 10 | 122333334
## 10 | 5566799
## 11 | 012223334
## 11 | 566667777889
## 12 | 01222333334444
## 12 | 5566788899
## 13 | 000000112223334444
## 13 | 556668999
## 14 | 00001111222333444
## 14 | 557888999
## 15 | 0
The mean and median are identical in a population with a (exact) symmetric
frequency curve. The histogram and stem-and-leaf displays for a sample selected
from a symmetric population will tend to be fairly symmetric. Further, the
sample means and medians will likely be close.
Unimodal, skewed right The distribution below is unimodal, and asym-

metric or skewed. The distribution is said to be skewed to the right, or
upper end, because the right tail is much longer than the left tail. The boxplot
also shows the skewness – the region between the minimum observation and the
median contains half the data in less than 1/5 the range of values. In addition,
the upper tail contains several outliers.
#### Unimodal, skewed right
# sample from exponential distribution
x4 <- rexp(250, rate = 1)
par(mfrow=c(3,1))
rug(x4)
# violin plot
library(vioplot)
# boxplot

, binwidth=0.5
, alpha = 1/5)
# violin plot
# boxplot
library(gridExtra)
Histogram of x4
0.6
density
0.6
0.4
Density
0.3
0.2
0.0
0.0
0 2 4 6
0 2 4 6 8
x4 x4
"x4"
●
x4 ● ● ● ● ●
1
0 1 2 3 4 5 6 7
0 2 4 6
x4
"x4"
● ● ● ● ●
x4 ● ● ● ● ●
0 1 2 3 4 5 6 7
0 2 4 6
x4
summary(x4)
## 0.005 0.333 0.744 1.080 1.590 7.060
sd(x4)
## [1] 1.044
skewness(x4)
## [1] 1.756
kurtosis(x4)
## [1] 4.708
stem(x4)
##
##
## 0 | 00000000000000111111111111111111111111112222222222222233333333334444
## 0 | 555555555555555555555666666666666666666677777778888999999999999
## 1 | 0000000000011111111222222333333344444
## 1 | 55556666666667778888899999
## 2 | 0000000111111222233344
## 2 | 66688
## 3 | 022333333444
## 3 | 569
## 4 |
## 4 | 6
## 5 | 2
## 5 |
## 6 |
## 6 |
## 7 | 1
Unimodal, skewed left The distribution below is unimodal and skewed
to the left. The two examples show that extremely skewed distributions often
contain outliers in the longer tail of the distribution.
#### Unimodal, skewed left
x5 <- 15 - rexp(250, rate = 0.5)
par(mfrow=c(3,1))
rug(x5)
# violin plot
library(vioplot)
# boxplot

, binwidth=0.5
, alpha = 1/5)
# violin plot
# boxplot
library(gridExtra)
Histogram of x5
0.4 0.4
0.3
density
Density
0.2
0.2
0.1
0.0
0.0
6 8 10 12 14
4 8 12 16
x5 x5
"x5"
●
x5 ●● ● ● ●
●●● ● ● ● ●●●
1
6 8 10 12 14
5.0 7.5 10.0 12.5 15.0
x5
"x5"
● ● ● ● ●● ● ● ●
● ● ● ● ●●
x5 ●● ● ● ●
●●● ● ● ● ●●●
6 8 10 12 14
5.0 7.5 10.0 12.5 15.0
x5
summary(x5)
## 4.99 12.30 13.60 13.00 14.40 15.00
sd(x5)
## [1] 2.03
skewness(x5)
## [1] -1.713
kurtosis(x5)
## [1] 2.96
stem(x5)
##
##
## 4 |
## 5 | 017
## 6 | 05567
## 7 | 55
## 8 | 15
## 9 | 0112467889
## 10 | 0012234566688
## 11 | 00122346667777788
## 12 | 001111222233444555555556667778888999
## 13 | 00000000111111122223333334445555666666666777888888999999
## 14 | 00000000000111111111112222223333333333444444444445555555555555666666+19
## 15 | 0000000
Bimodal (multi-modal) Not all distributions are unimodal. The distribu-
tion below has two modes or peaks, and is said to be bimodal. Distributions
with three or more peaks are called multi-modal.
#### Bimodal (multi-modal)
x6 <- c(rnorm(150, mean = 100, sd = 15), rnorm(150, mean = 150, sd = 15))
par(mfrow=c(3,1))
rug(x6)
# violin plot
library(vioplot)
# boxplot

, binwidth=5
, alpha = 1/5)
# violin plot
# boxplot
library(gridExtra)
Histogram of x6
0.015
density
0.010
0.010
Density
0.005
0.000
0.000
100 150 200
50 100 150 200
x6 x6
"x6"
●
x6
1
60 80 100 120 140 160 180 200

100 150 200
x6
"x6"
x6
60 80 100 120 140 160 180 200

100 150 200
x6
summary(x6)
## 61.1 102.0 124.0 126.0 153.0 204.0
sd(x6)
## [1] 28.84
skewness(x6)
## [1] -0.01518
kurtosis(x6)
## [1] -1.113
stem(x6)
##
##
## 6 | 139
## 7 | 0145677
## 8 | 001112233355677778899
## 9 | 001112223333344455566667777788899999
## 10 | 0001112222222333344444555555666677777777888999
## 11 | 00011111334556667777888899999
## 12 | 112333344556778889
## 13 | 0011122235555677777
## 14 | 000001122333344445666777777788999
## 15 | 000000012233333334444555555555555566666778889
## 16 | 00000011122222222333444555666789
## 17 | 0113355677
## 18 |
## 19 |
## 20 | 4
The boxplot and histogram or stem-and-leaf display (or dotplot) are used
1.5: Interpretations for examples 57
together to describe the distribution. The boxplot does not provide infor-
mation about modality – it only tells you about skewness and the presence of
outliers.
As noted earlier, many statistical methods assume the population frequency
curve is normal. Small deviations from normality usually do not dramatically
influence the operating characteristics of these methods. We worry most when
the deviations from normality are severe, such as extreme skewness or heavy
tails containing multiple outliers.
ClickerQ s — Graphical summaries
1.5 Interpretations for examples

The head breadth sample is slightly skewed to the left, unimodal, and has no
outliers. The distribution does not deviate substantially from normality. The
various measures of central location (Ȳ = 146.5, M = 147.5) are close, which
is common with fairly symmetric distributions containing no outliers.
The income sample is extremely skewed to the right due to the presence of
two extreme outliers at 46 and 1110. A normality assumption here is unrealistic.
It is important to recognize the influence that outliers can have on the values
of Ȳ and s. The median and interquartile range are more robust (less sensitive)
to the presence of outliers. For the income data Ȳ = 100.9 and s = 318, whereas
M = 7 and IQR = 8.3. If we omit the two outliers, then Ȳ = 5.5 and s = 3.8,
whereas M = 6 and IQR = 5.25.
The mean and median often have similar values in data sets without outliers,
so it does not matter much which one is used as the “typical value”. This issue
is important, however, in data sets with extreme outliers. In such instances,
the median is often more reasonable. For example, is Ȳ = 100.9 a reasonable
measure for a typical income in this sample, given that the second largest income
is only 46?
R Discussion I have included basic pointers on how to use R in these notes.

I find that copying an R code example from the internet, then modifying the
code to apply to my data is a productive strategy. When you’re first learning,
“pretty good” is often “good enough”, especially when it comes to plots (in other
words, spending 20 minutes vs 2 hours on a plot is fine). I will demonstrate
most of what you need and I will be happy to answer questions. You will learn
a lot more by using and experimenting with R than by watching.
Chapter 2
Estimation in
One-Sample Problems
Learning objectives
select graphical displays that meaningfully communicate properties of a
sample.
assess the assumptions of the one-sample t-test visually.
decide whether the mean of a population is different from a hypothesized
value.
recommend action based on a hypothesis test.
5. define parameters of interest and hypotheses in words and notation.
12. make evidence-based decisions.
2.1 Inference for a population mean

Suppose that you have identified a population of interest where individuals are
measured on a single quantitative characteristic, say, weight, height or IQ. You
select a random or representative sample from the population with the goal
60 Ch 2: Estimation in One-Sample Problems
of estimating the (unknown) population mean value, identified by µ. You
cannot see much of the population, but you would like to know what is typical
in the population (µ). The only information you can see is that in the sample.
This is a standard problem in statistical inference, and the first inferential
problem that we will tackle. For notational convenience, identify the measure-
ments on the sample as Y1, Y2, . . . , Yn, where n is the sample size. Given
P the
Y
data, our best guess, or estimate, of µ is the sample mean: Ȳ = ni i =
Y1 +Y2 +···+Yn
n .
Population
Huge set of values
Can see very little
Sample
Y1, Y2, …, Yn
Inference
Mean µ
Standard Deviation σ
µ and σ unknown
There are two main methods that are used for inferences on µ: confidence
intervals (CI) and hypothesis tests. The standard CI and test procedures
are based on the sample mean and the sample standard deviation, denoted by
s.
ClickerQ s — Inference for a population mean, 2
2.1.1 Standard error, LLN, and CLT

The standard error (SE) is the standard deviation of the sampling dis-
tribution of a statistic.
The sampling distribution of a statistic is the distribution of that statis-
tic, considered as a random variable, when derived from a random sample of
2.1: Inference for a population mean 61
size n.
The standard error of the mean (SEM) is the standard deviation of
the sample-mean’s estimate of a population mean. (It can also be viewed as the
standard deviation of the error in the sample mean relative to the true mean,
since the sample mean is an unbiased estimator.) SEM is usually estimated
by the sample estimate of the population standard deviation (sample standard
deviation) divided by the square root of the sample size (assuming statistical
independence of the values in the sample):
√
SEȲ = s/ n
where s is the sample standard deviation (i.e., the sample-based estimate of the
standard deviation of the population), and n is the size (number of observations)
of the sample.
In probability theory, the law of large numbers (LLN) is a theorem
that describes the result of performing the same experiment a large number of
times. According to the law, the average of the results obtained from a large
number of trials (the sample mean, Ȳ ) should be close to the expected value
(the population mean, µ), and will tend to become closer as more trials are
performed.
In probability theory, the central limit theorem (CLT) states that,
given certain conditions, the mean of a sufficiently large number of independent
random variables, each with finite mean and variance, will be approximately
normally distributed1.
As a joint illustration of these concepts, consider drawing random variables
following a Uniform(0,1) distribution, that is, any value in the interval [0, 1]
is equally likely. By definition, the mean of this distributionpis µ = 1/2 and
the variance is σ 2 = 1/12 (so the standard deviation is σ = 1/12 = 0.289).
Therefore, if we draw a sample of size n, then the standard error of the mean
will be σ/n, and as n gets larger the distribution of the mean will increasingly
follow a normal distribution. We illustrate this by drawing N = 10000 samples
of size n and plot those N means, computing the expected and observed SEM
1
The central limit theorem has a number of variants. In its common form, the random variables must
be identically distributed. In variants, convergence of the mean to the normal distribution also occurs for
non-identical distributions, given that they comply with certain conditions.
and how well the histogram of sampled means follows a normal distribution,
Notice, indeed, that even with samples as small as 2 and 6 that the properties
of the SEM and the distribution are as predicted.
#### Illustration of Central Limit Theorem, Uniform distribution
# demo.clt.unif(N, n)
# draws N samples of size n from Uniform(0,1)
# and plots the N means with a normal distribution overlay
demo.clt.unif <- function(N, n) {
# draw sample in a matrix with N columns and n rows
sam <- matrix(runif(N*n, 0, 1), ncol=N);
# calculate the mean of each column
sam.mean <- colMeans(sam)
# the sd of the mean is the SEM
sam.se <- sd(sam.mean)
# calculate the true SEM given the sample size n
true.se <- sqrt((1/12)/n)
# draw a histogram of the means
hist(sam.mean, freq = FALSE, breaks = 25
, main = paste("True SEM =", round(true.se, 4)
, ", Est SEM = ", round( sam.se, 4))
, xlab = paste("n =", n))
# overlay a density curve for the sample means
points(density(sam.mean), type = "l")
# overlay a normal distribution, bold and red
x <- seq(0, 1, length = 1000)
points(x, dnorm(x, mean = 0.5, sd = true.se), type = "l", lwd = 2, col = "red")
# place a rug of points under the plot
rug(sam.mean)
}
par(mfrow=c(2,2));
demo.clt.unif(10000, 1);
True SEM = 0.2887 , Est SEM = 0.2893 True SEM = 0.2041 , Est SEM = 0.2006
2.0
1.0
1.5
0.8
0.6
Density
Density
1.0
0.4
0.5
0.2
0.0
0.0
0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0
n=1 n=2
True SEM = 0.1179 , Est SEM = 0.1194 5

True SEM = 0.0833 , Est SEM = 0.0843
3.0
4
2.5
2.0
3
Density
Density
1.5
2
1.0
1
0.5
0.0
0.2 0.4 0.6 0.8 0.2 0.3 0.4 0.5 0.6 0.7 0.8
n=6 n = 12
In a more extreme example, we draw samples from an Exponential(1) dis-

tribution (µ = 1 and σ = 1), which is strongly skewed to the right. Notice that
the normality promised by the CLT requires larger samples sizes, about n ≥ 30,
than for the previous Uniform(0,1) example, which required about n ≥ 6.
#### Illustration of Central Limit Theorem, Exponential distribution
# demo.clt.exp(N, n) draws N samples of size n from Exponential(1)
# and plots the N means with a normal distribution overlay
demo.clt.exp <- function(N, n) {
# draw sample in a matrix with N columns and n rows
sam <- matrix(rexp(N*n, 1), ncol=N);
# calculate the mean of each column

# the sd of the mean is the SEM
sam.se <- sd(sam.mean)
# calculate the true SEM given the sample size n
true.se <- sqrt(1/n)
, main = paste("True SEM =", round(true.se, 4), ", Est SEM = ", round(sam.se, 4))
, xlab = paste("n =", n))
x <- seq(0, 5, length = 1000)
points(x, dnorm(x, mean = 1, sd = true.se), type = "l", lwd = 2, col = "red")
rug(sam.mean)
}
par(mfrow=c(2,2));
demo.clt.exp(10000, 1);
0.8
True SEM = 1 , Est SEM = 1.006 True SEM = 0.4082 , Est SEM = 0.4028
1.0
0.6
0.8
0.6
Density
Density
0.4
0.4
0.2
0.2
0.0
0.0
0 2 4 6 8 10 0.0 0.5 1.0 1.5 2.0 2.5 3.0
n=1 n=6
True SEM = 0.1826 , Est SEM = 0.1827 4

True SEM = 0.1 , Est SEM = 0.0995
2.0
3
1.5
Density
Density
2
1.0
0.5
1
0.0
0.5 1.0 1.5 2.0 0.8 1.0 1.2 1.4
n = 30 n = 100
Note well that the further the population distribution is from being nor-
mal, the larger the sample size is required to be for the sampling distribution
of the sample mean to be normal. If the population distribution is normal,
what’s the minimum sample size for the sampling distribution of the mean to
be normal?
For more examples, try:
#### More examples for Central Limit Theorem can be illustrated with this code
# install.packages("TeachingDemos")
library(TeachingDemos)
# look at examples at bottom of the help page
?clt.examp
2.1.2 t-distribution
The Student’s t-distribution is a family of continuous probability distribu-

tions that arises when estimating the mean of a normally distributed population
in situations where the sample size is small and population standard devi-
ation is unknown. The t-distribution is symmetric and bell-shaped, like the
normal distribution, but has heavier tails, meaning that it is more prone to pro-
ducing values that fall far from its mean. Effectively, the t-distribution is wider
than the normal distribution because in addition to estimating the mean µ with
Ȳ , we also have to estimate σ 2 with s2, so there’s some additional uncertainty.
The degrees-of-freedom (df) parameter of the t-distribution is the sample size
n minus the number of variance parameters estimated. Thus, df = n − 1 when
we have one sample and df = n − 2 when we have two samples. As n increases,
the t-distribution becomes close to the normal distribution, and when n = ∞
the distributions are equivalent.
#### Normal vs t-distributions with a range of degrees-of-freedom
x <- seq(-8, 8, length = 1000)
par(mfrow=c(1,1))
plot(x, dnorm(x), type = "l", lwd = 2, col = "red"
, main = "Normal (red) vs t-dist with df=1, 2, 6, 12, 30, 100")
points(x, dt(x, 1), type = "l")
points(x, dt(x,100), type = "l")
2.2: CI for µ 67
Normal (red) vs t−dist with df=1, 2, 6, 12, 30, 100
0.4
0.3
dnorm(x)
0.2
0.1
0.0
−5 0 5
2.2 CI for µ
Statistical inference provides methods for drawing conclusions about a pop-
ulation from sample data. In this chapter, we want to make a claim about
population mean µ given sample statistics Ȳ and s.
A CI for µ is a range of plausible values for the unknown population mean
µ, based on the observed data, of the form “Best Guess ± Reasonable Error of
the Guess”. To compute a CI for µ:
1. Define the population parameter, “Let µ = mean [characteristic] for
population of interest”.
2. Specify the confidence coefficient, which is a number between 0 and
100%, in the form 100(1 − α)%. Solve for α. (For example, 95% has
α = 0.05.)
3. Compute the t-critical value: tcrit = t0.5α such that the area under the
t-curve (df = n − 1) to the right of tcrit is 0.5α. See appendix or internet
for a t-table.
4. Report the CI in the form Ȳ ± tcritSEȲ or as an interval (L, U ). The
desired CI has lower and upper endpoints given by L = Ȳ − tcritSEȲ
√
and U = Ȳ + tcritSEȲ , respectively, where SEȲ = s/ n is the standard
error of the sample mean.
5. Assess method assumptions (see below).
In practice, the confidence coefficient is large, say 95% or 99%, which corre-
spond to α = 0.05 and 0.01, respectively. The value of α expressed as a percent
is known as the error rate of the CI.
The CI is determined once the confidence coefficient is specified and the data
are collected. Prior to collecting the data, the interval is unknown and is viewed
as random because it will depend on the actual sample selected. Different
samples give different CIs. The “confidence” in, say, the 95% CI (which has
a 5% error rate) can be interpreted as follows. If you repeatedly sample the
population and construct 95% CIs for µ, then 95% of the intervals will
contain µ, whereas 5% will not. The interval you construct from your data
will either cover µ, or it will not.
The length of the CI
U − L = 2tcritSEȲ
depends on the accuracy of our estimate Ȳ of µ, as measured by the standard

√
error of Ȳ , SEȲ = s/ n. Less precise estimates of µ lead to wider intervals
for a given level of confidence.
An example with 100 CIs Consider drawing a sample of 25 observations

from a normally distributed population with mean 10 and sd 2. Calculate the
95% t-CI. Now do that 100 times. The plot belows reflects the variability of
that process. We expect 95 of the 100 CIs to contain the true population mean
of 10, that is, on average 5 times out of 100 we draw the incorrect inference
that the population mean is in an interval when it does not contain the true
value of 10.
2.2: CI for µ 69
#### Illustration of Confidence Intervals (consistent with their interpretation)

ci.examp(mean.sim = 10, sd = 2, n = 25
, reps = 100, conf.level = 0.95, method = "t")
Confidence intervals based on t distribution

100
|
| | |
| |
| |
|| |
|
| | |
||
| | |
80
| |
|
|| |
| |
| |
| | |
| |
|
|| | |
60
||
| |
| |
| |
Index
| ||
| |
| | |
| |
| ||
40
| |
| | |
| | |
| | |
| |
| | | |
| |
20
| |
| |
| |
| |
| |
||
| |
| |
| ||
|
0
9 10 11 12
Confidence Interval
ClickerQ s — CI for µ, 2
2.2.1 Assumptions for procedures

I described the classical CI. The procedure is based on the assumptions that
the data are a random sample from the population of interest, and that the
population frequency curve is normal. The population frequency curve
can be viewed as a “smoothed histogram” created from the population data.
The normality assumption can never be completely verified without having
the entire population data. You can assess the reasonableness of this assumption
using a stem-and-leaf display or a boxplot of the sample data. The stem-and-
leaf display from the data should resemble a normal curve.
In fact, the assumptions are slightly looser than this, the population fre-
quency curve can be anything provided the sample size is large enough that
it’s reasonable to assume that the sampling distribution of the mean is
normal.
Assessing assumptions using the bootstrap

We will cover the bootstrap at the end of this course, but a brief introduction
here can help us check model assumptions. Recall in Section 2.1.1 the sampling
distribution examples. Assume the sample is representative of the population.
Let’s use our sample as a proxy for the population and repeatedly draw samples
(with replacement) of size n and calculate the mean, then plot the bootstrap
sampling distribution of means. If this bootstrap sampling distribution strongly
deviates from normal, then that’s evidence from the data that inference using
the t-distribution is not appropriate. Otherwise, if roughly normal, then the
t-distribution may be sensible.
#### Visual comparison of whether sampling distribution is close to Normal via Bootstrap
# a function to compare the bootstrap sampling distribution with
# a normal distribution with mean and SEM estimated from the data
bs.one.samp.dist <- function(dat, N = 1e4) {
n <- length(dat);
# resample from data
sam <- matrix(sample(dat, size = N * n, replace = TRUE), ncol=N);
sam.mean <- colMeans(sam);
2.2: CI for µ 71
# save par() settings

old.par <- par(no.readonly = TRUE)
# make smaller margins
par(mfrow=c(2,1), mar=c(3,2,2,1), oma=c(1,1,1,1))
hist(dat, freq = FALSE, breaks = 6
, main = "Plot of data with smoothed density curve")
points(density(dat), type = "l")
rug(dat)

, main = "Bootstrap sampling distribution of the mean"
, xlab = paste("Data: n =", n
, ", mean =", signif(mean(dat), digits = 5)
, ", se =", signif(sd(dat)/sqrt(n)), digits = 5))
x <- seq(min(sam.mean), max(sam.mean), length = 1000)
points(x, dnorm(x, mean = mean(dat), sd = sd(dat)/sqrt(n))
, type = "l", lwd = 2, col = "red")
rug(sam.mean)
# restore par() settings
par(old.par)
}
# example data, skewed --- try others datasets to develop your intuition
x <- rgamma(10, shape = .5, scale = 20)
bs.one.samp.dist(x)
Plot of data with smoothed density curve

0.06
0.04
Density
0.02
0.00
0 10 20 30 40 50
dat
Bootstrap sampling distribution of the mean
0.08
0.06
Density
0.04
0.02
0.00
0 5 10 15 20 25 30 35
Data: n = 10 , mean = 12.07 , se = 4.67812 5
Example: Age at First Heart Transplant Let us go through a hand-

calculation of a CI, using R to generate summary data. I’ll show you later how
to generate the CI in R.
We are interested in the mean age at first heart transplant for a population
of patients.
1. Define the population parameter
Let µ = mean age at the time of first heart transplant for population of
2.3: Hypothesis Testing for µ 73
patients.
2. Calculate summary statistics from sample
The ages (in years) at first transplant for a sample of 11 heart transplant
patients are as follows:
54, 42, 51, 54, 49, 56, 33, 58, 54, 64, 49.
Summaries for √ the data are: n = 11, Ȳ = 51.27, and s = 8.26 so that
SEȲ = 8.26/ 11 = 2.4904. The degrees of freedom are df = 11 − 1 = 10.
3. Specify confidence level, find critical value, calculate limits
Let us calculate a 95% CI for µ. For a 95% CI α = 0.05, so we need to
find tcrit = t0.025, which is 2.228. Now tcritSEȲ = 2.228 × 2.4904 = 5.55.
The lower limit on the CI is L = 51.27 − 5.55 = 45.72. The upper limit is
U = 51.27 + 5.55 = 56.82.
4. Summarize in words For example, I am 95% confident that the
population mean age at first transplant is 51.3 ± 5.55, that is, between 45.7
and 56.8 years (rounding off to 1 decimal place).
5. Check assumptions We will see this in several pages, sampling
distribution is reasonably normal.
2.2.2 The effect of α on a two-sided CI

A two-sided 100(1 − α)% CI for µ is given by Ȳ ± tcritSEȲ . The CI is centered
at Ȳ and has length 2tcritSEȲ . The confidence coefficient 100(1 − α)% is
increased by decreasing α, which increases tcrit. That is, increasing the
confidence coefficient makes the CI wider. This is sensible: to increase your
confidence that the interval captures µ you must pinpoint µ with less precision
by making the CI wider. For example, a 95% CI is wider than a 90% CI.
ClickerQ s — CI, quickie
2.3 Hypothesis Testing for µ

A hypothesis test is used to make a decision about a population parameter.
Suppose you are interested in checking whether the population mean µ is
equal to some prespecified value, say µ0. This question can be formulated as
a two-sided hypothesis test, where you are trying to decide which of two con-
tradictory claims or hypotheses about µ is more reasonable given the observed
data. The null hypothesis, or the hypothesis under test, is H0 : µ = µ0,
whereas the alternative hypothesis is HA : µ 6= µ0.
I will explore the ideas behind hypothesis testing later. At this point, I focus
on the mechanics behind the test. The steps in carrying out the test are:
1. Set up the null and alternative hypotheses in words and notation.
In words: “The population mean for [what is being studied] is different
from [value of µ0].” (Note that the statement in words is in terms of the
alternative hypothesis.)
In notation: H0 : µ = µ0 versus HA : µ 6= µ0 (where µ0 is specified by
the context of the problem).
2. Choose the size or significance level of the test, denoted by α. In
practice, α is set to a small value, say, 0.01 or 0.05, but theoretically can
be any value between 0 and 1.
3. Compute the test statistic
Ȳ − µ0
ts = ,
SEȲ
√
where SEȲ = s/ n is the standard error.
Note: I sometimes call the test statistic tobs to emphasize that the com-
puted value depends on the observed data.
4. Compute the critical value tcrit = t0.5α (or p-value from the test statis-
tic) in the direction of the alternative hypothesis from the t-distribution
table with degrees of freedom df = n − 1.
5. State the conclusion in terms of the problem.
Reject H0 in favor of HA (i.e., decide that H0 is false, based on the data)
if |ts| > tcrit or p-value < α, that is, reject if ts < −tcrit or if ts > tcrit.
Otherwise, Fail to reject H0.
(Note: We DO NOT accept H0 — more on this later.)
6. Check assumptions of the test, when possible (could do earlier to save
yourself some effort if they are not met).
The process is represented graphically below. The area under the t-probability
curve outside ±tcrit is the size of the test, α. One-half α is the area in each tail.
You reject H0 in favor of HA only if the test statistic is outside ±tcrit.
α 1−α α
2 2
Reject H0 Reject H0
− tcrit 0
tcrit
2.3.1 P-values
The p-value, or observed significance level for the test, provides a mea-
sure of plausibility for H0. Smaller values of the p-value imply that H0 is less
plausible. To compute the p-value for a two-sided test, you
1. Compute the test statistic ts as above.

2. Evaluate the area under the t-probability curve (with df = n − 1) outside
±|ts|.
p−value p−value
2 2
− ts 0
ts
The p-value is the total shaded area, or twice the area in either tail. A use-
ful interpretation of the p-value is that it is the chance of obtaining data
favoring HA by this much or more if H0 actually is true. Another interpre-
tation is that
the p-value is the probability of observing a sample mean at
least as extreme as the one observed assuming µ0 from H0 is
the true population mean.
If the p-value is small then the sample we obtained is pretty unusual to have
obtained if H0 is true — but we actually got the sample, so probably it is not
very unusual, so we would conclude H0 is false (it would not be unusual if HA
is true).
Most, if not all, statistical packages summarize hypothesis tests with a p-
value, rather than a decision (i.e., reject or not reject at a given α level). You
can make a decision to reject or not reject H0 for a size α test based on the
p-value as follows — reject H0 if the p-value is less than α. This decision is
identical to that obtained following the formal rejection procedure given earlier.
The reason for this is that the p-value can be interpreted as the smallest value
you can set the size of the test and still reject H0 given the observed data.
There are a lot of terms to keep straight here. α and tcrit are constants
we choose (actually, one determines the other so we really only choose one,
usually α) to set how rigorous evidence against H0 needs to be. ts and the
p-value (again, one determines the other) are random variables because they
are calculated from the random sample. They are the evidence against H0.
2.3.2 Assumptions for procedures

I described the classical t-test, which assumes that the data are a random sample
from the population and that the population frequency curve is normal. These
are the same assumptions as for the CI.
Example: Age at First Transplant (Revisited) The ages (in years)

at first transplant for a sample of 11 heart transplant patients are as follows:
54, 42, 51, 54, 49, 56, 33, 58, 54, 64, 49. Summaries for these data are: n = 11,
Ȳ = 51.27, s = 8.26 and SEȲ = 2.4904. Test the hypothesis that the mean
age at first transplant is 50. Use α = 0.05.
As in the earlier analysis, define
µ = mean age at time of first transplant for population of patients.
We are interested in testing H0 : µ = 50 against HA : µ 6= 50, so µ0 = 50.
The degrees of freedom are df = 11 − 1 = 10. The critical value for a 5%
test is tcrit = t0.025 = 2.228. (Note α/2 = 0.05/2 = 0.025). The same critical
value was used with the 95% CI.
For the test,
Ȳ − µ0 51.27 − 50
ts = = = 0.51.
SEȲ 2.4904
Since tcrit = 2.228, we do not reject H0 using a 5% test. Notice the placement
of ts relative to tcrit in the picture below. Equivalently, the p-value for the test
is 0.62, thus we fail to reject H0 because 0.62 > 0.05 = α. The results of
the hypothesis test should not be surprising, since the CI tells you that 50 is a
plausible value for the population mean age at transplant. Note: All you can
say is that the data could have come from a distribution with a mean of 50 —
this is not convincing evidence that µ actually is 50.
.95
.025 .025
Reject H0 0 Reject H0
−2.228 0.51 2.228 0
−.51 .51
ts in middle of distribution, so do not reject H0 Total shaded area is the p−value, .62
Example: Age at First Transplant R output for the heart transplant

problem is given below. Let us look at the output and find all of the summaries
we computed. Also, look at the graphical summaries to assess whether the
t-test and CI are reasonable here.
#### Example: Age at First Transplant
# enter data as a vector
age <- c(54, 42, 51, 54, 49, 56, 33, 58, 54, 64, 49)
The age data is unimodal, skewed left, no extreme outliers.

par(mfrow=c(2,1))
hist(age, freq = FALSE, breaks = 6)
points(density(age), type = "l")
rug(age)
# violin plot
library(vioplot)
vioplot(age, horizontal=TRUE, col="gray")
Histogram of age
Density
0.04
0.00
30 35 40 45 50 55 60 65
age
●
1
35 40 45 50 55 60 65
stem(age, scale=2)
##
##
## 3 | 3
## 3 |
## 4 | 2
## 4 | 99
## 5 | 1444
## 5 | 68
## 6 | 4
# t.crit
qt(1 - 0.05/2, df = length(age) - 1)
## [1] 2.228
# look at help for t.test
?t.test
# defaults include: alternative = "two.sided", conf.level = 0.95
t.summary <- t.test(age, mu = 50)
t.summary
##
## One Sample t-test
##
## data: age
## t = 0.5111, df = 10, p-value = 0.6204
## alternative hypothesis: true mean is not equal to 50
## 95 percent confidence interval:
## 45.72 56.82
## sample estimates:
## mean of x
## 51.27
summary(age)
## 33.0 49.0 54.0 51.3 55.0 64.0
The assumption of normality of the sampling distribution appears reason-
ablly close, using the bootstrap discussed earlier. Therefore, the results for the
t-test above can be trusted.
bs.one.samp.dist(age)

Density
0.04
0.00
30 35 40 45 50 55 60 65
dat
0.15
0.10
Density
0.05
0.00
40 45 50 55 60
Data: n = 11 , mean = 51.273 , se = 2.49031 5
Aside: To print the shaded region for the p-value, you can use the result of
t.test() with the function t.dist.pval() defined here.
# Function ot plot t-distribution with shaded p-value
t.dist.pval <- function(t.summary) {
par(mfrow=c(1,1))
lim.extreme <- max(4, abs(t.summary$statistic) + 0.5)
lim.lower <- -lim.extreme;
lim.upper <- lim.extreme;
x.curve <- seq(lim.lower, lim.upper, length=200)
y.curve <- dt(x.curve, df = t.summary$parameter)
plot(x.curve, y.curve, type = "n"
, ylab = paste("t-dist( df =", signif(t.summary$parameter, 3), ")")
, xlab = paste("t-stat =", signif(t.summary$statistic, 5)
, ", Shaded area is p-value =", signif(t.summary$p.value, 5)))
if ((t.summary$alternative == "less")
| (t.summary$alternative == "two.sided")) {
x.pval.l <- seq(lim.lower, -abs(t.summary$statistic), length=200)
y.pval.l <- dt(x.pval.l, df = t.summary$parameter)
polygon(c(lim.lower, x.pval.l, -abs(t.summary$statistic))
, c(0, y.pval.l, 0), col="gray")
}
if ((t.summary$alternative == "greater")
| (t.summary$alternative == "two.sided")) {
x.pval.u <- seq(abs(t.summary$statistic), lim.upper, length=200)

y.pval.u <- dt(x.pval.u, df = t.summary$parameter)
polygon(c(abs(t.summary$statistic), x.pval.u, lim.upper)
, c(0, y.pval.u, 0), col="gray")
}
points(x.curve, y.curve, type = "l", lwd = 2, col = "blue")
}
# for the age example

t.dist.pval(t.summary)
0.4
0.3
t−dist( df = 10 )
0.2
0.1
0.0
−4 −2 0 2 4
t−stat = 0.51107 , Shaded area is p−value = 0.62039
Aside: Note that the t.summary object returned from t.test() includes a
number of quantities that might be useful for additional calculations.
names(t.summary)
## [1] "statistic" "parameter" "p.value" "conf.int"
## [5] "estimate" "null.value" "alternative" "method"
## [9] "data.name"
t.summary$statistic
## t
## 0.5111
t.summary$parameter
## df
## 10
t.summary$p.value
## [1] 0.6204
t.summary$conf.int
## [1] 45.72 56.82
## attr(,"conf.level")
## [1] 0.95
t.summary$estimate
## mean of x
## 51.27
t.summary$null.value
## mean
## 50
t.summary$alternative
## [1] "two.sided"
t.summary$method
## [1] "One Sample t-test"
t.summary$data.name
## [1] "age"
Example: Meteorites One theory of the formation of the solar system

states that all solar system meteorites have the same evolutionary history and
thus have the same cooling rates. By a delicate analysis based on measure-
ments of phosphide crystal widths and phosphide-nickel content, the cooling
rates, in degrees Celsius per million years, were determined for samples taken
from meteorites named in the accompanying table after the places they were
found. The Walker2 County (Alabama, US), Uwet3 (Cross River, Nigeria), and
Tocopilla4 (Antofagasta, Chile) meteorite cooling rate data are below.
Suppose that a hypothesis of solar evolution predicted a mean cooling rate of
µ = 0.54 degrees per million year for the Tocopilla meteorite. Do the observed
cooling rates support this hypothesis? Test at the 5% level. The boxplot and
stem-and-leaf display (given below) show good symmetry. The assumption of
a normal distribution of observations basic to the t-test appears to be realistic.
Meteorite Cooling rates
Walker County 0.69 0.23 0.10 0.03 0.56 0.10 0.01 0.02 0.04 0.22
Uwet 0.21 0.25 0.16 0.23 0.47 1.20 0.29 1.10 0.16
Tocopilla 5.60 2.70 6.20 2.90 1.50 4.00 4.30 3.00 3.60 2.40 6.70 3.80
Let
µ = mean cooling rate over all pieces of the Tocopilla meteorite.
To answer the question of interest, we consider the test of H0 : µ = 0.54 against
HA : µ 6= 0.54. Let us go carry out the test, compute the p-value, and calculate
a 95% CI for µ. The sample summaries are n = 12, Ȳ = 3.892, s = 1.583.
√
The standard error is SEȲ = s/ n = 0.457.
R output for this problem is given below. For a 5% test (i.e., α = 0.05), you
would reject H0 in favor of HA because the p-value ≤ 0.05. The data strongly
2
http://www.lpi.usra.edu/meteor/metbull.php?code=24204
3
4
suggest that µ 6= 0.54. The 95% CI says that you are 95% confident that the
population mean cooling rate for the Tocopilla meteorite is between 2.89 and
4.90 degrees per million years. Note that the CI gives us a means to assess how
different µ is from the hypothesized value of 0.54.
#### Example: Meteorites
# enter data as a vector
toco <- c(5.6, 2.7, 6.2, 2.9, 1.5, 4.0, 4.3, 3.0, 3.6, 2.4, 6.7, 3.8)
The Tocopilla data is unimodal, skewed right, no extreme outliers.
par(mfrow=c(2,1))
hist(toco, freq = FALSE, breaks = 6)
points(density(toco), type = "l")
rug(toco)
# violin plot
library(vioplot)
vioplot(toco, horizontal=TRUE, col="gray")
Histogram of toco
0.00 0.15 0.30
Density
1 2 3 4 5 6 7
toco
●
1
2 3 4 5 6
stem(toco, scale=2)
##
##
## 1 | 5
## 2 | 479
## 3 | 068
## 4 | 03
## 5 | 6
## 6 | 27
# t.crit
qt(1 - 0.05/2, df = length(toco) - 1)

## [1] 2.201
t.summary <- t.test(toco, mu = 0.54)
t.summary
##
##
## data: toco
## t = 7.337, df = 11, p-value = 1.473e-05
## alternative hypothesis: true mean is not equal to 0.54
## 2.886 4.897
## mean of x
## 3.892
summary(toco)
## 1.50 2.85 3.70 3.89 4.62 6.70
able. Therefore, the results for the t-test above can be trusted.
bs.one.samp.dist(toco)

0.30
0.4
0.20
Density
0.10
0.3
0.00
t−dist( df = 11 )
1 2 3 4 5 6 7
0.2
dat
0.8
0.1
Density
0.4
0.0
0.0
−5 0 5
2.5 3.0 3.5 4.0 4.5 5.0 5.5
t−stat = 7.3366 , Shaded area is p−value = 1.473e−05
Data: n = 12 , mean = 3.8917 , se = 0.456843 5
2.3.3 The mechanics of setting up hypothesis tests

When setting up a test you should imagine you are the researcher conducting the
experiment. In many studies, the researcher wishes to establish that there has
been a change from the status quo, or that they have developed a method that
produces a change (possibly in a specified direction) in the typical response.
The researcher sets H0 to be the status quo and HA to be the research
hypothesis — the claim the researcher wishes to make. In some studies you
define the hypotheses so that HA is the take action hypothesis — rejecting
H0 in favor of HA leads one to take a radical action.
Some perspective on testing is gained by understanding the mechanics be-
hind the tests. A hypothesis test is a decision process in the face of uncertainty.
You are given data and asked which of two contradictory claims about a pop-
ulation parameter, say µ, is more reasonable. Two decisions are possible, but
whether you make the correct decision depends on the true state of nature
which is unknown to you.
State of nature
Decision H0 true HA true
Fail to reject [accept] H0 correct decision Type-II error
Reject H0 in favor of HA Type-I error correct decision
For a given problem, only one of these errors is possible. For example, if H0
is true you can make a Type-I error but not a Type-II error. Any reasonable
decision rule based on the data that tells us when to reject H0 and when to
not reject H0 will have a certain probability of making a Type-I error if H0 is
true, and a corresponding probability of making a Type-II error if H0 is false
and HA is true. For a given decision rule, define
α = Prob( Reject H0 given H0 is true ) = Prob( Type-I error )
and
β = Prob( Fail to reject H0 when HA true ) = Prob( Type-II error ).
The mathematics behind hypothesis tests allows you to prespecify or control
α. For a given α, the tests we use (typically) have the smallest possible value
of β. Given the researcher can control α, you set up the hypotheses so that
committing a Type-I error is more serious than committing a Type-II error.
The magnitude of α, also called the size or level of the test, should depend
on the seriousness of a Type-I error in the given problem. The more serious
the consequences of a Type-I error, the smaller α should be. In practice α is
often set to 0.10, 0.05, or 0.01, with α = 0.05 being the scientific standard. By
setting α to be a small value, you reject H0 in favor of HA only if the data
convincingly indicate that H0 is false.
Let us piece together these ideas for the meteorite problem. Evolutionary
history predicts µ = 0.54. A scientist examining the validity of the theory is
trying to decide whether µ = 0.54 or µ 6= 0.54. Good scientific practice dictates
that rejecting another’s claim when it is true is more serious than not being able
to reject it when it is false. This is consistent with defining H0 : µ = 0.54 (the
status quo) and HA : µ 6= 0.54. To convince yourself, note that the implications
of a Type-I error would be to claim the evolutionary theory is false when it is
true, whereas a Type-II error would correspond to not being able to refute the
evolutionary theory when it is false. With this setup, the scientist will refute
the theory only if the data overwhelmingly suggest that it is false.
2.3.4 The effect of α on the rejection region of a two-

sided test
For a size α test, you reject H0 : µ = µ0 if
Ȳ − µ0
ts =
SEȲ
satisfies |ts| > tcrit.

0
−3.106 3.106
−2.201 2.201
Rejection Regions for .05 and .01 level tests
The critical value is computed so that the area under the t-probability curve
(with df = n − 1) outside ±tcrit is α, with 0.5α in each tail. Reducing α
makes tcrit larger. That is, reducing the size of the test makes rejecting H0
harder because the rejection region is smaller. A pictorial representation is
given above for the Tocopilla data, where µ0 = 0.54, n = 12, and df = 11.
Note that tcrit = 2.201 and 3.106 for α = 0.05 and 0.01, respectively.
The mathematics behind the test presumes that H0 is true. Given the data,
you use
Ȳ − µ0
ts =
SEȲ
to measure how far Ȳ is from µ0, relative to the spread in the data given by
SEȲ . For ts to be in the rejection region, Ȳ must be significantly above or
below µ0, relative to the spread in the data. To see this, note that rejection
rule can be expressed as: Reject H0 if
Ȳ < µ0 − tcritSEȲ or Ȳ > µ0 + tcritSEȲ .
The rejection rule is sensible because Ȳ is our best guess for µ. You would
reject H0 : µ = µ0 only if Ȳ is so far from µ0 that you would question the
reasonableness of assuming µ = µ0. How far Ȳ must be from µ0 before you
reject H0 depends on α (i.e., how willing you are to reject H0 if it is true), and
on the value of SEȲ . For a given sample, reducing α forces Ȳ to be further
from µ0 before you reject H0. For a given value of α and s, increasing n allows
smaller differences between Ȳ and µ0 to be statistically significant (i.e.,
lead to rejecting H0). In problems where small differences between Ȳ and µ0
lead to rejecting H0, you need to consider whether the observed differences are
important.
In essence, the t-distribution provides an objective way to calibrate whether
the observed Ȳ is typical of what sample means look like when sampling from
a normal population where H0 is true. If all other assumptions are satisfied,
and Ȳ is inordinately far from µ0, then our only recourse is to conclude that
H0 must be incorrect.
2.4 Two-sided tests, CI and p-values

An important relationship among two-sided tests of H0 : µ = µ0, CI, and
p-values is that
size α test rejects H0 ⇔ 100(1 − α)% CI does not contain

µ0 ⇔ p-value ≤ α, and
size α test fails to reject H0 ⇔ 100(1 − α)% CI contains µ0 ⇔ p-value > α.
For example, an α = 0.05 test rejects H0 ⇔ 95% CI does not contain
µ0 ⇔ p-value ≤ 0.05. The picture below illustrates the connection between
p-values and rejection regions.
2.5: Statistical versus practical significance 89
0
− tcrit tcrit
If ts is here then p−value > α
If ts is here then p−value < α
Either a CI or a test can be used to decide the plausibility of the claim

that µ = µ0. Typically, you use the test to answer the question is there
a difference? If so, you use the CI to assess how much of a difference
exists. I believe that scientists place too much emphasis on hypothesis testing.
2.5 Statistical versus practical significance

Suppose in the Tocopilla meteorite example, you rejected H0 : µ = 0.54 at
the 5% level and found a 95% two-sided CI for µ to be 0.55 to 0.58. Although
you have sufficient evidence to conclude that the population mean cooling rate
µ differs from that suggested by evolutionary theory, the range of plausible
values for µ is small and contains only values close to 0.54. Although you have
shown statistical significance here, you need to ask yourself whether the actual
difference between µ and 0.54 is large enough to be important. The answer to
such questions is always problem specific.
2.6 Design issues and power

An experiment may not be sensitive enough to pick up true differences. For
example, in the Tocopilla meteorite example, suppose the true mean cooling
rate is µ = 1.00. To have a 50% chance of correctly rejecting H0 : µ = 0.54,
you would need about n = 48 observations. If the true mean is µ = 0.75, you
would need about 221 observations to have a 50% chance of correctly rejecting
H0. In general, the smaller the difference between the true and hypothesized
mean (relative to the spread in the population), the more data that is needed
to reject H0. If you have prior information on the expected difference between
the true and hypothesized mean, you can design an experiment appropriately
by choosing the sample size required to likely reject H0.
The power of a test is the probability of correctly rejecting H0 when it is
false. Equivalently,
power = 1 − Prob( failing to reject H0 when it is false ) = 1 − Prob( Type-II
error ).
For a given sample size, the tests I have discussed have maximum power (or
smallest probability of a Type-II error) among all tests with fixed size α. How-
ever, the actual power may be small, so sample size calculations, as briefly
highlighted above, are important prior to collecting data. See your local statis-
tician.
#### Power calculations (that you can do on your own)
# install.packages("pwr")
library(pwr)
?power.t.test
power.t.test(n = NULL, delta = 1.00 - 0.54, sd = sd(toco),
, sig.level = 0.05, power = 0.50
, type = "one.sample", alternative = "two.sided")
power.t.test(n = NULL, delta = 0.75 - 0.54, sd = sd(toco),
, sig.level = 0.05, power = 0.50
, type = "one.sample", alternative = "two.sided")
For more examples, try:
# install.packages("TeachingDemos")
# Demonstrate concepts of Power.
?power.examp
2.7: One-sided tests on µ 91
2.7 One-sided tests on µ

There are many studies where a one-sided test is appropriate. The two common
scenarios are the lower one-sided test H0 : µ = µ0 (or µ ≥ µ0) versus
HA : µ < µ0 and the upper one-sided test H0 : µ = µ0 (or µ ≤ µ0) versus
HA : µ > µ0. Regardless of the alternative hypothesis, the tests are based on
the t-statistic:
Ȳ − µ0
ts = .
SEȲ
For the upper one-sided test
1. Compute the critical value tcrit such that the area under the t-curve to
the right of tcrit is the desired size α, that is tcrit = tα .
2. Reject H0 if and only if ts ≥ tcrit.
3. The p-value for the test is the area under the t-curve to the right of the
test statistic ts.
The upper one-sided test uses the upper tail of the t-distribution for
a rejection region. The p-value calculation reflects the form of the rejection
region. You will reject H0 only for large positive values of ts which require Ȳ
to be significantly greater than µ0. Does this make sense?
For the lower one-sided test

1. Compute the critical value tcrit such that the area under the t-curve to
the right of tcrit is the desired size α, that is tcrit = tα .
2. Reject H0 if and only if ts ≤ −tcrit.
3. The p-value for the test is the area under the t-curve to the left of the
test statistic ts.
The lower one-sided test uses the lower tail of the t-distribution for
a rejection region. The calculation of the rejection region in terms of −tcrit is
awkward but is necessary for hand calculations because many statistical tables
only give upper tail percentiles. Note that here you will reject H0 only for large
negative values of ts which require Ȳ to be significantly less than µ0.
As with two-sided tests, the p-value can be used to decide between rejecting
or not rejecting H0 for a test with a given size α. A picture of the rejection
region and the p-value evaluation for one-sided tests is given below.
Upper One−Sided Rejection Region Upper One−Sided p−value
α p−value
0 0
tcrit ts
Lower One−Sided Rejection Region Lower One−Sided p−value
α p−value
0 0
− tcrit ts
ClickerQ s — One-sided tests on µ
Example: Weights of canned tomatoes A consumer group suspects

that the average weight of canned tomatoes being produced by a large cannery
is less than the advertised weight of 20 ounces. To check their conjecture, the
group purchases 14 cans of the canner’s tomatoes from various grocery stores.
The weights of the contents of the cans to the nearest half ounce were as follows:
20.5, 18.5, 20.0, 19.5, 19.5, 21.0, 17.5, 22.5, 20.0, 19.5, 18.5, 20.0, 18.0, 20.5. Do
the data confirm the group’s suspicions? Test at the 5% level.
Let µ = the population mean weight for advertised 20 ounce cans of toma-
toes produced by the cannery. The company claims that µ = 20, but the
consumer group believes that µ < 20. Hence the consumer group wishes to
test H0 : µ = 20 (or µ ≥ 20) against HA : µ < 20. The consumer group will
reject H0 only if the data overwhelmingly suggest that H0 is false.
You should assess the normality assumption prior to performing the t-test.
The stem-and-leaf display and the boxplot suggest that the distribution might
be slightly skewed to the left. However, the skewness is not severe and no
outliers are present, so the normality assumption is not unreasonable.
R output for the problem is given below. Let us do a hand calculation using
the summarized data. The sample size, mean, and standard deviation are 14,
√
19.679, and 1.295, respectively. The standard error is SEȲ = s/ n = 0.346.
We see that the sample mean is less than 20. But is it sufficiently less than 20
for us to be willing to publicly refute the canner’s claim? Let us carry out the
test, first using the rejection region approach, and then by evaluating a p-value.
The test statistic is
Ȳ − µ0 19.679 − 20
ts = = = −0.93.
SEȲ 0.346
The critical value for a 5% one-sided test is t0.05 = 1.771, so we reject H0

if ts < −1.771 (you can get that value from r or from the table). The test
statistic is not in the rejection region. Using the t-table, the p-value is between
0.15 and 0.20. I will draw a picture to illustrate the critical region and p-value
calculation. The exact p-value from R is 0.185, which exceeds 0.05.
Both approaches lead to the conclusion that we do not have sufficient evi-
dence to reject H0. That is, we do not have sufficient evidence to question the
accuracy of the canner’s claim. If you did reject H0, is there something about
how the data were recorded that might make you uncomfortable about your
conclusions?
#### Example: Weights of canned tomatoes
tomato <- c(20.5, 18.5, 20.0, 19.5, 19.5, 21.0, 17.5
, 22.5, 20.0, 19.5, 18.5, 20.0, 18.0, 20.5)
par(mfrow=c(2,1))
hist(tomato, freq = FALSE, breaks = 6)
points(density(tomato), type = "l")
rug(tomato)
# violin plot
library(vioplot)
vioplot(tomato, horizontal=TRUE, col="gray")
# t.crit
qt(1 - 0.05/2, df = length(tomato) - 1)
## [1] 2.16
t.summary <- t.test(tomato, mu = 20, alternative = "less")
t.summary
##
##
## data: tomato
## t = -0.9287, df = 13, p-value = 0.185
## alternative hypothesis: true mean is less than 20
## -Inf 20.29
## mean of x
## 19.68
summary(tomato)
## 17.5 18.8 19.8 19.7 20.4 22.5
Histogram of tomato
0.4
Density
0.2
0.0
17 18 19 20 21 22 23
tomato
●
1
18 19 20 21 22

able. Therefore, the results for the t-test above can be trusted.
bs.one.samp.dist(tomato)

0.0 0.1 0.2 0.3 0.4
0.4
Density
0.3
t−dist( df = 13 )
17 18 19 20 21 22 23
0.2
dat
1.2
0.1
0.8
Density
0.4
0.0
0.0
−4 −2 0 2 4
18.5 19.0 19.5 20.0 20.5 21.0
t−stat = −0.92866 , Shaded area is p−value = 0.18499
Data: n = 14 , mean = 19.679 , se = 0.346121 5
2.7.1 One-sided CIs

How should you couple a one-sided test with a CI procedure? For a lower one-
sided test, you are interested only in an upper bound on µ. Similarly,
with an upper one-sided test you are interested in a lower bound on
µ. Computing these type of bounds maintains the consistency between tests
and CI procedures. The general formulas for lower and upper 100(1 − α)%
confidence bounds on µ are given by
Ȳ − tcritSEȲ and Ȳ + tcritSEȲ
respectively, where tcrit = tα .
In the cannery problem, to get an upper 95% bound on µ, the critical value
is the same as we used for the one-sided 5% test: t0.05 = 1.771. The upper
bound on µ is
Ȳ + t0.05SEȲ = 19.679 + 1.771 × 0.346 = 19.679 + 0.613 = 20.292.
Thus, you are 95% confident that the population mean weight of the canner’s
20oz cans of tomatoes is less than or equal to 20.29. As expected, this interval
covers 20.
If you are doing a one-sided test in R, it will generate the correct one-sided
bound. That is, a lower one-sided test will generate an upper bound, whereas
an upper one-sided test generates a lower bound. If you only wish to compute
a one-sided bound without doing a test, you need to specify the direction of
the alternative which gives the type of bound you need. An upper bound was
generated by R as part of the test we performed earlier. The result agrees with
the hand calculation.
Quite a few packages, do not directly compute one-sided bounds so you have
to fudge a bit. In the cannery problem, to get an upper 95% bound on µ, you
take the upper limit from a 90% two-sided confidence limit on µ. The rational
for this is that with the 90% two-sided CI, µ will fall above the upper limit
5% of the time and fall below the lower limit 5% of the time. Thus, you are
95% confident that µ falls below the upper limit of this interval, which gives us
our one-sided bound. Here, you are 95% confident that the population mean
weight of the canner’s 20 oz cans of tomatoes is less than or equal to 20.29,
which agrees with our hand calculation.
One-Sample T: Cans
Variable N Mean StDev SE Mean 90% CI
Cans 14 19.6786 1.2951 0.3461 (19.0656, 20.2915)
The same logic applies if you want to generalize the one-sided confidence
bounds to arbitrary confidence levels and to lower one-sided bounds — always
double the error rate of the desired one-sided bound to get the error rate of the
required two-sided interval! For example, if you want a lower 99% bound on
µ (with a 1% error rate), use the lower limit on the 98% two-sided CI for µ
(which has a 2% error rate).
ClickerQ s — P-value
Chapter 3
Two-Sample Inferences
Learning objectives
select graphical displays that meaningfully compare independent popula-
tions.
assess the assumptions of the two-sample t-test visually.
decide whether the means between two populations are different.
3.1 Comparing Two Sets of Measurements

Suppose you have collected data on one variable from two (independent) sam-
ples and you are interested in “comparing” the samples. What tools are good
to use?
Example: Head Breadths In this analysis, we will compare a physical

feature of modern day Englishmen with the corresponding feature of some of
3.1: Comparing Two Sets of Measurements 99
their ancient countrymen. The Celts were a vigorous race of people who once
populated parts of England. It is not entirely clear whether they simply died
out or merged with other people who were the ancestors of those who live in
England today. A goal of this study might be to shed some light on possible
genetic links between the two groups.
The study is based on the comparison of maximum head breadths (in mil-
limeters) made on unearthed Celtic skulls and on a number of skulls of modern-
day Englishmen. The data are given below. We have a sample of 18 Englishmen
and an independent sample of 16 Celtic skulls.
#### Example: Head Breadths
# unstacked data as two vectors
english <- c(141, 148, 132, 138, 154, 142, 150, 146, 155, 158,
150, 140, 147, 148, 144, 150, 149, 145)
celts <- c(133, 138, 130, 138, 134, 127, 128, 138, 136, 131,
126, 120, 124, 132, 132, 125)
english
## [1] 141 148 132 138 154 142 150 146 155 158 150 140 147 148 144 150
## [17] 149 145
celts
## [1] 133 138 130 138 134 127 128 138 136 131 126 120 124 132 132 125
What features of these data would we likely be interested in comparing?

The centers of the distributions, the spreads within each distribution, the dis-
tributional shapes, etc.
These data can be analyzed in R as either unstacked separate vectors or as
stacked data where one column contains both samples, with a second column
of labels or subscripts to distinguish the samples. It is easy to create stacked
data from unstacked data and vice-versa. Many comparisons require the plots
for the two groups to have the same scale, which is easiest to control when the
data are stacked.
# stacked data as a vector of values and a vector of labels
HeadBreadth <- c(english, celts)
Group <- c(rep("English", length(english)), rep("Celts", length(celts)))
hb <- data.frame(HeadBreadth, Group)
hb
## HeadBreadth Group
## 1 141 English
## 2 148 English
## 3 132 English
## 4 138 English
## 5 154 English
100 Ch 3: Two-Sample Inferences
## 6 142 English
## 7 150 English
## 8 146 English
## 9 155 English
## 10 158 English
## 11 150 English
## 12 140 English
## 13 147 English
## 14 148 English
## 15 144 English
## 16 150 English
## 17 149 English
## 18 145 English
## 19 133 Celts
## 20 138 Celts
## 21 130 Celts
## 22 138 Celts
## 23 134 Celts
## 24 127 Celts
## 25 128 Celts
## 26 138 Celts
## 27 136 Celts
## 28 131 Celts
## 29 126 Celts
## 30 120 Celts
## 31 124 Celts
## 32 132 Celts
## 33 132 Celts
## 34 125 Celts
3.1.1 Plotting head breadth data:

1. A dotplot with the same scale for both samples is obtained easily from
the stacked data.
#### Plotting head breadth data
# stripchart (dotplot) using R base graphics
stripchart(HeadBreadth ~ Group, method = "stack", data = hb,
main = "Head breadth comparison", xlab = "head breadth (mm)")
# stripchart (dotplot) using ggplot

library(ggplot2)
p <- ggplot(hb, aes(x = HeadBreadth))
p <- p + geom_dotplot(binwidth = 2)
p <- p + facet_grid(Group ~ .) # rows are Group categories
p <- p + labs(title = "Head breadth comparison") + xlab("head breadth (mm)")
print(p)
Head breadth comparison
English
Celts
120 130 140 150
head breadth (mm)

1.00
0.75
Celts
0.50
● ●
0.25 ●● ●● ●
● ●●● ●●●●●
count
0.00
1.00
0.75
English
0.50
●●
0.25 ● ●●●● ●
0.00 ● ● ●● ●●●● ● ●
120 130 140 150 160
head breadth (mm)
2. Boxplots for comparison are most helpful when plotted in the same axes.

boxplot(HeadBreadth ~ Group, method = "stack", data = hb,
horizontal = TRUE,
main = "Head breadth comparison", xlab = "head breadth (mm)")
p <- ggplot(hb, aes(x = Group, y = HeadBreadth))

# add a "+" at the mean
p <- p + stat_summary(fun.y = mean, geom = "point", shape = 3, size = 2)
p <- p + labs(title = "Head breadth comparison")
print(p)
English
Celts
120 130 140 150
head breadth (mm)
English
Group
Celts
120 130 140 150

HeadBreadth
3. Histograms are hard to compare unless you make the scale and actual
bins the same for both. Why is the pair on the right clearly preferable?
# histogram using R base graphics

par(mfcol=c(2,2))
hist(hb$HeadBreadth[(hb$Group == "Celts")],
main = "Head breadth, Celts", xlab = "head breadth (mm)")
hist(hb$HeadBreadth[(hb$Group == "English")],
main = "Head breadth, English", xlab = "head breadth (mm)")
# common x-axis limits based on the range of the entire data set
hist(hb$HeadBreadth[(hb$Group == "Celts")], xlim = range(hb$HeadBreadth),
main = "Head breadth, Celts", xlab = "head breadth (mm)")
hist(hb$HeadBreadth[(hb$Group == "English")], xlim = range(hb$HeadBreadth),
main = "Head breadth, English", xlab = "head breadth (mm)")
Head breadth, Celts Head breadth, Celts
5
4
4
Frequency
Frequency
3
3
2
2
1
1
0
0
120 125 130 135 140 120 130 140 150
Head breadth, English Head breadth, English

8
8
6
6
Frequency
Frequency
4
4
2
2
0
130 135 140 145 150 155 160 120 130 140 150

p <- ggplot(hb, aes(x = HeadBreadth))
p <- p + facet_grid(Group ~ .)
print(p)
6
Celts
2
count
0
6
English
2
0
120 130 140 150 160
head breadth (mm)
p <- ggplot(hb, aes(x = HeadBreadth, fill=Group))

p <- p + geom_histogram(binwidth = 4, alpha = 0.5, position="identity")
print(p)
p <- ggplot(hb, aes(x = HeadBreadth, fill=Group))

p <- p + geom_histogram(binwidth = 4, alpha = 1, position="dodge")
print(p)
Head breadth comparison Head breadth comparison
6 6
4 4
Group Group
count
count
Celts Celts
English English
2 2
0 0
120 130 140 150 160 120 130 140 150 160
4. Stem-and-leaf displays for comparisons in R can be pretty useless. The

stems are not forced to match (just like with histograms). It is pretty
hard to make quick comparisons with the following:
stem(english, scale = 2)
##
##
## 13 | 2
## 13 | 8
## 14 | 0124
## 14 | 567889
## 15 | 0004
## 15 | 58
stem(celts, scale = 2)
##
##
## 120 | 0
## 122 |
## 124 | 00
## 126 | 00
## 128 | 0
## 130 | 00
## 132 | 000
## 134 | 0
## 136 | 0
## 138 | 000
Using the by() function, you can get summaries by group.

#### summaries by group
# summary for separate vectors
summary(english)
## 132 142 148 146 150 158
summary(celts)
## 120 127 132 131 134 138
# comparing spreads, an assumption of equal variances seems reasonable
sd(english)
## [1] 6.382
sd(celts)
## [1] 5.434
IQR(english)
## [1] 7.5
IQR(celts)
## [1] 7.75
# numerical summary of each column in data.frame hb by Group
by(hb, Group, summary)
## Group: Celts
## Min. :120 Celts :16
## 1st Qu.:127 English: 0
## Median :132
## Mean :131
## 3rd Qu.:134
## Max. :138
## ----------------------------------------------------
## Group: English
## Min. :132 Celts : 0
## 1st Qu.:142 English:18
## Median :148
## Mean :146
## 3rd Qu.:150
## Max. :158
3.1.2 Salient Features to Notice

The dotplots, boxplots, and histograms indicate that the English and Celt
samples are slightly skewed to the left. There are no outliers in either sample.
It is not unreasonable to operationally assume that the population frequency
curves (i.e., the histograms for the populations from which the samples were
selected) for the English and Celtic head breadths are normal. Therefore, the
sampling distribution of the means will be reasonably normal.
The sample means and medians are close to each other in each sample,
which is not surprising given the near symmetry and the lack of outliers.
The data suggest that the typical modern English head breadth is greater
than that for Celts. The data sets have comparable spreads, as measured by
either the standard deviation or the IQR.
3.2 Two-Sample Methods: Paired Versus

Independent Samples
Suppose you have two populations of interest, say populations 1 and 2, and
you are interested in comparing their (unknown) population means, µ1 and µ2.
Inferences on the unknown population means are based on samples from each
population. In practice, most problems fall into one of two categories.
Independent samples where the sample taken from population 1 has no
effect on which observations are selected from population 2, and vice versa.
Paired or dependent samples where experimental units are paired based on
factors related or unrelated to the variable measured.
The distinction between paired and independent samples may be made clear
through a series of examples.
Example The English and Celt head breadth samples are independent.
3.2: Two-Sample Methods: Paired Versus Independent Samples 107
Example Suppose you are interested in whether the CaCO3 (calcium car-
bonate) level in the Atrisco well field, which is the water source for Albuquerque,
is changing over time. To answer this question, the CaCO3 level was recorded
at each of 15 wells at two time points. These data are paired. The two samples
are the observations at Times 1 and 2.
Example To compare state incomes, a random sample of New Mexico house-

holds was selected, and an independent sample of Arizona households was ob-
tained. It is reasonable to assume independent samples.
Example Suppose you are interested in whether the husband or wife is typ-
ically the heavier smoker among couples where both adults smoke. Data are
collected on households. You measure the average number of cigarettes smoked
by each husband and wife within the sample of households. These data are
paired, i.e., you have selected husband wife pairs as the basis for the samples.
It is reasonable to believe that the responses within a pair are related, or cor-
related.
Although the focus here will be on comparing population means, you should
recognize that in paired samples you may also be interested, as in the problems
above, in how observations compare within a pair. That is, a paired compar-
ison might be interested in the difference between the two paired samples.
These goals need not agree, depending on the questions of interest. Note that
with paired data, the sample sizes are equal, and equal to the number of pairs.
ClickerQ s — Independent or paired 1, STT.08.02.030

ClickerQ s — Independent or paired 2, STT.08.02.040
3.3 Two Independent Samples: CI and Test

Using Pooled Variance
These methods assume that the populations have normal frequency curves,
with equal population standard deviations, i.e., σ1 = σ2. Let (n1, Ȳ1, s1) and
(n2, Ȳ2, s2) be the sample sizes, means and standard deviations from the two
samples. The standard CI for µ1 − µ2 is given by
CI = (Ȳ1 − Ȳ2) ± tcritSEȲ1−Ȳ2
The t-statistic for testing H0 : µ1 −µ2 = 0 (µ1 = µ2) against HA : µ1 −µ2 6= 0
(µ1 6= µ2) is given by
Ȳ1 − Ȳ2
ts = .
SEȲ1−Ȳ2
The standard error of Ȳ1 − Ȳ2 used in both the CI and the test is given by
r
1 1
SEȲ1−Ȳ2 = spooled + .
n1 n2
Here the pooled variance estimator,
(n1 − 1)s21 + (n2 − 1)s22
s2pooled= ,
n1 + n2 − 2
is our best estimate of the common population variance. The pooled estimator
of variance is a weighted average of the two sample variances, with more weight
given to the larger sample. If n1 = n2 then s2pooled is the average of s21 and s22.
The critical value tcrit for CI and tests is obtained in usual way from a t-table
with df = n1 + n2 − 2. For the test, follow the one-sample procedure, with the
new ts and tcrit.
The pooled CI and tests are sensitive to the normality and equal standard
deviation assumptions. The observed data can be used to assess the reason-
ableness of these assumptions. You should look at boxplots and histograms to
.
assess normality, and should check whether s1 = s2 to assess the assumption
σ1 = σ2. Formal tests of these assumptions will be discussed later.
3.4: Satterthwaite’s Method, unequal variances 109
3.4 Satterthwaite’s Method, unequal vari-
ances
Satterthwaite’s method assumes normality, but does not require equal
population standard deviations. Satterthwaite’s procedures are somewhat con-
servative, and adjust the SE and df to account for unequal population vari-
ances. Satterthwaite’s method uses the same CI and test statistic formula, with
a modified standard error:
s
s21 s22
SEȲ1−Ȳ2 = + ,
n1 n2
and degrees of freedom:

2
s21 s22

n1 + n2
df = .
s41 s42
n21 (n1 −1)
+ n22 (n2 −1)
Note that df = n1 + n2 − 2 when n1 = n2 and s1 = s2. The Satterthwaite and

.
pooled variance procedures usually give similar results when s1 = s2.
The df formula for Satterthwaite’s method is fairly complex, so when done
by hand some use a conservative df formula that uses the minimum of n1 − 1
and n2 − 1 instead.
3.4.1 R Implementation
R does the pooled and Satterthwaite (Welch) analyses, either on stacked or
unstacked data. The output will contain a p-value for a two-sided test of equal
population means and a CI for the difference in population means. If you
include var.equal = TRUE you will get the pooled method, otherwise the output
is for Satterthwaite’s method.
Example: Head Breadths The English and Celts are independent sam-
ples. We looked at boxplots and histograms, which suggested that the normality
assumption for the t-test is reasonable. The R output shows the English and
Celt sample standard deviations and IQRs are fairly close, so the pooled and
Satterthwaite results should be comparable. The pooled analysis is preferable
here, but either is appropriate.
We are interested in difference in mean head breadths between Celts and
English.
1. Define the population parameters and hypotheses in words
and notation
Let µ1 and µ2 be the mean head breadth for the Celts and English, respec-
tively.
In words: “The difference in population means between Celts and English is
different from zero mm.”
In notation: H0 : µ1 = µ2 versus HA : µ1 6= µ2.
Alternatively: H0 : µ1 − µ2 = 0 versus HA : µ1 − µ2 6= 0.
Mean, standard deviation, sample size:
#### Calculate summary statistics
m1 <- mean(celts)
s1 <- sd(celts)
n1 <- length(celts)
m2 <- mean(english)
s2 <- sd(english)
n2 <- length(english)
c(m1, s1, n1)
## [1] 130.750 5.434 16.000
c(m2, s2, n2)
## [1] 146.500 6.382 18.000
The pooled-standard devation, standard error, and degrees-of-freedom are:
sdpool <- sqrt(((n1 - 1) * s1^2 + (n2 - 1) * s2^2) / (n1 + n2 - 2))
sdpool
## [1] 5.957
SEpool <- sdpool * sqrt(1 / n1 + 1 / n2)
SEpool
## [1] 2.047
dfpool <- n1 + n2 - 2
dfpool
## [1] 32
t_pool <- (m1 - m2) / SEpool
t_pool
## [1] -7.695
The Satterthwaite SE and degrees-of-freedom are:
SE_Sat <- sqrt(s1^2 / n1 + s2^2 / n2)
SE_Sat
## [1] 2.027
df_Sat <- (SE_Sat^2)^2 / (s1^4 / (n1^2 * (n1 - 1)) + s2^4 / (n2^2 * (n2 - 1)))
df_Sat
## [1] 31.95
t_Sat <- (m1 - m2) / SE_Sat
t_Sat
## [1] -7.77
3. Specify confidence level, calculate t-stat, CI limits, p-value

Let us calculate a 95% CI for µ1 − µ2.
Assuming equal variances, using pooled-variance procedure:
## Equal variances
# var.equal = FALSE is the default
# two-sample t-test specifying two separate vectors

t.summary.eqvar <- t.test(celts, english, var.equal = TRUE)
t.summary.eqvar
##
## Two Sample t-test
##
## data: celts and english
## t = -7.695, df = 32, p-value = 9.003e-09
## alternative hypothesis: true difference in means is not equal to 0
## -19.92 -11.58
## mean of x mean of y
## 130.8 146.5
Not assuming equal variances, Satterthwaite (Welch):
# two-sample t-test with unequal variances (Welch = Satterthwaite)
# specified using data.frame and a formula, HeadBreadth by Group
t.summary.uneqvar <- t.test(HeadBreadth ~ Group, data = hb, var.equal = FALSE)
t.summary.uneqvar
##
## Welch Two Sample t-test
##
## data: HeadBreadth by Group
## t = -7.77, df = 31.95, p-value = 7.414e-09
## -19.88 -11.62
## mean in group Celts mean in group English
## 130.8 146.5
The form of the output will tell you which sample corresponds to population
1 and which corresponds to population 2.
4. Summarize in words (Using the pooled-variance results.)
The pooled analysis strongly suggests that H0 : µ1 − µ2 = 0 is false, given
the large t-statistic of −7.7 and two-sided p-value of 9 × 10−9. Because
the p-value < 0.05 we reject the Null hypothesis in favor of the Alternative
hypothesis concluding that the difference in population mean head breadths
between the Celts and English are different.
We are 95% confident that the difference in population means, µ1 − µ2, is
between −19.9 and −11.6 mm. That is, we are 95% confident that the
population mean head breadth for Englishmen (µ2) exceeds the population
mean head breadth for Celts (µ1) by between 11.6 and 19.9 mm.
The CI interpretation is made easier by recognizing that we concluded the
population means are different, so the direction of difference must be con-
sistent with that seen in the observed data, where the sample mean head
breadth for Englishmen exceeds that for the Celts. Thus, the limits on the
CI for µ1 − µ2 tells us how much smaller the mean is for the Celts (that is,
between −19.9 and −11.6 mm).
5. Check assumptions
The assumption of equal population variances will be left to a later chapter.
We can test the assumption that the distribution of Ȳ1 − Ȳ2 is normal using
the bootstrap in the following function.
#### Visual comparison of whether sampling distribution is close to Normal via Bootstrap
# a function to compare the bootstrap sampling distribution
# of the difference of means from two samples with
# a normal distribution with mean and SEM estimated from the data
bs.two.samp.diff.dist <- function(dat1, dat2, N = 1e4) {
n1 <- length(dat1);
n2 <- length(dat2);
sam1 <- matrix(sample(dat1, size = N * n1, replace = TRUE), ncol=N);
sam2 <- matrix(sample(dat2, size = N * n2, replace = TRUE), ncol=N);
# calculate the means and take difference between populations
sam1.mean <- colMeans(sam1);
sam2.mean <- colMeans(sam2);
diff.mean <- sam1.mean - sam2.mean;
hist(dat1, freq = FALSE, breaks = 6
, main = paste("Sample 1", "\n"
, "n =", n1
, ", mean =", signif(mean(dat1), digits = 5)
, ", sd =", signif(sd(dat1), digits = 5))
, xlim = range(c(dat1, dat2)))
points(density(dat1), type = "l")
rug(dat1)
hist(dat2, freq = FALSE, breaks = 6

, main = paste("Sample 2", "\n"
, "n =", n2
, ", mean =", signif(mean(dat2), digits = 5)
, ", sd =", signif(sd(dat2), digits = 5))
, xlim = range(c(dat1, dat2)))
rug(dat2)
hist(diff.mean, freq = FALSE, breaks = 25

, main = paste("Bootstrap sampling distribution of the difference in means", "\n"
, "mean =", signif(mean(diff.mean), digits = 5)
, ", se =", signif(sd(diff.mean), digits = 5)))
points(density(diff.mean), type = "l")
x <- seq(min(diff.mean), max(diff.mean), length = 1000)
points(x, dnorm(x, mean = mean(diff.mean), sd = sd(diff.mean))
, type = "l", lwd = 2, col = "red")
rug(diff.mean)
par(old.par)
}
The distribution of difference in means in the third plot looks very close to
normal.
bs.two.samp.diff.dist(celts, english)
Sample 1
n = 16 , mean = 130.75 , sd = 5.4345
0.06
0.04
Density
0.02
0.00
120 130 140 150
Sample
dat1 2
n = 18 , mean = 146.5 , sd = 6.3824
0.08
0.06
Density
0.04
0.02
0.00
120 130 140 150
Bootstrap sampling distribution

dat2 of the difference in means
mean = −15.749 , se = 1.9777
0.20
0.15
Density
0.10
0.05
0.00
−20 −15 −10
diff.mean
ClickerQ s — t-interval, STT.08.02.010
Example: Androstenedione levels in diabetics The data consist of

independent samples of diabetic men and women. For each individual, the
scientist recorded their androstenedione level (ng/dL) (a hormone, and Mark
McGwire’s favorite dietary supplement). Let µ1 = mean androstenedione level
for the population of diabetic men, and µ2 = mean androstenedione level for the
population of diabetic women. We are interested in comparing the population
means given the observed data.
The raw data and R output are given below. The boxplots suggest that the
distributions are reasonably symmetric. However, the normality assumption
for the women is unreasonable due to the presence of outliers. The equal pop-
ulation standard deviation assumption also appears unreasonable. The sample
standard deviation for men is noticeably larger than the women’s standard
deviation, even with outliers in the women’s sample.
#### Example: Androstenedione levels in diabetics
# Data and numerical summaries
men <- c(217, 123, 80, 140, 115, 135, 59, 126, 70, 63,
147, 122, 108, 70)
women <- c( 84, 87, 77, 84, 73, 66, 70, 35, 77, 73,
56, 112, 56, 84, 80, 101, 66, 84)
level <- c(men, women)
sex <- c(rep("men", length(men)), rep("women", length(women)))
andro <- data.frame(level, sex)
andro
## level sex
## 1 217 men
## 2 123 men
## 3 80 men
## 4 140 men
## 5 115 men
## 6 135 men
## 7 59 men
## 8 126 men
## 9 70 men
## 10 63 men
## 11 147 men
## 12 122 men
## 13 108 men
## 14 70 men
## 15 84 women
## 16 87 women
## 17 77 women
## 18 84 women
## 19 73 women
## 20 66 women
## 21 70 women
## 22 35 women
## 23 77 women
## 24 73 women
## 25 56 women
## 26 112 women
## 27 56 women
## 28 84 women
## 29 80 women
## 30 101 women
## 31 66 women
## 32 84 women
# numerical summaries
by(andro, sex, summary)
## sex: men
## level sex
## Min. : 59.0 men :14
## 1st Qu.: 72.5 women: 0
## Median :118.5
## Mean :112.5
## 3rd Qu.:132.8
## Max. :217.0
## ----------------------------------------------------
## sex: women
## level sex
## Min. : 35.0 men : 0
## 1st Qu.: 67.0 women:18
## Median : 77.0
## Mean : 75.8
## 3rd Qu.: 84.0
## Max. :112.0
c(sd(men), sd(women), IQR(men), IQR(women), length(men), length(women))
## [1] 42.75 17.24 60.25 17.00 14.00 18.00
p <- ggplot(andro, aes(x = sex, y = level, fill=sex))
#p <- p + coord_flip()
p <- p + labs(title = "Androstenedione Levels in Diabetics")
print(p)
p <- ggplot(andro, aes(x = level, fill=sex))

p <- p + geom_histogram(binwidth = 20, alpha = 0.5, position="identity")
p <- p + labs(title = "Androstenedione Levels in Diabetics")
print(p)
Androstenedione Levels in Diabetics Androstenedione Levels in Diabetics
200
6
150
sex sex
4
count
level
men men
● women women
100
2
50
● 0
men women 0 50 100 150 200 250

sex level
Because of the large difference in variances, I will be more comfortable

with the Satterthwaite analysis here than the pooled variance analysis. The
normality assumption of the difference in means appears to be met using the
bootstrap assessment. The distribution of difference in means in the third plot
looks very close to normal.
bs.two.samp.diff.dist(men, women)
Sample 1
n = 14 , mean = 112.5 , sd = 42.755
0.015
0.010
Density
0.005
0.000
50 100 150 200
Sample
dat1 2
n = 18 , mean = 75.833 , sd = 17.236
0.020
Density
0.010
0.000
50 100 150 200
Bootstrap sampling distribution

dat2 of the difference in means
mean = 36.509 , se = 11.669
0.030
0.020
Density
0.010
0.000
0 20 40 60 80
diff.mean
1. Define the population parameters and hypotheses in words

and notation
Let µ1 and µ2 be the mean androstenedione level for diabetic men and women,
respectively.
In words: “The difference in population mean androstenedione levels between
diabetic men and women is different from zero.”
In notation: H0 : µ1 − µ2 = 0 versus HA : µ1 − µ2 6= 0.
(see above)
3. Specify confidence level, calculate t-stat, CI limits, p-value
Not assuming equal variances, Satterthwaite (Welch):
# two-sample t-test with unequal variances (Welch = Satterthwaite)
# specified using data.frame and a formula, level by sex
t.summary <- t.test(level ~ sex, data = andro, var.equal = FALSE)
t.summary
##
##
## data: level by sex
## t = 3.023, df = 16.3, p-value = 0.007946
## 11.00 62.34
## mean in group men mean in group women
## 112.50 75.83
# plot t-distribution with shaded p-value

0.4
0.3
t−dist( df = 16.3 )
0.2
0.1
0.0
−4 −2 0 2 4

4. Summarize in words The unequal-variance analysis suggests that
H0 : µ1 − µ2 = 0 is false, given the large t-statistic of 3.02 and two-sided
p-value of 0.008. Because the p-value < 0.05 we reject the Null hypothesis
in favor of the Alternative hypothesis concluding that the difference in pop-
ulation mean androstenedione levels between diabetic men and women are
different.
We are 95% confident confident that the difference in population means,
µ1 − µ2, is between 11 and 62.3 ng/dL.
5. Check assumptions
As checked before, while the assumption of equal population variances is not
met, the assumption that the distribution of Ȳ1 − Ȳ2 is normal using the
bootstrap appeared reasonable.
As a comparison, let us examine the output for the pooled procedure (which
is inappropriate since variances are unequal). The p-value for the pooled t-test
is 0.002, whereas the 95% confidence limits are 14.1 and 59.2. That is, we are
95% confident that the population mean andro level for men exceeds that for
women by at least 14.1 but by no more than 59.2. These results are qualitatively
similar to the Satterthwaite conclusions.
3.5 One-Sided Tests
One-sided tests for two-sample problems are where the null hypothesis is H0 :
µ1 − µ2 = 0 but the alternative is directional, either HA : µ1 − µ2 < 0 (i.e.,
µ1 < µ2) or HA : µ1 − µ2 > 0 (i.e., µ1 > µ2). Once you understand the
general form of rejection regions and p-values for one-sample tests, the one-
sided two-sample tests do not pose any new problems. Use the t-statistic, with
the appropriate tail of the t-distribution to define critical values and p-values.
One-sided two-sample tests are directly implemented in R, by specifying the
type of test with alternative = "less" or alternative = "greater". One-sided
confidence bounds are given with the one-sided tests.
3.6: Paired Analysis 119
3.6 Paired Analysis
With paired data, inferences on µ1 − µ2 are based on the sample of differences
within pairs. By taking differences within pairs, two dependent samples are
transformed into one sample, which contains the relevant information for infer-
ences on µd = µ1 − µ2. To see this, suppose the observations within a pair are
Y1 and Y2. Then within each pair, compute the difference d = Y1 − Y2:
d1 = Y11 − Y21
d2 = Y12 − Y22
...
dn = Y1n − Y2n
If the Y1 data are from a population with mean µ1 and the Y2 data are from a
population with mean µ2, then the ds are a sample from a population with mean
µd = µ1 − µ2. Furthermore, if the sample of differences comes from a normal
population, then we can use standard one-sample techniques on d1, . . . , dn to
test µd = 0 (that is, µ1 = µ2), and to get a CI for µd = µ1 − µ2.
Let d¯ = n−1 i di = Ȳ1 − Ȳ2 be the sample mean of the differences (which
P
is also the mean difference), and let sd be the sample standard deviation of the
√
differences. The standard error of d¯ is SEd¯ = sd/ n, where n is the number
of pairs. The paired t-test (two-sided) CI for µd is given by d¯ ± tcritSEd¯. To
test H0 : µd = 0 (µ1 = µ2) against HA : µd 6= 0 (µ1 6= µ2), use
d¯ − 0
ts =
SEd¯
to compute a p-value as in a two-sided one-sample test. One-sided tests are

evaluated in the usual way for one-sample tests on means.
A graphical analysis of paired data focuses on the sample of differences,
and not on the original samples. In particular, the normality assumption is
assessed on the sample of differences.
3.6.1 R Analysis
The most natural way to enter paired data is as two columns, one for each
treatment group. You can then create a new column of differences, and do the
usual one-sample graphical and inferential analysis on this column of differences,
or you can do the paired analysis directly without this intermediate step.
Example: Paired Analysis of Data on Twins Burt (1966) presented

data on IQ scores for identical twins that were raised apart, one by foster parents
and one by the genetic parents. Assuming the data are a random sample of twin
pairs, consider comparing the population mean IQs for twins raised at home to
those raised by foster parents. Let µf=population mean IQ for twin raised by
foster parents, and µg=population mean IQ for twin raised by genetic parents.
I created the data in the worksheet (c1=foster; c2=genetic), and computed
the differences between the IQ scores for the children raised by the genetic and
foster parents (c3=diff=genetic-foster). I also made a scatter plot of the genetic
versus foster IQ scores.
#### Example: Paired Analysis of Data on Twins
foster <- c(82, 80, 88, 108, 116, 117, 132, 71, 75, 93,
95, 88, 111, 63, 77, 86, 83, 93, 97, 87,
94, 96, 112, 113, 106, 107, 98)
genetic <- c(82, 90, 91, 115, 115, 129, 131, 78, 79, 82,
97, 100, 107, 68, 73, 81, 85, 87, 87, 93,
94, 95, 97, 97, 103, 106, 111)
diff <- genetic - foster;
iq <- data.frame(foster, genetic, diff)
# scatterplot of foster and genetic IQs, with 1:1 line

p <- ggplot(iq, aes(x = foster, y = genetic))
# draw a 1:1 line, dots above line indicate "genetic > foster"
p <- p + geom_abline(intercept=0, slope=1, alpha=0.2)
# make the axes square so it's a fair visual comparison
p <- p + coord_equal()
p <- p + labs(title = "IQ of identical twins raised by genetic vs foster parents")
print(p)
IQ of identical twins raised by genetic vs foster parents

●
●
120
● ●
●
●
●
genetic
100 ●
● ●●
●
●
●
●
●
● ●
●
● ●
●
80 ●
●
60 80 100 120
foster
This plot of IQ scores shows that scores are related within pairs of twins.
This is consistent with the need for a paired analysis.
Given the sample of differences, I created a boxplot and a stem and leaf
display, neither which showed marked deviation from normality. The boxplot
is centered at zero, so one would not be too surprised if the test result is
insignificant.
p1 <- ggplot(iq, aes(x = diff))
p1 <- p1 + scale_x_continuous(limits=c(-20,+20))
# vertical line at 0
p1 <- p1 + geom_vline(xintercept=0, colour="#BB0000", linetype="dashed")
, binwidth=5
, alpha = 1/5)
# violin plot
p2 <- ggplot(iq, aes(x = "diff", y = diff))
p2 <- p2 + scale_y_continuous(limits=c(-20,+20))
p2 <- p2 + geom_hline(yintercept=0, colour="#BB0000", linetype="dashed")
p2 <- p2 + geom_violin(fill = "gray50", alpha=1/2)
# boxplot
p3 <- ggplot(iq, aes(x = "diff", y = diff))
library(gridExtra)
0.06
0.04
density
0.02
0.00
−20 −10 0 10 20
diff
"diff"
diff
−20 −10 0 10 20
diff
"diff"
diff
−20 −10 0 10 20
diff
The normality assumption of the sample mean for a one-sample test is

satisfied (below, left).
Given the sample of differences, I generated a one-sample CI and test. The
hypothesis under test is µd = µg − µf = 0. The p-value for this test is large. We
do not have sufficient evidence to claim that the population mean IQs for twins
raised apart are different. This is consistent with the CI for µd given below,
which covers zero.
bs.one.samp.dist(iq$diff)
# one-sample t-test of differences (paired t-test)

t.summary <- t.test(iq$diff)
t.summary
##
##
## data: iq$diff
## t = 0.1244, df = 26, p-value = 0.902
## -2.875 3.246
## mean of x
## 0.1852

0.06
0.4
0.04
Density
0.02
0.3
0.00
t−dist( df = 26 )
−20 −15 −10 −5 0 5 10 15

0.2
dat
0.1
0.20
Density
0.10
0.0
0.00
−4 −2 0 2 4
−6 −4 −2 0 2 4 6
Data: n = 27 , mean = 0.18519 , se = 1.48884 5
Alternatively, I can generate the test and CI directly from the raw data in
two columns, specifying paired=TRUE. This gives the following output, which
leads to identical conclusions to the earlier analysis.
# two-sample paired t-test
t.summary <- t.test(iq$genetic, iq$foster, paired=TRUE)
t.summary
##
## Paired t-test
##
## data: iq$genetic and iq$foster
## t = 0.1244, df = 26, p-value = 0.902
## -2.875 3.246
## mean of the differences
## 0.1852
You might ask why I tortured you by doing the first analysis, which re-
quired creating and analyzing the sample of differences, when the alternative
and equivalent second analysis is so much easier. (A later topic deals with
non-parametric analyses of paired data for which the differences must be first
computed.)
Remark: I could have defined the difference to be the foster IQ score

minus the genetic IQ score. How would this change the conclusions?
Example: Paired Comparisons of Two Sleep Remedies The fol-

lowing data give the amount of sleep gained in hours from two sleep remedies, A
and B, applied to 10 individuals who have trouble sleeping an adequate amount.
Negative values imply sleep loss. In 9 of the 10 individuals, the sleep gain on B
exceeded that on A.
Let µA = population mean sleep gain (among troubled sleepers) on remedy
A, and µB = population mean sleep gain (among troubled sleepers) on remedy
B. Consider testing H0 : µB − µA = 0 or equivalently µd = 0, where µd =
µB − µA .
The observed distribution of differences between B and A is slightly skewed
to the right, with a single outlier in the upper tail. The normality assumption
of the standard one-sample t-test and CI are suspect here. I will continue with
the analysis, anyways.
#### Example: Paired Comparisons of Two Sleep Remedies
a <- c( 0.7, -1.6, -0.2, -1.2, 0.1, 3.4, 3.7, 0.8, 0.0, 2.0)
b <- c( 1.9, 0.8, 1.1, 0.1, -0.1, 4.4, 5.5, 1.6, 4.6, 3.0)
d <- b - a;
sleep <- data.frame(a, b, d)
sleep
## a b d
## 1 0.7 1.9 1.2
## 2 -1.6 0.8 2.4
## 3 -0.2 1.1 1.3
## 4 -1.2 0.1 1.3
## 5 0.1 -0.1 -0.2
## 6 3.4 4.4 1.0
## 7 3.7 5.5 1.8
## 8 0.8 1.6 0.8
## 9 0.0 4.6 4.6
## 10 2.0 3.0 1.0
# scatterplot of a and b IQs, with 1:1 line
p <- ggplot(sleep, aes(x = a, y = b))
# draw a 1:1 line, dots above line indicate "b > a"
p <- p + geom_abline(intercept=0, slope=1, alpha=0.2)
# make the axes square so it's a fair visual comparison
p <- p + coord_equal()
p <- p + labs(title = "Sleep hours gained on two sleep remedies: a vs b")
print(p)
Sleep hours gained on two sleep remedies: a vs b

●
●
b
2 ●
●
0 ●
−1 0 1 2 3
a
There is evidence here against the normality assumption of the sample mean.
We’ll continue anyway (in practice we’d use a nonparametric method, instead,
in a later chapter).
p1 <- ggplot(sleep, aes(x = d))
p1 <- p1 + scale_x_continuous(limits=c(-5,+5))
# vertical line at 0
p1 <- p1 + geom_vline(xintercept=0, colour="#BB0000", linetype="dashed")
, binwidth=1
, alpha = 1/5)
p1 <- p1 + labs(title = "Difference of sleep hours gained: d = b - a")
# violin plot
p2 <- ggplot(sleep, aes(x = "d", y = d))
p2 <- p2 + geom_violin(fill = "gray50", alpha=1/2)

# boxplot
library(gridExtra)
bs.one.samp.dist(sleep$d)
Difference of sleep hours gained: d = b − a Plot of data with smoothed density curve
0.6 0.4
density
0.4
0.3
0.2
Density
0.0
0.2
−5.0 −2.5 0.0 2.5 5.0

d
0.1
0.0
−1 0 1 2 3 4 5
"d"
d ● ●
dat
−5.0 −2.5 0.0 2.5 5.0

0.8
d
Density
0.4
"d"
d ● ●
0.0
0.5 1.0 1.5 2.0 2.5 3.0

−5.0 −2.5 0.0 2.5 5.0
d
Data: n = 10 , mean = 1.52 , se = 0.402161 5
The p-value for testing H0 is 0.004. We’d reject H0 at the 5% or 1% level,

and conclude that the population mean sleep gains on the remedies are different.
We are 95% confident that µB exceeds µA by between 0.61 and 2.43 hours.
Again, these results must be reported with caution, because the normality
assumption is unreasonable. However, the presence of outliers tends to make
the t-test and CI conservative, so we’d expect to find similar conclusions if we
used the nonparametric methods discussed later in the semester.
# one-sample t-test of differences (paired t-test)
t.summary <- t.test(sleep$d)
t.summary
##
##
## data: sleep$d
## t = 3.78, df = 9, p-value = 0.004352
## 0.6102 2.4298
## mean of x
## 1.52
0.4
0.3
t−dist( df = 9 )
0.2
0.1
0.0
−4 −2 0 2 4
Question: In what order should the remedies be given to the patients?
ClickerQ s — Reporting results, STT.06.03.010
3.7 Should You Compare Means?

The mean is the most common feature on which two distributions are com-
pared. You should not, however, blindly apply the two-sample tests (paired or
unpaired) without asking yourself whether the means are the relevant feature to
compare. This issue is not a big concern when, as highlighted in the first graph
3.7: Should You Compare Means? 129
below, the two (normal) populations have equal spreads or standard deviations.
In such cases the difference between the two population means is equal to the
difference between any fixed percentile for the two distributions, so the mean
difference is a natural measure of difference.
Consider instead the hypothetical scenario depicted in the bottom pane
below, where the population mean lifetimes using two distinct drugs for a fatal
disease are µ1 = 16 months from time of diagnosis and µ2 = 22 months from
time of diagnosis, respectively. The standard deviations under the two drugs
are σ1 = 1 and σ2 = 6, respectively. The second drug has the higher mean
lifetime, but at the expense of greater risk. For example, the first drug gives
you a 97.7% chance of living at least 14 months, whereas the second drug only
gives you a 90.8% chance of living at least 14 months. Which drug is best? It
depends on what is important to you, a higher expected lifetime or a lower risk
of dying early.
Normal Distributions with Identical Variances
0 5 10 15
Normal Distributions with Different Variances
10 15 20 25 30 35 40
Chapter 4
Checking Assumptions
Learning objectives
assess the assumptions visually and via formal tests.
10. Model assumptions.
4.1 Introduction
Almost all statistical methods make assumptions about the data collection pro-
cess and the shape of the population distribution. If you reject the null hypoth-
esis in a test, then a reasonable conclusion is that the null hypothesis is false,
provided all the distributional assumptions made by the test are satisfied. If the
assumptions are not satisfied then that alone might be the cause of rejecting
H0. Additionally, if you fail to reject H0, that could be caused solely by failure
to satisfy assumptions also. Hence, you should always check assumptions to
the best of your abilities.
Two assumptions that underly the tests and CI procedures that I have
discussed are that the data are a random sample, and that the population fre-
quency curve is normal. For the pooled variance two-sample test the population
variances are also required to be equal.
The random sample assumption can often be assessed from an understand-
ing of the data collection process. Unfortunately, there are few general tests for
4.2: Testing Normality 131
checking this assumption. I have described exploratory (mostly visual) meth-
ods to assess the normality and equal variance assumption. I will now discuss
formal methods to assess these assumptions.
4.2 Testing Normality
An informal test of normality can be based on a normal scores plot, some-

times called a rankit plot or a normal probability plot or a normal
QQ plot (QQ = quantile-quantile). You plot the quantiles of the data against
the quantiles of the normal distribution, or expected normal order statis-
tics (in a standard normal distribution) for a sample with the given number of
observations. The normality assumption is plausible if the plot is fairly linear.
I give below several plots often seen with real data, and what they indicate
about the underlying distribution.
There are multiple ways to produce QQ plots in R. The shape can depend
upon whether you plot the normal scores on the x-axis or the y-axis. It is
conventional to plot the data on the y-axis and the normal scores on the x-axis.
Let’s start with some data from a normal distribution.
#### sample from normal distribution
x1 <- rnorm(150, mean = 100, sd = 15)
par(mfrow=c(3,1))
rug(x1)
# violin plot
library(vioplot)
# boxplot
132 Ch 4: Checking Assumptions
Histogram of x1
0.000 0.015 0.030

Density
70 80 90 100 110 120 130
x1
1
70 80 90 100 110 120 130
70 80 90 100 110 120 130
There are many ways to get adequate QQ plots. Consider how outliers
shows up in the QQ plot. There may be isolated points on ends of the QQ plot,
but only on the right side is there an outlier. How could you have identified
that the right tail looks longer than the left tail from the QQ plot?
#### QQ plots
# R base graphics
par(mfrow=c(1,1))
# plots the data vs their normal scores
qqnorm(x1)
# plots the reference line
qqline(x1)
# ggplot2 graphics
library(ggplot2)
# http://had.co.nz/ggplot2/stat_qq.html
df <- data.frame(x1)
# stat_qq() below requires "sample" to be assigned a data.frame column
p <- ggplot(df, aes(sample = x1))
# plots the data vs their normal scores
p <- p + stat_qq()
print(p)
Normal Q−Q Plot ● ●

●
●●
●
●
● ● ●
130
●
● ●●
●●
●● 120 ●●
●● ●
●●
●
120
●● ●●
●
●●
●● ●●
●● ●●
●
●●
● ●●●
●
● ●
●●
● ●
●●
●●
● ●
110
●
● ●
●●
●
● ●
Sample Quantiles
●
●
●●
●
●● ●●
●
●
●●
●
● ●
●●
●
● ●●
●●
●
●
●●
●
●●
●
●
●●
● ●
●
sample
● ●
●
●
●●
● ●
●●
●●
●
●
●●
●
●● ●
100
●●
● 100 ●●
●
●
●
●●
● ●
●
●
●
●
●
● ●
●
●
●● ●●
●
●
●
●● ●●
●
●●
●
●
●●
●
● ●●
●
●
● ●
●
●●
●●
●
●
●
●●
●
●●
●
● ●
●
●●
●
● ●
●●
●
●●
●●
●
90
● ●
●
●
● ●
●
●
●
●●
●● ●
●
●●
● ●●
●●
●
●
●
●●
●●
● ●
●
●
●● ●
●● ●●
●
●●
●● ●●
●●
●●●
●● ●
●
80
●
● ●
●● ●●
●●●
●●●●●● ●●●
80 ●
●
● ●●
● ● ●● ●●●●●
70
● ●
● ● ●●
●
−2 −1 0 1 2
Theoretical Quantiles −2 −1 0 1 2
theoretical
If you lay a straightedge along the bulk of the plot (putting in a regression
line is not the right way to do it, even if it is easy), you see that the most
extreme point on the right is a little below the line, and the last few points
on the left a little above the line. What does this mean? The point on the
right corresponds to a data value more extreme than expected from a normal
distribution (the straight line is where expected and actual coincide). Extreme
points on the right are above the line. What about the left? Extreme points
there should be above the line — since the deviations from the line are above
it on the left, those points are also more extreme than expected.
Even more useful is to add confidence intervals (point-wise, not family-wise
— you will learn the meaning of those terms in the ANOVA section). You
don’t expect a sample from a normally distributed population to have a normal
scores plot that falls exactly on the line, and the amount of deviation depends
upon the sample size.
The best QQ plot I could find is available in the car package called qqPlot.
Note that with the dist= option you can use this technique to see if the data
appear from lots of possible distributions, not just normal.
par(mfrow=c(1,1))
# Normality of Residuals
library(car)
# qq plot for studentized resid
# las = 1 : turns labels on y-axis to read horizontally
# id.n = n : labels n most extreme observations, and outputs to console
# id.cex = 1 : is the size of those labels
# lwd = 1 : line width

qqPlot(x1, las = 1, id.n = 6, id.cex = 1, lwd = 1, main="QQ Plot")
## 65 110 86 125 31 111
## 150 149 148 147 146 1
QQ Plot
110 ● 65 ●
130 31 ●86
125 ●●
●●
●●
●●
120 ●●
●●
●●
●●
●
●
●
●●
●
●
●●
●
110 ●
●●
●
●●
●●
●
●
●
●
●
●●
●
●
●
●●
●
●●
●
●●
●
●
●
●●
●
●●
●
x1
●
100 ●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●●
●
●●
●
●●
●
●●
●
●●
●
●●
●●
●
●●
●
●
90 ●
●
●
●
●●
●
●
●
●
●
●
●●
●●
●
●
●
●
●
●●
●
●●
●●
●
80 ●
●●
●●●●●
●
●
● ● ●●
70
● 111
−2 −1 0 1 2
norm quantiles
In this case the x-axis is labelled “norm quantiles”. You only see a couple
of data values outside the limits (in the tails, where it usually happens). You
expect around 5% outside the limits, so there is no indication of non-normality
here. I did sample from a normal population.
4.2.1 Normality tests on non-normal data

Let’s turn to examples of sampling from other, non-normal distributions to see
how the normal QQ plot identifies important features.
Light-tailed symmetric (Uniform)

#### Light-tailed symmetric (Uniform)

x2 <- runif(150, min = 50, max = 150)
par(mfrow=c(3,1))
rug(x2)
# violin plot
library(vioplot)
# boxplot
par(mfrow=c(1,1))
Histogram of x2
QQ Plot
0.010
Density
● ● ●
●●● ●
●●●
0.000
●●●
●●●●
140 ●
●
●
●
●
60 80 100 120 140 ●
●
●●
●●
●
x2 ●●
●
●●
●●
●●
●
●●
●
●
●
●●
●
●
120 ●
●●
●
●
●
●
●
●
●●
●
●●
●
●
●
●●
●●
●
●●
●
●
●
●
●
●●
●
●
●
●
●
●
x2
100 ●
●
● ●
●
●●
1
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
60 80 100 120 140 ●
●
●●
80 ●●
●
●
●●
●●
●●
●
●
●●
●●
●
●●
●●
●●
●●●
60 ●●●
●
●
●●●
● ●●
● ● ●
−2 −1 0 1 2
60 80 100 120 140
norm quantiles
Heavy-tailed (fairly) symmetric (Normal-squared)

#### Heavy-tailed (fairly) symmetric (Normal-squared)
x3.temp <- rnorm(150, mean = 0, sd = 1)
x3 <- sign(x3.temp)*x3.temp^2 * 15 + 100
par(mfrow=c(3,1))

rug(x3)
# violin plot
library(vioplot)
# boxplot
par(mfrow=c(1,1))
Histogram of x3
QQ Plot
0.000 0.010 0.020
Density
●
300
0 50 100 150 200 250 300

250
x3
200
●
●
●●
x3
● 150 ●●
1
●●●
●●●●●●
●
●
●
●●
●●
●●
●
●●
●
●●
●●
●
●●
●●
●
●●
●
●
●●
●
●●
●
●●
●
●●
●
●●
●
●●
●
●●
●
0 50 100 150 200 250 300 100 ●●
●
●●
●
●●
●
●●
●
●●
●
●●
●
●
●●
●
●●
●
●
●●
●
●●
●
●
●
●
●●
●●
●
●●
●
●●
●
●●
●
●●
●
●●
●
●●
●●
●
●●
●●
●●
●
●●
●●
●●
●
●
50 ●●●
●
●●●
●
0 ● ● ●
●● ● ● ●● ● ●
●● ● ● ● ●●●
●●● ● ● ●
● ● ● ●
−2 −1 0 1 2
0 50 100 150 200 250 300
norm quantiles
Right-skewed (Exponential)
#### Right-skewed (Exponential)
x4 <- rexp(150, rate = 1)
par(mfrow=c(3,1))
rug(x4)
# violin plot
library(vioplot)
# boxplot
par(mfrow=c(1,1))
Histogram of x4
QQ Plot
0.8
Density
8
0.4
●
0.0
0 2 4 6 8
x4
6
●
●
x4
● 4
1
●
●●●
0 2 4 6 8 ●●●
●●●●●●
●●
●
2 ●
●
●
●
●
●●
●●
●
●●
●
●●
●●
●
●●
●
●●
●
●●
●
●●
●
●
●●
●
●●
●
●
●
●●
●
●●
●
●●
●
●●
●
●●
●●
●
●●
●
●
●●
●
●●
●
●●
●
●●
●
●●
●
●●
●
●●
●
●●
●
●●
●
●●
●
●●
●
●●
●●
●●
●●
●
●●
●●
●
●●
●
●●●●●●●●
●●
0 ● ● ● ● ●●●●●●
● ● ● ● ●
−2 −1 0 1 2
0 2 4 6 8
norm quantiles
Left-skewed (Exponential, reversed)

#### Left-skewed (Exponential, reversed)
x5 <- 15 - rexp(150, rate = 0.5)
par(mfrow=c(3,1))
rug(x5)
# violin plot
library(vioplot)
# boxplot
par(mfrow=c(1,1))
Histogram of x5
0.4
QQ Plot
Density
0.2
●
●●●● ● ● ●
●●
●●
●●
●●
●●●●●●●●●
●●
●●
●●
●●
●
●●
●●
●
0.0
●
●●
●
●●
●
●
●
●●
●
●●
●
●
●●
●
●
8 10 12 14 14 ●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●●
●
●●
●
●●
●
x5 ●
●●
●●
●
●●
●
●●
●
●●
●●
●
●
●
●
●
●
●●
●●
12 ●
●
●
●
●
●
●
●
●
●
●●
x5
● ●
●●
1
●●
●
●
●●
●●●
●●
●●●
8 10 12 14 10 ●●
●
●
8 ●
●
● ● ●
−2 −1 0 1 2
8 10 12 14
norm quantiles
Notice how striking is the lack of linearity in the QQ plot for all the non-
normal distributions, particularly the symmetric light-tailed distribution where
the boxplot looks fairly good. The QQ plot is a sensitive measure of normality.
Let us summarize the patterns we see regarding tails in the plots:
Tail
Tail Weight Left Right
Light Left side of plot points left Right side of plot points right
Heavy Left side of plot points down Right side of plot points up
4.3 Formal Tests of Normality

A formal test of normality is based on the correlation between the data and
the normal scores. The correlation is a measure of the strength of a linear
relationship, with the sign of the correlation indicating the direction of the
relationship (that is, + for increasing relationship, and − for decreasing). The
correlation varies from −1 to +1. In a normal scores plot, you are looking for
a correlation close to +1. Normality is rejected if the correlation is too small.
Critical values for the correlation test of normality, which is commonly called
the Shapiro-Wilk test, can be found in many texts.
R has several tests of normality. The Shapiro-Wilk test shapiro.test() is
a base function. The R package nortest has five others: the Anderson-Darling
4.3: Formal Tests of Normality 139
test ad.test() is useful, related to the Kolmogorov-Smirnov (Lilliefors)
test lillie.test() which is commonly used in many scientific disciplines but is
essentially useless, the Cramer-von Mises test cvm.test(), and two more. Some
packages also have the Ryan-Joiner test (closely related to the Shapiro-Wilk
test).
Extreme outliers and skewness have the biggest effects on standard meth-
ods based on normality. The Shapiro-Wilk (SW) test is better at picking up
these problems than the Kolmogorov-Smirnov (KS) test. The KS test tends
to highlight deviations from normality in the center of the distribution. These
types of deviations are rarely important because they do not have a noticeable
effect on the operating characteristics of the standard methods. The AD and
RJ tests are modifications designed to handle some of these objections.
Tests for normality may have low power in small to moderate sized samples.
Visual assessment of normality is often more valuable than a formal test. The
tests for the distributions of data above are below and in Figure 4.1.
Normal distribution
#### Formal Tests of Normality
shapiro.test(x1)
##
## Shapiro-Wilk normality test
##
## data: x1
## W = 0.9858, p-value = 0.1289
library(nortest)
ad.test(x1)
##
## Anderson-Darling normality test
##
## data: x1
## A = 0.4073, p-value = 0.3446
# lillie.test(x1)
cvm.test(x1)
##
## Cramer-von Mises normality test
##
## data: x1
## W = 0.0567, p-value = 0.4159
Heavy-tailed (fairly) symmetric

shapiro.test(x3)
Light-tailed symmetric
##
shapiro.test(x2)
## ##
## Shapiro-Wilk normality test ## data: x3
## ## W = 0.7963, p-value = 3.587e-13
## data: x2
library(nortest)
## W = 0.9525, p-value = 5.336e-05 ad.test(x3)
library(nortest)
##
ad.test(x2)
## ##
## Anderson-Darling normality test ## data: x3
## ## A = 9.143, p-value < 2.2e-16
## data: x2
# lillie.test(x3)
## A = 1.943, p-value = 5.644e-05 cvm.test(x3)
# lillie.test(x2)
## Warning: p-value is smaller than 7.37e-10,
cvm.test(x2)
cannot be computed more accurately
## ##
## Cramer-von Mises normality test ## Cramer-von Mises normality test
## ##
## data: x2 ## data: x3
## W = 0.2957, p-value = 0.0003642 ## W = 1.825, p-value = 7.37e-10
Right-skewed
shapiro.test(x4)
Left-skewed
##
## Shapiro-Wilk normality test shapiro.test(x5)
## ##
## data: x4 ## Shapiro-Wilk normality test
## W = 0.7412, p-value = 5.872e-15 ##
library(nortest) ## data: x5
ad.test(x4) ## W = 0.8743, p-value = 5.933e-10
## library(nortest)
## Anderson-Darling normality test ad.test(x5)
## ##
## data: x4 ## Anderson-Darling normality test
## A = 9.371, p-value < 2.2e-16 ##
# lillie.test(x4) ## data: x5
cvm.test(x4) ## A = 6.002, p-value = 7.938e-15
## Warning: p-value is smaller than 7.37e-10, # lillie.test(x5)

cannot be computed more accurately cvm.test(x5)
## ##
## ##
## data: x4 ## data: x5
## W = 1.654, p-value = 7.37e-10 ## W = 1.055, p-value = 9.648e-10
Figure 4.1: Normality tests for non-normal distributions

Example: Paired Differences on Sleep Remedies The following box-
plot and normal scores plots suggest that the underlying distribution of differ-
ences (for the paired sleep data taken from the previous chapter) is reasonably
symmetric, but heavy tailed. The p-value for the SW test of normality is 0.042,
and for the AD test is 0.029, both of which call into question a normality as-
sumption. A non-parametric test comparing the sleep remedies (one that does
not assume normality) is probably more appropriate here. We will return to
these data later.
# Normality tests
shapiro.test(sleep$d)
##
##
## data: sleep$d
## W = 0.838, p-value = 0.04173
library(nortest)
ad.test(sleep$d)
##
##
## data: sleep$d
## A = 0.7738, p-value = 0.02898
# lillie.test(sleep£d)
cvm.test(sleep$d)
##
##
## data: sleep$d
## W = 0.1382, p-value = 0.02769
# plot of data
par(mfrow=c(3,1))
hist(sleep$d, freq = FALSE, breaks = 20)
points(density(sleep$d), type = "l")
rug(sleep$d)
# violin plot
library(vioplot)
vioplot(sleep$d, horizontal=TRUE, col="gray")
# boxplot
boxplot(sleep$d, horizontal=TRUE)
# QQ plot
par(mfrow=c(1,1))
qqPlot(sleep$d, las = 1, id.n = 4, id.cex = 1, lwd = 1, main="QQ Plot")
## 9 5 2 8
## 10 1 9 2
Histogram of sleep$d
QQ Plot
0.8
Density
9●
0.4
0.0
0 1 2 3 4 4
sleep$d
sleep$d
2●
●
1
2
●
0 1 2 3 4
● ●
●
1 ● ●
● 8
0
● 5
● ●
−1.5 −1.0 −0.5 0.0 0.5 1.0 1.5

0 1 2 3 4
norm quantiles
Example: Androstenedione Levels This is an independent two-sample

problem, so you must look at normal scores plots for males and females.
The AD test p-value and the SW test p-value for testing normality exceeds
0.10 in each sample. Thus, given the sample sizes (14 for men, 18 for women), we
have insufficient evidence (at α = 0.05) to reject normality in either population.
The women’s boxplot contains two mild outliers, which is highly unusual
when sampling from a normal distribution. The tests are possibly not powerful
enough to pick up this type of deviation from normality in such a small sample.
In practice, this may not be a big concern. The two mild outliers probably
have a small effect on inferences in the sense that non-parametric methods
would probably lead to similar conclusions here.
library(ggplot2)
p1 <- ggplot(andro, aes(x = sex, y = level, fill=sex))
p1 <- p1 + stat_summary(fun.y = mean, geom = "point", shape = 3, size = 2)
#p1 <- p1 + coord_flip()
p1 <- p1 + labs(title = "Androstenedione Levels in Diabetics")
#print(p1)
Men Women
shapiro.test(men) shapiro.test(women)
## ##
## Shapiro-Wilk normality test ## Shapiro-Wilk normality test
## ##
## data: men ## data: women
## W = 0.9059, p-value = 0.1376 ## W = 0.9598, p-value = 0.5969
library(nortest) library(nortest)
ad.test(men) ad.test(women)
## ##
## Anderson-Darling normality test ## Anderson-Darling normality test
## ##
## A = 0.4718, p-value = 0.2058 ## A = 0.3947, p-value = 0.3364
# lillie.test(men) # lillie.test(women)
cvm.test(men) cvm.test(women)
## ##
## ##
## W = 0.0631, p-value = 0.3221 ## W = 0.0652, p-value = 0.3057
p2 <- ggplot(andro, aes(x = level, fill=sex))

p2 <- p2 + geom_histogram(binwidth = 20, alpha = 0.5, position="identity")
p2 <- p2 + labs(title = "Androstenedione Levels in Diabetics")
#print(p2)
library(gridExtra)
grid.arrange(p1, p2, ncol=1)
# QQ plot
par(mfrow=c(2,1))
qqPlot(men, las = 1, id.n = 0, id.cex = 1, lwd = 1, main="QQ Plot, Men")
qqPlot(women, las = 1, id.n = 0, id.cex = 1, lwd = 1, main="QQ Plot, Women")
Androstenedione Levels in Diabetics QQ Plot, Men
200 ●
200
150
sex 150 ●
men
level
●
men ●
● ● ●
● women ●
●
100 100
●
● ●
● ●
50
● −1 0 1
men women norm quantiles

sex
Androstenedione Levels in Diabetics QQ Plot, Women
●
6
100 ●
●
● ● ● ●
women
sex 80 ●
4
count
● ●
men ● ●
●
● ●
women
60 ● ●
2
40
●
0 −2 −1 0 1 2
0 50 100 150 200 250 norm quantiles

level
Most statisticians use graphical methods (boxplot, normal scores plot) to

assess normality, and do not carry out formal tests.
You may be surprised at how variable 10 observations from a Normal(0,1)

distribution looks like; here are 25 samples.
n = 10
r = 5
norm.many <- data.frame(id = rep(seq(1:r^2), n)
, x = rnorm(r^2 * n)
)
library(ggplot2)
p <- ggplot(norm.many, aes(x = x))
p <- p + geom_histogram(binwidth = 0.4)
p <- p + facet_wrap(~ id, ncol = r)
p <- p + labs(title = "Twenty-five samples of size n=10 from a Normal(0,1) distribution")
print(p)
Twenty−five samples of size n=10 from a Normal(0,1) distribution

1 2 3 4 5
4
0
6 7 8 9 10
4
0
11 12 13 14 15
4
3
count
0
16 17 18 19 20
4
0
21 22 23 24 25
4
0
−2 0 2 −2 0 2 −2 0 2 −2 0 2 −2 0 2
x
. . . and here are samples of size n = 30.

n = 30
r = 5
norm.many <- data.frame(id = rep(seq(1:r^2), n)
, x = rnorm(r^2 * n)
)
library(ggplot2)
p <- ggplot(norm.many, aes(x = x))
p <- p + geom_histogram(binwidth = 0.4)
p <- p + facet_wrap(~ id, ncol = r)
p <- p + labs(title = "Twenty-five samples of size n=30 from a Normal(0,1) distribution")
print(p)
Twenty−five samples of size n=30 from a Normal(0,1) distribution

1 2 3 4 5
7.5
5.0
2.5
0.0
6 7 8 9 10
7.5
5.0
2.5
0.0
11 12 13 14 15
7.5
count
5.0
2.5
0.0
16 17 18 19 20
7.5
5.0
2.5
0.0
21 22 23 24 25
7.5
5.0
2.5
0.0
−2 0 2 4 −2 0 2 4 −2 0 2 4 −2 0 2 4 −2 0 2 4
x
By viewing many versions of this of varying samples sizes you’ll develop

your intuition about what a normal sample looks like.
4.4 Testing Equal Population Variances

In the independent two sample t-test, some researchers test H0 : σ12 = σ22 as a
means to decide between using the pooled-variance procedure or Satterthwaite’s
methods. They suggest the pooled t-test and CI if H0 is not rejected, and
4.4: Testing Equal Population Variances 147
Satterthwaite’s methods otherwise.
There are a number of well-known tests for equal population variances, of
which Bartlett’s test and Levene’s test are probably the best known. Bartlett’s
test assumes the population distributions are normal, and is the best test when
this is true. In practice, unequal variances and non-normality often go hand-in-
hand, so you should check normality prior to using Bartlett’s test. It is sensitive
to data which is not non-normally distributed, thus it is more likely to return
a “false positive” (reject H0 of equal variances) when the data is non-normal.
Levene’s test is more robust to departures from normality than Bartlett’s test;
it is in the car package. Fligner-Killeen test is a non-parametric test which is
very robust against departures from normality.
I will now define Bartlett’s test, which assumes normally distributed
data. As above, let n∗ = n1 + n2 + · · · + nk , where the nis are the sample sizes
from the k groups, and define
k
!
1 X 1 1
v =1+ − ∗ .
3(k − 1) i=1 ni − 1 n − k
Bartlett’s statistic for testing H0 : σ12 = · · · = σk2 is given by

( k
)
2.303 X
Bobs = (n − k) log(s2pooled) − (ni − 1) log(s2i ) ,
v i=1
where s2pooled is the pooled estimator of variance and s2i is the estimated variance
based on the ith sample.
Large values of Bobs suggest that the population variances are unequal. For
a size α test, we reject H0 if Bobs ≥ χ2k−1,crit, where χ2k−1,crit is the upper-α
percentile for the χ2k−1 (chi-squared) probability distribution with k − 1 degrees
of freedom. A generic plot of the χ2 distribution is given below. A p-value for
the test is given by the area under the chi-squared curve to the right of Bobs.
Example: Androstenedione Levels The sample standard deviations

and samples sizes are: s1 = 42.8 and n1 = 14 for men and s2 = 17.2 and
n2 = 18 for women. The sample standard deviations appear to be very different,
so I would not be surprised if the test of equal population variances is highly
significant. The output below confirms this: the p-values for Bartlett’s F-test,
Levene’s Test, and Fligner-Killeen test are all much smaller than 0.05. An
implication is that the standard pooled-CI and test on the population means is
inappropriate.
#### Testing Equal Population Variances
c(mean(men), mean(women), sd(men), sd(women))
## [1] 112.50 75.83 42.75 17.24
c(IQR(men), IQR(women), length(men), length(women))
## [1] 60.25 17.00 14.00 18.00
## Test equal variance
# assumes populations are normal
bartlett.test(level ~ sex, data = andro)
##
## Bartlett test of homogeneity of variances
##
## Bartlett's K-squared = 11.2, df = 1, p-value = 0.0008183
# does not assume normality, requires car package
library(car)
leveneTest(level ~ sex, data = andro)
## Levene's Test for Homogeneity of Variance (center = median)
## Df F value Pr(>F)
## group 1 7.2 0.012 *
## 30
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
# nonparametric test
fligner.test(level ~ sex, data = andro)
##
## Fligner-Killeen test of homogeneity of variances
##
## Fligner-Killeen:med chi-squared = 5.892, df = 1, p-value =
## 0.01521
4.5 Small sample sizes, a comment

In Daniel Kahneman’s “Thinking, Fast and Slow” (Ch 10), he discusses “The
Law of Small Numbers” (in contrast to the Law of Large Numbers). As an
example from statisticians Howard Wainer and Harris Zwerling, he makes this
4.5: Small sample sizes, a comment 149
observation about the incidence of kidney cancer in the 3,141 counties of the
United States. “The counties in which the incidence of kidney cancer is lowest
are mostly rural, sparsely populated, and located in traditionally Republican
states in the Midwest, the South, and the West. What do you make of this?”
The statistians comment: “It is both easy and tempting to infer that their
low cancer rates are directly due to the clean living of the rural lifestyleno air
pollution, no water pollution, access to fresh food without additives.” This
makes perfect sense.
“Now consider the counties in which the incidence of kidney cancer is high-
est. These ailing counties tend to be mostly rural, sparsely populated, and
located in traditionally Republican states in the Midwest, the South, and the
West.” Tongue-in-cheek, Wainer and Zwerling comment: “It is easy to infer
that their high cancer rates might be directly due to the poverty of the rural
lifestyle — no access to good medical care, a high-fat diet, and too much al-
cohol, too much tobacco.” Something is wrong, of course. The rural lifestyle
cannot explain both very high and very low incidence of kidney cancer.
The key factor is not that the counties were rural or predominantly Re-
publican. It is that rural counties have small populations. The law of large
numbers says that as sample sizes increase that the sample statistic converges
to the population proportion, that is, large samples are more precise than small
samples. What Kahneman is calling the law of small numbers warns that small
samples yield extreme results more often than large samples do.
Chapter 5
One-Way Analysis of
Variance
Learning objectives
select graphical displays that meaningfully compare independent popula-
tions.
assess the assumptions of the ANOVA visually and by formal tests.
decide whether the means between populations are different, and how.
5.1 ANOVA
The one-way analysis of variance (ANOVA) is a generalization of the two
sample t-test to k ≥ 2 groups. Assume that the populations of interest have
the following (unknown) population means and standard deviations:
5.1: ANOVA 151
population 1 population 2 · · · population k
mean µ1 µ2 ··· µk
std dev σ1 σ2 ··· σk
A usual interest in ANOVA is whether µ1 = µ2 = · · · = µk . If not, then we
wish to know which means differ, and by how much. To answer these questions
we select samples from each of the k populations, leading to the following data
summary:
sample 1 sample 2 · · · sample k
size n1 n2 ··· nk
mean Ȳ1 Ȳ2 ··· Ȳk
std dev s1 s2 ··· sk
A little more notation is needed for the discussion. Let Yij denote the j th
observation in the ith sample and define the total sample size n∗ = n1 + n2 +
· · · + nk . Finally, let Ȳ¯ be the average response over all samples (combined),
that is P
Y
P
ij ij i ni Ȳi
Ȳ¯ = = .
n∗ n∗
Note that Ȳ¯ is not the average of the sample means, unless the samples sizes
ni are equal.
An F -statistic is used to test H0 : µ1 = µ2 = · · · = µk against HA : not H0
(that is, at least two means are different). The assumptions needed for the
standard ANOVA F -test are analogous to the independent pooled two-sample
t-test assumptions: (1) Independent random samples from each population. (2)
The population frequency curves are normal. (3) The populations have equal
standard deviations, σ1 = σ2 = · · · = σk .
The F -test is computed from the ANOVA table, which breaks the spread in
the combined data set into two components, or Sums of Squares (SS). The
Within SS, often called the Residual SS or the Error SS, is the portion
of the total spread due to variability within samples:
SS(Within) = (n1 − 1)s1 + (n2 − 1)s2 + · · · + (nk − 1)sk = ij (Yij − Ȳi)2.
2 2 2
P
The Between SS, often called the Model SS, measures the spread between
the sample means
SS(Between) =
n1(Ȳ1 − Ȳ¯ )2 + n2(Ȳ2 − Ȳ¯ )2 + · · · + nk (Ȳk − Ȳ¯ )2 = i ni(Ȳi − Ȳ¯ )2,
P
152 Ch 5: One-Way Analysis of Variance
weighted by the sample sizes. These two SS add to give
SS(Total) = SS(Between) + SS(Within) = ij (Yij − Ȳ¯ )2.
P
Each SS has its own degrees of freedom (df ). The df (Between) is the number
of groups minus one, k − 1. The df (Within) is the total number of observations
minus the number of groups: (n1 − 1) + (n2 − 1) + · · · + (nk − 1) = n∗ − k.
These two df add to give df (Total) = (k − 1) + (n∗ − k) = n∗ − 1.
The Sums of Squares and df are neatly arranged in a table, called the
ANOVA table:
Source df SS MS F
SSM =P i ni (Ȳi − Ȳ¯ )2
P
Between Groups (Model) df M = k − 1 M SM = SSM/df M M SM/M SE
Within Groups (Error) df E = n∗ − k SSE = i (ni − 1)s2i M SE = SSE/df E
df T = n∗ − 1 SST = ij (Yij − Ȳ¯ )2
P
Total M ST = SST /df T
The Mean Square for each source of variation is the corresponding SS divided
by its df . The Mean Squares can be easily interpreted.
The MS(Within)
(n1 − 1)s21 + (n2 − 1)s22 + · · · + (nk − 1)s2k

MS(Within) = ∗
= s2pooled
n −k
is a weighted average of the sample variances. The MS(Within) is known as the

pooled estimator of variance, and estimates the assumed common population
variance. If all the sample sizes are equal, the MS(Within) is the average sample
variance. The MS(Within) is identical to the pooled variance estimator
in a two-sample problem when k = 2.
The MS(Between)
− Ȳ¯ )2
P
i ni (Ȳi
MS(Between) =
k−1
is a measure of variability among the sample means. This MS is a multiple of

the sample variance of Ȳ1, Ȳ2, . . . , Ȳk when all the sample sizes are equal.
The MS(Total)
P ¯ 2
ij (Yij − Ȳ )
MS(Total) =
n∗ − 1
is the variance in the combined data set.
5.1: ANOVA 153
The decision on whether to reject H0 : µ1 = µ2 = · · · = µk is based on the
ratio of the MS(Between) and the MS(Within):
MS(Between)
Fs = .
MS(Within)
Large values of Fs indicate large variability among the sample means Ȳ1, Ȳ2, . . . , Ȳk
relative to the spread of the data within samples. That is, large values of Fs
suggest that H0 is false.
Formally, for a size α test, reject H0 if Fs ≥ Fcrit, where Fcrit is the upper-α
percentile from an F distribution with numerator degrees of freedom k − 1 and
denominator degrees of freedom n∗ − k (i.e., the df for the numerators and
denominators in the F -ratio). The p-value for the test is the area under the
F -probability curve to the right of Fs:
F with 4 and 20 degrees of freedom F with 4 and 20 degrees of freedom

FS not significant
α = .05 (fixed) p−value (random)
0 1 2 3 4 5 6 0 1 2 3 4 5 6
FCrit Reject H0 for FS here FS FCrit
For k = 2 the ANOVA F -test is equivalent to the pooled two-sample t-test.

The specification of a one-way analysis of variance is analogous to a re-
gression analysis (discussed later, though we’ve seen the specification in some
plotting functions). The only difference is that the descriptive (x) variable
needs to be a factor and not a numeric variable. We calculate a model object
using lm() and extract the analysis of variance table with anova().
Example: Comparison of Fats During cooking, doughnuts absorb fat
in various amounts. A scientist wished to learn whether the amount absorbed
depends on the type of fat. For each of 4 fats, 6 batches of 24 doughnuts were
prepared. The data are grams of fat absorbed per batch.
Let
µi = pop mean grams of fat i absorbed per batch of 24 doughnuts (−100).
The scientist wishes to test H0 : µ1 = µ2 = µ3 = µ4 against HA : not H0.
There is no strong evidence against normality here. Furthermore the sample
standard deviations (see output below) are close. The standard ANOVA ap-
pears to be appropriate here.
Row fat1 fat2 fat3 fat4
1 164 178 175 155
2 172 191 186 166
3 168 197 178 149
4 177 182 171 164
5 190 185 163 170
6 176 177 176 168
Let’s take a short detour to read the wide table and convert it into long
format.
R skills: wide to long table format Many functions in R expect

data to be in a long format rather than a wide format. Let’s use the fat data
as an example of how to read a table as text, convert the wide format to long,
and then back to wide format.
#### Example: Comparison of Fats
fat <- read.table(text="
Row fat1 fat2 fat3 fat4
1 164 178 175 155
2 172 191 186 166
3 168 197 178 149
4 177 182 171 164
5 190 185 163 170
6 176 177 176 168
", header=TRUE)
fat
## Row fat1 fat2 fat3 fat4
## 1 1 164 178 175 155
## 2 2 172 191 186 166
## 3 3 168 197 178 149
## 4 4 177 182 171 164
## 5 5 190 185 163 170
## 6 6 176 177 176 168
From wide to long: Use melt() from the reshape2 package.
5.1: ANOVA 155
#### From wide to long format

library(reshape2)
fat.long <- melt(fat,
# id.vars: ID variables
# all variables to keep but not split apart on
id.vars=c("Row"),
# measure.vars: The source columns
# (if unspecified then all other variables are measure.vars)
measure.vars = c("fat1", "fat2", "fat3", "fat4"),
# variable.name: Name of the destination column identifying each
# original column that the measurement came from
variable.name = "type",
# value.name: column name for values in table
value.name = "amount"
)
## naming variables manually, the variable.name and value.name not working 11/2012
#names(fat.long) <- c("Row", "type", "amount")
fat.long
## Row type amount
## 1 1 fat1 164
## 2 2 fat1 172
## 3 3 fat1 168
## 4 4 fat1 177
## 5 5 fat1 190
## 6 6 fat1 176
## 7 1 fat2 178
## 8 2 fat2 191
## 9 3 fat2 197
## 10 4 fat2 182
## 11 5 fat2 185
## 12 6 fat2 177
## 13 1 fat3 175
## 14 2 fat3 186
## 15 3 fat3 178
## 16 4 fat3 171
## 17 5 fat3 163
## 18 6 fat3 176
## 19 1 fat4 155
## 20 2 fat4 166
## 21 3 fat4 149
## 22 4 fat4 164
## 23 5 fat4 170
## 24 6 fat4 168
# or as simple as:
# melt(fat, "Row")
If you don’t specify variable.name, it will name that column “variable”, and
if you leave out value.name, it will name that column “value”.
From long to wide: Use dcast() from the reshape2 package.
#### From long to wide format

fat.wide <- dcast(fat.long, Row ~ type, value.var = "amount")
fat.wide
## Row fat1 fat2 fat3 fat4
## 1 1 164 178 175 155
## 2 2 172 191 186 166
## 3 3 168 197 178 149
## 4 4 177 182 171 164
## 5 5 190 185 163 170
## 6 6 176 177 176 168
Now that we’ve got our data in the long format, let’s return to the ANOVA.
Back to ANOVA: Let’s look at the numerical summaries. We’ve seen

other ways of computing these so I’ll show you another way.
#### Back to ANOVA
# Calculate the mean, sd, n, and se for the four fats
# The plyr package is an advanced way to apply a function to subsets of data

# "Tools for splitting, applying and combining data"
library(plyr)
# ddply "dd" means the input and output are both data.frames
fat.summary <- ddply(fat.long,
"type",
function(X) {
data.frame( m = mean(X$amount),
s = sd(X$amount),
n = length(X$amount)
)
}
)
# standard errors
fat.summary$se <- fat.summary$s/sqrt(fat.summary$n)
# individual confidence limits
fat.summary$ci.l <- fat.summary$m - qt(1-.05/2, df=fat.summary$n-1) * fat.summary$se
fat.summary$ci.u <- fat.summary$m + qt(1-.05/2, df=fat.summary$n-1) * fat.summary$se
fat.summary
## type m s n se ci.l ci.u
## 1 fat1 174.5 9.028 6 3.686 165.0 184.0
## 2 fat2 185.0 7.772 6 3.173 176.8 193.2
## 3 fat3 174.8 7.627 6 3.114 166.8 182.8
## 4 fat4 162.0 8.222 6 3.357 153.4 170.6
Let’s plot the data with boxplots, individual points, mean, and CI by fat
type.
# Plot the data using ggplot
library(ggplot2)
p <- ggplot(fat.long, aes(x = type, y = amount))
# plot a reference line for the global mean (assuming no groups)
p <- p + geom_hline(yintercept = mean(fat.long$amount),
5.1: ANOVA 157
colour = "black", linetype = "dashed", size = 0.3, alpha = 0.5)

# boxplot, size=.75 to stand out behind CI
p <- p + geom_boxplot(size = 0.75, alpha = 0.5)
# points for observed data
p <- p + geom_point(position = position_jitter(w = 0.05, h = 0), alpha = 0.5)
# diamond at mean for each group
p <- p + stat_summary(fun.y = mean, geom = "point", shape = 18, size = 6,
aes(colour = type), alpha = 0.8)
# confidence limits based on normal distribution
p <- p + stat_summary(fun.data = "mean_cl_normal", geom = "errorbar",
width = .2, aes(colour=type), alpha = 0.8)
p <- p + labs(title = "Doughnut fat absorption") + ylab("amount absorbed (g)")
print(p)
Doughnut fat absorption
190 ●
●
amount absorbed (g)
180
type
fat1
fat2
fat3
170
fat4
160
150
fat1 fat2 fat3 fat4

type
The p-value for the F -test is 0.001. The scientist would reject H0 at any
of the usual test levels (such as, 0.05 or 0.01). The data suggest that the
population mean absorption rates differ across fats in some way. The F -test
does not say how they differ. The pooled standard deviation spooled = 8.18 is
the “Residual standard error”. We’ll ignore the rest of this output for now.
fit.f <- aov(amount ~ type, data = fat.long)
summary(fit.f)
## Df Sum Sq Mean Sq F value Pr(>F)
## type 3 1595 532 7.95 0.0011 **

## Residuals 20 1338 67
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
fit.f
## Call:
## aov(formula = amount ~ type, data = fat.long)
##
## Terms:
## type Residuals
## Sum of Squares 1596 1338
## Deg. of Freedom 3 20
##
## Residual standard error: 8.18
## Estimated effects may be unbalanced
ClickerQ s — ANOVA, Fat 1/2
ClickerQ s — ANOVA, Fat 1/2
5.2 Multiple Comparison Methods: Fisher’s

Method
The ANOVA F -test checks whether all the population means are equal. Mul-
tiple comparisons are often used as a follow-up to a significant ANOVA F -
test to determine which population means are different. I will discuss Fisher’s,
Bonferroni’s, and Tukey’s methods for comparing all pairs of means.
Fisher’s least significant difference method (LSD or FSD) is a two-step

process:
1. Carry out the ANOVA F -test of H0 : µ1 = µ2 = · · · = µk at the α level.
If H0 is not rejected, stop and conclude that there is insufficient evidence
to claim differences among population means. If H0 is rejected, go to step
2.
2. Compare each pair of means using a pooled two sample t-test at the α
level. Use spooled from the ANOVA table and df = df E (Residual).
5.2: Multiple Comparison Methods: Fisher’s Method 159
To see where the name LSD originated, consider the t-test of H0 : µi = µj (i.e.,
populations i and j have same mean). The t-statistic is
Ȳi − Ȳj
ts = q .
1 1
spooled ni + nj
You reject H0 if |ts| ≥ tcrit, or equivalently, if

s
1 1
|Ȳi − Ȳj | ≥ tcritspooled + .
ni nj
The minimum absolute difference between Ȳi and Ȳj needed to reject H0 is the
LSD, the quantity on the right hand side of this inequality. If all the sample sizes
are equal n1 = n2 = · · · = nk then the LSD is the same for each comparison:
r
2
LSD = tcritspooled ,
n1
where n1 is the common sample size.
I will illustrate Fisher’s method on the doughnut data, using α = 0.05. At
the first step, you reject the hypothesis that the population mean absorptions
are equal because p-value= 0.001. At the second step, compare all pairs of fats
at the 5% level. Here, spooled = 8.18 and tcrit = 2.086 for a two-sided test based
on 20 df (the df E for Residual SS). Each sample has six observations, so the
LSD for each comparison is
r
2
LSD = 2.086 × 8.18 × = 9.85.
6
Any two sample means that differ by at least 9.85 in magnitude are signifi-
cantly different at the 5% level.
An easy way to compare all pairs of fats is to order the samples by their
sample means. The samples can then be grouped easily, noting that two fats
are in the same group if the absolute difference between their sample means is
smaller than the LSD.
Fats Sample Mean
2 185.00
3 174.83
1 174.50
4 162.00
There are six comparisons of two fats. From this table, you can visually
assess which sample means differ by at least the LSD=9.85, and which ones do
not. For completeness, the table below summarizes each comparison:
Comparison Absolute difference in means Exceeds LSD?

Fats 2 and 3 10.17 Yes
2 and 1 10.50 Yes
2 and 4 23.00 Yes
Fats 3 and 1 0.33 No
3 and 4 12.83 Yes
Fats 1 and 4 12.50 Yes
The end product of the multiple comparisons is usually presented as a col-

lection of groups, where a group is defined to be a set of populations with
sample means that are not significantly different from each other. Overlap
among groups is common, and occurs when one or more populations appears
in two or more groups. Any overlap requires a more careful interpretation of
the analysis.
There are three groups for the doughnut data, with no overlap. Fat 2 is in
a group by itself, and so is Fat 4. Fats 3 and 1 are in a group together. This
information can be summarized by ordering the samples from lowest to highest
average, and then connecting the fats in the same group using an underscore:
FAT 4 FAT 1 FAT 3 FAT 2
----- -------------- -----
The results of a multiple comparisons must be interpreted carefully. At the
5% level, you have sufficient evidence to conclude that the population mean
absorption for Fat 2 and Fat 4 are each different than the other population
means. However, there is insufficient evidence to conclude that the population
mean absorptions for Fats 1 and 3 differ.
Be Careful with Interpreting Groups in Multiple Comparisons!
To see why you must be careful when interpreting groupings, suppose you obtain
two groups in a three sample problem. One group has samples 1 and 3. The
other group has samples 3 and 2:
1 3 2
-----------
-----------
This occurs, for example, when |Ȳ1 − Ȳ2| ≥ LSD, but both |Ȳ1 − Ȳ3| and
|Ȳ3 − Ȳ2| are less than the LSD. There is a tendency to conclude, and please
try to avoid this line of attack, that populations 1 and 3 have the same mean,
populations 2 and 3 have the same mean, but populations 1 and 2 have different
means. This conclusion is illogical. The groupings imply that we have sufficient
evidence to conclude that population means 1 and 2 are different, but insufficient
evidence to conclude that population mean 3 differs from either of the other
population means.
5.2.1 FSD Multiple Comparisons in R

One way to get Fisher comparisons in R uses pairwise.t.test() with p.adjust.method
The resulting summary of the multiple comparisons is in terms of p-values for
all pairwise two-sample t-tests using the pooled standard deviation from the
ANOVA using pool.sd = TRUE. This output can be used to generate groupings.
A summary of the p-values is given below. Let us see that we can recover the
groups from this output.
#### Multiple Comparisons
# all pairwise comparisons among levels of fat
# Fisher's LSD (FSD) uses "none"
pairwise.t.test(fat.long$amount, fat.long$type,
pool.sd = TRUE, p.adjust.method = "none")
##
## Pairwise comparisons using t tests with pooled SD
##
## data: fat.long$amount and fat.long$type
##
## fat1 fat2 fat3
## fat2 0.038 - -
## fat3 0.944 0.044 -
## fat4 0.015 9.3e-05 0.013
##
## P value adjustment method: none
Discussion of the FSD Method: family error rate
There are c = k(k − 1)/2 pairs of means to compare in the second step of
the FSD method. Each comparison is done at the α level, where for a generic
comparison of the ith and j th populations
α = probability of rejecting H0 : µi = µj when H0 is true.
The individual error rate is not the only error rate that is important in mul-
tiple comparisons. The family error rate (FER), or the experimentwise
error rate, is defined to be the probability of at least one false rejection of
a true hypothesis H0 : µi = µj over all comparisons. When many compar-
isons are made, you may have a large probability of making one or more false
rejections of true null hypotheses. In particular, when all c comparisons of two
population means are performed, each at the α level, then
α < F ER < cα.
For example, in the doughnut problem where k = 4, there are c = 4(3)/2 =

6 possible comparisons of pairs of fats. If each comparison is carried out at the
5% level, then 0.05 < F ER < 0.30. At the second step of the FSD method,
you could have up to a 30% chance of claiming one or more pairs of population
means are different if no differences existed between population means. Most
statistical packages do not evaluate the FER, so the upper bound is used.
The first step of the FSD method is the ANOVA “screening” test. The
multiple comparisons are carried out only if the F -test suggests that not all
population means are equal. This screening test tends to deflate the FER for
the two-step FSD procedure. However, the FSD method is commonly criticized
for being extremely liberal (too many false rejections of true null hypotheses)
when some, but not many, differences exist — especially when the number of
comparisons is large. This conclusion is fairly intuitive. When you do a large
number of tests, each, say, at the 5% level, then sampling variation alone will
suggest differences in 5% of the comparisons where the H0 is true. The number
of false rejections could be enormous with a large number of comparisons. For
example, chance variation alone would account for an average of 50 significant
differences in 1000 comparisons (about 45 populations) each at the 5% level.
5.2.2 Bonferroni Comparisons
The Bonferroni method controls the FER by reducing the individual comparison
error rate. The FER is guaranteed to be no larger than a prespecified amount,
say α, by setting the individual error rate for each of the c comparisons of
interest to α/c. Therefore, α/c < F ER < cα/c = α, thus the upper bound for
FER is α. Larger differences in the sample means are needed before declaring
statistical significance using the Bonferroni adjustment than when using the
FSD method at the α level.
ClickerQ s — ANOVA, Bonferroni
Assuming all comparisons are of interest, you can implement the Bonferroni
adjustment in R by specifying p.adjust.method = "bonf" A by-product of the
Bonferroni adjustment is that we have at least 100(1 − α)% confidence that all
pairwise t-test statements hold simultaneously!
# Bonferroni 95% Individual p-values
# All Pairwise Comparisons among Levels of fat
pool.sd = TRUE, p.adjust.method = "bonf")
##
##
##
## fat1 fat2 fat3
## fat2 0.22733 - -
## fat3 1.00000 0.26241 -
## fat4 0.09286 0.00056 0.07960
##
## P value adjustment method: bonferroni
Looking at the output, can you create the groups? You should get the
groups given below, which implies you have sufficient evidence to conclude that
the population mean absorption for Fat 2 is different than that for Fat 4.
FAT 4 FAT 1 FAT 3 FAT 2
-----------------------
-----------------------
The Bonferroni method tends to produce “coarser” groups than the FSD method,
because the individual comparisons are conducted at a lower (alpha/error) level.
Equivalently, the minimum significant difference is inflated for the Bonferroni
method. For example, in the doughnut problem with F ER ≤ 0.05, the critical
value for the individual comparisons at the 0.0083 level is tcrit = 2.929 with
df = 20. The minimum significant difference for the Bonferroni comparisons is
r
2
LSD = 2.929 × 8.18 × = 13.824
6
versus an LSD=9.85 for the FSD method. Referring back to our table of sam-
ple means on page 160, we see that the sole comparison where the absolute
difference between sample means exceeds 13.824 involves Fats 2 and 4.
Example from Koopmans: glabella facial tissue thickness In an

anthropological study of facial tissue thickness for different racial groups, data
were taken during autopsy at several points on the faces of deceased individuals.
The Glabella measurements taken at the bony ridge for samples of individuals
from three racial groups (cauc = Caucasian, afam = African American, and
naaa = Native American and Asian) follow. The data values are in mm.
Bregma
temporal line
r
rio
pe
Zygomatic tubercle Su
Zygomaticofrontal Lambda
suture rio
n
Supraorbital foramen Pte
Glabella
22mm
35mm
Nasion
Asterion
Inion
Zygomatic
Zygomatic
arch
hal
bone nuc Reid's base
line line
Maxilla Median nuchal crest
Mandible
Auricular point
Pre-auricular point
#### Example from Koopmans: glabella facial tissue thickness

glabella <- read.table(text="
Row cauc afam naaa
1 5.75 6.00 8.00
2 5.50 6.25 7.00
3 6.75 6.75 6.00
4 5.75 7.00 6.25
5 5.00 7.25 5.50
6 5.75 6.75 4.00
7 5.75 8.00 5.00
8 7.75 6.50 6.00
9 5.75 7.50 7.25
10 5.25 6.25 6.00
11 4.50 5.00 6.00
12 6.25 5.75 4.25
13 NA 5.00 4.75
14 NA NA 6.00
", header=TRUE)
glabella.long <- melt(glabella,

id.vars=c("Row"),
variable.name = "pop",
value.name = "thickness",
# remove NAs
na.rm = TRUE
)
# naming variables manually, the variable.name and value.name not working 11/2012
names(glabella.long) <- c("Row", "pop", "thickness")
# another way to remove NAs:
#glabella.long <- subset(glabella.long, !is.na(thickness))
library(ggplot2)
p <- ggplot(glabella.long, aes(x = pop, y = thickness))
p <- p + geom_hline(yintercept = mean(glabella.long$thickness),
aes(colour=pop), alpha = 0.8)
width = .2, aes(colour=pop), alpha = 0.8)
p <- p + labs(title = "Glabella thickness") + ylab("thickness (mm)")
print(p)
Glabella thickness
8 ●
7
●
thickness (mm)
pop
cauc
6
afam
naaa
cauc afam naaa

pop
There are 3 groups, so there are 3 possible pairwise comparisons. If you

want a Bonferroni analysis with FER of no greater than 0.05, you should do
the individual comparisons at the 0.05/3 = 0.0167 level. Except for the mild
outlier in the Caucasian sample, the observed distributions are fairly symmetric,
with similar spreads. I would expect the standard ANOVA to perform well here.
Let µc = population mean Glabella measurement for Caucasians, µa =
population mean Glabella measurement for African Americans, and µn = pop-
ulation mean Glabella measurement for Native Americans and Asians.
glabella.summary <- ddply(glabella.long, "pop",
function(X) { data.frame( m = mean(X$thickness),
s = sd(X$thickness),
n = length(X$thickness) ) } )
glabella.summary
## pop m s n
## 1 cauc 5.812 0.8334 12
## 2 afam 6.462 0.8947 13
## 3 naaa 5.857 1.1168 14
fit.g <- aov(thickness ~ pop, data = glabella.long)

summary(fit.g)
## pop 2 3.4 1.699 1.83 0.18
## Residuals 36 33.5 0.929
fit.g
## Call:
## aov(formula = thickness ~ pop, data = glabella.long)
##
## Terms:
## pop Residuals
## Sum of Squares 3.40 33.46
##
At the 5% level, you would not reject the hypothesis that the population
mean Glabella measurements are identical. That is, you do not have sufficient
evidence to conclude that these racial groups differ with respect to their average
Glabella measurement. This is the end of the analysis!
The Bonferroni intervals reinforce this conclusion, all the p-values are greater
than 0.05. If you were to calculate CIs for the difference in population means,
each would contain zero. You can think of the Bonferroni intervals as simul-
taneous CI. We’re (at least) 95% confident that all of the following statements
hold simultaneously: −1.62 ≤ µc − µa ≤ 0.32, −0.91 ≤ µn − µc ≤ 1.00, and
−1.54 ≤ µn − µa ≤ 0.33. The individual CIs have level 100(1 − 0.0167)% =
98.33%.
# Bonferroni 95% Individual p-values
# All Pairwise Comparisons among Levels of glabella
pairwise.t.test(glabella.long$thickness, glabella.long$pop,
pool.sd = TRUE, p.adjust.method = "bonf")
##
##
## data: glabella.long$thickness and glabella.long$pop
##
## cauc afam
## afam 0.30 -
## naaa 1.00 0.34
##
## P value adjustment method: bonferroni
5.3 Further Discussion of Multiple Compar-
isons
The FSD and Bonferroni methods comprise the ends of the spectrum of mul-
tiple comparisons methods. Among multiple comparisons procedures, the FSD
method is most likely to find differences, whether real or due to sampling vari-
ation, whereas Bonferroni is often the most conservative method. You can
be reasonably sure that differences suggested by the Bonferroni method will
be suggested by almost all other methods, whereas differences not significant
under FSD will not be picked up using other approaches.
The Bonferroni method is conservative, but tends to work well when the
number of comparisons is small, say 4 or less. A smart way to use the Bonferroni
adjustment is to focus attention only on the comparisons of interest (generated
independently of looking at the data!), and ignore the rest. I will return to this
point later.
A commonly-used alternative is Tukey’s honest significant difference method
(HSD). John Tukey’s honest significant difference method is to reject the equal-
ity of a pair of means based, not on the t-distribution, but the studentized
range distribution. To implement Tukey’s method with a FER of α, reject
H0 : µi = µj when
s
qcrit 1 1
|Ȳi − Ȳj | ≥ √ spooled + ,
2 ni nj
where qcrit is the α level critical value of the studentized range distribution. For
the doughnut fats, the groupings based on Tukey and Bonferroni comparisons
are identical.
#### Tukey's honest significant difference method (HSD)
## Fat
# Tukey 95% Individual p-values
# All Pairwise Comparisons among Levels of fat
TukeyHSD(fit.f)
## Tukey multiple comparisons of means
## 95% family-wise confidence level
##
## Fit: aov(formula = amount ~ type, data = fat.long)
##
5.3: Further Discussion of Multiple Comparisons 169
## $type
## diff lwr upr p adj
## fat2-fat1 10.5000 -2.719 23.7190 0.1511
## fat3-fat1 0.3333 -12.886 13.5524 0.9999
## fat4-fat1 -12.5000 -25.719 0.7190 0.0679
## fat3-fat2 -10.1667 -23.386 3.0524 0.1710
## fat4-fat2 -23.0000 -36.219 -9.7810 0.0005
## fat4-fat3 -12.8333 -26.052 0.3857 0.0590
## Glabella
# All Pairwise Comparisons among Levels of pop
TukeyHSD(fit.g)
##
## Fit: aov(formula = thickness ~ pop, data = glabella.long)
##
## $pop
## afam-cauc 0.64904 -0.2943 1.5924 0.2260
## naaa-cauc 0.04464 -0.8824 0.9717 0.9924
## naaa-afam -0.60440 -1.5120 0.3033 0.2473
Another popular method controls the false discovery rate (FDR) instead
of the FER. The FDR is the expected proportion of false discoveries amongst
the rejected hypotheses. The false discovery rate is a less stringent condition
than the family-wise error rate, so these methods are more powerful than the
others, though with a higher FER. I encourage you to learn more about the
methods by Benjamini, Hochberg, and Yekutieli.
#### false discovery rate (FDR)
## Fat
# FDR
pool.sd = TRUE, p.adjust.method = "BH")
##
##
##
## fat1 fat2 fat3
## fat2 0.05248 - -
## fat3 0.94443 0.05248 -
## fat4 0.03095 0.00056 0.03095
##
## P value adjustment method: BH
## Glabella
# FDR
pairwise.t.test(glabella.long$thickness, glabella.long$pop,
pool.sd = TRUE, p.adjust.method = "BH")
##
##
## data: glabella.long$thickness and glabella.long$pop
##
## cauc afam
## afam 0.17 -
## naaa 0.91 0.17
##
## P value adjustment method: BH
ClickerQ s — ANOVA, multiple comparisons
5.4 Checking Assumptions in ANOVA Prob-

lems
The classical ANOVA assumes that the populations have normal frequency
curves and the populations have equal variances (or spreads). You can test
the normality assumption using multiple normal QQ-plots and normal scores
tests, which we discussed in Chapter 4. An alternative approach that is useful
with three or more samples is to make a single normal scores plot for the entire
data set. The samples must be centered at the same location for this to be
meaningful. (WHY?) This is done by subtracting the sample mean from each
observation in the sample, giving the so-called residuals. A normal scores
plot or histogram of the residuals should resemble a sample from a normal
population. These two plots can be generated with output in $residuals from
the anova() procedure.
For the glabella residuals, there are a few observations outside the confidence
bands, but the formal normality tests each have p-values > 0.2, so there’s weak
but unconvincing evidence of nonnormality.
#### Checking Assumptions in ANOVA Problems
# plot of data
par(mfrow=c(3,1))
hist(fit.g$residuals, freq = FALSE, breaks = 20)
5.4: Checking Assumptions in ANOVA Problems 171
points(density(fit.g$residuals), type = "l")

rug(fit.g$residuals)
# violin plot
library(vioplot)
vioplot(fit.g$residuals, horizontal=TRUE, col="gray")
# boxplot
boxplot(fit.g$residuals, horizontal=TRUE)
# QQ plot
par(mfrow=c(1,1))
library(car)
qqPlot(fit.g$residuals, las = 1, id.n = 8, id.cex = 1, lwd = 1, main="QQ Plot")
## 29 8 34 40 21 25 27 37
## 39 38 1 2 37 3 4 36
Histogram of fit.g$residuals
QQ Plot
0.8
Density
0.4
29 ●
2 8●
0.0
−2 −1 0 1 2 21 ●
37 ●
fit.g$residuals
●
●
1 ●
●
fit.g$residuals
●
●●
●●
● ●●●●●
1
●
0 ●●●●●
●●
●●
−2 −1 0 1 2 ●
●
●
●●
−1 ●
●
● ● 2527
● 40
● 34
● ●
−2
−2 −1 0 1 2
−2 −1 0 1 2
norm quantiles
shapiro.test(fit.g$residuals)
##
##
## data: fit.g$residuals
## W = 0.9769, p-value = 0.5927
library(nortest)
ad.test(fit.g$residuals)
##
##
## A = 0.3773, p-value = 0.3926
# lillie.test(fit.g£residuals)
cvm.test(fit.g$residuals)
##
##
## W = 0.0709, p-value = 0.2648
In Chapter 4, I illustrated the use of Bartlett’s test and Levene’s test for
equal population variances, and showed how to evaluate these tests in R.
χ2 distribution with 3 degrees of freedom
α = .05 (fixed)
0 4
χ2Crit Reject H0 for χ2S here
R does the calculation for us, as illustrated below. Because the p-value
> 0.5, we fail to reject the null hypothesis that the population variances are
equal. This result is not surprising given how close the sample variances are to
each other.
# Barlett assumes populations are normal
bartlett.test(thickness ~ pop, data = glabella.long)
##
##
5.5: Example from the Child Health and Development Study (CHDS) 173
## data: thickness by pop

Levene’s and Flinger’s tests are consistent with Bartlett’s.
# Levene does not assume normality, requires car package
library(car)
leveneTest(thickness ~ pop, data = glabella.long)
## group 2 0.53 0.59
## 36
# Fligner is a nonparametric test
fligner.test(thickness ~ pop, data = glabella.long)
##
##
## data: thickness by pop
## 0.5972
5.5 Example from the Child Health and De-

velopment Study (CHDS)
We consider data from the birth records of 680 live-born white male infants.
The infants were born to mothers who reported for pre-natal care to three
clinics of the Kaiser hospitals in northern California. As an initial analysis, we
will examine whether maternal smoking has an effect on the birth weights of
these children. To answer this question, we define 3 groups based on mother’s
smoking history: (1) mother does not currently smoke or never smoked, (2)
mother smoked less than one pack of cigarettes a day during pregnancy, and
(3) mother smoked at least one pack of cigarettes a day during pregnancy.
Let µi = pop mean birth weight (lb) for children in group i, (i = 1, 2, 3).
We wish to test H0 : µ1 = µ2 = µ3 against HA : not H0.
We read in the data, create a smoke factor variable, and plot the data by
smoking group.
#### Example from the Child Health and Development Study (CHDS)
# description at http://statacumen.com/teach/ADA1/ADA1_notes_05-CHDS_desc.txt
# read data from website
chds <- read.csv("http://statacumen.com/teach/ADA1/ADA1_notes_05-CHDS.csv")
chds$smoke <- rep(NA, nrow(chds));

# no cigs
chds[(chds$m_smok == 0), "smoke"] <- "0 cigs";
# less than 1 pack (20 cigs = 1 pack)
chds[(chds$m_smok > 0) & (chds$m_smok < 20),"smoke"] <- "1-19 cigs";
# at least 1 pack (20 cigs = 1 pack)
chds[(chds$m_smok >= 20),"smoke"] <- "20+ cigs";
chds$smoke <- factor(chds$smoke)
p1 <- ggplot(chds, aes(x = c_bwt))
p1 <- p1 + geom_histogram(binwidth = .4)
p1 <- p1 + facet_grid(smoke ~ .)
p1 <- p1 + labs(title = "Child birthweight vs maternal smoking") +
xlab("child birthweight (lb)")
#print(p1)
p2 <- ggplot(chds, aes(x = c_bwt, fill=smoke))

p2 <- p2 + geom_histogram(binwidth = .4, alpha = 1/3, position="identity")
p2 <- p2 + labs(title = "Child birthweight vs maternal smoking") +
xlab("child birthweight (lb)")
#print(p2)
library(gridExtra)
grid.arrange(p1, p2, ncol=1)

library(ggplot2)
p <- ggplot(chds, aes(x = smoke, y = c_bwt))
p <- p + geom_hline(yintercept = mean(chds$c_bwt),
aes(colour=smoke), alpha = 0.8)
width = .2, aes(colour=smoke), alpha = 0.8)
p <- p + labs(title = "Child birthweight vs maternal smoking") +
ylab("child birthweight (lb)")
print(p)
Child birthweight vs maternal smoking Child birthweight vs maternal smoking
60
0 cigs
●
40
20 ●
0 ●
1−19 cigs 20+ cigs

60
count
40 10.0 ●
20
0
60
40
child birthweight (lb)

20
0
2.5 5.0 7.5 10.0 smoke
child birthweight (lb) 7.5
0 cigs
1−19 cigs
Child birthweight vs maternal smoking 20+ cigs
60
smoke
0 cigs 5.0
count
40
1−19 cigs
20+ cigs
20 ●
●
0
2.5 5.0 7.5 10.0 0 cigs 1−19 cigs 20+ cigs
child birthweight (lb) smoke
Looking at the boxplots, there is some evidence of non-normality here. Al-

though there are outliers in the no smoking group, we need to recognize that
the sample size for this group is fairly large (381). Given that boxplots are cal-
ibrated in such a way that 7 outliers per 1000 observations are expected when
sampling from a normal population, 5 outliers (only 4 are visible!) out of 381
seems a bit excessive. A formal test rejects the hypothesis of normality in the
no and low smoker groups. The normal probability plot and the histogram of
the residuals also suggests that the population distributions are heavy tailed.
library(car)
par(mfrow=c(1,3))
qqPlot(subset(chds, smoke == "0 cigs" )$c_bwt, las = 1, id.n = 0,
id.cex = 1, lwd = 1, main="QQ Plot, 0 cigs")
qqPlot(subset(chds, smoke == "1-19 cigs")$c_bwt, las = 1, id.n = 0,
id.cex = 1, lwd = 1, main="QQ Plot, 1-19 cigs")
qqPlot(subset(chds, smoke == "20+ cigs" )$c_bwt, las = 1, id.n = 0,
id.cex = 1, lwd = 1, main="QQ Plot, 20+ cigs")
QQ Plot, 0 cigs QQ Plot, 1−19 cigs QQ Plot, 20+ cigs
●
10 ● ●
● ●
●
●●
●●● 9 ●●●●
subset(chds, smoke == "1−19 cigs")$c_bwt

● ● ●
●●●●
subset(chds, smoke == "20+ cigs")$c_bwt

●
● ●● ●
subset(chds, smoke == "0 cigs")$c_bwt

● ●
● ●
●●
●
10 ●
●
●
●●
● ●
●
●
●
●●
●
●
●
●
● 9 ●
●
●●
●
●
●
●
●
●● ●
●
●●
●● ●
●
●
●●
●
● ●
●
●
●
●
●
●
●
●
●● ●● 8 ●●
●
●
●
●
●
●
●
●●
●●
●
● ●●
● ●
●●
●
●
●
● ●
● ●●
●
●
●●
●
●
●●
●
●
●
●●
●
●●
● ●
●●
●
●
●●
●
●
●
● ●
●
●●
●●
●
●
● ●
●
●
●●
●
● ●●
●
●●
●●
●
●
●
8 ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
● 8 ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●
●●
● ●●
●
●● ●
●
●●
●
●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
7 ●
●
●
●●
●
●●
●
● ●
●
● ●
●
●
●●
●
●●
●
●
●●
●
●
●●
●
●
● ● ●
●
●● ●●
●
●● ●
●●
●●
●
●
●●
●
●
●●
● ●●
● ●
●●
●
●●
●●
●
● ●
●●
●
●●
●●
● ●
●
●●
●●
●
●●
●
●
●
●
●
●
●
● 7 ●
●
●
●
●
● ●
●
●
6 ●
●
●
●
●●
●
●
●
●
●●
6 ●
●●
●
●
●
●
●
●
● ●
●● ●
●●
● ●
●● ●
●
●●
●
●● ●
●
●● ●
●●
●
●
●●
●
●
● ●●
● ●
●
●●
●
●●
●
●●
●
● ●
6 ●
●
●●
●●
●
●●
5
4 ●
●
●●●
●●
●● ●
● ●● ●
● ● ● ●
−3 −2 −1 0 1 2 3 −2 −1 0 1 2 −2 −1 0 1 2
norm quantiles norm quantiles norm quantiles
library(nortest)
# 0 cigs --------------------
shapiro.test(subset(chds, smoke == "0 cigs" )$c_bwt)
##
##
## data: subset(chds, smoke == "0 cigs")$c_bwt
## W = 0.9872, p-value = 0.00199
ad.test( subset(chds, smoke == "0 cigs" )$c_bwt)
##
##
## A = 0.9283, p-value = 0.01831
cvm.test( subset(chds, smoke == "0 cigs" )$c_bwt)
##
##
## W = 0.1384, p-value = 0.03374
# 1-19 cigs --------------------
shapiro.test(subset(chds, smoke == "1-19 cigs")$c_bwt)
##
##
## data: subset(chds, smoke == "1-19 cigs")$c_bwt
## W = 0.9785, p-value = 0.009926
ad.test( subset(chds, smoke == "1-19 cigs")$c_bwt)
##
##
## A = 0.8308, p-value = 0.03149
cvm.test( subset(chds, smoke == "1-19 cigs")$c_bwt)

##
##
## W = 0.1133, p-value = 0.07317
# 20+ cigs --------------------
shapiro.test(subset(chds, smoke == "20+ cigs" )$c_bwt)
##
##
## data: subset(chds, smoke == "20+ cigs")$c_bwt
## W = 0.9813, p-value = 0.06962
ad.test( subset(chds, smoke == "20+ cigs" )$c_bwt)
##
##
## A = 0.4001, p-value = 0.3578
cvm.test( subset(chds, smoke == "20+ cigs" )$c_bwt)
##
##
## W = 0.0405, p-value = 0.6694
Fit the ANOVA (we’ll look at the table below).

fit.c <- aov(c_bwt ~ smoke, data = chds)
A formal test of normality on the residuals of the combined sample is

marginally significant (SW p-value= 0.047, others > 0.10). However, I am
not overly concerned about this for the following reasons: in large samples,
small deviations from normality are often statistically significant and in my
experience, the small deviations we are seeing here are not likely to impact
our conclusions, in the sense that non-parametric methods that do not require
normality will lead to the same conclusions.
# plot of data
par(mfrow=c(3,1))
hist(fit.c$residuals, freq = FALSE, breaks = 20)
points(density(fit.c$residuals), type = "l")
rug(fit.c$residuals)
# violin plot
library(vioplot)
vioplot(fit.c$residuals, horizontal=TRUE, col="gray")
# boxplot
boxplot(fit.c$residuals, horizontal=TRUE)
# QQ plot
par(mfrow=c(1,1))
library(car)
qqPlot(fit.c$residuals, las = 1, id.n = 0, id.cex = 1, lwd = 1, main="QQ Plot")
Histogram of fit.c$residuals
QQ Plot
0.4
Density
0.2
●
●
0.0
●●●
●●
−4 −2 0 2 4
●
●●●●
●
●●
●●
●
●
●
●●
fit.c$residuals 2 ●
●
●●
●
●
●
●
●
●
●
●
●●
●●
●
●
●●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●●
●
●
●●
●
●
●
●
●●
●
●
●
●●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
fit.c$residuals ●
●
●●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●●
●
●
●●
●●
●
●●
●
●
●●
0 ●●
●
●●
●
●
●
●
●
●●
●
●●
●
●
●
●
●
●●
●
●
●●
●
● ●
●
●
●●
●
●●
●
1
●
●
●●
●
●●
●
●
●●
●
●●
●
●
●
●●
●
●●
●
●
●●
●
●
●
●
●●
●
●●
●
●
●●
●●
●
●
●●
●
●
●●
●
●●
●
●●
●
●
●●
−4 −2 0 2 ●
●
●●
●●
●
●
●
●
●●
●
●
●
●●
●●
●
●
−2 ●
●
●
●
●
●●●
●
●●
●●●
●
−4
●
● ● ● ●
−3 −2 −1 0 1 2 3
−4 −2 0 2
norm quantiles
shapiro.test(fit.c$residuals)
##
##
## data: fit.c$residuals
## W = 0.9955, p-value = 0.04758
library(nortest)
ad.test(fit.c$residuals)
##
##
## A = 0.6218, p-value = 0.1051
cvm.test(fit.c$residuals)
##
##
## W = 0.092, p-value = 0.1449
Looking at the summaries, we see that the sample standard deviations are
close. Formal tests of equal population variances are far from significant. The
p-values for Bartlett’s test and Levene’s test are greater than 0.4. Thus, the
standard ANOVA appears to be appropriate here.
# calculate summaries
chds.summary <- ddply(chds, "smoke",
function(X) { data.frame( m = mean(X$c_bwt),
s = sd(X$c_bwt),
n = length(X$c_bwt) ) } )
chds.summary
## smoke m s n
## 1 0 cigs 7.733 1.052 381
## 2 1-19 cigs 7.221 1.078 169
## 3 20+ cigs 7.266 1.091 130
# assumes populations are normal
bartlett.test(c_bwt ~ smoke, data = chds)
##
##
## data: c_bwt by smoke
# does not assume normality, requires car package
library(car)
leveneTest(c_bwt ~ smoke, data = chds)
## group 2 0.76 0.47
## 677
# nonparametric test
fligner.test(c_bwt ~ smoke, data = chds)
##
##
## data: c_bwt by smoke
## 0.3512
The p-value for the F -test is less than 0.0001. We would reject H0 at any of
the usual test levels (such as 0.05 or 0.01). The data suggest that the population
mean birth weights differ across smoking status groups.
summary(fit.c)
## smoke 2 41 20.35 17.9 2.6e-08 ***
## Residuals 677 769 1.14
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
fit.c
## Call:
## aov(formula = c_bwt ~ smoke, data = chds)
##
## Terms:
## smoke Residuals
##
The Tukey multiple comparisons suggest that the mean birth weights are
different (higher) for children born to mothers that did not smoke during preg-
nancy.
## CHDS
TukeyHSD(fit.c)
##
## Fit: aov(formula = c_bwt ~ smoke, data = chds)
##
## $smoke
## 1-19 cigs-0 cigs -0.51151 -0.7429 -0.2801 0.0000
## 20+ cigs-0 cigs -0.46665 -0.7210 -0.2123 0.0001
## 20+ cigs-1-19 cigs 0.04485 -0.2473 0.3370 0.9308
Chapter 6
Nonparametric Methods
Learning objectives
select the appropriate procedure based on assumptions.
explain reason for using one procedure over another.
decide whether the medians between multiple populations are different.
3. select correct statistical procedure.
10. identify and explain statistical methods, assumptions, and limitations.
6.1 Introduction
Nonparametric methods do not require the normality assumption of classical
techniques. When the normality assumption is met, the ANOVA and t-test are
most powerful, in that if the alternative is true these methods will make the
correct decision with highest probability. However, if the normality assumption
is not met, results from the ANOVA and t-test can be misleading and too
liberal. I will describe and illustrate selected non-parametric methods,
and compare them with classical methods. Some motivation and discussion of
182 Ch 6: Nonparametric Methods
the strengths and weaknesses of non-parametric methods is given.
6.2 The Sign Test and CI for a Population

Median
The sign test assumes that you have a random sample from a population,
but makes no assumption about the population shape. The standard t-test
provides inferences on a population mean. The sign test, in contrast, provides
inferences about a population median.
If the population frequency curve is symmetric (see below), then the popu-
lation median, identified by η, and the population mean µ are identical. In this
case the sign procedures provide inferences for the population mean, though
less powerfully than the t-test.
The idea behind the sign test is straightforward. Suppose you have a sample
of size m from the population, and you wish to test H0 : η = η0 (a given value).
Let S be the number of sampled observations above η0. If H0 is true, you expect
S to be approximately one-half the sample size, 0.5m. If S is much greater than
0.5m, the data suggests that η > η0. If S is much less than 0.5m, the data
suggests that η < η0.
Mean and Median differ with skewed distributions Mean and Median are the same with symmetric distributions
50%
Median = η Mean = µ Mean = Median
S has a Binomial distribution when H0 is true. The Binomial distri-

6.2: The Sign Test and CI for a Population Median 183
bution is used to construct a test with size α (approximately). For a two-sided
alternative HA : η 6= η0, the test rejects H0 when S is significantly different
from 0.5m, as determined from the reference Binomial distribution. One-sided
tests use the corresponding lower or upper tail of the distribution. To generate
a CI for η, you can exploit the duality between CI and tests. A 100(1 − α)%
CI for η consists of all values η0 not rejected by a two-sided size α test of
H0 : η = η 0 .
Not all test sizes and confidence levels are possible because the test statistic
S is discrete valued. R’s SIGN.test() in the BSDA package gives an exact p-
value for the test, and approximates the desired confidence level using a linear
interpolation algorithm.
Example: Income Data Recall that the income distribution is extremely

skewed, with two extreme outliers at 46 and 1110.
#### Example: Income Data
income <- c(7, 1110, 7, 5, 8, 12, 0, 5, 2, 2, 46, 7)
income <- sort(income, decreasing = TRUE)
income
## [1] 1110 46 12 8 7 7 7 5 5 2 2 0
summary(income)
## 0.0 4.2 7.0 101.0 9.0 1110.0
sd(income)
## [1] 318
The income data is unimodal, skewed right, with two extreme outliers.
par(mfrow=c(3,1))
hist(income, freq = FALSE, breaks = 1000)
points(density(income), type = "l")
rug(income)
# violin plot
library(vioplot)
vioplot(income, horizontal=TRUE, col="gray")
# boxplot
boxplot(income, horizontal=TRUE)
Histogram of income
Density
0.15
0.00
0 200 400 600 800 1000
income
●
1
0 200 400 600 800 1000
● ●
0 200 400 600 800 1000
The normal QQ-plot of the sample data indicates strong deviation from
normality, and the CLT can’t save us: even the bootstrap sampling distribution
of the mean indicates strong deviation from normality.
library(car)
qqPlot(income, las = 1, id.n = 0, id.cex = 1, lwd = 1, main="QQ Plot, Income")
bs.one.samp.dist(income)
QQ Plot, Income Plot of data with smoothed density curve

0.004
●
Density
1000
0.002
800
0.000
income
600 0 200 400 600 800 1000 1200
dat
400
0.010
Density
200
●
● ● ● ● ● ● ●
0 ● ● ●
0.000
−1.5 −1.0 −0.5 0.0 0.5 1.0 1.5

0 100 200 300 400 500 600
norm quantiles
Data: n = 12 , mean = 100.92 , se = 91.801 5
The presence of the outliers has a dramatic effect on the 95% CI for the
population mean income µ, which goes from −101 to 303 (in 1000 dollar units).
This t-CI is suspect because the normality assumption is unreasonable. A CI
for the population median income η is more sensible because the median is likely
to be a more reasonable measure of typical value. Using the sign procedure,
you are 95% confident that the population median income is between 2.32 and
11.57 (times $1000).
library(BSDA)
t.test(income)
##
##
## data: income
## t = 1.099, df = 11, p-value = 0.2951
## -101.1 303.0
## mean of x
## 100.9
SIGN.test(income)
##
## One-sample Sign-Test
##
## data: income
## s = 11, p-value = 0.0009766
## alternative hypothesis: true median is not equal to 0
## 2.319 11.575
## median of x
## 7
## Conf.Level L.E.pt U.E.pt
## Lower Achieved CI 0.8540 5.000 8.00
## Interpolated CI 0.9500 2.319 11.57
## Upper Achieved CI 0.9614 2.000 12.00
Example: Age at First Heart Transplant Recall that the distribution

of ages is skewed to the left with a lower outlier. A question of interest is whether
the “typical age” at first transplant is 50. This can be formulated as a test about
the population median η or as a test about the population mean µ, depending
on the interpretation.
#### Example: Age at First Heart Transplant
age <- c(54, 42, 51, 54, 49, 56, 33, 58, 54, 64, 49)
age <- sort(age, decreasing = TRUE)
age
## [1] 64 58 56 54 54 54 51 49 49 42 33
summary(age)
## 33.0 49.0 54.0 51.3 55.0 64.0
sd(age)
## [1] 8.259
The age data is unimodal, skewed left, no extreme outliers.

par(mfrow=c(3,1))
hist(age, freq = FALSE, breaks = 10)
points(density(age), type = "l")
rug(age)
# violin plot
library(vioplot)
vioplot(age, horizontal=TRUE, col="gray")
# boxplot
boxplot(age, horizontal=TRUE)
Histogram of age
Density
0.04
0.00
30 35 40 45 50 55 60 65
age
●
1
35 40 45 50 55 60 65
35 40 45 50 55 60 65
The normal QQ-plot of the sample data indicates mild deviation from nor-
mality in the left tail (2 points of 11 outside the bands), and the bootstrap
sampling distribution of the mean indicates weak deviation from normality. It
is good practice in this case to use the nonparametric test as a double-check of
the t-test, with the nonparametric test being the more conservative test.
library(car)
qqPlot(age, las = 1, id.n = 0, id.cex = 1, lwd = 1, main="QQ Plot, Income")
bs.one.samp.dist(age)
65
●
Density
0.04
60
●
●
55
● ● ●
0.00
●
50 30 35 40 45 50 55 60 65
age
● ●
dat
45
0.15
●
40
0.10
Density
0.05
35
●
0.00
−1.5 −1.0 −0.5 0.0 0.5 1.0 1.5
40 45 50 55 60
norm quantiles
Data: n = 11 , mean = 51.273 , se = 2.49031 5
The sign test for H0 : η = 50 against HA : η 6= 50 has a p-value of 0.549,

which is not sufficient to reject H0. A 95% CI for η is 47.0 to 56.6 years, which
includes the hypothesized median age of 50. Similar conclusions are reached
with the t-CI and the test on µ, but you should have less confidence in these
results because the normality assumption is tenuous.
library(BSDA)
t.test(age, mu=50)
##
##
## data: age
## t = 0.5111, df = 10, p-value = 0.6204
## 45.72 56.82
## mean of x
## 51.27
SIGN.test(age, md=50)
##
##
## data: age
## s = 7, p-value = 0.5488
## 46.99 56.57
## median of x
## 54
## Conf.Level L.E.pt U.E.pt
## Lower Achieved CI 0.9346 49.00 56.00

## Interpolated CI 0.9500 46.99 56.57
## Upper Achieved CI 0.9883 42.00 58.00
6.3 Wilcoxon Signed-Rank Procedures

The Wilcoxon procedure assumes you have a random sample from a popula-
tion with a symmetric frequency curve. The curve need not be normal. The
test and CI can be viewed as procedures for either the population median or
mean.
To illustrate the computation of the Wilcoxon statistic W , suppose you
wish to test H0 : µ = µ0 = 10 on the made-up data below. The test statistic
requires us to compute the signs of Xi − µ0 and the ranks of |Xi − µ0|. Ties
in |Xi − µ0| get the average rank and observations at µ0 (here 10) are always
discarded. The Wilcoxon statistic is the sum of the signed ranks for
observations above µ0 = 10. For us
W = 6 + 4.5 + 8 + 2 + 4.5 + 7 = 32.

Xi Xi − 10 sign |Xi − 10| rank sign× rank
20 10 + 10 6 6
18 8 + 8 4.5 4.5
23 13 + 13 8 8
5 −5 − 5 3 −3
14 4 + 4 2 2
8 −2 − 2 1 −1
18 8 + 8 4.5 4.5
22 12 + 12 7 7
The sum of all ranks is always 0.5m(m+1), where m is the sample size. If H0
is true, you expect W to be approximately 0.5 × 0.5m(m + 1) = 0.25m(m + 1).
Why? Recall that W adds up the ranks for observations above µ0. If H0 is
true, you expect 1/2 of all observations to be above µ0, assuming the population
distribution is symmetric. The ranks of observations above µ0 should add
to approximately 1/2 times the sum of all ranks. You reject H0 in favor of
6.3: Wilcoxon Signed-Rank Procedures 189
HA : µ 6= µ0 if W is much larger than, or much smaller than 0.25m(m + 1).
One sided tests can also be constructed. The Wilcoxon CI for µ is computed
in a manner analogous to that described for the sign CI.
Here, m = 8 so the sum of all ranks is 0.5 × 8 × 9 = 36 (check yourself).
The expected value of W is 0.5 × 0.5 × 8 × 9 = 18. Is the observed value of
W = 32 far from the expected value of 18? To formally answer this question,
we need to use the Wilcoxon procedures, which are implemented in R with
wilcox.test().
Example: Made-up Data The boxplot indicates that the distribution is

fairly symmetric, so the Wilcoxon method is reasonable (so is a t-CI and test).
#### Example: Made-up Data
dat <- c(20, 18, 23, 5, 14, 8, 18, 22)
dat <- sort(dat, decreasing = TRUE)
dat
## [1] 23 22 20 18 18 14 8 5
summary(dat)
## 5.0 12.5 18.0 16.0 20.5 23.0
sd(dat)
## [1] 6.525
The dat data is unimodal, skewed left, no extreme outliers.

par(mfrow=c(3,1))
hist(dat, freq = FALSE, breaks = 10)
rug(dat)
# violin plot
library(vioplot)
vioplot(dat, horizontal=TRUE, col="gray")
# boxplot
boxplot(dat, horizontal=TRUE)
Histogram of dat
0.12
Density
0.06
0.00
5 10 15 20
dat
1
5 10 15 20
5 10 15 20
The normal QQ-plot of the sample data indicates insufficient evidence of

deviation from normality though both the QQ-plot and the bootstrap sampling
distribution of the mean indicates weak left-skewness. Either the Wilcoxon or
t-test are appropriate.
par(mfrow=c(1,1))
library(car)
qqPlot(dat, las = 1, id.n = 0, id.cex = 1, lwd = 1, main="QQ Plot, Income")
bs.one.samp.dist(dat)
●
Density
0.04
20 ●
● ●
0.00
15 5 10 15 20 25
dat
dat
10
Density
0.10
5 ●
0.00
−1.5 −1.0 −0.5 0.0 0.5 1.0 1.5

8 10 12 14 16 18 20 22
norm quantiles
Data: n = 8 , mean = 16 , se = 2.30682 5
The Wilcoxon p-value with continuity correction for testing H0 : µ = 10

against a two-sided alternative is 0.058. This would not lead to rejecting H0 at
the 5% level.
t.test(dat, mu=10)
##
##
## data: dat
## t = 2.601, df = 7, p-value = 0.03537
## 10.55 21.45
## mean of x
## 16
# with continuity correction in the normal approximation for the p-value
wilcox.test(dat, mu=10, conf.int=TRUE)
## Warning: cannot compute exact p-value with ties
## Warning: cannot compute exact confidence interval with ties
##
## Wilcoxon signed rank test with continuity correction
##
## data: dat
## V = 32, p-value = 0.0584
## alternative hypothesis: true location is not equal to 10
## 9.5 21.5
## (pseudo)median
## 16.01
# without continuity correction
wilcox.test(dat, mu=10, conf.int=TRUE, correct=FALSE)
##
## Wilcoxon signed rank test
##
## data: dat
## V = 32, p-value = 0.04967
## 11 21
## (pseudo)median
## 16.01
6.3.1 Nonparametric Analyses of Paired Data
Nonparametric methods for single samples can be used to analyze paired data
because the difference between responses within pairs is the unit of analysis.
Example: Sleep Remedies I will illustrate Wilcoxon methods on the

paired comparison of two remedies A and B for insomnia. The number of hours
of sleep gained on each method was recorded.
#### Example: Sleep Remedies
a <- c( 0.7, -1.6, -0.2, -1.2, 0.1, 3.4, 3.7, 0.8, 0.0, 2.0)
b <- c( 1.9, 0.8, 1.1, 0.1, -0.1, 4.4, 5.5, 1.6, 4.6, 3.0)
d <- b - a;
sleep <- data.frame(a, b, d)
summary(sleep$d)
## -0.20 1.00 1.25 1.52 1.67 4.60
shapiro.test(sleep$d)
##
##
## data: sleep$d
## W = 0.838, p-value = 0.04173
# boxplot
library(ggplot2)
p3 <- p3 + geom_point()
p3 <- p3 + stat_summary(fun.y = mean, geom = "point", shape = 18,
size = 4, alpha = 0.3)
print(p3)
"d"
d ● ● ● ● ● ● ● ●
0 1 2 3 4
d
The boxplot shows that distribution of differences is reasonably symmetric

but not normal. Recall that the Shapiro-Wilk test of normality was significant
at the 5% level (p-value=0.042). It is sensible to use the Wilcoxon procedure
on the differences. Let µB be the population mean sleep gain on remedy B,
and µA be the population mean sleep gain on remedy A. You are 95% confident
that µB − µA is between 0.8 and 2.8 hours. Putting this another way, you are
95% confident that µB exceeds µA by between 0.8 and 2.8 hours. The p-value
for testing H0 : µB − µA = 0 against a two-sided alternative is 0.008, which
strongly suggests that µB 6= µA. This agrees with the CI. Note that the t-CI
and test give qualitatively similar conclusions as the Wilcoxon methods, but
the t-test p-value is about half as large.
If you are uncomfortable with the symmetry assumption, you could use the
sign CI for the population median difference between B and A. I will note that
a 95% CI for the median difference goes from 0.86 to 2.2 hours.
t.test(sleep$d, mu=0)
##
##
## data: sleep$d
## t = 3.78, df = 9, p-value = 0.004352
## 0.6102 2.4298
## mean of x
## 1.52
wilcox.test(sleep$d, mu=0, conf.int=TRUE)
##
## Wilcoxon signed rank test with continuity correction
##
## data: sleep$d
## V = 54, p-value = 0.008004
## 0.7999 2.8000
## (pseudo)median
## 1.3
# can use the paired= option
#wilcox.test(sleep£b, sleep£a, paired=TRUE, mu=0, conf.int=TRUE)
# if don't assume symmetry, can use sign test
#SIGN.test(sleep£d)
6.3.2 Comments on One-Sample Nonparametric Meth-
ods
For this discussion, I will assume that the underlying population distribution
is (approximately) symmetric, which implies that population means and me-
dians are equal (approximately). For symmetric distributions the t, sign, and
Wilcoxon procedures are all appropriate.
If the underlying population distribution is extremely skewed, you can use
the sign procedure to get a CI for the population median. Alternatively, as
illustrated on HW 2, you can transform the data to a scale where the underlying
distribution is nearly normal, and then use the classical t-methods. Moderate
degrees of skewness will not likely have a big impact on the standard t-test and
CI.
The one-sample t-test and CI are optimal when the underlying population
frequency curve is normal. Essentially this means that the t-CI is, on average,
narrowest among all CI procedures with given level, or that the t-test has the
highest power among all tests with a given size. The width of a CI provides a
measure of the sensitivity of the estimation method. For a given level CI, the
narrower CI better pinpoints the unknown population mean.
With heavy-tailed symmetric distributions, the t-test and CI tend to be
conservative. Thus, for example, a nominal 95% t-CI has actual coverage rates
higher than 95%, and the nominal 5% t-test has an actual size smaller than 5%.
The t-test and CI possess a property that is commonly called robustness of
validity. However, data from heavy-tailed distributions can have a profound
effect on the sensitivity of the t-test and CI. Outliers can dramatically inflate
the standard error of the mean, causing the CI to be needlessly wide, and
tests to have diminished power (outliers typically inflate p-values for the t-
test). The sign and Wilcoxon procedures downweight the influence of outliers
by looking at sign or signed-ranks instead of the actual data values. These
two nonparametric methods are somewhat less efficient than the t-methods
when the population is normal (efficiency is about 0.64 and 0.96 for the sign
and Wilcoxon methods relative to the normal t-methods, where efficiency is the
ratio of sample sizes needed for equal power), but can be infinitely more efficient
with heavier than normal tailed distributions. In essence, the t-methods do not
6.4: (Wilcoxon-)Mann-Whitney Two-Sample Procedure 195
have a robustness of sensitivity.
Nonparametric methods have gained widespread acceptance in many sci-
entific disciplines, but not all. Scientists in some disciplines continue to use
classical t-methods because they believe that the methods are robust to non-
normality. As noted above, this is a robustness of validity, not sensitivity. This
misconception is unfortunate, and results in the routine use of methods that
are less powerful than the non-parametric techniques. Scientists need to
be flexible and adapt their tools to the problem at hand, rather
than use the same tool indiscriminately! I have run into suspicion
that use of nonparametric methods was an attempt to “cheat” in some way —
properly applied, they are excellent tools that should be used.
A minor weakness of nonparametric methods is that they do not easily
generalize to complex modelling problems. A great deal of progress has been
made in this area, but most software packages have not included the more
advanced techniques (R is among the forerunners).
Nonparametric statistics used to refer almost exclusively to the set of meth-
ods such as we have been discussing that provided analogs like tests and CIs
to the normal theory methods without requiring the assumption of sampling
from normal distributions. There is now a large area of statistics also called
nonparametric methods not focused on these goals at all. In our department
we (used to) have a course titled “Nonparametric Curve Estimation & Image
Reconstruction”, where the focus is much more general than relaxing an as-
sumption of normality. In that sense, what we are covering in this course could
be considered “classical” nonparametrics.
6.4 (Wilcoxon-)Mann-Whitney Two-Sample

Procedure
The WMW procedure assumes you have independent random samples from the
two populations, and assumes that the populations have the same shapes
and spreads (the frequency curves for the two populations are “shifted” ver-
sions of each other — see below). The frequency curves are not required to be
symmetric. The WMW procedures give a CI and tests on the difference η1 − η2
between the two population medians. If the populations are symmetric, then
the methods apply to µ1 − µ2.
0.20
0.15
0.15
0.10
0.10
0.05
0.05
0.0
0.0
-5 0 5 10 15 20 0 5 10 15
The R help on ?wilcox.test gives references to how the exact WMW proce-
dure is actually calculated; here is a good approximation to the exact method
that is easier to understand. The WMW procedure is based on ranks. The
two samples are combined, ranked from smallest to largest (1=smallest) and
separated back into the original samples. If the two populations have equal me-
dians, you expect the average rank in the two samples to be roughly equal. The
WMW test computes a classical two sample t-test using the pooled variance on
the ranks to assess whether the sample mean ranks are significantly different.
Example: Comparison of Cooling Rates of Uwet and Walker

Co. Meteorites The Uwet1 (Cross River, Nigeria) and Walker2 County
(Alabama, US) meteorite cooling rate data are below. A primary interest is
comparing the population “typical” cooling rate measurements.
1
2
#### Example: Comparison of Cooling Rates of Uwet and Walker Co. Meteorites
Uwet <- c(0.21, 0.25, 0.16, 0.23, 0.47, 1.20, 0.29, 1.10, 0.16)
Walker <- c(0.69, 0.23, 0.10, 0.03, 0.56, 0.10, 0.01, 0.02, 0.04, 0.22)
The boxplots and normal QQ-plots show that the distributions are rather
skewed to the right. The AD test of normality indicate that a normality as-
sumption is unreasonable for each population.
met <- data.frame(Uwet=c(Uwet,NA), Walker)
library(reshape2)
met.long <- melt(met, variable.name = "site", value.name = "cool", na.rm=TRUE)
## Using as id variables
names(met.long) <- c("site", "cool")
library(ggplot2)
p <- ggplot(met.long, aes(x = site, y = cool, fill=site))
p <- p + labs(title = "Cooling rates for samples of meteorites at two locations")
p <- p + theme(legend.position="none")
print(p)
par(mfrow=c(1,2))
library(car)
qqPlot(Walker, las = 1, id.n = 0, id.cex = 1, lwd = 1, main="QQ Plot, Walker")
qqPlot(Uwet, las = 1, id.n = 0, id.cex = 1, lwd = 1, main="QQ Plot, Uwet")
Cooling rates for samples of meteorites at two locations QQ Plot, Walker QQ Plot, Uwet
0.7 ● 1.2 ●
●
0.6
● 1.0
Walker ● ●
0.5
0.8
0.4
Walker
Uwet
site
0.3 0.6
●
● ●
0.2
0.4
Uwet ● ●
0.1 ● ● ●
●
●
● ●
●
● 0.2
● ● ●
0.0
−1.5 −0.5 0.5 1.5 −1.5 −0.5 0.5 1.5
0.00 0.25 0.50 0.75 1.00 1.25 norm quantiles norm quantiles
cool
I carried out the standard two-sample procedures to see what happens. The
pooled-variance and Satterthwaithe results are comparable, which is expected
because the sample standard deviations and sample sizes are roughly equal.
Both tests indicate that the mean cooling rates for Uwet and Walker Co. me-
teorites are not significantly different at the 10% level. You are 95% confident
that the mean cooling rate for Uwet is at most 0.1 less, and no more than 0.6
greater than that for Walker Co. (in degrees per million years).
summary(Uwet)
## 0.160 0.210 0.250 0.452 0.470 1.200
c(sd(Uwet), IQR(Uwet), length(Uwet))
## [1] 0.407 0.260 9.000
summary(Walker)
## 0.0100 0.0325 0.1000 0.2000 0.2280 0.6900
c(sd(Walker), IQR(Walker), length(Walker))
## [1] 0.239 0.195 10.000
t.test(Uwet, Walker, var.equal = TRUE)
##
##
## data: Uwet and Walker
## t = 1.669, df = 17, p-value = 0.1134
## -0.06663 0.57107
## 0.4522 0.2000
t.test(Uwet, Walker)
##
##
## t = 1.624, df = 12.65, p-value = 0.129
## -0.08421 0.58865
## 0.4522 0.2000
Given the marked skewness, a nonparametric procedure is more appropriate.

The Wilcoxon-Mann-Whitney comparison of population medians is reasonable.
Why? The WMW test of equal population medians is significant (barely) at
the 5% level. You are 95% confident that median cooling rate for Uwet exceeds
that for Walker by between 0+ and 0.45 degrees per million years.
wilcox.test(Uwet, Walker, conf.int = TRUE)
## Warning: cannot compute exact confidence intervals with ties
##
## Wilcoxon rank sum test with continuity correction
##
## W = 69.5, p-value = 0.04974
## alternative hypothesis: true location shift is not equal to 0
## 4.497e-05 4.500e-01
## difference in location
## 0.1703
The difference between the WMW and t-test p-values and CI lengths (i.e.
the WMW CI is narrower and the p-value smaller) reflects the effect of the
outliers on the sensitivity of the standard tests and CI.
I conducted a pooled-variance two-sample t-test on ranks to show you that

the p-value is close to the WMW p-value, as expected.
rank(met.long$cool)
## [1] 9.0 13.0 7.5 11.5 15.0 19.0 14.0 18.0 7.5 17.0 11.5 5.5 3.0
## [14] 16.0 5.5 1.0 2.0 4.0 10.0
by(rank(met.long$cool), met.long$site, summary)
## met.long$site: Uwet
## 7.5 9.0 13.0 12.7 15.0 19.0
## ----------------------------------------------------
## met.long$site: Walker
## 1.00 3.25 5.50 7.55 11.10 17.00
# note: the CI for ranks is not interpretable
t.test(rank(met.long$cool) ~ met.long$site, var.equal = TRUE)
##
##
## data: rank(met.long$cool) by met.long$site
## t = 2.208, df = 17, p-value = 0.04125
## 0.2305 10.1140
## mean in group Uwet mean in group Walker
## 12.72 7.55
Example: Newcombe’s Data Experiments of historical importance were
performed beginning in the eighteenth century to determine physical constants,
such as the mean density of the earth, the distance from the earth to the sun,
and the velocity of light. An interesting series of experiments to determine
the velocity of light was begun in 1875. The first method used, and reused
with refinements several times thereafter, was the rotating mirror method3. In
this method a beam of light is reflected off a rapidly rotating mirror to a fixed
mirror at a carefully measured distance from the source. The returning light
is re-reflected from the rotating mirror at a different angle, because the mir-
ror has turned slightly during the passage of the corresponding light pulses.
From the speed of rotation of the mirror and from careful measurements of
the angular difference between the outward-bound and returning light beams,
the passage time of light can be calculated for the given distance. After av-
eraging several calculations and applying various corrections, the experimenter
can combine mean passage time and distance for a determination of the veloc-
ity of light. Simon Newcombe, a distinguished American scientist, used this
method during the year 1882 to generate the passage time measurements given
below, in microseconds. The travel path for this experiment was 3721 meters
in length, extending from Ft. Meyer, on the west bank of the Potomac River in
Washington, D.C., to a fixed mirror at the base of the Washington Monument.
3
http://en.wikipedia.org/wiki/File:Speed_of_light_(foucault).PNG
The problem is to determine a 95% CI for the “true” passage time, which
is taken to be the typical time (mean or median) of the population of measure-
ments that were or could have been taken by this experiment.
#### Example: Newcombe's Data
time <- c(24.828, 24.833, 24.834, 24.826, 24.824, 24.756
, 24.827, 24.840, 24.829, 24.816, 24.798, 24.822
, 24.824, 24.825, 24.823, 24.821, 24.830, 24.829
, 24.831, 24.824, 24.836, 24.819, 24.820, 24.832
, 24.836, 24.825, 24.828, 24.828, 24.821, 24.829
, 24.837, 24.828, 24.830, 24.825, 24.826, 24.832
, 24.836, 24.830, 24.836, 24.826, 24.822, 24.823
, 24.827, 24.828, 24.831, 24.827, 24.827, 24.827
, 24.826, 24.826, 24.832, 24.833, 24.832, 24.824
, 24.839, 24.824, 24.832, 24.828, 24.825, 24.825
, 24.829, 24.828, 24.816, 24.827, 24.829, 24.823)
library(nortest)
ad.test(time)
##
##
## data: time
## A = 5.884, p-value = 1.217e-14
Passage_df <- data.frame(time)
p1 <- ggplot(Passage_df, aes(x = time))
, binwidth=0.001

p1 <- p1 + geom_point(aes(y = -1)
, position = position_jitter(height = .5)
, alpha = 1/5)
# violin plot
p2 <- ggplot(Passage_df, aes(x = "t", y = time))
# boxplot
p3 <- ggplot(Passage_df, aes(x = "t", y = time))
library(gridExtra)
## Warning: position stack requires constant width: output may be incorrect
90
density
60
30
0
24.77 24.80 24.82
time
"t"
t ● ●
24.77 24.80 24.82

time
"t"
t ● ●
24.77 24.80 24.82

time
par(mfrow=c(1,1))
library(car)
qqPlot(time, las = 1, id.n = 0, id.cex = 1, lwd = 1, main="QQ Plot, Time")
bs.one.samp.dist(time)
QQ Plot, Time Plot of data with smoothed density curve
60
24.84 ● ●
●● ● ● ●
●
40
Density
●●●●●●●●●
●●●●●●●●
●●●●●●●●●●●●●
●●●●●●●●●●●●●●●
●●●●●
20
24.82 ● ●●●
● ●
0
24.76 24.78 24.80 24.82 24.84
time
24.80 ●
dat
300
24.78
200
Density
100
24.76
●
0
−2 −1 0 1 2
24.820 24.822 24.824 24.826 24.828 24.830
norm quantiles
Data: n = 66 , mean = 24.826 , se = 0.00132266 5
The data set is skewed to the left, due to the presence of two extreme
outliers that could potentially be misrecorded observations. Without additional
information I would be hesitant to apply normal theory methods (the t-test),
even though the sample size is “large” (bootstrap sampling distribution is still
left-skewed). Furthermore, the t-test still suffers from a lack of robustness of
sensitivity, even in large samples. A formal QQ-plot and normal test rejects, at
the 0.01 level, the normality assumption needed for the standard methods.
The table below gives 95% t, sign, and Wilcoxon CIs. I am more comfortable
with the sign CI for the population median than the Wilcoxon method, which
assumes symmetry.
t.sum <- t.test(time)
t.sum$conf.int
## [1] 24.82 24.83
## [1] 0.95
diff(t.test(time)$conf.int)
## [1] 0.005283
s.sum <- SIGN.test(time)
##
##
## data: time
## s = 66, p-value < 2.2e-16
## 24.83 24.83
## median of x
## 24.83
s.sum[2,c(2,3)]
## L.E.pt U.E.pt
## 24.83 24.83
diff(s.sum[2,c(2,3)])
## U.E.pt
## 0.0025
w.sum <- wilcox.test(time, conf.int=TRUE)
w.sum$conf.int
## [1] 24.83 24.83
## [1] 0.95
diff(w.sum$conf.int)
## [1] 0.002488
parameter Method CI Limits Width

mean t (24.8236, 24.8289) 0.0053
median sign (24.8260, 24.8285) 0.0025
median Wilcoxon (24.8260, 24.8285) 0.0025
Note the big difference between the nonparametric and the t-CI. The nonpara-
metric CIs are about 1/2 as wide as the t-CI. This reflects the impact that
outliers have on the standard deviation, which directly influences the CI width.
6.5 Alternatives for ANOVA and Planned

Comparisons
The classical ANOVA assumes that the populations have normal frequency
curves and the populations have equal variances (or spreads). You learned
formal tests for these assumptions in Chapter 5. When the assumptions do
not hold, you can try one of the following two approaches. Before describing
alternative methods, I will note that deviations from normality in one or more
samples might be expected in a comparison involving many samples. You
should downplay small deviations from normality in problems involving many
samples.
6.5: Alternatives for ANOVA and Planned Comparisons 205
6.5.1 Kruskal-Wallis ANOVA
The Kruskal-Wallis (KW) test is a non-parametric method for testing the
hypothesis of equal population medians against the alternative that not all pop-
ulation medians are equal. The procedure assumes you have independent ran-
dom samples from populations with frequency curves having identical shapes
and spreads. The KW ANOVA is essentially the standard ANOVA based
on ranked data. That is, we combine the samples, rank the observations from
smallest to largest, and then return the ranks to the original samples and do
the standard ANOVA using the ranks. The KW ANOVA is a multiple sample
analog of the Wilcoxon-Mann-Whitney two sample procedure. Hence, multiple
comparisons for a KW analysis, be they FSD or Bonferroni comparisons, are
based on the two sample WMW procedure.
6.5.2 Transforming Data

The distributions in many data sets are skewed to the right with outliers. If the
sample spreads, say s and IQR, increase with an increasing mean or median, you
can often transform data to a scale where the normality and the constant
spread assumption are more nearly satisfied. The transformed data are analyzed
using the standard ANOVA. The two most commonly used transforms for this
problem are the square root and natural logarithm, provided the data are non-
negative4.
If the original distributions are nearly symmetric, but heavy-tailed, non-
linear transformations will tend to destroy the symmetry. Many statisticians
recommend methods based on trimmed means for such data. These methods
are not commonly used by other researchers.
4
The aim behind the choice of a variance-stabilizing transformation is to find a simple function f
to apply to values y in a data set to create new values y 0 = f (y) such that the variability of the values y 0
is not related to their mean value. For example, suppose that the values y are realizations from a Poisson
distribution. Because for the Poisson distribution the variance is identical to the mean, the variance varies
√
with the mean. However, if the simple variance-stabilizing transformation y 0 = y is applied, the sampling
variance will be independent of the mean. A few distributional examples are provided in the table below.
Distribution Variance=g(mean) Transformation y 0 = f (y)
√
Poisson σ2 = µ y0 = y p
binomial σ 2 = µ(1 − µ) y 0 = arcsin( (y))
lognormal σ 2 = µ2 y 0 = log(y)
Example: Hydrocarbon (HC) Emissions Data These data are the
HC emissions at idling speed, in ppm, for automobiles of different years of
manufacture. The data are a random sample of all automobiles tested at an
Albuquerque shopping center. (It looks like we need to find some newer cars!)
#### Example: Hydrocarbon (HC) Emissions Data
emis <- read.table(text="
Pre-y63 y63-7 y68-9 y70-1 y72-4
2351 620 1088 141 140
1293 940 388 359 160
541 350 111 247 20
1058 700 558 940 20
411 1150 294 882 223
570 2000 211 494 60
800 823 460 306 20
630 1058 470 200 95
905 423 353 100 360
347 900 71 300 70
NA 405 241 223 220
NA 780 2999 190 400
NA 270 199 140 217
NA NA 188 880 58
NA NA 353 200 235
NA NA 117 223 1880
NA NA NA 188 200
NA NA NA 435 175
NA NA NA 940 85
NA NA NA 241 NA
", header=TRUE)
#emis
# convert to long format

emis.long <- melt(emis,
variable.name = "year",
value.name = "hc",
na.rm = TRUE
)
names(emis.long) <- c("year", "hc")
# summary of each year
by(emis.long$hc, emis.long$year, summary)
## emis.long$year: Pre.y63
## 347 548 715 891 1020 2350
## ----------------------------------------------------
## emis.long$year: y63.7
## 270 423 780 801 940 2000
## ----------------------------------------------------
## 71 196 324 506 462 3000
## ----------------------------------------------------
## 100 198 244 381 450 940
## ----------------------------------------------------
## 20 65 160 244 222 1880
# IQR and sd of each year
by(emis.long$hc, emis.long$year, function(X) { c(IQR(X), sd(X), length(X)) })
## [1] 471.5 591.6 10.0
## ----------------------------------------------------
## [1] 517.0 454.9 13.0
## ----------------------------------------------------
## [1] 266.2 707.8 16.0
## ----------------------------------------------------
## [1] 252.2 287.9 20.0
## ----------------------------------------------------
## [1] 156.5 410.8 19.0
library(ggplot2)
p <- ggplot(emis.long, aes(x = year, y = hc))
p <- p + geom_hline(yintercept = mean(emis.long$hc),
aes(colour=year), alpha = 0.8)
width = .2, aes(colour=year), alpha = 0.8)
p <- p + labs(title = "Albuquerque automobile hydrocarbon emissions data") + ylab("hc (ppm)")
# to reverse order that years print, so oldest is first on top
p <- p + scale_x_discrete(limits = rev(levels(emis.long$year)) )
print(p)
Albuquerque automobile hydrocarbon emissions data
Pre.y63 ●
y63.7 ●
year
y68.9 ● ●
y70.1 ● ●
y72.4 ●
0 1000 2000 3000

hc (ppm)
The standard ANOVA shows significant differences among the mean HC

emissions. However, the standard ANOVA is inappropriate because the distri-
butions are extremely skewed to the right due to presence of outliers in each
sample.
fit.e <- aov(hc ~ year, data = emis.long)
summary(fit.e)
## year 4 4226834 1056709 4.34 0.0033 **
## Residuals 73 17759968 243287
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
fit.e
## Call:
## aov(formula = hc ~ year, data = emis.long)
##
## Terms:
## year Residuals
## Sum of Squares 4226834 17759968
##
The boxplots show that the typical HC emissions appear to decrease as

the age of car increases (the simplest description). Although the spread in the
samples, as measured by the IQR, also decreases as age increases, I am more
comfortable with the KW ANOVA, in part because the KW analysis is not too
sensitive to differences in spreads among samples. This point is elaborated upon
later. As described earlier, the KW ANOVA is essentially an ANOVA based
on the ranks. I give below the ANOVA based on ranks and the output from
the KW procedure. They give similar p-values, and lead to the conclusion that
there are significant differences among the population median HC emissions.
A simple description is that the population median emission tends to decrease
with the age of the car. You should follow up this analysis with Mann-Whitney
multiple comparisons.
# ANOVA of rank, for illustration that this is similar to what KW is doing
fit.er <- aov(rank(hc) ~ year, data = emis.long)
summary(fit.er)
## year 4 16329 4082 12.8 5.7e-08 ***
## Residuals 73 23200 318
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
fit.er
## Call:
## aov(formula = rank(hc) ~ year, data = emis.long)
##
## Terms:
## year Residuals
## Sum of Squares 16329 23200
##
# KW ANOVA
fit.ek <- kruskal.test(hc ~ year, data = emis.long)
fit.ek
##
## Kruskal-Wallis rank sum test
##
## data: hc by year
## Kruskal-Wallis chi-squared = 31.81, df = 4, p-value =
## 2.093e-06
It is common to transform the data to a log scale when the spread increases
as the median or mean increases.
# log scale
emis.long$loghc <- log(emis.long$hc)
# summary of each year
by(emis.long$loghc, emis.long$year, summary)

## 5.85 6.31 6.57 6.63 6.93 7.76
## ----------------------------------------------------
## 5.60 6.05 6.66 6.55 6.85 7.60
## ----------------------------------------------------
## 4.26 5.28 5.78 5.76 6.14 8.01
## ----------------------------------------------------
## 4.61 5.29 5.50 5.71 6.11 6.85
## ----------------------------------------------------
## 3.00 4.17 5.08 4.84 5.40 7.54
# IQR and sd of each year
by(emis.long$loghc, emis.long$year, function(X) { c(IQR(X), sd(X), length(X)) })
## [1] 0.6186 0.5702 10.0000
## ----------------------------------------------------
## [1] 0.7985 0.5525 13.0000
## ----------------------------------------------------
## [1] 0.8575 0.9062 16.0000
## ----------------------------------------------------
## [1] 0.8216 0.6776 20.0000
## ----------------------------------------------------
## [1] 1.229 1.139 19.000
library(ggplot2)
p <- ggplot(emis.long, aes(x = year, y = loghc))
p <- p + geom_hline(yintercept = mean(emis.long$loghc),
aes(colour=year), alpha = 0.8)
width = .2, aes(colour=year), alpha = 0.8)
p <- p + labs(title = "Albuquerque automobile hydrocarbon emissions data (log scale)")

p <- p + ylab("log(hc) (log(ppm))")
p <- p + scale_x_discrete(limits = rev(levels(emis.long$year)) )
print(p)
Albuquerque automobile hydrocarbon emissions data (log scale)
Pre.y63
y63.7
year
y68.9 ●
y70.1
y72.4 ●
3 4 5 6 7 8
log(hc) (log(ppm))
After transformation, the samples have roughly the same spread (IQR and
s) and shape. The transformation does not completely eliminate the outliers.
However, I am more comfortable with a standard ANOVA on this scale than
with the original data. A difficulty here is that the ANOVA is comparing
population mean log HC emission (so interpretations are on the log ppm scale,
instead of the natural ppm scale). Summaries for the ANOVA on the log
hydrocarbon emissions levels are given below.
# ANOVA of rank, for illustration that this is similar to what KW is doing
fit.le <- aov(loghc ~ year, data = emis.long)
summary(fit.le)
## year 4 31.9 7.97 11.4 3e-07 ***

## Residuals 73 51.0 0.70
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
fit.le
## Call:
## aov(formula = loghc ~ year, data = emis.long)
##
## Terms:
## year Residuals
##
# KW ANOVA -- same conclusions as original scale, since based on ranks
fit.lek <- kruskal.test(loghc ~ year, data = emis.long)
fit.lek
##
##
## data: loghc by year
## 2.093e-06
The boxplot of the log-transformed data reinforces the reasonableness of the

original KW analysis. Why? The log-transformed distributions have fairly sim-
ilar shapes and spreads, so a KW analysis on these data is sensible. The ranks
for the original and log-transformed data are identical, so the KW analyses on
the log-transformed data and the original data must lead to the same conclu-
sions. This suggests that the KW ANOVA is not overly sensitive to differences
in spreads among the samples.
There are two reasonable analyses here: the standard ANOVA using log HC
emissions, and the KW analysis of the original data. The first analysis gives a
comparison of mean log HC emissions. The second involves a comparison of me-
dian HC emissions. A statistician would present both analyses to the scientist
who collected the data to make a decision on which was more meaningful (inde-
pendently of the results5!). Multiple comparisons would be performed relative
to the selected analysis (t-tests for ANOVA or WMW-tests for KW ANOVA).
5
It is unethical to choose a method based on the results it gives.
Example: Hodgkin’s Disease Study Plasma bradykininogen levels
were measured in normal subjects, in patients with active Hodgkin’s disease,
and in patients with inactive Hodgkin’s disease. The globulin bradykininogen
is the precursor substance for bradykinin, which is thought to be a chemical
mediator of inflammation. The data (in micrograms of bradykininogen per
milliliter of plasma) are displayed below. The three samples are denoted by
nc for normal controls, ahd for active Hodgkin’s disease patients, and ihd for
inactive Hodgkin’s disease patients.
The medical investigators wanted to know if the three samples differed in
their bradykininogen levels. Carry out the statistical analysis you consider to
be most appropriate, and state your conclusions to this question.
Read in the data, look at summaries on the original scale, and create a plot.
Also, look at summaries on the log scale and create a plot.
#### Example: Hodgkin's Disease Study
hd <- read.table(text="
nc ahd ihd
5.37 3.96 5.37
5.80 3.04 10.60
4.70 5.28 5.02
5.70 3.40 14.30
3.40 4.10 9.90
8.60 3.61 4.27
7.48 6.16 5.75
5.77 3.22 5.03
7.15 7.48 5.74
6.49 3.87 7.85
4.09 4.27 6.82
5.94 4.05 7.90
6.38 2.40 8.36
9.24 5.81 5.72
5.66 4.29 6.00
4.53 2.77 4.75
6.51 4.40 5.83
7.00 NA 7.30
6.20 NA 7.52
7.04 NA 5.32
4.82 NA 6.05
6.73 NA 5.68
5.26 NA 7.57
NA NA 5.68
NA NA 8.91
NA NA 5.39
NA NA 4.40
NA NA 7.13
", header=TRUE)
#hd
# convert to long format

hd.long <- melt(hd,
variable.name = "patient",
value.name = "level",
na.rm = TRUE
)
names(hd.long) <- c("patient", "level")
# summary of each patient
by(hd.long$level, hd.long$patient, summary)
## hd.long$patient: nc
## 3.40 5.32 5.94 6.08 6.86 9.24
## ----------------------------------------------------
## hd.long$patient: ahd
## 2.40 3.40 4.05 4.24 4.40 7.48
## ----------------------------------------------------
## hd.long$patient: ihd
## 4.27 5.38 5.92 6.79 7.64 14.30
# IQR and sd of each patient
by(hd.long$level, hd.long$patient, function(X) { c(IQR(X), sd(X), length(X)) })
## [1] 1.550 1.362 23.000
## ----------------------------------------------------
## [1] 1.000 1.303 17.000
## ----------------------------------------------------
## [1] 2.255 2.176 28.000
# log scale
hd.long$loglevel <- log(hd.long$level)
# summary of each patient
by(hd.long$loglevel, hd.long$patient, summary)
## 1.22 1.67 1.78 1.78 1.93 2.22
## ----------------------------------------------------
## 0.875 1.220 1.400 1.400 1.480 2.010
## ----------------------------------------------------
## 1.45 1.68 1.78 1.87 2.03 2.66
# IQR and sd of each patient
by(hd.long$loglevel, hd.long$patient, function(X) { c(IQR(X), sd(X), length(X)) })
## [1] 0.2558 0.2303 23.0000
## ----------------------------------------------------
## [1] 0.2578 0.2921 17.0000
## ----------------------------------------------------
## [1] 0.3497 0.2803 28.0000

library(ggplot2)
p <- ggplot(hd.long, aes(x = patient, y = level))
p <- p + geom_hline(yintercept = mean(hd.long$level),
aes(colour=patient), alpha = 0.8)
width = .2, aes(colour=patient), alpha = 0.8)
p <- p + labs(title = "Plasma bradykininogen levels for three patient groups")
p <- p + ylab("level (mg/ml)")
p <- p + scale_x_discrete(limits = rev(levels(hd.long$patient)) )
p <- p + ylim(c(0,max(hd.long$level)))
print(p)
## log scale
library(ggplot2)
p <- ggplot(hd.long, aes(x = patient, y = loglevel))
p <- p + geom_hline(yintercept = mean(hd.long$loglevel),
aes(colour=patient), alpha = 0.8)
width = .2, aes(colour=patient), alpha = 0.8)
p <- p + labs(title = "Plasma bradykininogen levels for three patient groups (log scale)")
p <- p + ylab("log(level) (log(mg/ml))")
p <- p + scale_x_discrete(limits = rev(levels(hd.long$patient)) )
p <- p + ylim(c(0,max(hd.long$loglevel)))
print(p)
Plasma bradykininogen levels for three patient groups Plasma bradykininogen levels for three patient groups (log scale)
nc ● nc ●
patient
patient
ahd ● ● ahd ●
ihd ● ihd ●
0 5 10 15 0 1 2
level (mg/ml) log(level) (log(mg/ml))
Although the spread (IQR, s) in the ihd sample is somewhat greater than
the spread in the other samples, the presence of skewness and outliers in the
boxplots is a greater concern regarding the use of the classical ANOVA. The
shapes and spreads in the three samples are roughly identical, so a Kruskal-
Wallis nonparametric ANOVA appears ideal. As a sidelight, I transformed
plasma levels to a log scale to reduce the skewness and eliminate the outliers.
The boxplots of the transformed data show reasonable symmetry across groups,
but outliers are still present. I will stick with the Kruskal-Wallis ANOVA
(although it would not be much of a problem to use the classical ANOVA on
transformed data).
Let ηnc = population median plasma level for normal controls, ηahd = pop-
ulation median plasma level for active Hodgkin’s disease patients, and ηihd =
population median plasma level for inactive Hodgkin’s disease patients. The
KW test of H0 : ηnc = ηahd = ηihd versus HA : not H0 is highly significant (p-
value= 0.00003), suggesting differences among the population median plasma
levels. The Kruskal-Wallis ANOVA summary is given below.
# KW ANOVA
fit.h <- kruskal.test(level ~ patient, data = hd.long)
fit.h
##
##
## data: level by patient
## 3.421e-05
I followed up the KW ANOVA with Bonferroni comparisons of the samples,

using the Mann-Whitney two sample procedure. There are three comparisons,
so an overall FER of 0.05 is achieved by doing the individual tests at the
0.05/3=0.0167 level. Alternatively, you can use 98.33% CI for differences in
population medians.
wilcox.test(hd$nc , hd$ahd, conf.int=TRUE, conf.level = 0.9833)
##
##
## data: hd$nc and hd$ahd
## W = 329, p-value = 0.0002735
## 98.33 percent confidence interval:
## 0.8599 2.9001
## 1.91
wilcox.test(hd$nc , hd$ihd, conf.int=TRUE, conf.level = 0.9833)
##
##
## data: hd$nc and hd$ihd
## W = 276.5, p-value = 0.3943
## -1.56 0.68
## -0.3414
wilcox.test(hd$ahd, hd$ihd, conf.int=TRUE, conf.level = 0.9833)
##
##
## data: hd$ahd and hd$ihd
## W = 56, p-value = 2.143e-05
## -3.50 -1.32
## -2.147
The only comparison with a p-value greater than 0.0167 involved the nc
and ihd samples. The comparison leads to two groups, and is consistent with
what we see in the boxplots.
ahd nc ihd
--- --------
You have sufficient evidence to conclude that the plasma bradykininogen levels
for active Hodgkin’s disease patients (ahd) is lower than the population median
levels for normal controls (nc) and for patients with inactive Hodgkin’s disease
(ihd). You do not have sufficient evidence to conclude that the population
median levels for normal controls (nc) and for patients with inactive Hodgkin’s
disease (ihd) are different. The CIs give an indication of size of differences in
the population medians.
6.5.3 Planned Comparisons

Bonferroni multiple comparisons are generally preferred to Fisher’s least sig-
nificant difference approach. Fisher’s method does not control the familywise
error rate and produces too many spurious significant differences (claims of sig-
nificant differences that are due solely to chance variation and not to actual
differences in population means). However, Bonferroni’s method is usually very
conservative when a large number of comparisons is performed — large differ-
ences in sample means are needed to claim significance. A way to reduce this
conservatism is to avoid doing all possible comparisons. Instead, one should,
when possible, decide a priori (before looking at the data) which comparisons
are of primary interest, and then perform only those comparisons.
For example, suppose a medical study compares five new treatments with a
control (a six group problem). The medical investigator may not be interested
in all 15 possible comparisons, but only in which of the five treatments differ
on average from the control. Rather than performing the 15 comparisons, each
at the say 0.05/15 = 0.0033 level, she could examine the five comparisons of
interest at the 0.05/5 = 0.01 level. By deciding beforehand which comparisons
are of interest, she can justify using a 0.01 level for the comparisons, instead of
the more conservative 0.0033 level needed when doing all possible comparisons.
To illustrate this idea, consider the KW analysis of HC emissions. We
saw that there are significant differences among the population median HC
emissions. Given that the samples have a natural ordering
Sample Year of manufacture
1 Pre-1963
2 63 – 67
3 68 – 69
4 70 – 71
5 72 – 74
you may primarily be interested in whether the population medians for cars
manufactured in consecutive samples are identical. That is, you may be pri-
marily interested in the following 4 comparisons:
Pre-1963 vs 63 – 67
63 – 67 vs 68 – 69
68 – 69 vs 70 – 71
70 – 71 vs 72 – 74
A Bonferroni analysis would carry out each comparison at the 0.05/4 = 0.0125
level versus the 0.05/10 = 0.005 level when all comparisons are done.
The following output was obtained for doing these four comparisons, based
on Wilcoxon-Mann-Whitney two-sample tests (why?6). Two-year groups are
claimed to be different if the p-value is 0.0125 or below, or equivalently, if a
98.75% CI for the difference in population medians does not contain zero.
#### Planned Comparisons
wilcox.test(emis$y63.7, emis$Pre.y63, conf.int=TRUE, conf.level = 0.9875)
##
6
The ANOVA is the multi-sample analog to the two-sample t-test for the mean, and the KW ANOVA is
the multi-sample analog to the WMW two-sample test for the median. Thus, we follow up a KW ANOVA
with WMW two-sample tests at the chosen multiple comparison adjusted error rate.
##
## data: emis$y63.7 and emis$Pre.y63
## W = 61.5, p-value = 0.8524
## -530 428
## -15.48
wilcox.test(emis$y68.9, emis$y63.7 , conf.int=TRUE, conf.level = 0.9875)
##
##
## data: emis$y68.9 and emis$y63.7
## W = 43, p-value = 0.007968
## -709 -52
## -397.4
##
##
## W = 156, p-value = 0.9112
## -206 171
## -11
##
##
## W = 92.5, p-value = 0.006384
## -286 -6
6.6: Permutation tests 221
## -130
There are significant differences between the 1963-67 and 1968-69 samples,
and between the 1970-71 and 1972-74 samples. You are 98.75% confident that
the population median HC emissions for 1963-67 year cars is between 52 and
708.8 ppm greater than the population median for 1968-69 cars. Similarly, you
are 98.75% confident that the population median HC emissions for 1970-71 year
cars is between 6.1 and 285.9 ppm greater than the population median for 1972-
74 cars. Overall, you are 95% confident among the four pairwise comparisons
that you have not declared a difference significant when it isn’t.
6.5.4 Two final ANOVA comments

It is not uncommon for researchers to combine data from groups not found to
be significantly different. This is not, in general, a good practice. Just because
you do not have sufficient evidence to show differences does not imply that you
should treat the groups as if they are the same!
If the data distributions do not substantially deviate from normality, but the
spreads are different across samples, you might consider the standard ANOVA
followed with multiple comparisons using two-sample tests based on Satterth-
waite’s approximation.
6.6 Permutation tests

Permutation tests7 are a subset of non-parametric statistics. The basic premise
is to use only the assumption that it is possible that all of the treatment groups
are equivalent, and that every member of them is the same before sampling
began (i.e., the position in the group to which they belong is not differentiable
from other position before the positions are filled). From this, one can calculate
a statistic and then see to what extent this statistic is special by seeing how
likely it would be if the group assignments had been jumbled.
A permutation test (also called a randomization test, re-randomization test,
or an exact test) is a type of statistical significance test in which the distribution
7
http://en.wikipedia.org/wiki/Resampling_(statistics)
of the test statistic under the null hypothesis is obtained by calculating all
possible values of the test statistic under rearrangements of the labels on
the observed data points. In other words, the method by which treatments are
allocated to subjects in an experimental design is mirrored in the analysis of
that design. If the labels are exchangeable under the null hypothesis, then the
resulting tests yield exact significance levels. Confidence intervals can then be
derived from the tests. The theory has evolved from the works of R.A. Fisher
and E.J.G. Pitman in the 1930s.
Let’s illustrate the basic idea of a permutation test using the Meteorites
example. Suppose we have two groups Uwet and Walker whose sample means
are ȲU and ȲW, and that we want to test, at 5% significance level, whether they
come from the same distribution. Let nU = 9 and nW = 10 be the sample size
corresponding to each group. The permutation test is designed to determine
whether the observed difference between the sample means is large enough to
reject the null hypothesis H0 : µU = µW, that the two groups have identical
means.
The test proceeds as follows. First, the difference in means between the
two samples is calculated: this is the observed value of the test statistic, T(obs).
Then the observations of groups Uwet and Walker are pooled.
#### Permutation tests
# Calculated the observed difference in means
# met.long includes both Uwet and Walker groups
Tobs <- mean(met.long[(met.long$site == "Uwet" ), 2]) -
mean(met.long[(met.long$site == "Walker"), 2])
Tobs
## [1] 0.2522
Next, the difference in sample means is calculated and recorded for every
possible way of dividing these pooled values into two groups of size nU =
9 and nW = 10 (i.e., for every permutation of the group labels Uwet and
Walker). The set of these calculated differences is the exact distribution of
possible differences under the null hypothesis that group label does not matter.
This exact distribution can be approximated by drawing a large number of
random permutations.
# Plan:
# Initialize a vector in which to store the R number of difference of means.
# Calculate R differences in means for R permutations, storing the results.
# Note that there are prod(1:19) = 10^17 total permutations,
# but the R repetitions will serve as a good approximation.
# Plot the permutation null distribution with an indication of the Tobs.
# R = a large number of repetitions

R <- 1e4
# initialize the vector of difference of means from the permutations
Tperm <- rep(NA, R)
# For each of R repetitions, permute the Uwet and Walker labels,
# calculate the difference of means with the permuted labels,
# and store the result in the i.R'th position of Tperm.
for (i.R in 1:R) {
# permutation of 19 = 9+10 integers 1, 2, ..., 19
ind.perm <- sample.int(nrow(met.long))
# identify as "TRUE" numbers 1, ..., 9 (the number of Uwet labels)
lab.U <- (ind.perm <= sum(met.long$site == "Uwet")) #£
# identify as "TRUE" numbers 10, ..., 19 (the number of Walker labels)
# that is, all the non-Uwet labels
lab.W <- !lab.U
# calculate the difference in means and store in Tperm at index i.R

Tperm[i.R] <- mean(met.long[lab.U, 2]) - mean(met.long[lab.W, 2])
}
# Plot the permutation null distribution with an indication of the Tobs.

dat <- data.frame(Tperm)
library(ggplot2)
p <- ggplot(dat, aes(x = Tperm))
#p <- p + scale_x_continuous(limits=c(-20,+20))
p <- p + geom_histogram(aes(y=..density..)
, binwidth=0.01
p <- p + geom_density(alpha=0.2, fill="#FF6666")
#p <- p + geom_point(aes(y = -0.05)
# , position = position_jitter(height = 0.01)
# , alpha = 1/5)
# vertical line at Tobs
p <- p + geom_vline(aes(xintercept=Tobs), colour="#BB0000", linetype="dashed")
p <- p + labs(title = "Permutation distribution of difference in means, Uwet and Walker Meteor
p <- p + xlab("difference in means (red line = observed difference in means)")
print(p)
Permutation distribution of difference in means, Uwet and Walker Meteorites
2
density
−0.25 0.00 0.25 0.50

difference in means (red line = observed difference in means)
Notice the contrast in this permutation distribution of the difference in

means from a normal distribution.
The one-sided p-value of the test is calculated as the proportion of sampled
permutations where the difference in means was at least as extreme as T(obs).
The two-sided p-value of the test is calculated as the proportion of sampled
permutations where the absolute difference was at least as extreme as |T(obs)|.
# Calculate a two-sided p-value.
p.upper <- sum((Tperm >= abs(Tobs))) / R
p.upper
## [1] 0.0619
p.lower <- sum((Tperm <= -abs(Tobs))) / R
p.lower
## [1] 0.061
p.twosided <- p.lower + p.upper
p.twosided
## [1] 0.1229
Note that the two-sided p-value of 0.1229 is consistent, in this case, with
the two-sample t-test p-values of 0.1134 (pooled) and 0.1290 (Satterthwaite),
but different from 0.0497 (WMW). The permutation is a comparison of means
without the normality assumption, though requires that the observations are
exchangable between populations under H0.
If the only purpose of the test is reject or not reject the null hypothesis, we
can as an alternative sort the recorded differences, and then observe if T(obs) is
contained within the middle 95% of them. If it is not, we reject the hypothesis
of equal means at the 5% significance level.
6.6.1 Automated in R
The lmPerm package provides permutation tests for linear models (including t-
tests and ANOVA) and is particularly easy to impliment. You can use it for
all manner of ANOVA/ANCOVA designs, as well as simple, polynomial, and
multiple regression. Simply use lmp() and aovp() where you would have used
lm() and aov(). Note that the t-test can be performed with lm().
Below I calculate the standard t-test for the Meteorite data using t.test()
and lm(), then compare that with lmp() and what we calculated using our
calculation of the permutation test.
# standard two-sample t-test with equal variances
t.summary <- t.test(cool ~ site, data = met.long, var.equal = TRUE)
t.summary
##
##
## data: cool by site
## t = 1.669, df = 17, p-value = 0.1134
## -0.06663 0.57107
## mean in group Uwet mean in group Walker
## 0.4522 0.2000
# linear model form of t-test, "siteWalker" has estimate, se, t-stat, and p-value
lm.summary <- lm(cool ~ site, data = met.long)
summary(lm.summary)
##
## Call:
## lm(formula = cool ~ site, data = met.long)
##
## Residuals:
## Min 1Q Median 3Q Max
## -0.292 -0.196 -0.160 0.025 0.748

##
## Coefficients:
## Estimate Std. Error t value Pr(>|t|)
## (Intercept) 0.452 0.110 4.12 0.00071 ***
## siteWalker -0.252 0.151 -1.67 0.11344
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## Residual standard error: 0.329 on 17 degrees of freedom
## Multiple R-squared: 0.141,Adjusted R-squared: 0.0902
## F-statistic: 2.79 on 1 and 17 DF, p-value: 0.113
# permutation test version
library(lmPerm)
lmp.summary <- lmp(cool ~ site, data = met.long)
## [1] "Settings: unique SS "
summary(lmp.summary)
##
## Call:
## lmp(formula = cool ~ site, data = met.long)
##
## Residuals:
## -0.292 -0.196 -0.160 0.025 0.748
##
## Coefficients:
## Estimate Iter Pr(Prob)
## site1 0.126 604 0.14
##
## Multiple R-Squared: 0.141,Adjusted R-squared: 0.0902
The permutation test gives a p-value of 0.1424 which is a bit different from
our previously calculated 0.1229, and is different each time you run the routine
(since it is based on a sampling approximation, just as we coded). The function
also provides, as a reference, the pooled two-sample test at the bottom.
For the emisions data, the permutation test gives this result.
library(lmPerm)
aovp.summary <- aovp(hc ~ year, data = emis.long)
aovp.summary
## Call:
## aovp(formula = hc ~ year, data = emis.long)
##
## Terms:
## year Residuals
## Sum of Squares 4226834 17759968
##
summary(aovp.summary)
## Component 1 :
## Df R Sum Sq R Mean Sq Iter Pr(Prob)
## year 4 4226834 1056709 5000 0.0084 **
## Residuals 73 17759968 243287
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
Thus the overall ANOVA rejects the null hypothesis of all equal means. A
followup set of pairwise tests can be done with lmp().
# these are the levels of the factor, the first is the baseline group
levels(emis.long$year)
## [1] "Pre.y63" "y63.7" "y68.9" "y70.1" "y72.4"
library(lmPerm)
lmp.summary <- lmp(hc ~ year, data = emis.long)
summary(lmp.summary)
##
## Call:
## lmp(formula = hc ~ year, data = emis.long)
##
## Residuals:
## -543.6 -224.1 -126.4 44.2 2492.7
##
## Coefficients:
## Estimate Iter Pr(Prob)
## year1 325.8 5000 <2e-16 ***
## year2 236.7 798 0.112
## year3 -58.5 221 0.312
## year4 -183.3 1793 0.053 .
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## Residual standard error: 493 on 73 degrees of freedom
## Multiple R-Squared: 0.192,Adjusted R-squared: 0.148
The year1, year2, year3, and year4 correspond to the 2nd through 5th
groups. The Pr(Prob) column gives the p-values for the permutation test com-
paring each of the 2nd through 5th groups to the 1st group (year0). For the
other comparisons, reorder the factor levels and rerun.
(Note, I thought lmp() would use the baseline factor level as the intercept,
which would result in the year1 . . . year4 estimates being the pairwise differ-
ences from the intercept group. Therefore, by changing which group is the
baseline intercept group, I thought you could get pairwise tests between all
pairs. However, lmp() must be performing some internal factor level ordering
ordering because it always seems to use the third group in the emis.long$year
factor as the baseline intercept. Unfortunately, I don’t know how to use lmp()
to get the other pairwise comparisons.)
6.7 Density estimation
Density estimation is like a histogram: It is a method for visualizing the shape

of a univariate distribution (there are methods for doing multivariate density
estimation as well, but we will ignore those for the time being). In fact, I snuck
in density estimation in the first chapter and have been using it all along! Let’s
experiment with Newcombe’s speed-of-light data (excluding the two outliers).
Consider the shape of the histogram for different numbers of bins.
#### Density estimation
# include time ranks 3 and above, that is, remove the lowest two values
time2 <- time[(rank(time) >= 3)]

hist(time2, breaks=1 , main="1 break" , xlim=c(24.80,24.84), xlab=""); rug(time2)

hist(time2, main="default" , xlim=c(24.80,24.84), xlab=""); rug(time2)
hist(time2, breaks=10 , main="10 breaks" , xlim=c(24.80,24.84), xlab=""); rug(time2)
hist(time2, breaks=20 , main="20 breaks" , xlim=c(24.80,24.84), xlab=""); rug(time2)
hist(time2, breaks=100 , main="100 breaks", xlim=c(24.80,24.84), xlab=""); rug(time2)

par(old.par)
6.7: Density estimation 229
1 break
60
Frequency
40
20
0
24.80 24.81 24.82 24.83 24.84
default
25
Frequency
15
0 5
24.80 24.81 24.82 24.83 24.84
10 breaks
12
Frequency
8
4
0
24.80 24.81 24.82 24.83 24.84
20 breaks
6
Frequency
4
2
0
24.80 24.81 24.82 24.83 24.84
100 breaks
6
Frequency
4
2
0
24.80 24.81 24.82 24.83 24.84
Notice that we are starting to see more and more bins that include only a
single observation (or multiple observations at the precision of measurement).
Taken to its extreme, this type of exercise gives in some sense a “perfect” fit to
the data but is useless as an estimator of shape.
On the other hand, it is obvious that a single bin would also be completely
useless. So we try in some sense to find a middle ground between these two
extremes: “Oversmoothing” by using only one bin and “undersmooting” by
using too many. This same paradigm occurs for density estimation, in which
the amount of smoothing is determined by a quantity called the bandwidth.
By default, R uses an optimal (in some sense) choice of bandwidth.
We’ve already used the density() function to provide a smooth curve to
our histograms. So far, we’ve taken the default “bandwidth”. Let’s see what
happens when we use different bandwidths.
par(mfrow=c(3,1))
# prob=TRUE scales the y-axis like a density function, total area = 1

hist(time2, prob=TRUE, main="")
# apply a density function, store the result
den = density(time2)
# plot density line over histogram
lines(den, col=2, lty=2, lwd=2)
# extract the bandwidth (bw) from the density line
b = round(den$bw, 4)
title(main=paste("Default =", b), col.main=2)
# undersmooth
lines(density(time2, bw=0.0004), col=3, lwd=2)
text(17.5, .35, "", col=3, cex=1.4)
title(main=paste("Undersmooth, BW = 0.0004"), col.main=3)
# oversmooth
lines(density(time2, bw=0.008), col=4, lwd=2)
title(main=paste("Oversmooth, BW = 0.008"), col.main=4)
Default = 0.0018
80
60
Density
40
20
0
24.815 24.820 24.825 24.830 24.835 24.840
time2
Undersmooth, BW = 0.0004
80
60
Density
40
20
0
24.815 24.820 24.825 24.830 24.835 24.840
time2
Oversmooth, BW = 0.008
80
60
Density
40
20
0
24.815 24.820 24.825 24.830 24.835 24.840
time2
6.7: Density estimation 231
The other determining factor is the kernel, which is the shape each individual
point takes before all the shapes are added up for a final density line. While
the choice of bandwidth is very important, the choice of kernel is not. Choosing
a kernel with hard edges (such as ”rect”) will result in jagged artifacts, so
smoother kernels are often preferred.
par(mfrow=c(1,1))
# default kernel is Gaussian ("Normal")

lines(density(time2) , col=2, lty=1, lwd=2)
lines(density(time2, ker="epan"), col=3, lty=1, lwd=2)
lines(density(time2, ker="rect"), col=4, lty=1, lwd=2)
title(main="Gaussian, Epanechnikov, Rectangular")
# other kernels include: "triangular", "biweight", "cosine", "optcosine"
Gaussian, Epanechnikov, Rectangular

80
60
Density
40
20
0
24.815 24.820 24.825 24.830 24.835 24.840
time2
Chapter 7
Categorical Data
Analysis
Learning objectives
select the appropriate statistical method to compare summaries from cate-
gorical variables.
assess the assumptions of one-way and two-way tests of proportions and
independence.
decide whether the proportions between populations are different, including
in stratified and cross-sectional studies.
7.1 Categorical data

When the response variable is categorical, the interesting questions are often
about the probability of one possible outcome versus another, and whether
7.1: Categorical data 233
these probabilities depend on other variables (continuous or categorical).
Example: Titanic The sinking of the Titanic is a famous event, and new
books are still being published about it. Many well-known facts — from the
proportions of first-class passengers to the “women and children first” policy,
and the fact that that policy was not entirely successful in saving the women
and children in the third class — are reflected in the survival rates for various
classes of passenger. The source provides a data set recording class, sex, age,
and survival status for each person on board of the Titanic, and is based on
data originally collected by the British Board of Trade1.
# The Titanic dataset is a 4-dimensional table: Class, Sex, Age, Survived
library(datasets)
data(Titanic)
Titanic
## , , Age = Child, Survived = No
##
## Sex
## Class Male Female
## 1st 0 0
## 2nd 0 0
## 3rd 35 17
## Crew 0 0
##
## , , Age = Adult, Survived = No
##
## Sex
## 1st 118 4
## 2nd 154 13
## 3rd 387 89
## Crew 670 3
##
## , , Age = Child, Survived = Yes
##
## Sex
## 1st 5 1
## 2nd 11 13
## 3rd 13 14
## Crew 0 0
##
1
British Board of Trade (1990), Report on the Loss of the “Titanic” (S.S.). British Board of Trade
Inquiry Report (reprint). Gloucester, UK: Allan Sutton Publishing. Note that there is not complete
agreement among primary sources as to the exact numbers on board, rescued, or lost.
234 Ch 7: Categorical Data Analysis
## , , Age = Adult, Survived = Yes

##
## Sex
## 1st 57 140
## 2nd 14 80
## 3rd 75 76
## Crew 192 20
# reshape into long data.frame
library(reshape2)
df.titanic <- melt(Titanic, value.name = "Freq")
df.titanic
## Class Sex Age Survived Freq
## 1 1st Male Child No 0
## 2 2nd Male Child No 0
## 3 3rd Male Child No 35
## 4 Crew Male Child No 0
## 5 1st Female Child No 0
## 6 2nd Female Child No 0
## 7 3rd Female Child No 17
## 8 Crew Female Child No 0
## 9 1st Male Adult No 118
## 10 2nd Male Adult No 154
## 11 3rd Male Adult No 387
## 12 Crew Male Adult No 670
## 13 1st Female Adult No 4
## 14 2nd Female Adult No 13
## 15 3rd Female Adult No 89
## 16 Crew Female Adult No 3
## 17 1st Male Child Yes 5
## 18 2nd Male Child Yes 11
## 19 3rd Male Child Yes 13
## 20 Crew Male Child Yes 0
## 21 1st Female Child Yes 1
## 22 2nd Female Child Yes 13
## 23 3rd Female Child Yes 14
## 24 Crew Female Child Yes 0
## 25 1st Male Adult Yes 57
## 26 2nd Male Adult Yes 14
## 27 3rd Male Adult Yes 75
## 28 Crew Male Adult Yes 192
## 29 1st Female Adult Yes 140
## 30 2nd Female Adult Yes 80
## 31 3rd Female Adult Yes 76
## 32 Crew Female Adult Yes 20
# Total number of people
sum(df.titanic$Freq)
## [1] 2201
# create colors based on survival
df.titanic$Color <- ifelse(df.titanic$Survived == "Yes", "#008888", "#330066")
7.2: Single Proportion Problems 235
# subset only the adults (since there were so few children)

df.titanic.adult <- subset(df.titanic, Age == "Adult")
# see R code on website for function parallelset()

# see help for with(), it allows temporary direct reference to columns in a data.frame
# otherwise, we'd need to specify df.titanic.adult£Survived, ...
with(df.titanic.adult
, parallelset(Survived, Sex, Class, freq = Freq, col = Color, alpha=0.2)
)
Survived
No Yes
Sex
Female Male
Class
1st 2nd 3rd Crew
There are many questions that can be asked of this dataset. How likely were
people to survive such a ship sinking in cold water? Is the survival proportion
dependent on sex, class, or age, or a combination of these? How different is the
survival proportions for 1st class females versus 3rd class males?
7.2 Single Proportion Problems

Assume that you are interested in estimating the proportion p of individuals
in a population with a certain characteristic or attribute based on a random or
representative sample of size n from the population. The sample proportion
p̂ =(# with attribute in the sample)/n is the best guess for p based on the data.
This is the simplest categorical data problem. Each response falls into
one of two exclusive and exhaustive categories, called “success” and “failure”.
Individuals with the attribute of interest are in the success category. The rest
fall into the failure category. Knowledge of the population proportion p
of successes characterizes the distribution across both categories because the
population proportion of failures is 1 − p.
As an aside, note that the probability that a randomly selected individual
has the attribute of interest is the population proportion p with the attribute,
so the terms population proportion and probability can be used interchangeably
with random sampling.
7.2.1 A CI for p
A two-sided CI for p is a range of plausible values for the unknown population
proportion p, based on the observed data. To compute a two-sided CI for p:
1. Specify the confidence level as the percent 100(1 − α)% and solve for the
error rate α of the CI.
2. Compute zcrit = z0.5α (i.e., area under the standard normal curve to
the left and to the right of zcrit are 1 − 0.5α and 0.5α, respectively).
qnorm(1-0.05/2)=1.96.
3. The 100(1 − α)% CI for p has endpoints L = p̂ − zcritSE and U =
p̂ + zcritSE, respectively, where the “CI standard error” is
r
p̂(1 − p̂)
SE = .
n
The CI is often written as p̂ ± zcritSE.
Reminder of CI interpretation. The CI is determined once the confi-

dence level is specified and the data are collected. Prior to collecting data, the
CI is unknown and can be viewed as random because it will depend on the
actual sample selected. Different samples give different CIs. The “confidence”
in, say, the 95% CI (which has a 0.05 or 5% error rate) can be interpreted as
follows. If you repeatedly sample the population and construct 95% CIs for p,
then 95% of the intervals will contain p, whereas 5% (the error rate) will not.
The CI you get from your data either covers p, or it does not.
The length of the CI
U − L = 2zcritSE
depends on the accuracy of the estimate p̂, as measured by the standard error
SE. For a given p̂, this standard error decreases as the sample size n increases,
yielding a narrower CI. For a fixed sample size, this standard error is maxi-
mized at p̂ = 0.5, and decreases as p̂ moves towards either 0 or 1. In essence,
sample proportions near 0 or 1 give narrower CIs for p. However, the normal
approximation used in the CI construction is less reliable for extreme values of
p̂.
ClickerQ s — CI for proportions STT.08.01.010
Example: Tamper resistant packaging The 1983 Tylenol poisoning

episode highlighted the desirability of using tamper-resistant packaging. The
article “Tamper Resistant Packaging: Is it Really?” (Packaging Engineering,
June 1983) reported the results of a survey on consumer attitudes towards
tamper-resistant packaging. A sample of 270 consumers was asked the ques-
tion: “Would you be willing to pay extra for tamper resistant packaging?” The
number of yes respondents was 189. Construct a 95% CI for the proportion p
of all consumers who were willing in 1983 to pay extra for such packaging.
Here n = 270 and p̂ = 189/270 = 0.700. The critical value for a 95% CI for
p is z0.025 = 1.96. The CI standard error is given by
r
0.7 × 0.3
SE = = 0.028,
270
so zcritSE = 1.96 × 0.028 = 0.055. The 95% CI for p is 0.700 ± 0.055. You are
95% confident that the proportion of consumers willing to pay extra for better
packaging is between 0.645 and 0.755. (Willing to pay how much extra?)
Appropriateness of the CI
The standard CI is based on a large-sample standard normal approximation

to
p̂ − p
z= .
SE
A simple rule of thumb requires np̂ ≥ 5 and n(1 − p̂) ≥ 5 for the method to
be suitable. Given that np̂ and n(1 − p̂) are the observed numbers of successes
and failures, you should have at least 5 of each to apply the large-sample CI.
In the packaging example, np̂ = 270 × (0.700) = 189 (the number who
support the new packaging) and n(1 − p̂) = 270 × (0.300) = 81 (the number
who oppose) both exceed 5. The normal approximation is appropriate here.
7.2.2 Hypothesis Tests on Proportions

The following example is typical of questions that can be answered using a
hypothesis test for a population proportion.
Example Environmental problems associated with leaded gasolines are well-

known. Many motorists have tampered with the emission control devices on
their cars to save money by purchasing leaded rather than unleaded gasoline.
A Los Angeles Times article on March 17, 1984 reported that 15% of all
California motorists have engaged in emissions tampering. A random sample
of 200 cars from L.A. county was obtained, and the emissions devices on 21 are
found to be tampered with. Does this suggest that the proportion of cars in
L.A. county with tampered devices differs from the statewide proportion?
Two-Sided Hypothesis Test for p
Suppose you are interested in whether the population proportion p is equal to

a prespecified value, say p0. This question can be formulated as a two-sided
test. To carry out the test:
1. Define the null hypothesis H0 : p = p0 and alternative hypothesis HA :
p 6= p0.
2. Choose the size or significance level of the test, denoted by α.
3. Using the standard normal probability table, find the critical value zcrit
such that the areas under the normal curve to the left and right of zcrit
are 1 − 0.5α and 0.5α, respectively. That is, zcrit = z0.5α .
4. Compute the test statistic (often to be labelled zobs)
p̂ − p0
zs = ,
SE
where the “test standard error” (based on the hypothesized value) is
r
p0(1 − p0)
SE = .
n
5. Reject H0 in favor of HA if |zobs| ≥ zcrit. Otherwise, do not reject H0.
The rejection rule is easily understood visually. The area under the normal
curve outside ±zcrit is the size α of the test. One-half of α is the area in
each tail. You reject H0 in favor of HA if the test statistic exceeds ±zcrit. This
occurs when p̂ is significantly different from p0, as measured by the standardized
distance zobs between p̂ and p0.
Z−distribution with two−sided size α = .05 critical region Z−distribution with two−sided p−value
Total area is p−value

and is random, not fixed
α α p − value p − value
(fixed)
2 2 2 2
−4 Rej H0 − z 0 4 −4 − zs 0 zs 4
Crit zCrit Rej H0 − zCrit zCrit
ClickerQ s — Test statistic STT.07.01.057
7.2.3 The p-value for a two-sided test

To compute the p-value (not to be confused with the value of the proportion
p) for a two-sided test:
1. Compute the test statistic zs = zobs.
2. Evaluate the area under the normal probability curve outside ±|zs|.
Recall that the null hypothesis for a size α test is rejected if and only if the
p-value is less than or equal to α.
Example: Emissions data Each car in the target population (L.A. county)
either has been tampered with (a success) or has not been tampered with (a
failure). Let p = the proportion of cars in L.A. county with tampered emissions
control devices. You want to test H0 : p = 0.15 against HA : p 6= 0.15 (here
p0 = 0.15). The critical value for a two-sided test of size α = 0.05 is zcrit = 1.96.
The data are a sample of n = 200 cars. The sample proportion of cars that
have been tampered with is p̂ = 21/200 = 0.105. The test statistic is
0.105 − 0.15
zs = = −1.78,
0.02525
where the test standard error satisfies
r
0.15 × 0.85
SE = = 0.02525.
200
Given that |zs| = 1.78 < 1.96, you have insufficient evidence to reject H0 at the
5% level. That is, you have insufficient evidence to conclude that the proportion
of cars in L.A. county that have been tampered with differs from the statewide
proportion.
This decision is reinforced by the p-value calculation. The p-value is the area
under the standard normal curve outside ±1.78. This is 2 × 0.0375 = 0.075,
which exceeds the test size of 0.05.
Emissions data p−value
Total area is p−value

= .075
.0375 .0375
−4 −1.78 0 1.78 4
Remark The SE used in the test and CI are different. This implies that a
hypothesis test and CI could potentially lead to different decisions. That is, a
95% CI for a population proportion might cover p0 when the p-value for testing
H0 : p = p0 is smaller than 0.05. This will happen, typically, only in cases
where the decision is “borderline.”
7.2.4 Appropriateness of Test

The z-test is based on a large-sample normal approximation, which works better
for a given sample size when p0 is closer to 0.5. The sample size needed for an
accurate approximation increases dramatically the closer p0 gets to 0 or to 1. A
simple rule of thumb is that the test is appropriate when (the expected number
of successes) np0 ≥ 5 and (the expected number of failures) n(1 − p0) ≥ 5.
In the emissions example, np0 = 200 × (0.15) = 30 and n(1 − p0) =
200 × (0.85) = 170 exceed 5, so the normal approximation is appropriate.
#### Single Proportion Problems
# Approximate normal test for proportion, without Yates' continuity correction
prop.test(21, 200, p = 0.15, correct = FALSE)
##
## 1-sample proportions test without continuity correction
##
## data: 21 out of 200, null probability 0.15
## X-squared = 3.176, df = 1, p-value = 0.07471
## alternative hypothesis: true p is not equal to 0.15
## 0.06971 0.15518
## p
## 0.105
# Approximate normal test for proportion, with Yates' continuity correction
#prop.test(21, 200, p = 0.15)
ClickerQ s — Parachute null hypothesis STT.08.01.040
ClickerQ s — Parachute conclusion STT.08.01.050
ClickerQ s — Parachute p-value STT.08.01.060
7.2.6 One-Sided Tests and One-Sided Confidence Bounds

The mechanics of tests on proportions are similar to tests on means, except we
use a different test statistic and a different probability distribution for critical
values. This applies to one-sided and two-sided procedures. The example below
illustrates a one-sided test and bound.
Example: brain hemispheres An article in the April 6, 1983 edition of

The Los Angeles Times reported on a study of 53 learning-impaired youngsters
at the Massachusetts General Hospital. The right side of the brain was found
to be larger than the left side in 22 of the children. The proportion of the
general population with brains having larger right sides is known to be 0.25.
Does the data provide strong evidence for concluding, as the article claims, that
the proportion of learning impaired youngsters with brains having larger right
sides exceeds the proportion in the general population?
I will answer this question by computing a p-value for a one-sided test. Let
p be the population proportion of learning disabled children with brains having
larger right sides. I am interested in testing H0 : p = 0.25 against HA : p > 0.25
(here p0 = 0.25).
The proportion of children sampled with brains having larger right sides is
p̂ = 22/53 = 0.415. The test statistic is
0.415 − 0.25
zs = = 2.78,
0.0595
where the test standard error satisfies
r
0.25 × 0.75
SE = = 0.0595.
53
The p-value for an upper one-sided test is the area under the standard normal
curve to the right of 2.78, which is approximately .003; see the picture below.
I would reject H0 in favor of HA using any of the standard test levels, say 0.05
or 0.01. The newspaper’s claim is reasonable.
Right brain upper one−sided p−value
p−value is area in
right tail only
.003
−4 −2 0 zs = 2.78 4
A sensible next step in the analysis would be to compute a lower confi-

dence bound p̂ − zcritSE for p. For illustration, consider a 95% bound. The
CI standard error is
r r
p̂(1 − p̂) 0.415 × 0.585
SE = = = 0.0677.
n 53
The critical value for a one-sided 5% test is zcrit = 1.645, so a lower 95% bound
on p is 0.415 − 1.645 × 0.0677 = 0.304. I am 95% confident that the population
proportion of learning disabled children with brains having larger right sides is
at least 0.304. Values of p smaller than 0.304 are not supported by the data.
You should verify that the sample size is sufficiently large to use the approx-
imate methods in this example.
#### Example: brain hemispheres
prop.test(22, 53, p = 0.25, alternative = "greater", correct = FALSE)
##
## 1-sample proportions test without continuity correction

##
## data: 22 out of 53, null probability 0.25
## alternative hypothesis: true p is greater than 0.25
## 0.3105 1.0000
## p
## 0.4151
7.2.7 Small Sample Procedures

Large sample tests and CIs for p should be interpreted with caution in small
sized samples because the true error rate usually exceeds the assumed (nominal)
value. For example, an assumed 95% CI, with a nominal error rate of 5%, may
be only an 80% CI, with a 20% error rate. The large-sample CIs are usually
overly optimistic (i.e., too narrow) when the sample size is too small to use the
normal approximation.
Alan Agresti suggests the following method for a 95% CI. The standard
method computes the sample proportion as p̂ = x/n where x is the number of
successes in the sample and n is the sample size. Agresti suggested using the
estimated proportion p̃ = (x + 2)/(n + 4) with the standard error
r
p̃(1 − p̃)
SE = ,
n+4
in the “usual 95% interval” formula: p̃ ± 1.96SE. This appears odd, but
amounts to adding two successes and two failures to the observed data, and
then computing the standard CI.
This adjustment has little effect when n is large and p̂ is not near either 0
or 1, as in the Tylenol example.
Example: swimming segregation This example is based on a case heard

before the U.S. Supreme Court. A racially segregated swimming club was
ordered to admit minority members. However, it is unclear whether the club
has been following the spirit of the mandate. Historically, 85% of the white
applicants were approved. Since the mandate, only 1 of 6 minority applicants
has been approved. Is there evidence of continued discrimination?
I examine this issue by constructing a CI and a test for the probability p (or
population proportion) that a minority applicant is approved. Before examining
the question of primary interest, let me show that the two approximate CIs are
very different, due to the small sample size. One minority applicant (x = 1)
was approved out of n = 6 candidates, giving p̂ = 1/6 = 0.167.
A 95% large-sample CI for p is (−0.14, 0.46). Since a negative proportion is
not possible, the CI should be reported as (0.00, 0.46). Agresti’s 95% CI (based
on 3 successes and 7 failures) is (0.02, 0.58). The big difference between the two
intervals coupled with the negative lower endpoint on the standard CI suggests
that the normal approximation used with the standard method is inappropriate.
This view is reinforced by the rule-of-thumb calculation for using the standard
interval. Agresti’s CI is wider, which is consistent with my comment that the
standard CI is too narrow in small samples. As a comparison, the exact 95%
CI is (0.004, 0.64), which agrees more closely with Agresti’s interval.
I should emphasize that the exact CI is best to use, but is not available in
all statistical packages, so methods based on approximations may be required,
and if so, then Agresti’s method is clearly better than the standard normal
approximation in small sized samples.
Recall that the results of the asymptotic 95% CIs may disagree with the
hypothesis test results. Exact methods will not contradict each other this way
(neither do these asymptotic methods, usually).
#### Example: swimming segregation
## The prop.test() does an additional adjustment, so does not match precisely
## the results in the above paragraphs

prop.test(1, 6, p = 0.85, correct = FALSE)$conf.int
## Warning: Chi-squared approximation may be incorrect
## [1] 0.03005 0.56350
## [1] 0.95
# Agresti's method
prop.test(1+2, 6+4, p = 0.85, correct = FALSE)$conf.int
## Warning: Chi-squared approximation may be incorrect
## [1] 0.1078 0.6032
7.3: Analyzing Raw Data 247
## [1] 0.95
# Exact binomial test for proportion
binom.test(1, 6, p = 0.85)$conf.int
## [1] 0.004211 0.641235
## [1] 0.95
Returning to the problem, you might check for discrimination by testing
H0 : p = 0.85 against HA : p < 0.85 using an exact test. The exact test
p-value is 0.000 to three decimal places, and an exact upper bound for p is
0.582. What does this suggest to you?
# Exact binomial test for proportion
binom.test(1, 6, alternative = "less", p = 0.85)
##
## Exact binomial test
##
## data: 1 and 6
## number of successes = 1, number of trials = 6, p-value =
## 0.0003987
## alternative hypothesis: true probability of success is less than 0.85
## 0.0000 0.5818
## probability of success
## 0.1667
7.3 Analyzing Raw Data

In most studies, your data will be stored in a spreadsheet with one observation
per case or individual. For example, the data below give the individual responses
to the applicants of the swim club.
#### Example: swimming segregation, raw data
## read.table options
# sep = default is any white space, but our strings contain a space,
# so I changed this to a comma
# header = there are no column headers
# stringsAsFActors = default converts strings to factors, but I want them
# to just be the plain character text
swim <- read.table(text="
not approved
not approved
not approved
approved
not approved
not approved
", sep = ",", header=FALSE, stringsAsFactors=FALSE)
# name the column

names(swim) <- c("application")
# show the structure of the data.frame
str(swim)
## 'data.frame': 6 obs. of 1 variable:
## $ application: chr "not approved" "not approved" "not approved" "approved" ...
# display the data.frame
swim
## application
## 1 not approved
## 2 not approved
## 3 not approved
## 4 approved
## 5 not approved
## 6 not approved
The data were entered as alphabetic strings. We can use table() to count
frequencies of categorical variables.
# count the frequency of each categorical variable
table(swim)
## swim
## approved not approved
## 1 5
You can compute a CI and test for a proportion using raw data, provided
the data column includes only two distinct values. The levels can be numerical
or alphanumeric.
# use the counts from table() for input in binom.text()
# the help for binom.test() says x can be a vector of length 2
# giving the numbers of successes and failures, respectively
# that's exactly what table(swim) gave us
binom.test(table(swim), p = 0.85, alternative = "less")
##
## Exact binomial test
##
## data: table(swim)
## number of successes = 1, number of trials = 6, p-value =
## 0.0003987
## alternative hypothesis: true probability of success is less than 0.85
## 0.0000 0.5818
## probability of success
## 0.1667
It is possible that the order (alphabetically) is the wrong order, failures and
successes, in which case we’d need to reorder the input to binom.test().
In Chapter 6 we looked at the binomial distribution to obtain an exact Sign
7.4: Goodness-of-Fit Tests (Multinomial) 249
Test confidence interval for the median. Examine the following to see where
the exact p-value for this test comes from.
n <- 6
x <- 0:n
p0 <- 0.85
bincdf <- pbinom(x, n, p0)
cdf <- data.frame(x, bincdf)
cdf
## x bincdf
## 1 0 1.139e-05
## 2 1 3.987e-04
## 3 2 5.885e-03
## 4 3 4.734e-02
## 5 4 2.235e-01
## 6 5 6.229e-01
## 7 6 1.000e+00
ClickerQ s — Excess successes STT.05.01.030
7.4 Goodness-of-Fit Tests (Multinomial)

Example: jury pool The following data set was used as evidence in a
court case. The data represent a sample of 1336 individuals from the jury pool
of a large municipal court district for the years 1975–1977. The fairness of the
representation of various age groups on juries was being contested. The strategy
for doing this was to challenge the representativeness of the pool of individuals
from which the juries are drawn. This was done by comparing the age group
distribution within the jury pool against the age distribution in the district as
a whole, which was available from census figures.
Age group (yrs) Obs. Counts Obs. Prop. Census Prop.
18-19 23 0.017 0.061
20-24 96 0.072 0.150
25-29 134 0.100 0.135
30-39 293 0.219 0.217
40-49 297 0.222 0.153
50-64 380 0.284 0.182
65-99 113 0.085 0.102
Total: 1336 1.000 1.000
A statistical question here is whether the jury pool population proportions
are equal to the census proportions across the age categories. This comparison
can be formulated as a goodness-of-fit test, which generalizes the large-
sample test on a single proportion to a categorical variable (here age) with
r > 2 levels. For r = 2 categories, the goodness-of-fit test and large-sample
test on a single proportion are identical. Although this problem compares two
populations, only one sample is involved because the census data is a population
summary!
In general, suppose each individual in a population is categorized into one
and only one of r levels or categories. Let p1, p2, . . ., pr , be the population
proportions in the r categories, where each pi ≥ 0 and p1 + p2 + · · · + pr = 1.
The hypotheses of interest in a goodness-of-fit problem are H0 : p1 = p01,
p2 = p02, . . ., pr = p0r and HA : not H0, where p01, p02, . . ., p0r are given
category proportions.
The plausibility of H0 is evaluated by comparing the hypothesized category
proportions to estimated (i.e., observed) category proportions p̂1, p̂2, . . ., p̂r
from a random or representative sample of n individuals selected from the pop-
ulation. The discrepancy between the hypothesized and observed proportions
is measured by the Pearson chi-squared statistic:
r
X (Oi − Ei)2
χ2s = ,
i=1
Ei
where Oi is the observed number in the sample that fall into the ith category
(Oi = np̂i), and Ei = np0i is the number of individuals expected to be in the
ith category when H0 is true.
The Pearson statistic can also be computed as the sum of the squared resid-
uals: r
X
χ2s = Zi2,
i=1
√
where Zi = (Oi − Ei)/ Ei, or in terms of the observed and hypothesized
category proportions
r
2
X (p̂i − p0i)2
χs = n .
i=1
p 0i
The Pearson statistic χ2s is “small” when all of the observed counts (propor-
tions) are close to the expected counts (proportions). The Pearson χ2 is “large”
when one or more observed counts (proportions) differs noticeably from what
is expected when H0 is true. Put another way, large values of χ2s suggest that
H0 is false.
The critical value χ2crit for the test is obtained from a chi-squared probability
table with r − 1 degrees of freedom. The picture below shows the form of the
rejection region. For example, if r = 5 and α = 0.05, then you reject H0 when
χ2s ≥ χ2crit = 9.49 (qchisq(0.95, 5-1)). The p-value for the test is the area
under the chi-squared curve with df = r − 1 to the right of the observed χ2s
value.
χ2 with 4 degrees of freedom χ2 with 4 degrees of freedom

χ2S significant
α = .05 (fixed) p − value (random)
0 5 10 15 0 5 10 15
χ2Crit Reject H0 for χ2S here χ2Crit χ2S
Example: jury pool Let p18 be the proportion in the jury pool population
between ages 18 and 19. Define p20, p25, p30, p40, p50, and p65 analogously.
You are interested in testing that the true jury proportions equal the census
proportions, H0 : p18 = 0.061, p20 = 0.150, p25 = 0.135, p30 = 0.217, p40 =
0.153, p50 = 0.182, and p65 = 0.102 against HA : not H0, using the sample of
1336 from the jury pool.
The observed counts, the expected counts, and the category residuals are
given in the table √below. For example, E18 = 1336 × (0.061) = 81.5 and
Z18 = (23 − 81.5)/ 81.5 = −6.48 in the 18-19 year category.
The Pearson statistic is
χ2s = (−6.48)2 +(−7.38)2 +(−3.45)2 +0.182 +6.482 +8.782 +(−1.99)2 = 231.26
on r − 1 = 7 − 1 = 6 degrees of freedom. Here χ2crit = 12.59 at α = 0.05. The

p-value for the goodness-of-fit test is less than 0.001, which suggests that H0 is
false.
Age group (yrs) Obs. Counts Exp. Counts Residual
18-19 23 81.5 −6.48
20-24 96 200.4 −7.38
25-29 134 180.4 −3.45
30-39 293 289.9 0.18
40-49 297 204.4 6.48
50-64 380 243.2 8.78
65-99 113 136.3 −1.99
7.4.1 Adequacy of the Goodness-of-Fit Test

The chi-squared goodness-of-fit test is a large-sample test. A conservative rule
of thumb is that the test is suitable when each expected count is at least five.
This holds in the jury pool example. There is no widely available alternative
method for testing goodness-of-fit with smaller sample sizes. There is evidence,
however, that the chi-squared test is slightly conservative (the p-values are
too large, on average) when the expected counts are smaller. Some statisticians
recommend that the chi-squared approximation be used when the minimum
expected count is at least one, provided the expected counts are not too variable.
#### Example: jury pool
jury <- read.table(text="
Age Count CensusProp
18-19 23 0.061
20-24 96 0.150
25-29 134 0.135
30-39 293 0.217
40-49 297 0.153
50-64 380 0.182
65-99 113 0.102

", header=TRUE)

str(jury)
## $ Age : Factor w/ 7 levels "18-19","20-24",..: 1 2 3 4 5 6 7
## $ Count : int 23 96 134 293 297 380 113
## $ CensusProp: num 0.061 0.15 0.135 0.217 0.153 0.182 0.102
jury
## Age Count CensusProp
## 1 18-19 23 0.061
## 2 20-24 96 0.150
## 3 25-29 134 0.135
## 4 30-39 293 0.217
## 5 40-49 297 0.153
## 6 50-64 380 0.182
## 7 65-99 113 0.102
# calculate chi-square goodness-of-fit test
x.summary <- chisq.test(jury$Count, correct = FALSE, p = jury$CensusProp)
# print result of test
x.summary
##
## Chi-squared test for given probabilities
##
## data: jury$Count
## X-squared = 231.3, df = 6, p-value < 2.2e-16
# use output in x.summary and create table
x.table <- data.frame(age = jury$Age
, obs = x.summary$observed
, exp = x.summary$expected
, res = x.summary$residuals
, chisq = x.summary$residuals^2
, stdres = x.summary$stdres)
x.table
## age obs exp res chisq stdres
## 1 18-19 23 81.5 -6.4797 41.98712 -6.687
## 2 20-24 96 200.4 -7.3748 54.38802 -7.999
## 3 25-29 134 180.4 -3.4520 11.91644 -3.712
## 4 30-39 293 289.9 0.1814 0.03289 0.205
## 5 40-49 297 204.4 6.4763 41.94199 7.037
## 6 50-64 380 243.2 8.7761 77.01921 9.703
## 7 65-99 113 136.3 -1.9936 3.97430 -2.104
Plot observed vs expected values to help identify age groups that deviate
the most. Plot contribution to chi-square values to help identify age groups
that deviate the most. The term “Contribution to Chi-Square” (chisq) refers
(O−E)2
to the values of E for each category. χ2s is the sum of those contributions.
library(reshape2)
x.table.obsexp <- melt(x.table,
# id.vars: ID variables
# all variables to keep but not split apart on
id.vars = c("age"),
# measure.vars: The source columns
# (if unspecified then all other variables are measure.vars)
measure.vars = c("obs", "exp"),
# variable.name: Name of the destination column identifying each
# original column that the measurement came from
variable.name = "stat",
# value.name: column name for values in table
value.name = "value"
)
names(x.table.obsexp) <- c("age", "stat", "value")
# Observed vs Expected counts

library(ggplot2)
p <- ggplot(x.table.obsexp, aes(x = age, fill = stat, weight=value))
p <- p + geom_bar(position="dodge")
p <- p + labs(title = "Observed and Expected frequencies")
p <- p + xlab("Age category (years)")
print(p)
# Contribution to chi-sq
# pull out only the age and chisq columns
x.table.chisq <- x.table[, c("age","chisq")]
# reorder the age categories to be descending relative to the chisq statistic
x.table.chisq$age <- with(x.table, reorder(age, -chisq))
p <- ggplot(x.table.chisq, aes(x = age, weight = chisq))

p <- p + geom_bar()
p <- p + labs(title = "Contribution to Chi-sq statistic")
p <- p + xlab("Sorted age category (years)")
p <- p + ylab("Chi-sq")
print(p)
Observed and Expected frequencies Contribution to Chi−sq statistic
80
300 60
stat
Chi−sq
count
200 40
obs
exp
100 20
0 0
18−19 20−24 25−29 30−39 40−49 50−64 65−99 50−64 20−24 18−19 40−49 25−29 65−99 30−39
Age category (years) Sorted age category (years)
7.4.3 Multiple Comparisons in a Goodness-of-Fit Prob-

lem
The goodness-of-fit test suggests that at least one of the age category propor-
tions for the jury pool population differs from the census figures. A reasonable
next step in the analysis would be to separately test the seven hypotheses:
H0 : p18 = 0.061, H0 : p20 = 0.150, H0 : p25 = 0.135, H0 : p30 = 0.217,
H0 : p40 = 0.153, H0 : p50 = 0.182, and H0 : p65 = 0.102 to see which age
categories led to this conclusion.
A Bonferroni comparison with a Family Error Rate ≤ 0.05 will be considered
for this multiple comparisons problem. The error rates for the seven individual
tests are set to α = 0.05/7 = 0.0071, which corresponds to 99.29% two-sided
CIs for the individual jury pool proportions. The area under the standard
normal curve to the right of 2.69 is 0.05/2/7 = 0.0036, about one-half the
error rate for the individual CIs, so the critical value for the CIs, or for the
tests, is zcrit ≈ 2.69. The next table gives individual 99.29% CIs based on
the large sample approximation. You can get the individual CIs in R using
the binom.test() or prop.test() function. For example, the CI for Age Group
18-19 is obtained by specifying 23 successes in 1336 trials.
Below I perform exact binomial tests of proportion for each of the seven age
categories at the Bonferroni-adjusted significance level. I save the p-values and
confidence intervals in a table along with the observed and census proportions
in order to display the table below.
b.sum1 <- binom.test(jury$Count[1], sum(jury$Count), p = jury$CensusProp[1], alternative = "two.sided", conf.level = 1-0.05/7)
# put the p-value and CI into a data.frame
b.sum <- data.frame(
rbind( c(b.sum1$p.value, b.sum1$conf.int)
, c(b.sum2$p.value, b.sum2$conf.int)
)
)
names(b.sum) <- c("p.value", "CI.lower", "CI.upper")
b.sum$Age <- jury$Age
b.sum$Observed <- x.table$obs/sum(x.table$obs)
b.sum$CensusProp <- jury$CensusProp
b.sum
## p.value CI.lower CI.upper Age Observed CensusProp
## 1 8.815e-15 0.009137 0.02920 18-19 0.01722 0.061
## 2 2.695e-18 0.054160 0.09294 20-24 0.07186 0.150
## 3 1.394e-04 0.079398 0.12435 25-29 0.10030 0.135
## 4 8.422e-01 0.189621 0.25120 30-39 0.21931 0.217
## 5 2.383e-11 0.192456 0.25433 40-49 0.22231 0.153
## 6 5.916e-20 0.251744 0.31881 50-64 0.28443 0.182
## 7 3.742e-02 0.065366 0.10708 65-99 0.08458 0.102
The CIs for the 30-39 and 65-99 year categories contain the census propor-
tions. In the other five age categories, there are significant differences between
the jury pool proportions and the census proportions. In general, young adults
appear to be underrepresented in the jury pool whereas older age groups are
overrepresented.
Age p.value CI.lower CI.upper Observed CensusProp
1 18-19 0.000 0.009 0.029 0.017 0.061
2 20-24 0.000 0.054 0.093 0.072 0.150
3 25-29 0.000 0.079 0.124 0.100 0.135
4 30-39 0.842 0.190 0.251 0.219 0.217
5 40-49 0.000 0.192 0.254 0.222 0.153
6 50-64 0.000 0.252 0.319 0.284 0.182
7 65-99 0.037 0.065 0.107 0.085 0.102
The residuals also highlight significant differences because the largest resid-
uals correspond to the categories that contribute most to the value of χ2s. Some
researchers use the residuals for the multiple comparisons, treating the Zis as
standard normal variables. Following this approach, you would conclude that
the jury pool proportions differ from the proportions in the general population
7.5: Comparing Two Proportions: Independent Samples 257
in every age category where |Zi| ≥ 2.70 (using the same Bonferroni correction).
This gives the same conclusion as before.
The two multiple comparison methods are similar, but not identical. The
residuals
Oi − Ei p̂i − p0i
Zi = √ = p p0i
Ei n
agree with the large-sample statistic for testing H0 : pi = p0i, except that the
divisor in Zi omits a 1 − p0i term. The Zis are not standard normal random
variables as assumed, and the value of Zi underestimates the significance of the
observed differences. Multiple comparisons using the Zis will find, on average,
fewer significant differences than the preferred method based on the large sample
tests. However, the differences between the two methods are usually minor when
all of the hypothesized proportions are small.
7.5 Comparing Two Proportions: Indepen-

dent Samples
The New Mexico state legislature is interested in how the proportion of reg-
istered voters that support Indian gaming differs between New Mexico and
Colorado. Assuming neither population proportion is known, the state’s statis-
tician might recommend that the state conduct a survey of registered voters
sampled independently from the two states, followed by a comparison of the
sample proportions in favor of Indian gaming.
Statistical methods for comparing two proportions using independent sam-
ples can be formulated as follows. Let p1 and p2 be the proportion of populations
1 and 2, respectively, with the attribute of interest. Let p̂1 and p̂2 be the corre-
sponding sample proportions, based on independent random or representative
samples of size n1 and n2 from the two populations.
7.5.1 Large Sample CI and Tests for p1 − p2
A large-sample CI for p1 − p2 is (p̂1 − p̂2) ± zcritSECI (p̂1 − p̂2), where zcrit is
the standard normal critical value for the desired confidence level, and
s
p̂1(1 − p̂1) p̂2(1 − p̂2)
SECI (p̂1 − p̂2) = +
n1 n2
is the CI standard error.

A large-sample p-value for a test of the null hypothesis H0 : p1 − p2 = 0
against the two-sided alternative HA : p1 − p2 6= 0 is evaluated using tail
areas of the standard normal distribution (identical to one sample evaluation)
in conjunction with the test statistic
p̂1 − p̂2
zs = ,
SEtest(p̂1 − p̂2)
where
s s
p̄(1 − p̄) p̄(1 − p̄) 1 1
SEtest(p̂1 − p̂2) = + = p̄(1 − p̄) +
n1 n2 n1 n2
is the test standard error for p̂1 − p̂2. The pooled proportion
n1p̂1 + n2p̂2
p̄ =
n1 + n2
is the proportion of successes in the two samples combined. The test standard
error has the same functional form as the CI standard error, with p̄ replacing
the individual sample proportions.
The pooled proportion is the best guess at the common population propor-
tion when H0 : p1 = p2 is true. The test standard error estimates the standard
deviation of p̂1 − p̂2 assuming H0 is true.
Remark: As in the one-sample proportion problem, the test and CI SE’s are
different. This can (but usually does not) lead to some contradiction between
the test and CI.
Example, vitamin C Two hundred and seventy nine (279) French skiers
were studied during two one-week periods in 1961. One group of 140 skiers
receiving a placebo each day, and the other 139 receiving 1 gram of ascorbic
acid (Vitamin C) per day. The study was double blind — neither the subjects
nor the researchers knew who received which treatment. Let p1 be the prob-
ability that a member of the ascorbic acid group contracts a cold during the
study period, and p2 be the corresponding probability for the placebo group.
Linus Pauling (Chemistry and Peace Nobel prize winner) and I are interested in
testing whether H0 : p1 = p2. The data are summarized below as a two-by-two
table of counts (a contingency table)
Outcome Ascorbic Acid Placebo

# with cold 17 31
# with no cold 122 109
Totals 139 140
The sample sizes are n1 = 139 and n2 = 140. The sample proportion of
skiers developing colds in the placebo and treatment groups are p̂2 = 31/140 =
0.221 and p̂1 = 17/139 = 0.122, respectively. The difference is p̂1 − p̂2 = 0.122−
0.221 = −0.099. The pooled proportion is the number of skiers that developed
colds divided by the number of skiers in the study: p̄ = 48/279 = 0.172.
The test standard error is
s
1 1
SEtest(p̂1 − p̂2) = 0.172 × (1 − 0.172) + = 0.0452.
139 140
The test statistic is

0.122 − 0.221
zs = = −2.19.
0.0452
The p-value for a two-sided test is twice the area under the standard normal
curve to the right of 2.19 (or twice the area to the left of −2.19), which is 2 ×
(0.014) = 0.028. At the 5% level, we reject the hypothesis that the probability
of contracting a cold is the same whether you are given a placebo or Vitamin
C.
A CI for p1 − p2 provides a measure of the size of the treatment effect. For
a 95% CI
r
0.221 × (1 − 0.221) 0.122 × (1 − 0.122)
zcritSECI (p̂1 − p̂2) = 1.96 +
140 139
= 1.96 × (0.04472) = 0.088.
The 95% CI for p1 − p2 is −0.099 ± 0.088, or (−0.187, −0.011). We are 95%

confident that p2 exceeds p1 by at least 0.011 but not by more than 0.187.
On the surface, we would conclude that a daily dose of Vitamin C decreases
a French skier’s chance of developing a cold by between 0.011 and 0.187 (with
95% confidence). This conclusion was somewhat controversial. Several reviews
of the study felt that the experimenter’s evaluations of cold symptoms were
unreliable. Many other studies refute the benefit of Vitamin C as a treatment
for the common cold.
#### Example, vitamin C
# Approximate normal test for two-proportions, without Yates' continuity correction
prop.test(c(17, 31), c(139, 140), correct = FALSE)
##
## 2-sample test for equality of proportions without continuity
## correction
##
## data: c(17, 31) out of c(139, 140)
## alternative hypothesis: two.sided
## -0.18686 -0.01139
## prop 1 prop 2
## 0.1223 0.2214
Conditional probability
In probability theory, a conditional probability is the probability that an event
will occur, when another event is known to occur or to have occurred. If
the events are A and B respectively, this is said to be “the probability of A
given B”. It is commonly denoted by Pr(A|B). Pr(A|B) may or may not be
equal to Pr(A), the probability of A. If they are equal, A and B are said to
be independent. For example, if a coin is flipped twice, “the outcome of the
second flip” is independent of “the outcome of the first flip”.
In the Vitamin C example above, the unconditional observed probability of
contracting a cold is Pr(cold) = (17 + 31)/(139 + 140) = 0.172. The condi-
tional observed probabilities are Pr(cold|ascorbic acid) = 17/139 = 0.1223 and
Pr(cold|placebo) = 31/140 = 0.2214. The two-sample test of H0 : p1 = p2
where p1 = Pr(cold|ascorbic acid) and p2 = Pr(cold|placebo) is effectively test-
ing whether Pr(cold) = Pr(cold|ascorbic acid) = Pr(cold|placebo). This tests
whether contracting a cold is independent of the vitamin C treatment.
Example, cervical dysplasia A case-control study was designed to ex-

amine risk factors for cervical dysplasia2 (Becker et al. 1994). All the women
in the study were patients at UNM clinics. The 175 cases were women, aged
18-40, who had cervical dysplasia. The 308 controls were women aged 18-40
who did not have cervical dysplasia. Each women was classified as positive or
negative, depending on the presence of HPV (human papilloma virus). The
data are summarized below.
HPV Outcome Cases Controls
Positive 164 130
Negative 11 178
Sample size 175 308
Let p1 be the probability that a case is HPV positive and let p2 be the
probability that a control is HPV positive. The sample sizes are n1 = 175
and n2 = 308. The sample proportions of positive cases and controls are
p̂1 = 164/175 = 0.937 and p̂2 = 130/308 = 0.422.
For a 95% CI
r
0.937 × (1 − 0.937) 0.422 × (1 − 0.422)
zcritSECI (p̂1 − p̂2) = 1.96 +
175 308
= 1.96 × (0.03336) = 0.0659.
A 95% CI for p1 −p2 is (0.937−0.422)±0.066, or 0.515±0.066, or (0.449, 0.581).

I am 95% confident that p1 exceeds p2 by at least 0.45 but not by more than
0.58.
2
http://www.ncbi.nlm.nih.gov/pubmedhealth/PMH0002461/
# Approximate normal test for two-proportions, without Yates' continuity correction

prop.test(c(164, 130), c(175, 308), correct = FALSE)
##
## correction
##
## data: c(164, 130) out of c(175, 308)
## alternative hypothesis: two.sided
## 0.4492 0.5809
## prop 1 prop 2
## 0.9371 0.4221
Not surprisingly, a two-sided test at the 5% level would reject H0 : p1 = p2.

In this problem one might wish to do a one-sided test, instead of a two-sided
test. Let us carry out this test, as a refresher on how to conduct one-sided tests.
# one-sided test, are cases more likely to be HPV positive?
prop.test(c(164, 130), c(175, 308), correct = FALSE, alternative = "greater")
##
## correction
##
## data: c(164, 130) out of c(175, 308)
## alternative hypothesis: greater
## 0.4598 1.0000
## prop 1 prop 2
## 0.9371 0.4221
Appropriateness of Large Sample Test and CI
The standard two-sample CI and test used above are appropriate when each
sample is large. A rule of thumb suggests a minimum of at least five successes
(i.e., observations with the characteristic of interest) and failures (i.e., observa-
tions without the characteristic of interest) in each sample before using these
methods. This condition is satisfied in our two examples.
7.6: Effect Measures in Two-by-Two Tables 263
ClickerQ s — Comparing two proportions STT.08.02.010
7.6 Effect Measures in Two-by-Two Tables

Consider a study of a particular disease, where each individual is either exposed
or not-exposed to a risk factor. Let p1 be the proportion diseased among the
individuals in the exposed population, and p2 be the proportion diseased among
the non-exposed population. This population information can be summarized
as a two-by-two table of population proportions:
Outcome Exposed population Non-Exposed population
Diseased p1 p2
Non-Diseased 1 − p1 1 − p2
A standard measure of the difference between the exposed and non-exposed
populations is the absolute difference: p1 −p2. We have discussed statistical
methods for assessing this difference.
In many epidemiological and biostatistical settings, other measures of the
difference between populations are considered. For example, the relative risk
p1
RR =
p2
is commonly reported when the individual risks p1 and p2 are small. The odds
ratio
p1/(1 − p1)
OR =
p2/(1 − p2)
is another standard measure. Here p1/(1 − p1) is the odds of being diseased
in the exposed group, whereas p2/(1 − p2) is the odds of being diseased in the
non-exposed group.
I mention these measures because you may see them or hear about them.
Note that each of these measures can be easily estimated from data, using the
sample proportions as estimates of the unknown population proportions. For
example, in the vitamin C study:
Outcome Ascorbic Acid Placebo
# with cold 17 31
# with no cold 122 109
Totals 139 140
the proportion with colds in the placebo group is p̂2 = 31/140 = 0.221. The
proportion with colds in the vitamin C group is p̂1 = 17/139 = 0.122.
The estimated absolute difference in risk is p̂1 −p̂2 = 0.122−0.221 = −0.099.
The estimated risk ratio and odds ratio are
d = 0.122 = 0.55
RR
0.221
and
d = 0.122/(1 − 0.122) = 0.49,
OR
0.221/(1 − 0.221)
respectively.
Interpretting odds ratios, two examples Let’s begin with probabil-

ity3. Let’s say that the probability of success is 0.8, thus p = 0.8 Then the
probability of failure is q = 1 − p = 0.2 The odds of success are defined
as odds(success) = p/q = 0.8/0.2 = 4, that is, the odds of success are 4
to 1. The odds of failure would be odds(failure) = q/p = 0.2/0.8 = 0.25,
that is, the odds of failure are 1 to 4. Next, let’s compute the odds ratio
by OR = odds(success)/odds(failure) = 4/0.25 = 16 The interpretation of
this odds ratio would be that the odds of success are 16 times greater than
for failure. Now if we had formed the odds ratio the other way around with
odds of failure in the numerator, we would have gotten something like this,
OR = odds(failure)/odds(success) = 0.25/4 = 0.0625 Interestingly enough, the
interpretation of this odds ratio is nearly the same as the one above. Here the
interpretation is that the odds of failure are one-sixteenth the odds of success.
In fact, if you take the reciprocal of the first odds ratio you get 1/16 = 0.0625.
3
Borrowed graciously from UCLA Academic Technology Services at http://www.ats.ucla.edu/stat/
sas/faq/oratio.htm
7.7: Analysis of Paired Samples: Dependent Proportions 265
Another example This example is adapted from Pedhazur (1997). Sup-
pose that seven out of 10 males are admitted to an engineering school while
three of 10 females are admitted. The probabilities for admitting a male are,
p = 7/10 = 0.7 and q = 1 − 0.7 = 0.3. Here are the same probabilities
for females, p = 3/10 = 0.3 and q = 1 − 0.3 = 0.7. Now we can use
the probabilities to compute the admission odds for both males and females,
odds(male) = 0.7/0.3 = 2.33333 and odds(female) = 0.3/0.7 = 0.42857. Next,
we compute the odds ratio for admission, OR = 2.3333/0.42857 = 5.44. Thus,
the odds of a male being admitted are 5.44 times greater than for a female.
7.7 Analysis of Paired Samples: Dependent

Proportions
Paired and more general block analyses are appropriate with longitudinal
data collected over time and in medical studies where several treatments are
given to the same patient over time. A key feature of these designs that inval-
idates the two-sample method discussed earlier is that repeated observations
within a unit or individual are likely to be correlated, and not independent.
Example, President performance For example, in a random sample of

n = 1600 voter-age Americans, 944 said that they approved of the President’s
performance. One month later, only 880 of the original 1600 sampled approved.
The following two-by-two table gives the numbers of individuals with each of
the four possible sequences of responses over time. Thus, 150 voter-age Amer-
icans approved of the President’s performance when initially asked but then
disapproved one month later. The row and column totals are the numbers of
approvals and disapprovals for the two surveys (Agresti, 1990, p. 350).
(Obs Counts) Second survey
First Survey Approve Disapprove Total
Approve 794 150 944
Disapprove 86 570 656
Total 880 720 1600
Let pAA, pAD , pDA, pDD be the population proportion of voter-age Amer-
icans that fall into the four categories, where the subscripts preserve the time
ordering and indicate Approval or Disapproval. For example, pAD is the pop-
ulation proportion that approved of the President’s performance initially and
disapproved one month later. The population proportion that initially approved
is pA+ = pAA + pAD . The population proportion that approved at the time
of the second survey is p+A = pAA + pDA. The “+” sign used as a subscript
means that the replaced subscript has been summed over.
Similarly, let p̂AA, p̂AD , p̂DA, p̂DD be the sample proportion of voter-age
Americans that fall into the four categories, and let p̂A+ = p̂AA + p̂AD and
p̂+A = p̂AA + p̂DA be the sample proportion that approves the first month and
the sample proportion that approves the second month, respectively. The table
below summarizes the observed proportions. For example, p̂AA = 794/1600 =
0.496 and p̂A+ = 944/1600 = 0.496 + 0.094 = 0.590. The sample proportion
of voting-age Americans that approve of the President’s performance decreased
from one month to the next.
(Obs Proportions) Second survey
First Survey Approve Disapprove Total
Approve 0.496 0.094 0.590
Disapprove 0.054 0.356 0.410
Total 0.550 0.450 1.000
The difference in the population proportions from one month to the next
can be assessed by a large-sample CI for pA+ − p+A, given by (p̂A+ − p̂+A) ±
zcritSECI (p̂A+ − p̂+A), where the CI standard error satisfies
r
p̂A+(1 − p̂A+) + p̂+A(1 − p̂+A) − 2(p̂AAp̂DD − p̂AD p̂DA)
SECI (p̂A+−p̂+A) =
n
One-sided bounds are constructed in the usual way.

The −2(·) term in the standard error accounts for the dependence between
the samples at the two time points. If independent samples of size n had been
selected for the two surveys, then this term would be omitted from the standard
error, giving the usual two-sample CI.
7.7: Analysis of Paired Samples: Dependent Proportions 267
For a 95% CI in the Presidential survey,
r
0.590 × 0.410 + 0.550 × 0.450 − 2(0.496 × 0.356 − 0.094 × 0.054)
zcrit SECI (p̂A+ − p̂+A ) = 1.96
1600
= 1.96 × (0.0095) = 0.0186.
A 95% CI for pA+ − p+A is (0.590 − 0.550) ± 0.019, or (0.021, 0.059). You
are 95% confident that the population proportion of voter-age Americans that
approved of the President’s performance the first month was between 0.021
and 0.059 larger than the proportion that approved one month later. This
gives evidence of a decrease in the President’s approval rating.
A test of H0 : pA+ = p+A can be based on the CI for pA+ − p+A, or on a
standard normal approximation to the test statistic
p̂A+ − p̂+A
zs = ,
SEtest(p̂A+ − p̂+A)
where the test standard error is given by
r
p̂A+p̂+A − 2p̂AA
SEtest(p̂A+ − p̂+A) = .
n
The test statistic is often written in the simplified form
nAD − nDA
zs = √ ,
nAD + nDA
where the nij s are the observed cell counts. An equivalent form of this test,
based on comparing the square of zs to a chi-squared distribution with 1 degree
of freedom, is the well-known McNemar’s test for marginal homogeneity (or
symmetry) in the two-by-two table.
For example, in the Presidential survey
150 − 86
zs = √ = 4.17.
150 + 86
The p-value for a two-sided test is, as usual, the area under the standard normal
curve outside ±4.17. The p-value is less than 0.001, suggesting that H0 is false.
R can perform this test as McNemar’s test.
#### Example, President performance

# McNemar's test needs data as a matrix
# Presidential Approval Ratings.

# Approval of the President's performance in office in two surveys,
# one month apart, for a random sample of 1600 voting-age Americans.
pres <-
matrix(c(794, 150, 86, 570),
nrow = 2, byrow = TRUE,
dimnames = list("1st Survey" = c("Approve", "Disapprove"),
"2nd Survey" = c("Approve", "Disapprove")))
pres
## 2nd Survey
## 1st Survey Approve Disapprove
## Approve 794 150
## Disapprove 86 570
mcnemar.test(pres, correct=FALSE)
##
## McNemar's Chi-squared test
##
## data: pres
## McNemar's chi-squared = 17.36, df = 1, p-value = 3.099e-05
# => significant change (in fact, drop) in approval ratings
7.8 Testing for Homogeneity of Proportions

Example, cancer deaths The following two-way table of counts summa-
rizes the location of death and age at death from a study of 2989 cancer deaths
(Public Health Reports, 1983).
(Obs Counts) Location of death
Age Home Acute Care Chronic care Row Total
15-54 94 418 23 535
55-64 116 524 34 674
65-74 156 581 109 846
75+ 138 558 238 934
Col Total 504 2081 404 2989
The researchers want to compare the age distributions across locations. A
one-way ANOVA would be ideal if the actual ages were given. Because the ages
are grouped, the data should be treated as categorical. Given the differences in
7.8: Testing for Homogeneity of Proportions 269
numbers that died at the three types of facilities, a comparison of proportions
or percentages in the age groups is appropriate. A comparison of counts is not.
The table below summarizes the proportion in the four age groups by loca-
tion. For example, in the acute care facility 418/2081 = 0.201 and 558/2081 =
0.268. The pooled proportions are the Row Totals divided by the total
sample size of 2989. The pooled summary gives the proportions in the four age
categories, ignoring location of death.
The age distributions for home and for the acute care facilities are similar,
but are very different from the age distribution at chronic care facilities.
To formally compare the observed proportions, one might view the data
as representative sample of ages at death from the three locations. Assum-
ing independent samples from the three locations (populations), a chi-squared
statistic is used to test whether the population proportions of ages at death are
identical (homogeneous) across locations. The chi-squared test for homo-
geneity of population proportions can be defined in terms of proportions, but
is traditionally defined in terms of counts.
(Proportions) Location of death
Age Home Acute Care Chronic care Pooled
15-54 0.187 0.201 0.057 0.179
55-64 0.230 0.252 0.084 0.226
65-74 0.310 0.279 0.270 0.283
75+ 0.273 0.268 0.589 0.312
Total 1.000 1.000 1.000 1.000
In general, assume that the data are independent samples from c populations
(strata, groups, sub-populations), and that each individual is placed into one of
r levels of a categorical variable. The raw data will be summarized as a r × c
contingency table of counts, where the columns correspond to the samples,
and the rows are the levels of the categorical variable. In the age distribution
problem, r = 4 and c = 3.
To implement the test:
1. Compute the (estimated) expected count for each cell in the table as
follows:
Row Total × Column Total
E= .
Total Sample Size
2. Compute the Pearson test statistic
X (O − E)2
χ2s = ,
E
all cells
where O is the observed count.

3. For a size α test, reject the hypothesis of homogeneity if χ2s ≥ χ2crit, where
χ2crit is the upper α critical value from the chi-squared distribution with
df = (r − 1)(c − 1).
The p-value for the chi-squared test of homogeneity is equal to the area under
the chi-squared curve to the right of χ2s.
For a two-by-two table of counts, the chi-squared test of homo-
geneity of proportions is identical to the two-sample proportion
test we discussed earlier.
The (estimated) expected counts for the (15-54, Home) cell and for the (75+,
Acute Care) cell in the age distribution data are E = 535 × 504/2989 = 90.21
and E = 934 × 2081/2989 = 650.27, respectively. The other expected counts
were computed similarly, and are summarized below. The row and column
sums on the tables of observed and expected counts always agree.
(Exp Counts) Location of death
Age Home Acute Care Chronic care Row Total
15-54 90.21 372.48 72.31 535
55-64 113.65 469.25 91.10 674
65-74 142.65 589 114.35 846
75- 157.49 650.27 126.24 934
Col Total 504 2081 404 2989
Why is a comparison of the observed and expected counts relevant for testing
homogeneity? To answer this question, first note that the expected cell count
can be expressed
E = Col Total × Pooled proportion for category.
For example, E = 504×0.179 = 90.21 in the (15-54, Home) cell. A comparison

of the observed and expected counts is a comparison of the observed category
proportions in a location with the pooled proportions, taking the size of each
sample into consideration. Thinking of the pooled proportion as a weighted
average of the sample proportions for a category, the Pearson χ2s statistic is
an aggregate measure of variation in the observed proportions across samples.
If the category proportions are similar across samples then the category and
pooled proportions are similar, resulting in a “small” value of χ2s. Large values of
χ2s occur when there is substantial variation in the observed proportions across
samples, in one or more categories. In this regard, the Pearson statistic is similar
to the F -statistic in a one-way ANOVA (where large differences between groups
result in a large F -statistic).
#### Example, cancer deaths
candeath <-
matrix(c( 94, 418, 23,
116, 524, 34,
156, 581, 109,
138, 558, 238),
dimnames = list("Age" = c("15-54", "55-64", "65-74", "75+"),
"Location of death" = c("Home", "Acute Care", "Chronic care")))
candeath
## Location of death
## Age Home Acute Care Chronic care
## 15-54 94 418 23
## 55-64 116 524 34
## 65-74 156 581 109
## 75+ 138 558 238
chisq.summary <- chisq.test(candeath, correct=FALSE)
chisq.summary
##
## Pearson's Chi-squared test
##
## data: candeath
# The Pearson residuals
chisq.summary$residuals
## 15-54 0.3990 2.3587 -5.7989
## 55-64 0.2206 2.5274 -5.9824
## 65-74 1.1177 -0.3297 -0.5001
## 75+ -1.5530 -3.6183 9.9467
# The sum of the squared residuals is the chi-squared statistic:
chisq.summary$residuals^2
## 15-54 0.15916 5.5636 33.6274
## 55-64 0.04865 6.3875 35.7888

## 65-74 1.24916 0.1087 0.2501
## 75+ 2.41184 13.0924 98.9369
sum(chisq.summary$residuals^2)
## [1] 197.6
A visualization of the Pearson residuals is available with a mosaic() plot in

the vcd package. Extended mosaic and association plots are each helpful meth-
ods of visualing complex data and evaluating deviations from a specified inde-
pendence model. For extended mosaic plots, use mosaic(x, condvar=, data=)
where x is a table or formula, condvar= is an optional conditioning variable, and
data= specifies a data frame or a table. Include shade=TRUE to color the figure,
and legend=TRUE to display a legend for the Pearson residuals.
# mosaic plot
library(vcd)
mosaic(candeath, shade=TRUE, legend=TRUE)
# association plot
library(vcd)
assoc(candeath, shade=TRUE)
Location of death
Home Acute Care Chronic care
Location of death Pearson
Home Acute Care Chronic care residuals:
15−54
Pearson 9.9
15−54
residuals:
9.9
55−64
55−64
4.0
4.0
Age
Age
2.0 2.0
65−74
65−74
0.0
0.0
−2.0
−2.0
−4.0
75+
75+
−6.0 −4.0
p−value =
<2e−16 −6.0
p−value =
<2e−16
The vcd package provides a variety of methods for visualizing multivariate

categorical data, inspired by Michael Friendly’s wonderful “Visualizing Cate-
gorical Data”. For more details, see The Strucplot Framework4.
4
http://cran.r-project.org/web/packages/vcd/vignettes/strucplot.pdf
For example, a sieve plot for an n-way contingency table plots rectangles
with areas proportional to the expected cell frequencies and filled with a number
of squares equal to the observed frequencies. Thus, the densities visualize the
deviations of the observed from the expected values.
# sieve plot
library(vcd)
# plot expected values

sieve(candeath, sievetype = "expected", shade = TRUE)
# plot observed table, then label cells with observed values in the cells
sieve(candeath, pop = FALSE, shade = TRUE)
labeling_cells(text = candeath, gp_text = gpar(fontface = 2))(as.table(candeath))
Location of death Location of death
Home Acute Care Chronic care Home Acute Care Chronic care
15−54
15−54
94 418 23
55−64
55−64
116 524 34
Age
Age
65−74
65−74
156 581 109

75+
75+
138 558 238
7.8.1 Adequacy of the Chi-Square Approximation

The chi-squared tests for homogeneity and independence are large-sample tests.
As with the goodness-of-fit test, a simple rule-of-thumb is that the approxima-
tion is adequate when the expected (not observed) cell counts are 5 or more.
This rule is conservative, and some statisticians argue that the approximation
is valid for expected counts as small as one.
In practice, the chi-squared approximation to χ2s tends to be a bit conserva-
tive, meaning that statistically significant results would likely retain significance
had a more accurate approximation been used.
R may not print out a warning message whenever a noticeable percentage
of cells have expected counts less than 5. Ideally, one would use Fisher’s exact
test (fisher.test()) for tables with small counts, and can be used for larger
than 2-by-2 tables when the frequencies are not too large.
7.9 Testing for Homogeneity in Cross-Sectional

and Stratified Studies
Two-way tables of counts are often collected using either stratified sampling
or cross-sectional sampling.
In a stratified design, distinct groups, strata, or sub-populations are
identified. Independent samples are selected from each group, and the sampled
individuals are classified into categories. The Indian gaming example is an
illustration of a stratified design (where the two strata were NM and CO voters).
Stratified designs provide estimates for the strata (population) proportion in
each of the categories. A test for homogeneity of proportions is used to
compare the strata.
In a cross-sectional design, individuals are randomly selected from a
population and classified by the levels of two categorical variables. With cross-
sectional samples you can test homogeneity of proportions by comparing either
the row proportions or by comparing the column proportions.
Example, antismoking adverts The following data (The Journal of

Advertising, 1983, pp. 34–42) are from a cross-sectional study that involved
soliciting opinions on anti-smoking advertisements. Each subject was asked
whether they smoked and their reaction (on a five-point ordinal scale) to the
ad. The data are summarized as a two-way table of counts, given below:
Str. Dislike Dislike Neutral Like Str. Like Row Tot
Smoker 8 14 35 21 19 97
Non-smoker 31 42 78 61 69 281
Col Total 39 56 113 82 88 378
7.9: Testing for Homogeneity in Cross-Sectional and Stratified Studies 275
The row proportions (i.e., fix a row and compute the proportions for the column
categories) are
(Row Prop) Str. Dislike Dislike Neutral Like Str. Like Row Tot
Smoker 0.082 0.144 0.361 0.216 0.196 1.000
Non-smoker 0.110 0.149 0.278 0.217 0.245 1.000
For example, the entry for the (Smoker, Str. Dislike) cell is: 8/97 = 0.082.
Similarly, the column proportions are
(Col Prop) Str. Dislike Dislike Neutral Like Str. Like
Smoker 0.205 0.250 0.310 0.256 0.216
Non-smoker 0.795 0.750 0.690 0.744 0.784
Total 1.000 1.000 1.000 1.000 1.000
Although it may be more natural to compare the smoker and non-smoker
row proportions, the column proportions can be compared across ad responses.
There is no advantage to comparing “rows” instead of “columns” in a formal
test of homogeneity of proportions with cross-sectional data. The Pearson chi-
squared test treats the rows and columns interchangeably, so you get the same
result regardless of how you view the comparison. However, one of the two
comparisons may be more natural to interpret.
Note that checking for homogeneity of proportions is mean-
ingful in stratified studies only when the comparison is between
strata! Further, if the strata correspond to columns of the table, then the
column proportions or percentages are meaningful whereas the row proportions
are not.
7.9.1 Testing for Independence in a Two-Way Con-

tingency Table
The row and column classifications for a population where each individual is
cross-classified by two categorical variables are said to be independent if each
population cell proportion in the two-way table is the product of the propor-
tion in a given row and the proportion in a given column. One can show that
independence is equivalent to homogeneity of proportions. In particular, the
two-way table of population cell proportions satisfies independence if and only if
the population column proportions are homogeneous. If the population column
proportions are homogeneous then so are the population row proportions.
This suggests that a test for independence or no association between
two variables based on a cross-sectional study can be implemented using the
chi-squared test for homogeneity of proportions. This suggestion is correct.
If independence is not plausible, I interpret the dependence as a deviation
from homogeneity, using the classification for which the interpretation is most
natural.
The Pearson chi-squared test of independence is not significant (p-value =
0.56). The observed association between smoking status and the ad reaction is
not significant. This suggests, for example, that the smoker’s reactions to the
ad were not statistically significantly different from the non-smoker’s reactions,
which is consistent with the smokers and non-smokers attitudes being fairly
similar.
#### Example, antismoking adverts
antismokead <-
matrix(c( 8, 14, 35, 21, 19,
31, 42, 78, 61, 69),
dimnames = list(
"Status" = c("Smoker", "Non-smoker"),
"Reaction" = c("Str. Dislike", "Dislike", "Neutral", "Like", "Str. Like")))
antismokead
## Reaction
## Status Str. Dislike Dislike Neutral Like Str. Like
## Smoker 8 14 35 21 19
## Non-smoker 31 42 78 61 69
chisq.summary <- chisq.test(antismokead, correct=FALSE)
chisq.summary
##
##
## data: antismokead
# All expected frequencies are at least 5
chisq.summary$expected
## Reaction
## Smoker 10.01 14.37 29 21.04 22.58
## Non-smoker 28.99 41.63 84 60.96 65.42
# Contribution to chi-squared statistic
chisq.summary$residuals^2
## Reaction
## Smoker 0.4029 0.009546 1.2426 8.515e-05 0.5682

## Non-smoker 0.1391 0.003295 0.4289 2.939e-05 0.1961
7.9.2 Further Analyses in Two-Way Tables

The χ2s statistic is a summary measure of independence or homogeneity. A
careful look at the data usually reveals the nature of the association or het-
erogeneity when the test is significant. There are numerous meaningful ways
to explore two-way tables to identify sources of association or heterogeneity.
For example, in the comparison of age distributions across locations, you might
consider the 4 × 2 tables comparing all possible pairs of locations. Another
possibility would be to compare the proportion in the 75+ age category across
locations. For the second comparison you need a 2 × 3 table of counts, where
the two rows correspond to the individuals less than 75 years old and those
75+ years old, respectively (i.e., collapse the first three rows of the original
4 × 2 table). The possibilities are almost limitless in large tables. Of course,
theoretically generated comparisons are preferred to data dredging.
Example: drugs and nausea, testing for homogeneity A random-

ized double-blind experiment compared the effectiveness of several drugs in
reducing postoperative nausea. All patients were anesthetized with nitrous ox-
ide and ether. The following table shows the incidence of nausea during the
first four postoperative hours of four drugs and a placebo. Compare the drugs
to each other and to the placebo.
Drug # with Nausea # without Nausea Sample Size
Placebo 96 70 166
Chlorpromazine 52 100 152
Dimenhydrinate 52 33 85
Pentobarbitol (100mg) 35 32 67
Pentobarbitol (150mg) 37 48 85
Let pP L be the probability of developing nausea given a placebo, and define
pCH , pDI , pP E100, and pP E150 analogously. A simple initial analysis would be
to test homogeneity of proportions: H0 : pP L = pCH = pDI = pP E100 = pP E150
against HA : not H0.
The data were entered as frequencies. The output shows that the proportion
of patients exhibiting nausea (see the column percents — the cell and row
percentages are not interpretable, so they are omitted) is noticeably different
across drugs. In particular, Chlorpromazine is the most effective treatment with
p̂CH = 0.34 and Dimenhydrinate is the least effective with p̂DI = 0.61.
The p-value for the chi-squared test is 0.00005, which leads to rejecting H0
at the 0.05 or 0.01 levels. The data strongly suggest there are differences in the
effectiveness of the various treatments for postoperative nausea.
#### Example: drugs and nausea, testing for homogeneity
nausea <-
matrix(c(96, 70, 52, 100, 52, 33, 35, 32, 37, 48),
dimnames = list("Drug" = c("PL", "CH", "DI", "PE100", "PE150"),
"Result" = c("Nausea", "No Nausea")))
nausea
## Result
## Drug Nausea No Nausea
## PL 96 70
## CH 52 100
## DI 52 33
## PE100 35 32
## PE150 37 48
# Sorted proportions of nausea by drug
nausea.prop <- sort(nausea[,1]/rowSums(nausea))
nausea.prop
## CH PE150 PE100 PL DI
## 0.3421 0.4353 0.5224 0.5783 0.6118
# chi-sq test of association
chisq.summary <- chisq.test(nausea, correct=FALSE)
chisq.summary
##
##
## data: nausea
## X-squared = 24.83, df = 4, p-value = 5.451e-05
# All expected frequencies are at least 5
chisq.summary$expected
## Result
## Drug Nausea No Nausea
## PL 81.35 84.65
## CH 74.49 77.51
## DI 41.66 43.34
## PE100 32.84 34.16
## PE150 41.66 43.34
A sensible follow-up analysis is to identify which treatments were responsible
for the significant differences. For example, the placebo and chlorpromazine can
be compared using a test of pP L = pCH or with a CI for pP L − pCH .
In certain experiments, specific comparisons are of interest, for example a
comparison of the drugs with the placebo. Alternatively, all possible compar-
isons might be deemed relevant. The second case is suggested here based on
the problem description. I will use a Bonferroni adjustment to account for the
multiple comparisons. The Bonferroni adjustment accounts for data dredging,
but at a cost of less sensitive comparisons.
There are 10 possible comparisons here. The Bonferroni analysis with an
overall Family Error Rate of 0.05 (or less) tests the 10 individual hypotheses at
the 0.05/10=0.005 level.
nausea.table <- data.frame(Interval = rep(NA,10)
, CI.lower = rep(NA,10)
, CI.upper = rep(NA,10)
, Z = rep(NA,10)
, p.value = rep(NA,10)
, sig.temp = rep(NA,10)
, sig = rep(NA,10))
# row names for table
nausea.table[,1] <- c("p_PL - p_CH"
, "p_PL - p_DI"
, "p_PL - p_PE100"
, "p_PL - p_PE150"
, "p_CH - p_DI"
, "p_CH - p_PE100"
, "p_CH - p_PE150"
, "p_DI - p_PE100"
, "p_DI - p_PE150"
, "p_PE100 - p_PE150")
# test results together in a table
i.tab <- 0
for (i in 1:4) {
for (j in (i+1):5) {
i.tab <- i.tab + 1
nausea.summary <- prop.test(nausea[c(i,j),], correct = FALSE, conf.level = 1-0.05/10)
nausea.table[i.tab, 2:6] <- c(nausea.summary$conf.int[1]
, nausea.summary$conf.int[2]
, sign(-diff(nausea.summary$estimate)) * nausea.summary$statistic^0.5
, nausea.summary$p.value
, (nausea.summary$p.value < 0.05/10))
if (nausea.table$sig.temp[i.tab] == 1) { nausea.table$sig[i.tab] <- "*" }
else { nausea.table$sig[i.tab] <- " " }
}
}
# remove temporary sig indicator

nausea.table <- subset(nausea.table, select = -sig.temp)
#nausea.table
The following table gives two-sample tests of proportions with nausea and
99.5% CIs for the differences between the ten pairs of proportions. The only
two p-values are less than 0.005 corresponding to pP L − pCH and pCH − pDI .
I am 99.5% confident that pCH is between 0.084 and 0.389 less than pP L, and
I am 99.5% confident that pCH is between 0.086 and 0.453 less than pDI . The
other differences are not significant.
Interval CI.lower CI.upper Z p.value sig
1 p PL - p CH 0.0838 0.3887 4.2182 0.0000 *
2 p PL - p DI −0.2167 0.1498 −0.5099 0.6101
3 p PL - p PE100 −0.1464 0.2582 0.7788 0.4361
4 p PL - p PE150 −0.0424 0.3284 2.1485 0.0317
5 p CH - p DI −0.4532 −0.0861 −4.0122 0.0001 *
6 p CH - p PE100 −0.3828 0.0222 −2.5124 0.0120
7 p CH - p PE150 −0.2788 0.0924 −1.4208 0.1554
8 p DI - p PE100 −0.1372 0.3160 1.1058 0.2688
9 p DI - p PE150 −0.0352 0.3881 2.3034 0.0213
10 p PE100 - p PE150 −0.1412 0.3154 1.0677 0.2857
Using ANOVA-type groupings, and arranging the treatments from most to
least effective (low proportions to high), we get:
CH (0.34) PE150 (0.44) PE100 (0.52) PL (0.58) DI (0.61)
---------------------------------------
---------------------------------------------------
Chapter 8
Correlation and
Regression
Learning objectives
select graphical displays that reveal the relationship between two continu-
ous variables.
summarize model fit.
interpret model parameters, such as slope and R2.
assess the model assumptions visually and numerically.
8.1 Introduction
Suppose we select n = 10 people from the population of college seniors who
plan to take the medical college admission test (MCAT) exam. Each takes the
test, is coached, and then retakes the exam. Let Xi be the pre-coaching score
and let Yi be the post-coaching score for the ith individual, i = 1, 2, . . . , n.
282 Ch 8: Correlation and Regression
There are several questions of potential interest here, for example: Are Y and
X related (associated), and how? Does coaching improve your MCAT score?
Can we use the data to develop a mathematical model (formula) for predicting
post-coaching scores from the pre-coaching scores? These questions can be
addressed using correlation and regression models.
The correlation coefficient is a standard measure of association or
relationship between two features Y and X. Most scientists equate Y and X
being correlated to mean that Y and X are associated, related, or dependent
upon each other. However, correlation is only a measure of the strength of a
linear relationship. For later reference, let ρ be the correlation between Y
and X in the population and let r be the sample correlation. I define r below.
The population correlation is defined analogously from population data.
Suppose each of n sampled individuals is measured on two quantitative
characteristics called Y and X. The data are pairs of observations (X1, Y1),
(X2, Y2), . . ., (Xn, Yn), where (Xi, Yi) is the (X, Y ) pair for the ith individual in
the sample. The sample correlation between Y and X, also called the Pearson
product moment correlation coefficient, is
P
SXY (Xi − X̄)(Yi − Ȳ )
r= = pP i ,
SX SY 2
P 2
i (Xi − X̄) i (Yi − Ȳ )
where Pn
i=1 (Xi − X̄)(Yi − Ȳ )
SXY =
n−1
pP
2
is the sample
pPcovariance between Y and X, and SY = i (Yi − Ȳ ) /(n − 1)
2
and SX = i (Xi − X̄) /(n − 1) are the standard deviations for the Y and
X samples.
Important properties of r:
1. −1 ≤ r ≤ 1.
2. If Yi tends to increase linearly with Xi then r > 0.
3. If Yi tends to decrease linearly with Xi then r < 0.
4. If there is a perfect linear relationship between Yi and Xi with a positive
slope then r = +1.
8.1: Introduction 283
5. If there is a perfect linear relationship between Yi and Xi with a negative
slope then r = −1.
6. The closer the points (Xi, Yi) come to forming a straight line, the closer
r is to ±1.
7. The magnitude of r is unchanged if either the X or Y sample is trans-
formed linearly (such as feet to inches, pounds to kilograms, Celsius to
Fahrenheit).
8. The correlation does not depend on which variable is called Y and which
is called X.
If r is near ±1, then there is a strong linear relationship between Y and X
in the sample. This suggests we might be able to accurately predict Y from X
with a linear equation (i.e., linear regression). If r is near 0, there is a weak
linear relationship between Y and X, which suggests that a linear equation
provides little help for predicting Y from X. The pictures below should help
you develop a sense about the size of r.
Note that r = 0 does not imply that Y and X are not related in the sample.
It only implies they are not linearly related. For example, in the last plot r = 0
yet Yi = Xi2, exactly.
Correlation=1 Correlation=-1 Correlation=.3 Correlation=-.3
• • • •
•
•• ••• •• • •
••• • • • •• •
••• • ••••
•••• •
• • • • • • • •
•••••••
• •••••• • •• •• •• ••••• • • • • •• •• •
•
••••••
•
• •••• • • • • ••• • • •• ••• • •• ••• • • ••
• •
•••••• • ••• •• •• • • •• ••••• •••• • • • • • • •
•••••• ••••• •• ••• ••• •• • ••• • • • •••• ••• •• • • •
•••••• •• •
•••••• • • ••• • • ••• •• • • • • •• • • • •
•
•
•• •
•• •••• • • • ••• •
• • •• •• ••• •••• •••••• • • •
• •
•• •
• ••••
• •• • • • • • • •• • •• • •
• •• • • • • •
Correlation=.7 Correlation=-.7 Correlation=0 Correlation=0

0 20 40 60 80 100
• • • • • • • • •
• •
• •• • • •
• • ••
• • •• • • • • •••• • • •• • • • •
• • • ••• •• • • • • • • • • ••• • • •
•
• • • •• ••• •• • • • • • • •
• • • •••••••• •• •••••• •• •• • ••• • •• •• • ••••• ••• • •
• •• •••••••• • • • •• • •••• •••••••••• •••••••••••• •••••• • • •• •• •••••••• •••••• • • •
• •• •••• • • • • •• •••• • • • • • • •
• •• •• • • • •••• •••• •• •• •
••• • ••
• • • ••• • • •• • •• • • •
• •• •••• •
•
• •
•
•
•
• •• • •
• •• •• • •
• • •• •
• • • ••••
-10 -5 0 5 10
ClickerQ s — Correlation coefficients STT.02.02.010

ClickerQ s — Strong correlation STT.02.02.020
8.2 Testing that ρ = 0

Suppose you want to test H0 : ρ = 0 against HA : ρ 6= 0, where ρ is the
population correlation between Y and X. This test is usually interpreted as
a test of no association, or relationship, between Y and X in the population.
Keep in mind, however, that ρ measures the strength of a linear relationship.
The standard test of H0 : ρ = 0 is based on the magnitude of r. If we let
r
n−2
ts = r ,
1 − r2
then the test rejects H0 in favor of HA if |ts| ≥ tcrit, where tcrit is the two-sided
test critical value from a t-distribution with df = n − 2. The p-value for the
test is the area under the t-curve outside ±ts (i.e., two-tailed test p-value).
This test assumes that the data are a random sample from a bivariate
normal population for (X, Y ). This assumption implies that all linear com-
binations of X and Y , say aX + bY , are normal. In particular, the (marginal)
population frequency curves for X and Y are normal. At a minimum, you
should make boxplots of the X and Y samples to check marginal normality.
For large-sized samples, a plot of Y against X should be roughly an elliptical
cloud, with the density of the points decreasing as the points move away from
the center of the cloud.
8.2.1 The Spearman Correlation Coefficient

The Pearson correlation r can be highly influenced by outliers in one or both
samples. For example, r ≈ −1 in the plot below. If you delete the one extreme
case with the largest X and smallest Y value then r ≈ 0. The two analyses
are contradictory. The first analysis (ignoring the plot) suggests a strong linear
relationship, whereas the second suggests the lack of a linear relationship. I
will not strongly argue that you should (must?) delete the extreme case, but I
am concerned about any conclusion that depends heavily on the presence of a
single observation in the data set.
8.2: Testing that ρ = 0 285
• • •• • •
••••••••••••••••••
• ••••••••••••••••••••••••••••
• •••• •
•• •••••••••••••••••••••••• •
0
• ••••••••••••••••••••••••••••
• •• •• •
-2
Y
-4
-6
•
-8
0 2 4 6 8 10
X
Spearman’s rank correlation coefficient rS is a sensible alternative

to r when normality is unreasonable or outliers are present. Most books give
a computational formula for rS . I will verbally describe how to compute rS .
First, order the Xis and assign them ranks. Then do the same for the Yis and
replace the original data pairs by the pairs of ranked values. The Spearman
rank correlation is the Pearson correlation computed from the pairs of ranks.
The Spearman correlation rS estimates the population rank correla-
tion coefficient, which is a measure of the strength of linear relationship
between population ranks. The Spearman correlation, as with other rank-
based methods, is not sensitive to the presence of outliers in the data (or any
information about the marginal distribution of X or Y ). In the plot above,
rS ≈ 0 whether the unusual point is included or excluded from the analysis. In
samples without unusual observations and a linear trend, you often find that
the Spearman and Pearson correlations are similar, rS ≈ r.
An important point to note is that the magnitude of the Spearman correla-
tion does not change if either X or Y or both are transformed (monotonically).
Thus, if rS is noticeably greater than r, a transformation of the data might
provide a stronger linear relationship.
Example: Blood loss Eight patients underwent a thyroid operation. Three
variables were measured on each patient: weight in kg, time of operation in min-
utes, and blood loss in ml. The scientists were interested in the factors that
influence blood loss.
#### Example: Blood loss
thyroid <- read.table(text="
weight time blood_loss
44.3 105 503
40.6 80 490
69.0 86 471
43.7 112 505
50.3 109 482
50.2 100 490
35.4 96 513
52.2 120 464
", header=TRUE)

str(thyroid)
## $ weight : num 44.3 40.6 69 43.7 50.3 50.2 35.4 52.2
## $ time : int 105 80 86 112 109 100 96 120
## $ blood_loss: int 503 490 471 505 482 490 513 464
#thyroid
weight time blood loss

1 44.3 105 503
2 40.6 80 490
3 69.0 86 471
4 43.7 112 505
5 50.3 109 482
6 50.2 100 490
7 35.4 96 513
8 52.2 120 464
Below, we calculate the Pearson correlations between all pairs of variables

(left), as well as the p-values (right) for testing whether the correlation is equal
to zero.
p.corr <- cor(thyroid);
#p.corr
# initialize pvalue table with dim names

p.corr.pval <- p.corr;
for (i1 in 1:ncol(thyroid)) {
p.corr.pval[i1,i2] <- cor.test(thyroid[, i1], thyroid[, i2])$p.value
}
}
#p.corr.pval
weight time blood loss weight time blood loss

weight 1.000 −0.066 −0.772 weight 0.0000 0.8761 0.0247
time −0.066 1.000 −0.107 time 0.8761 0.0000 0.8003
blood loss −0.772 −0.107 1.000 blood loss 0.0247 0.8003 0.0000
8.2: Testing that ρ = 0 287
Similarly, we calculate the Spearman (rank) correlation table (left), as well
as the p-values (right) for testing whether the correlation is equal to zero.
s.corr <- cor(thyroid, method = "spearman");
#s.corr
# initialize pvalue table with dim names

s.corr.pval <- p.corr;
s.corr.pval[i1,i2] <- cor.test(thyroid[, i1], thyroid[, i2],
method = "spearman")$p.value
}
}
## Warning: Cannot compute exact p-value with ties
#s.corr.pval
weight time blood loss weight time blood loss

weight 1.000 0.286 −0.874 weight 0.0000 0.5008 0.0045
time 0.286 1.000 −0.156 time 0.5008 0.0000 0.7128
blood loss −0.874 −0.156 1.000 blood loss 0.0045 0.7128 0.0000
Here are scatterplots for the original data and the ranks of the data using
ggpairs() from the GGally package with ggplot2.
library(ggplot2)
library(GGally)
## Loading required package: reshape
##
## Attaching package: ’reshape’
##
## The following object is masked from ’package:class’:
##
## condense
##
## The following objects are masked from ’package:plyr’:
##
## rename, round any
##
## The following objects are masked from ’package:reshape2’:
##
## colsplit, melt, recast
p1 <- ggpairs(thyroid)
print(p1)
thyroid$rank_weight <- rank(thyroid$weight )

thyroid$rank_time <- rank(thyroid$time )
thyroid$rank_blood_loss <- rank(thyroid$blood_loss)
p2 <- ggpairs(thyroid[,4:6])
print(p2)
# detach package after use so reshape2 works (old reshape (v.1) conflicts)
detach("package:GGally", unload=TRUE)
detach("package:reshape", unload=TRUE)
70 8
60 6
weight Corr: Corr: rank_weight Corr: Corr:
50 −0.0663 −0.772 4 0.286 −0.874
40 50 60 70 2 4 6 8
●
120 ●
8
●
●
● 110 ● 6
●
time Corr: ●
rank_time Corr:
● 100
● −0.107 ● 4 −0.156
●
90
● ● 2 4 6 8
● 80 90 100 110 120 ●
● ● ● ●
8
510
● ● ● ●
● ●
500 ● ● 6
● ● ● ● 490 blood_loss ● ● ● ● rank_blood_loss
● ●
4
480 ● ●
● ●
470 480 490 500 510 ● ● 2 4 6 8
● ● ● ●
Comments:
1. (Pearson correlations). Blood loss tends to decrease linearly as weight
increases, so r should be negative. The output gives r = −0.77. There is
not much of a linear relationship between blood loss and time, so r should
be close to 0. The output gives r = −0.11. Similarly, weight and time
have a weak negative correlation, r = −0.07.
2. The Pearson and Spearman correlations are fairly consistent here. Only
the correlation between blood loss and weight is significantly different from
zero at the α = 0.05 level (the p-values are given below the correlations).
3. (Spearman p-values) R gives the correct p-values. Calculating the p-value
using the Pearson correlation on the ranks is not correct, strictly speaking.
8.3: Simple Linear Regression 289
ClickerQ s — Coney Island STT.02.02.050
8.3 Simple Linear Regression
In linear regression, we are interested in developing a linear equation that best

summarizes the relationship in a sample between the response variable Y
and the predictor variable (or independent variable) X. The equation
is also used to predict Y from X. The variables are not treated symmetrically in
regression, but the appropriate choice for the response and predictor is usually
apparent.
8.3.1 Linear Equation
If there is a perfect linear relationship between Y and X then Y = β0 + β1X

for some β0 and β1, where β0 is the Y-intercept and β1 is the slope of the line.
Two plots of linear relationships are given below. The left plot has β0 = 5 and
β1 = 3. The slope is positive, which indicates that Y increases linearly when
X increases. The right plot has β0 = 7 and β1 = −2. The slope is negative,
which indicates that Y decreases linearly when X increases.
The line Y = 5 + 3X The line Y = 7 - 2X
8
15
6
1
10
Y
Y
4
-2
2
5
0
-1 0 1 2 3 4 -1 0 1 2 3 4
X X
ClickerQ s — Equation STT.02.03.010
8.3.2 Least Squares

Data rarely, if ever, fall on a straight line. However, a straight line will often
describe the trend for a set of data. Given a data set, (Xi, Yi), i = 1, . . . , n,
with a linear trend, what linear equation “best” summarizes the observed
relationship between Y and X? There is no universally accepted definition of
“best”, but many researchers accept the Least Squares line (LS line) as a
reasonable summary.
Mathematically, the LS line chooses the values of β0 and β1 that minimize
n
X
{Yi − (β0 + β1Xi)}2
i=1
over all possible choices of β0 and β1. These values can be obtained using
calculus. Rather than worry about this calculation, note that the LS line makes
the sum of squared (vertical) deviations between the responses Yi and the line
as small as possible, over all possible lines. The LS line goes through the mean
8.3: Simple Linear Regression 291
point, (X̄, Ȳ ), which is typically in the “the heart” of the data, and is often
closely approximated by an eye-ball fit to the data.
19
•
18
•
17
Y
16
•
15
•
4.5 5.0 5.5
X
The equation of the LS line is
ŷ = b0 + b1X
where the intercept b0 satisfies
b0 = Ȳ − b1X̄
and the slope is P

i (Y − Ȳ )(Xi − X̄) SY
b1 = Pi 2
= r .
i (X i − X̄) SX
As before, r is the Pearson correlation between Y and X, whereas SY and SX

are the sample standard deviations for the Y and X samples, respectively. The
sign of the slope and the sign of the correlation are identical (i.e., +
correlation implies + slope).
Special symbols b0 and b1 identify the LS intercept and slope to distinguish
the LS line from the generic line Y = β0 + β1X. You should think of Ŷ as the
fitted value at X, or the value of the LS line at X.
Fit a regression for the equation estimates from summary(). Note that we’ll
reuse the output of lm() over and over again.
lm.blood.wt <- lm(blood_loss ~ weight, data = thyroid)
lm.blood.wt
##
## Call:
## lm(formula = blood_loss ~ weight, data = thyroid)
##
## Coefficients:
## (Intercept) weight
## 552.4 -1.3
# use summary() to get t-tests of parameters (slope, intercept)
summary(lm.blood.wt)
##
## Call:
##
## Residuals:
## -20.57 -6.19 4.71 8.19 9.38
##
## Coefficients:
## (Intercept) 552.442 21.441 25.77 2.3e-07 ***
## weight -1.300 0.436 -2.98 0.025 *
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
Create a scatterplot with regression fit.

# ggplot: Plot the data with linear regression fit and confidence bands
library(ggplot2)
p <- ggplot(thyroid, aes(x = weight, y = blood_loss))
p <- p + geom_smooth(method = lm, se = FALSE)
print(p)
# Base graphics: Plot the data with linear regression fit and confidence bands
# scatterplot
plot(thyroid$weight, thyroid$blood_loss)
# regression line from lm() fit
abline(lm.blood.wt)
8.4: ANOVA Table for Regression 293
510
●
510
●
●
500 ●
500
thyroid$blood_loss
blood_loss
490 ● ●
490
● ●
● ●
480
480
470
●
●
470
●
35 40 45 50 55 60 65 70
40 50 60 70 thyroid$weight
weight
For the thyroid operation data with Y = Blood loss in ml and X =

Weight in kg, the LS line is Ŷ = 552.44 − 1.30X, or Predicted Blood Loss =
552.44 − 1.30 Weight. For an 86kg individual, the Predicted Blood Loss =
552.44 − 1.30 × 86 = 440.64ml.
The LS regression coefficients for this model are interpreted as follows. The
intercept b0 is the predicted blood loss for a 0 kg individual. The intercept has
no meaning here. The slope b1 is the predicted increase in blood loss for each
additional kg of weight. The slope is −1.30, so the predicted decrease in blood
loss is 1.30 ml for each increase of 1 kg in weight.
Any fitted linear relationship holds only approximately and does not neces-
sarily extend outside the range of the data. In particular, nonsensical predicted
blood losses of less than zero are obtained at very large weights outside the
range of data.
8.4 ANOVA Table for Regression

The LS line minimizes
n
X
{Yi − (β0 + β1Xi)}2
i=1
over all choices for β0 and β1. Inserting the LS estimates b0 and b1 into this
expression gives
n
X
Residual Sums of Squares = {Yi − (b0 + b1Xi)}2.
i=1
Several bits of notation are needed. Let
Ŷi = b0 + b1Xi
be the predicted or fitted Y -value for an X-value of Xi and let ei = Yi − Ŷi.

The fitted value Ŷi is the value of the LS line at Xi whereas the residual ei
is the distance that the observed response Yi is from the LS line. Given this
notation,
n
X n
X
Residual Sums of Squares = Res SS = (Yi − ŷi)2 = e2i .
i=1 i=1
Here is a picture to clarify matters:
•
19
•
18
•
17
Fitted
residual
16
Response •
15
•
4.5 X-val 5.0 5.5
The Residual SS, or sum of squared residuals, is small if each Ŷi is close to
Yi (i.e., the line closely fits the data). It can be shown that
n
X
Total SS in Y = (Yi − Ȳ )2 ≥ Res SS ≥ 0.
i=1
Also define
n
X
Regression SS = Reg SS = Total SS − Res SS = b1 (Yi − Ȳ )(Xi − X̄).
i=1
The Total SS measures the variability in the Y -sample. Note that
0 ≤ Regression SS ≤ Total SS.
The percentage of the variability in the Y -sample that is explained by

the linear relationship between Y and X is
Reg SS
R2 = coefficient of determination = .
Total SS
Given the definitions of the Sums of Squares, we can show 0 ≤ R2 ≤ 1 and
R2 = square of Pearson correlation coefficient = r2.
To understand the interpretation of R2, at least in two extreme cases, note that
Reg SS = Total SS ⇔ Res SS = 0

⇔ all the data points fall on a straight line
⇔ all the variability in Y is explained by the linear relationship
(which has variation)
⇔ R2 = 1. (see the picture below)
8
•
6
Variation in Y
•
4
•
2
•
0
•
-2
-1 0 1 2 3 4
Variation in X
Furthermore,
Reg SS = 0 ⇔ Total SS = Res SS

⇔ b1 = 0
⇔ LS line is Ŷ = Ȳ
⇔ none of the variability in Y is explained by a linear relationship
⇔ R2 = 0.
5
• •
• •
4
• ••
• • • ••
•
•• • •• •• • • •
• • • • • • • ••• • • • •
• • ••
3
• •• • • •• • • • •
• • ••• •• • •• •• •• •
• • • • • •• • • ••
Y
• • • • ••• • • • • •
• • • • • • •• •• • •
••
• • • • •• • • ••• • • •• • • •
2
• • • • • • •••• • • • •
• • •• •
•• •
• • • •• • • • •
• ••• •• •
• • •
• ••
•• • • •• •
1
• ••
• • • • •• •
•
• •• • •
•
0
• •
•
-3 -2 -1 0 1 2
X
Each Sum of Squares has a corresponding df (degrees of freedom). The

Sums of Squares and df are arranged in an analysis of variance (ANOVA)
table:
Source df SS MS
Regression 1 SSR MSR
Residual (Error) n − 2 SSE MSE
Total n−1
The Total df is n−1. The Residual df is n minus the number of parameters
(2) estimated by the LS line. The Regression df is the number of predictor
variables (1) in the model. A Mean Square is always equal to the Sum of
Squares divided by the df . Sometime the following notation is used for the
Residual MS: s2Y |X = Resid(SS)/(n − 2).
# ANOVA table of the simple linear regression fit
anova(lm.blood.wt)
## Analysis of Variance Table
##
## Response: blood_loss
## weight 1 1207 1207 8.88 0.025 *
## Residuals 6 816 136
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
8.4.1 Brief discussion of the output for blood loss

problem
1. Identify fitted line: Blood Loss = 552.44 − 1.30 Weight (i.e., b0 = 552.44
and b1 = −1.30).
What is the line of best fit? What is the direction of the relationship?
2. Locate Analysis of Variance Table.
This tests the the hypothesis H0 : βj = 0, j = 0, 1, . . . , p (for all p + 1
beta coefficients), against HA : not H0, i.e., at least one βj 6= 0. More on
this later.
3. Locate Parameter Estimates Table.
Given that not all the betas are zero from the ANOVA, which parameter
betas are different from zero, and what are their estimated values and
standard errors? More on this later.
4. Note that R2 = 0.5967 = (−0.77247)2 = r2.
R2 indicates the proportion of the variance explained by the regression
model. This indicates the predictive ability of the model, and is not a
indication of model fit.
ClickerQ s — Salaries STT.02.02.060
8.5 The regression model

The following statistical model is assumed as a means to provide error estimates
for the LS line, regression coefficients, and predictions. Assume that the data
(Xi, Yi), i = 1, . . . , n, are a sample of (X, Y ) values from the population of
interest, and
1. The mean in the population of all responses Y at a given X value (some-
times called µY |X ) falls on a straight line, β0 + β1X, called the population
regression line.
8.5: The regression model 299
2. The variation among responses Y at a given X value is the same for each
X, and is denoted by σY2 |X .
3. The population of responses Y at a given X is normally distributed.
4. The pairs (Xi, Yi) are a random sample from the population. Alterna-
tively, we can think that the Xis were fixed by the experimenter, and that
the Yi are random responses at the selected predictor values.
The model is usually written in the form
Yi = β0 + β1Xi + εi
(i.e., Response = Mean Response + Residual), where the εis are, by virtue of
assumptions 2, 3, and 4, independent normal random variables with mean 0
and variance σY2 |X . The following picture might help see this. Note that the
population regression line is unknown, and is estimated from the data using the
LS line.
14
14
12
12
10
10
Y
epsilon_i
8
Y_i
6
1 2 3 4 5 1 2 X_i 3 4 5
X X
In the plot below, data are simulated where yi = 4 − 2xi + ei, where
xi ∼ Gamma(3, 0.5) and ei ∼ Normal(0, 32). The data are plotted and a
linear regression is fit and the mean regression line is overlayed. Select normal
distributions with variance estimated from the linear model fit are overlayed,
one which indicates limits at two standard deviations. See the R code to create
this image.
5
●
●
●
● ●
●
● ●
●
0
●
●● ●
●
● ●●
● ● ●
●● ●
●
● ● ● ●
● ● ● ●●
● ●
−5
●
● ● ●
● ● ●
●● ●
● ●
● ●
● ● ●
●
● ● ●
● ●
●●
●
● ● ●
−10
●
●
● ●
● ● ●
●
●
● ● ●
● ●
● ● ●
●
●
−15
●
● ●
● ●
●
● ●
●
−20
● ●
●
−25
2 4 6 8 10 12 14
Model assumptions In decreasing order of importance, the model assump-

tions are
1. Validity. Most importantly, the data you are analyzing should map to
the research question you are trying to answer. This sounds obvious but
is often overlooked or ignored because it can be inconvenient.
2. Additivity and linearity. The most important mathematical as-

sumption of the regression model is that its deterministic component is a
linear function of the separate predictors.
3. Independence of errors. This assumption depends on how the data

were collected.
4. Equal variance of errors.
5. Normality of errors.
Normality and equal variance are typically minor concerns, unless you’re using
the model to make predictions for individual data points.
8.6: CI and tests for β1 301
8.5.1 Back to the Data
There are three unknown population parameters in the model: β0, β1 and
σY2 |X . Given the data, the LS line
Ŷ = b0 + b1X
estimates the population regression line Y = β0 + β1X. The LS line is our

best guess about the unknown population regression line. Here b0 estimates the
intercept β0 of the population regression line and b1 estimates the slope β1 of
the population regression line.
The ith observed residual ei = Yi − Ŷi, where Ŷi = b0 + b1Xi is the ith
fitted value, estimates the unobservable residual εi (εi is unobservable
because β0 and β1 are unknown). The Residual MS from the ANOVA table is
used to estimate σY2 |X :
− Ŷi)2
P
Res SS i (Yi
s2Y |X = Res MS = = .
Res df n−2
The denominator df = n − 2 is the number of observations minus the number
of beta parameters in the model, i.e., β0 and β1.
8.6 CI and tests for β1

A CI for β1 is given by b1 ± tcritSEb1 , where the standard error of b1 under the
model is
sY |X
SEb1 = pP ,
(X − X̄) 2
i i
and where tcrit is the appropriate critical value for the desired CI level from a
t-distribution with df =Res df .
To test H0 : β1 = β10 (a given value) against HA : β1 6= β10, reject H0 if
|ts| ≥ tcrit, where
b1 − β10
ts = ,
SEb1
and tcrit is the t-critical value for a two-sided test, with the desired size and
df =Res df . Alternatively, you can evaluate a p-value in the usual manner to
make a decision about H0.
# CI for beta1
sum.lm.blood.wt <- summary(lm.blood.wt)
sum.lm.blood.wt$coefficients
## (Intercept) 552.4 21.4409 25.77 2.253e-07
## weight -1.3 0.4364 -2.98 2.465e-02
est.beta1 <- sum.lm.blood.wt$coefficients[2,1]
se.beta1 <- sum.lm.blood.wt$coefficients[2,2]
sum.lm.blood.wt$fstatistic
## value numdf dendf
## 8.878 1.000 6.000
df.beta1 <- sum.lm.blood.wt$fstatistic[3]
t.crit <- qt(1-0.05/2, df.beta1)
t.crit
## [1] 2.447
CI.lower <- est.beta1 - t.crit * se.beta1
CI.upper <- est.beta1 + t.crit * se.beta1
c(CI.lower, CI.upper)
## [1] -2.3682 -0.2325
The parameter estimates table gives the standard error, t-statistic, and p-
value for testing H0 : β1 = 0. Analogous summaries are given for the intercept,
β0, but these are typically of less interest.
8.6.1 Testing β1 = 0
Assuming the mean relationship is linear, consider testing H0 : β1 = 0 against
HA : β1 6= 0. This test can be conducted using a t-statistic, as outlined above,
or with an ANOVA F -test, as outlined below.
For the analysis of variance (ANOVA) F -test, compute
Reg MS
Fs =
Res MS
and reject H0 when Fs exceeds the critical value (for the desired size test) from
an F -table with numerator df = 1 and denominator df = n − 2 (see qf()).
The hypothesis of zero slope (or no relationship) is rejected when Fs is large,
which happens when a significant portion of the variation in Y is explained by
8.7: A CI for the population regression line 303
the linear relationship with X.
The p-values from the t-test and the F -test are always equal. Furthermore
this p-value is equal to the p-value for testing no correlation between Y and X,
using the t-test described earlier. Is this important, obvious, or disconcerting?
8.7 A CI for the population regression line

I can not overemphasize the power of the regression model. The model allows
you to estimate the mean response at any X value in the range for which the
model is reasonable, even if little or no data is observed at that location.
We estimate the mean population response among individuals with X = Xp
µp = β0 + β1Xp,
with the fitted value, or the value of the least squares line at Xp:
Ŷp = b0 + b1Xp.
Xp is not necessarily one of the observed Xis in the data. To get a CI for µp,
use Ŷp ± tcritSE(Ŷp), where the standard error of Ŷp is
s
1 (Xp − X̄)2
SE(Ŷp) = sY |X +P 2
.
n i (X i − X̄)
The t-critical value is identical to that used in the subsection on CI for β1.
8.7.1 CI for predictions

Suppose a future individual (i.e., someone not used to compute the LS line) has
X = Xp. The best prediction for the response Y of this individual is the value
of the least squares line at Xp:
Ŷp = b0 + b1Xp.
To get a CI (prediction interval) for an individual response, use Ŷp±tcritSEpred(Ŷp),
where s
1 (Xp − X̄)2
SEpred(Ŷp) = sY |X 1 + + P 2
,
n i (X i − X̄)
and tcrit is identical to the critical value used for a CI on β1. The prediction vari-
ance has two parts: (1) the 1 indicates the variability associated with the data
around the mean (regression line), and (2) the rest is the variability associated
with estimating the mean.
For example, in the blood loss problem you may want to estimates the blood
loss for an 50kg individual, and to get a CI for this prediction. This problem is
different from computing a CI for the mean blood loss of all 50kg individuals!
# CI for the mean and PI for a new observation at weight=50
predict(lm.blood.wt, data.frame(weight=50), interval = "confidence", level = 0.95)
## fit lwr upr
## 1 487.4 477.2 497.7
predict(lm.blood.wt, data.frame(weight=50), interval = "prediction", level = 0.95)
## fit lwr upr
## 1 487.4 457.1 517.8
Comments
1. The prediction interval is wider than the CI for the mean response. This
is reasonable because you are less confident in predicting an individual
response than the mean response for all individuals.
2. The CI for the mean response and the prediction interval for an individual
response become wider as Xp moves away from X̄. That is, you get a
more sensitive CI and prediction interval for Xps near the center of the
data.
3. In plots below include confidence and prediction bands along with the
fitted LS line.
library(ggplot2)
p <- ggplot(thyroid, aes(x = weight, y = blood_loss))
p <- p + geom_smooth(method = lm, se = TRUE)
print(p)
# Base graphics: Plot the data with linear regression fit and confidence bands
# scatterplot
plot(thyroid$weight, thyroid$blood_loss)
# regression line from lm() fit
abline(lm.blood.wt)
# x values of weight for predictions of confidence bands
x.pred <- data.frame(weight = seq(min(thyroid$weight), max(thyroid$weight),
8.7: A CI for the population regression line 305
length = 20))
# draw upper and lower confidence bands
lines(x.pred$weight, predict(lm.blood.wt, x.pred,
interval = "confidence")[, "upr"], col = "blue")
lines(x.pred$weight, predict(lm.blood.wt, x.pred,
interval = "confidence")[, "lwr"], col = "blue")
525
●
●
510
●
● ●
500 ●
500
thyroid$blood_loss
● ●
blood_loss
490
● ● ●
475
● ●
480
●
470
●
450
●
35 40 45 50 55 60 65 70
40 50 60 70 thyroid$weight
weight
8.7.2 A further look at the blood loss data

The LS line is: Predicted Blood Loss = 552.442 - 1.30 Weight.
The R2 is 0.597 (i.e., 59.7%).
The F -statistic for testing H0 : β1 = 0 is Fobs = 8.88 with a p-value=0.025.
The Error MS is s2Y |X = 136.0; see ANOVA table.
The Parameter Estimates table gives b0 and b1 , their standard errors, and
t-statistics and p-values for testing H0 : β0 = 0 and H0 : β1 = 0. The t-
test and F -test p-values for testing that the slope is zero are identical. We
could calculate a 95% CI for β0 and β1. If we did so (using the t critical
value) we find we are 95% confident that the slope of the population
regression line is between −2.37 and −0.23.
Suppose we are interested in estimating the average blood loss among all
50kg individuals. The estimated mean blood loss is 552.442 − 1.30033 ×
50 = 487.43. Reading off the plot, we are 95% confident that the mean
blood loss of all 50kg individuals is between (approximately) 477 and 498
ml. A 95% prediction interval for the blood loss of a single 50 kg person
is less precise (about 457 to 518 ml).
As a summary we might say that weight is important for explaining the
variation in blood loss. In particular, the estimated slope of the least squares
line (Predicted Blood loss = 552.442 - 1.30 Weight) is significantly different from
zero (p-value = 0.0247), with weight explaining approximately 60% (59.7%) of
the variation in blood loss for this sample of 8 thyroid operation patients.
8.8 Model Checking and Regression Diag-

nostics
8.8.1 Introduction
The simple linear regression model is usually written as
Yi = β0 + β1Xi + εi
where the εis are independent normal random variables with mean 0 and vari-
ance σ 2. The model implies (1) The average Y -value at a given X-value is
linearly related to X. (2) The variation in responses Y at a given X value
is constant. (3) The population of responses Y at a given X is normally dis-
tributed. (4) The observed data are a random sample.
A regression analysis is never complete until these assumptions have been
checked. In addition, you need to evaluate whether individual observations, or
groups of observations, are unduly influencing the analysis. A first step in any
analysis is to plot the data. The plot provides information on the linearity and
constant variance assumption.
8.8: Model Checking and Regression Diagnostics 307
(a) (b) (c) (d)
800 1000 1200 1400

400
• • • •
•
20
• •
25
•
18
300
•
•
• • •
16
• •• •
20
•
200
••
14
Y
Y
• ••
600
• •
• •
15
• •
12
• •
400
• • •
100
•••
10
• • • •
•
200
• 10 • ••
• • •• • •
• • •
8
• • • ••
0
• • •• • • •
0
1 2 3 4 5 1 2 3 4 5 1 2 3 4 5 1 2 3 4 5
X X X X
Figure (a) is the only plot that is consistent with the assumptions. The plot
shows a linear relationship with constant variance. The other figures show one
or more deviations. Figure (b) shows a linear relationship but the variability
increases as the mean level increases. In Figure (c) we see a nonlinear relation-
ship with constant variance, whereas (d) exhibits a nonlinear relationship with
non-constant variance.
In many examples, nonlinearity or non-constant variability can be addressed
by transforming Y or X (or both), or by fitting polynomial models.
These issues will be addressed later.
8.8.2 Residual Analysis

A variety of methods for assessing model adequacy are based on the observed
residuals,
ei = Yi − Ŷi i.e., Observed − Fitted values.
The residual is the difference between the observed values and predicted or
fitted values. The residual is the part of the observation that is not explained
by the fitted model. You can analyze residuals to determine the adequacy of
the model. A large residual identifies an observation poorly fit by the model.
The residuals are usually plotted in various ways to assess potential inad-
equacies. The observed residuals have different variances, depending on Xi.
Recall that the standard error of Ŷi (and therefore ei) is
s
1 (Xi − X̄)2
SE(Ŷi) = SE(ei) = sY |X +P 2
.
n j (X j − X̄)
So many statisticians prefer to plot the studentized residuals (sometimes

called the standardized residuals or internally Studentized residual)
ei
ri = .
SE(ei)
The standardized residual is the residual, ei, divided by an estimate of its stan-
dard deviation. This form of the residual takes into account that the residuals
may have different variances, which can make it easier to detect outliers. The
studentized residuals have a constant variance of 1 (approximately). Standard-
ized residuals greater than 2 and less than −2 are usually considered large. I
will focus on diagnostic methods using the studentized residuals.
A plot of the studentized residuals ri against the fitted values Ŷi often reveals
inadequacies with the model. The real power of this plot is with multiple
predictor problems (multiple regression). The information contained in this
plot with simple linear regression is similar to the information contained in the
original data plot, except it is scaled better and eliminates the effect of the
trend on your perceptions of model adequacy. The residual plot should exhibit
no systematic dependence of the sign or the magnitude of the residuals on the
fitted values:
2
• • •
25
•
•
• •
1
•• • • •
•• • • •
• • • •• •
••
20
• • • • •
•• • •
•
Resids
• • •
•• • • • • •••
0
Y
• •••• • • • •• •
• • •
• •••••• •
15
•••
••
• • •
••••
• •
-1
• •• •• •
•• •
•
10
•
•
• •
-2
3 4 5 6 7 10 15 20 25
X Fitted
The following sequence of plots show how inadequacies in the data plot
appear in a residual plot. The first plot shows a roughly linear relationship
between Y and X with non-constant variance. The residual plot shows a
megaphone shape rather than the ideal horizontal band. A possible remedy is
a weighted least squares analysis to handle the non-constant variance (see
end of chapter for an example), or to transform Y to stabilize the variance.
Transforming the data may destroy the linearity.
• • • •
12
120
60
•
100
10
40
• • •
• • •• ••
• • •
80
• •• • • • •
20
• • • •
8
• • • • • •• •
2
•
Resids
Resids
•
• • •
60
••
Y
• • •
• •• • • •
0
• • •
• • • ••
•• • • • •• • • • •
6
• • • • ••
40
• • • •
• • • • •• ••
-20
• •• ••• •
• • • • • • • •
0
•• • • • •
• •• •
• • ••• ••• • •• • • ••
• •
20
• • • •
• •• • •• •
4
••• ••• • •• •• ••
-40
• • • •• • •
• • • •• • • ••• • • ••
•• • •
• •• •• • •
0
• • • •
-60
• • • • •
2
-2
3 4 5 6 7 25 30 35 40 45 50 55 2 3 4 5 6 2 3 4 5 6 7
X Fitted X Fitted
The plot above shows a nonlinear relationship between Y and X. The

residual plot shows a systematic dependence of the sign of the residual on the
fitted value. Possible remedies were mentioned earlier.
The plot below shows an outlier. This case has a large residual and large
studentized residual. A sensible approach here is to refit the model after holding
out the case to see if any conclusions change.
• • •• •
2
40
• • •
• • ••• • • • •• • • •
• • •• • •
•• ••••• • ••• •
0
• • •
• • • •
35
• •• •
• •• •
-2
• •
• •• •
•
••
30
Resids
• • ••
-4
• ••
• •••••• •
Y
•••
•
-6
• •••
25
• •
•• •
-8
••
20
•• •
-10
•
• •
-12
3 4 5 6 7 20 25 30 35 40
X Fitted
A third type of residual is called an externally Studentized residual

(or Studentized deleted residual or deleted t-residual). The problem with resid-
uals is that a highly influential value can force the residual to have a very small
value. This measure tries to correct for that by looking at how well the model
fits this observation without using this observation to construct the fit. It is
quite possible for the deleted residual to be huge when the raw residual is tiny.
The studentized deleted residual for observation ith is calculated by fitting
the regression based on all of the cases except the ith one. The residual is then
divided by its estimated standard deviation. Since the Studentized deleted
residual for the ith observation estimates all quantities with this observation
deleted from the data set, the ith observation cannot influence these estimates.
Therefore, unusual Y values clearly stand out. Studentized deleted residuals
with large absolute values are considered large. If the regression model is appro-
priate, with no outlying observations, each Studentized deleted residual follows
the t-distribution with n − 1 − p degrees of freedom.
Nonconstant variance vs sample size Because more extreme observa-

tions are more likely to occur with larger sample sizes, sometimes when sample
size depends on X it can appear as noncontant variance. Below sample sizes
are either all 25 or (3, 5, 10, 25, 125), and standard deviations are either all 1
or (0.1, 0.5, 1, 1.5, 3). Note that constant variance and different sample sizes
appears as though it has nonconstant variance.
#### Nonconstant variance vs sample size
dat.var.sam <- function(n, s) {
dat <- data.frame(
x = c(rep(0, n[1]),
rep(1, n[2]),
rep(2, n[3]),
rep(3, n[4]),
rep(4, n[5])),
y = c(rnorm(n[1], mean = 0, sd = s[1]),
rnorm(n[2], mean = 0, sd = s[2]),
rnorm(n[3], mean = 0, sd = s[3]),
rnorm(n[4], mean = 0, sd = s[4]),
rnorm(n[5], mean = 0, sd = s[5]))
)
return(dat)
}
library(ggplot2)
n <- c(25, 25, 25, 25, 25)

s <- c(1, 1, 1, 1, 1)
dat <- dat.var.sam(n, s)
p1 <- ggplot(dat, aes(x = x, y = y))
p1 <- p1 + geom_point(position = position_jitter(width = 0.1))
p1 <- p1 + labs(title = "Constant variance, constant sample size")
#print(p1)
n <- c(3, 5, 10, 25, 125)

s <- c(1, 1, 1, 1, 1)
p2 <- p2 + labs(title = "Constant variance, different sample sizes")
#print(p2)
n <- c(25, 25, 25, 25, 25)

s <- c(0.1, 0.5, 1, 1.5, 3)

p3 <- p3 + labs(title = "Different variance, constant sample size")
#print(p3)
n <- c(3, 5, 10, 25, 125)

s <- c(0.1, 0.5, 1, 1.5, 3)
p4 <- p4 + labs(title = "Different variance, different sample sizes")
#print(p4)
library(gridExtra)
grid.arrange(p1, p2, p3, p4, nrow=2, ncol=2)
Constant variance, constant sample size Constant variance, different sample sizes
● ● ● ●
● ●
● ●
●
2 ● ● ●
●
2 ● ●
●●●
● ● ● ●
●
● ● ●
●● ● ●●● ●●
●● ● ● ● ●
●●
●
● ●● ●
●
● ●● ●●● ●●●
● ●● ●
1 ● ● ● ● ●
● ●●●● ●●
● ●●●
● ● ● ●● ● ●●
● ●● ● ●●
●●●●
●
●
● ● ● ● ● ●● ● ●●
● ● ● ● ●
●● 0 ● ●●
●●●● ●
● ● ●● ● ●●●● ●
● ● ●● ● ●●
● ●● ●●●●●
● ● ● ● ● ●
●●● ●
0 ●● ● ●
y
● ●● ● ● ● ● ●●
●● ● ●● ● ● ● ● ● ●● ●
● ●●
●●
●
● ●
● ● ●● ● ●
●
●●●
● ● ● ●●●●
● ● ●
● ●● ●● ●● ● ●
●
● ●● ●
● ● ●●
●
● ●
●
● ●
● ● ●
●● ● ●●
−1 ● ● ●●
●●
● ●● ● ●
−2 ● ●
● ●●
●
●
● ●
●●
●
● ●
●
−2
●
● ● ●
0 1 2 3 4 0 1 2 3 4
x x
Different variance, constant sample size Different variance, different sample sizes
● ●
●
6
●●
● ●
●
● 4 ● ●●●
●
●
●●●
● ● ● ●●
● ●● ●●●
●●
3 ●
●
● ●●●
●
●●
●●
●●●●●●
● ● ● ●●
● ●●
● ● ●● ● ● ●●● ● ●● ●
●●
●●
●
●
● ● ●●● ●
●● ● ●● ● ● ● ● ●
●●
●
●
● ● ● ● ●● ●●● ●
0 ●●
● ● ●● ● ● ●
● ●● ●● ● ● ● ● ● ● ● ● ● ●● ●
● ● ●● ●●●
y
●● ●
●
●
●● ●● ●● ● ●
● ●● ●●●
●
●●●
●● ●●●●● ●● ● ● ●●●●
0 ●
●●
●
●
●●●
●
●●
●●
● ●●
●● ●●●
●
●●●
●●
●● ●
●
●
● ●●●
● ● ● ●● ●● ●
●●● ● ● ● ●
● ● ● ● ●
● ●● ●● ● ●●
● ● ● ●●
●●● ●
●
● ●●
●● ● ●●
● ●●● ●
−4 ●● ●
●
●
−3
●
●
●
●● ●
●
● ●
−8
0 1 2 3 4 0 1 2 3 4
x x
8.8.3 Checking Normality

The normality assumption for the εis can be evaluated visually with a boxplot or
a normal probability plot (rankit plot) of the ri, or formally with a Shapiro-Wilk
test. The normal probability plot often highlights outliers, or poorly fitted
cases. If an outlier is held out of the data and a new analysis is performed,
the resulting normal scores plot may be roughly linear, but often will show a
short-tailed distribution. (Why?).
You must interpret regression tests and CI with caution with non-normal
data. Statisticians developed robust regression methods for non-normal data
which are available in R packages.
8.8.4 Checking Independence

Diagnosing dependence among observations requires an understanding of the
data collection process. There are a variety of graphical and inferential tools for
checking independence for data collected over time (called a time series). The
easiest check is to plot the ri against time index and look for any suggestive
patterns.
8.8.5 Outliers
Outliers are observations that are poorly fitted by the regression model. The
response for an outlier is far from the fitted line, so outliers have large positive
or negative values of the studentized residual ri. Usually, |ri| > 2 is considered
large. Outliers are often highlighted in residual plots.
What do you do with outliers? Outliers may be due to incorrect recordings
of the data or failure of the measuring device, or indications or a change in the
mean or variance structure for one or more cases. Incorrect recordings should
be fixed if possible, but otherwise deleted from the analysis.
Routine deletion of outliers from the analysis is not recommended. This
practice can have a dramatic effect on the fit of the model and the perceived
precision of parameter estimates and predictions. Analysts who routinely omit
outliers without cause tend to overstate the significance of their findings and get
a false sense of precision in their estimates and predictions. To assess effects of
outliers, a data analyst should repeat the analysis holding out the outliers to see
whether any substantive conclusions are changed. Very often the only real effect
of an outlier is to inflate MSE and hence make p-values a little larger and CIs
a little wider than necessary, but without substantively changing conclusions.
They can completely mask underlying patterns, however.
ClickerQ s — Outliers STT.02.04.040
8.8.6 Influential Observations

Certain data points can play an important role in determining the position of
the LS line. These data points may or may not be outliers. In the plots below,
the solid line is the LS line from the full data set, whereas the dashed line is
the LS line after omitting the unusual point. For example, the observation
with Y > 45 in the first plot is an outlier relative to the LS fit. The extreme
observation in the second plot has a very small ri. Both points are highly
influential observations — the LS line changes dramatically when these
observations are held out.
10
• •
45
8
40
•
6
35
•
Y
•
4
•
30
• •
••
2
•
25
• • • •
••••• •
0
• • •
20
4 5 6 0 2 4 6 8 10
X X
In the second plot, the extreme value is a high leverage value, which is
basically an outlier among the X values; Y does not enter in this calculation.
This influential observation is not an outlier because its presence in the analysis
determines that the LS line will essentially pass through it! These are values
with the potential of greatly distorting the fitted model. They may or may not
actually have distorted it.
The hat variable from the influence() function on the object returned from
lm() fit will give the leverages: influence(lm.output)$hat. Leverage values
fall between 0 and 1. Experts consider a leverage value greater than 2p/n or
3p/n, where p is the number of predictors or factors plus the constant and n is
the number of observations, large and suggest you examine the corresponding
observation. A rule-of-thumb is to identify observations with leverage over 3p/n
or 0.99, whichever is smaller.
Dennis Cook developed a measure of the impact that individual cases have
on the placement of the LS line. His measure, called Cook’s distance or
Cook’s D, provides a summary of how far the LS line changes when each
individual point is held out (one at a time) from the analysis. While high
leverage values indicate observations that have the potential of causing trouble,
those with high Cook’s D values actually do disproportionately affect the overall
fit. The case with the largest D has the greatest impact on the placement of
the LS line. However, the actual influence of this case may be small. In the
plots above, the observations I focussed on have the largest Cook’s Ds.
A simple, but not unique, expression for Cook’s distance for the j th case is
X
Dj ∝ (Ŷi − Ŷi[−j])2,
i
where Ŷi[−j] is the fitted value for the ith case when the LS line is computed from
all the data except case j. Here ∝ means that Dj is a multiple of i(Ŷi −Ŷi[−j])2
P
where the multiplier does not depend on the case. This expression implies that
Dj is also an overall measure of how much the fitted values change when case
j is deleted.
Observations with large D values may be outliers. Because D is calculated
using leverage values and standardized residuals, it considers whether an ob-
servation is unusual with respect to both x- and y-values. To interpret D,
compare it to the F -distribution with (p, n − p) degrees-of-freedom to deter-
mine the corresponding percentile. If the percentile value is less than 10% or
20%, the observation has little influence on the fitted values. If the percentile
value is greater than 50%, the observation has a major influence on the fitted
values and should be examined.
Many statisticians make it a lot simpler than this sounds and use 1 as a
cutoff value for large Cook’s D (when D is on the appropriate scale). Using
the cutoff of 1 can simplify an analysis, since frequently one or two values will
have noticeably larger D values than other observations without actually having
much effect, but it can be important to explore any observations that stand
out. Cook’s distance values for each observation from a linear regression fit are
given with cooks.distance(lm.output).
Given a regression problem, you should locate the points with the largest
Dj s and see whether holding these cases out has a decisive influence on the fit
of the model or the conclusions of the analysis. You can examine the relative
magnitudes of the Dj s across cases without paying much attention to the actual
value of Dj , but there are guidelines (see below) on how large Dj needs to be
before you worry about it.
It is difficult to define a good strategy for dealing with outliers and influential
observations. Experience is the best guide. I will show you a few examples that
highlight some standard phenomena. One difficulty you will find is that certain
observations may be outliers because other observations are influential, or vice-
versa. If an influential observation is held out, an outlier may remain an outlier,
may become influential, or both, or neither. Observations of moderate influence
may become more, or less influential, when the most influential observation is
held out.
Thus, any sequential refitting of models holding out of observations should

not be based on the original (full-data) summaries, but rather on the summaries
that result as individual observations are omitted. I tend to focus more on
influential observations than outliers.
In the plots below, which cases do you think are most influential, and which
are outliers. What happens in each analysis if I delete the most influential case?
Are any of the remaining cases influential or poorly fitted?
• •
10 12 14 16
15
•
• •
10
Y
Y
• •
•
5
8
• •
0
2 4 6 8 10 12 1.0 1.5 2.0 2.5 3.0 3.5 4.0

X X
• •
15
15
•
• •
• •
10
10
Y
• •
5
•
• •
0
0 2 4 6 8 10 12 2 4 6 8 10 12
X X
Many researchers are hesitant to delete points from an analysis. I think

this view is myopic, and in certain instances, such as the Gesell example to be
discussed, can not be empirically supported. Being rigid about this can lead to
some silly analyses of data, but one needs a very good reason and full disclosure
if any points are deleted.
8.8.7 Summary of diagnostic measures

The various measures discussed above often flag the same observations as un-
usual, but they certainly can flag different observations. At the very least I
examine standardized residuals and Cooks’s D values. They are invaluable
diagnostic measures, but nothing is perfect. Observations can be unusual in
groups — a pair of unusual high leverage values close to each other will not
necessarily be flagged by Cook’s D since removing just one may not affect the
fit very much. Any analysis takes some careful thought.
These measures and techniques really are designed for multiple regression
problems where several predictor variables are used. We are using them in
simple linear regression to learn what they do and see what they have to say
about data, but in truth it is fairly simple with one variable to see what may
be an outlier in the x-direction, to see if the data are poorly fit, etc. With more
variables all that becomes quite difficult and these diagnostics are essential parts
of those analyses.
8.9 Regression analysis suggestion

There are a lot options allowed in R. I will make a few suggestions here on how
to start an analysis. What you find once you get started determines what more
you might wish to explore.
1. Plot the data. With lots of variables the matrix plot is valuable as a
quick screen. If you want to see better resolution on an individual scatter
plot, do the individual scatter plot.
2. Do any obvious transformations of the data. We will discuss this in a
lot more detail later. Re-plot with transformations.
3. Fit the least squares equation.
4. Examine the residual plots and results. Check for the patterns
discussed earlier.
(a) Plotting several diagnostic plots together seems convenient to me.
This gives you the essential plot (residuals vs. fitted values) plus a
quick check on normality and possible violations of independence
and influence. If you see something that needs closer investigation,
you may need to recreate one of the plots larger by itself.
(b) Do you see curvature in the standardized residual plot? If the sign of
the residuals has a distinct pattern vs. the fitted values, the linear fit
is not adequate and you need some remedy, such as transformations.
Note that standardized residuals are more conventional and show
you what actually happened, while deleted residuals are probably
the better diagnostic tool for identifying problem cases.
8.9: Regression analysis suggestion 319
(c) Does it appear σY |X depends upon X (we are assuming it does not)?
A megaphone pattern in residuals vs. fits is the classic (not the only)
pattern to look for. Weighted least squares or transformations may
be called for.
(d) Do you see obvious outliers? Make sure you do not have a mis-
recorded data value. It might be worth refitting the equation without
the outlier to see if it affects conclusions substantially.
(e) Is the normality assumption reasonable? This can be very closely
related to the preceding points.
(f) Is there a striking pattern in residuals vs. order of the data? This
can be an indication that the independence assumption is not valid.
5. Check the Cook’s D values. The Cook’s distance plot and Residuals
vs. Leverage (with Cook’s D) plot are both helpful.
6. If you found problem observations, omit them from the analysis and see
if any conclusions change substantially. There are two good ways to do
this.
(a) Subset the data.frame using subset().
(b) Use lm() with the weights= option with weights of 0 for the excluded
observations, weights of 1 for those included.
You may need to repeat all these steps many times for a complete analysis.
8.9.1 Residual and Diagnostic Analysis of the Blood

Loss Data
We looked at much of this before, but let us go through the above steps sys-
tematically. Recall the data set (we want to predict blood loss from weight):
weight time blood loss
1 44.3 105 503
2 40.6 80 490
3 69.0 86 471
4 43.7 112 505
5 50.3 109 482
6 50.2 100 490
7 35.4 96 513
8 52.2 120 464
1. Plot the data. Plot blood loss vs. weight.
# create data ids
thyroid$id <- 1:nrow(thyroid)
library(ggplot2)
p <- ggplot(thyroid, aes(x = weight, y = blood_loss, label = id))
# plot labels next to points
p <- p + geom_text(hjust = 0.5, vjust = -0.5)
# plot regression line and confidence band
p <- p + geom_smooth(method = lm)
print(p)
525
7
●
4
●1
●
500
2 6
● ●
blood_loss
5
●
475
3
●
8
●
450
40 50 60 70
weight
Clearly the heaviest individual is an unusual value that warrants a closer

look (maybe data recording error). I might be inclined to try a transforma-
tion here (such as log(weight)) to make that point a little less influential.
2. Do any obvious transformations of the data. We will look at transfor-
mations later.
3. Fit the least squares equation. Blood Loss appears significantly nega-
tively associated with weight.
lm.blood.wt <- lm(blood_loss ~ weight, data = thyroid)
summary(lm.blood.wt)
##
## Call:
##
## Residuals:

## -20.57 -6.19 4.71 8.19 9.38
##
## Coefficients:
## (Intercept) 552.442 21.441 25.77 2.3e-07 ***
## weight -1.300 0.436 -2.98 0.025 *
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
(a) Graphs: Check Standardized Residuals (or the Deleted Residu-

als). The residual plots:
# plot diagnistics
par(mfrow=c(2,3))
plot(lm.blood.wt, which = c(1,4,6))
# residuals vs weight
plot(thyroid$weight, lm.blood.wt$residuals, main="Residuals vs weight")
# horizontal line at zero
abline(h = 0, col = "gray75")
library(car)
qqPlot(lm.blood.wt$residuals, las = 1, id.n = 3, main="QQ Plot")
## 8 2 4
## 1 2 8
# residuals vs order of data
plot(lm.blood.wt$residuals, main="Residuals vs Order of data")
Residuals vs Fitted Cook's distance Cook's dist vs Leverage hii (1 − hii)

2 1.5 1 3●
2.5
3
10
2.5
4●
● ●
●
2.0
2.0
●
Cook's distance
Cook's distance
0
Residuals
1.5
1.5
●
1.0
−10
1.0
2●
0.5
0.5
0.5
8
●8
−20
●8 7
● ●7
0.0
0.0
●
● 0
470 480 490 500 1 2 3 4 5 6 7 8 0.1 0.4 0.6 0.7
Fitted values Obs. number Leverage hii
Residuals vs weight QQ Plot Residuals vs Order of data

10
10
● 10 4● ●
● ● ● ● ● ●
● ● ●
5
5
5
lm.blood.wt$residuals
● ● ●
0
0
0
−5
−5
● −5 ● ●
−10
−10
●
−10 ● 2 ●
−15
−20
−20
● −20 ● 8 ●
35 40 45 50 55 60 65 70 −1.5 −1.0 −0.5 0.0 0.5 1.0 1.5 1 2 3 4 5 6 7 8
thyroid$weight norm quantiles Index
4. Examine the residual plots and results.

(a) Do you see curvature? There does not appear to be curvature (and
it could be hard to detect with so few points).
(b) Does it appear σY |X depends upon X? Not much evidence for this.
(c) Do you see obvious outliers? Observation 3 is an outlier in the x
direction, and therefore possibly a high leverage point and influential
on the model fit.
(d) Is the normality assumption reasonable? There appears to be
some skewness, but with so few points normality may be reasonable.
(e) Is there a striking pattern in residuals vs. order of the data? No
striking pattern.
5. Check the Cook’s D values. We anticipated that the 3rd observation is
affecting the fit by a lot more than any other values. The D-value is much
larger than 1. Note that the residual is not large for this value.
6. Omit problem observations from the analysis and see if any conclu-
sions change substantially. Let us refit the equation without observa-
tion 3 to see if anything changes drastically. I will use the weighted least
squares approach discussed earlier on this example. Define a variable wt
that is 1 for all observations except obs. 3, and make it 0 for that one.
# wt = 1 for all except obs 3 where wt = 0

thyroid$wt <- as.numeric(!(thyroid$id == 3))
thyroid$wt
## [1] 1 1 0 1 1 1 1 1
What changes by deleting case 3? The fitted line gets steeper (slope
changes from −1.30 to −2.19), adjusted R2 gets larger (up to 58% from
53%), and S changes from 11.7 to 10.6. Because the Weight values are
much less spread out, SE(βˆ1) becomes quite a bit larger (to 0.714, up
from 0.436) and we lose a degree of freedom for MS Error (which will
penalize us on tests and CIs). Just about any quantitative statement
we would want to make using CIs would be about the same either way
since CIs will overlap a great deal, and our qualitative interpretations are
unchanged (Blood Loss drops with Weight). Unless something shows up
in the plots, I don’t see any very important changes here.
lm.blood.wt.no3 <- lm(blood_loss ~ weight, data = thyroid, weights = wt)
summary(lm.blood.wt.no3)
##
## Call:
## lm(formula = blood_loss ~ weight, data = thyroid, weights = wt)
##
## Weighted Residuals:
## 1 2 3 4 5 6 7 8
## 8.503 -12.613 0.000 9.187 0.664 8.445 -1.019 -13.168
##
## Coefficients:
## (Intercept) 591.668 32.567 18.17 9.3e-06 ***
## weight -2.193 0.714 -3.07 0.028 *
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
# exclude obs 3
thyroid.no3 <- subset(thyroid, wt == 1)
library(ggplot2)
p <- ggplot(thyroid.no3, aes(x = weight, y = blood_loss, label = id))
print(p)
520
7
●
4
● 1
blood_loss
●
500
2 6
● ●
5
●
480
8
●
460
35 40 45 50
weight
Nothing very striking shows up in the residual plots, and no Cook’s D

values are very large among the remaining observations.
# plot diagnistics
par(mfrow=c(2,3))
plot(lm.blood.wt.no3, which = c(1,4,6))
plot(thyroid.no3$weight, lm.blood.wt.no3$residuals[(thyroid$wt == 1)]
, main="Residuals vs weight")
library(car)
qqPlot(lm.blood.wt.no3$residuals, las = 1, id.n = 3, main="QQ Plot")
## 3 8 2
## 8 1 2
plot(lm.blood.wt.no3$residuals, main="Residuals vs Order of data")

2 1.5 1
10
8 8●
4●
● ●
0.6
0.6
5
Cook's distance
Cook's distance
Residuals
●
0.4
0
0.4
●
0.5
2 ●2
−5
0.2
0.2
6
−10
●6
●
2● ●
●8
−15
0.0
0.0
●
● 0
480 490 500 510 1 2 3 4 5 6 7 0.1 0.3 0.4 0.5
Residuals vs weight QQ Plot Residuals vs Order of data

lm.blood.wt.no3$residuals[(thyroid$wt == 1)]
10
3●
30
● ●
● ● 30
lm.blood.wt.no3$residuals
lm.blood.wt.no3$residuals
5
20
20
●
0
10
●
10 ● ●
●
● ● ●
−5
● ●
0
0
● ●
−10
−10
−10
●
● ● 8 ● 2 ● ●
35 40 45 50 −1.5 −1.0 −0.5 0.0 0.5 1.0 1.5 1 2 3 4 5 6 7 8
thyroid.no3$weight norm quantiles Index
How much difference is there in a practical sense? Examine the 95% predic-
tion interval for a new observation at Weight = 50kg. Previously we saw that
interval based on all 8 observations was from 457.1 to 517.8 ml of Blood Loss.
Based on just the 7 observations the prediction interval is 451.6 to 512.4 ml.
There really is no practical difference here.
# CI for the mean and PI for a new observation at weight=50
predict(lm.blood.wt , data.frame(weight=50), interval = "prediction")
## fit lwr upr
## 1 487.4 457.1 517.8
predict(lm.blood.wt.no3, data.frame(weight=50), interval = "prediction")
## Warning: Assuming constant prediction variance even though model fit is weighted
## fit lwr upr
## 1 482 451.6 512.4
Therefore, while obs. 3 was potentially influential, whether the value is

included or not makes very little difference in the model fit or relationship
between Weight and BloodLoss.
8.9.2 Gesell data
These data are from a UCLA study of cyanotic heart disease in children. The
predictor is the age of the child in months at first word and the response variable
is the Gesell adaptive score, for each of 21 children.
id age score
1 1 15 95
2 2 26 71
3 3 10 83
4 4 9 91
5 5 15 102
6 6 20 87
7 7 18 93
8 8 11 100
9 9 8 104
10 10 20 94
11 11 7 113
12 12 9 96
13 13 10 83
14 14 11 84
15 15 11 102
16 16 10 100
17 17 12 105
18 18 42 57
19 19 17 121
20 20 11 86
21 21 10 100
Let us go through the same steps as before.

1. Plot Score versus Age. Comment on the relationship between Score and
Age.
library(ggplot2)
p <- ggplot(gesell, aes(x = age, y = score, label = id))
print(p)
19
●
120
11
●
9 17
● 15 ● 5
16 8
21 ● ●
100 ● ●
12 1
●
● 7 10
●
4 ●
●
20 6
●
3 14
13 ●
score
●
●
80
2
●
60 18
●
10 20 30 40
age
2. There are no obvious transformations to try here.

3. Fit a simple linear regression model. Provide an equation for the LS line.
Does age at first word appear to be an “important predictor” of Gesell
adaptive score? (i.e., is the estimated slope significantly different from
zero?)
lm.score.age <- lm(score ~ age, data = gesell)
summary(lm.score.age)
##
## Call:
## lm(formula = score ~ age, data = gesell)
##
## Residuals:
## -15.60 -8.73 1.40 4.52 30.29
##
## Coefficients:
## (Intercept) 109.87 5.07 21.68 7.3e-15 ***
## age -1.13 0.31 -3.63 0.0018 **
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## Residual standard error: 11 on 19 degrees of freedom
4. Do these plots suggest any inadequacies with the model?
# plot diagnistics
par(mfrow=c(2,3))
plot(lm.score.age, which = c(1,4,6))
plot(gesell$age, lm.score.age$residuals, main="Residuals vs age")
library(car)
qqPlot(lm.score.age$residuals, las = 1, id.n = 3, main="QQ Plot")
## 19 3 13
## 21 1 2
plot(lm.score.age$residuals, main="Residuals vs Order of data")

18
2.5
3 2 1.5 1 18 ●
30
● 19
0.6
0.6
20
Cook's distance
Cook's distance
0.5
Residuals
0.4
0.4
●
10
● ●
●
●
● ● ●
● ●
0
● 19 ● 19
0.2
0.2
●
●
−10
● ●
● 2
● ● ●2
3●
13 ●●
●
●
0.0
0.0
●
−20
●
●●
●
●● 0
70 80 90 100 5 10 15 20 0 0.2 0.4 0.5 0.6
Residuals vs age QQ Plot Residuals vs Order of data
19 ●
30
30
● 30 ●
lm.score.age$residuals
20
20
20
● ●
10
10
● ● 10 ● ● ● ●
● ● ●
● ● ●
● ● ● ● ● ● ● ●
● ● ● ●
● ● ● ● ●
0
0
● ● ●
● ● ●
● ● ●
−10
−10
● ●
●
● ●
●
−10 ●
●
●
● ● ●
● ● 3 ● 13 ● ●
10 15 20 25 30 35 40 −2 −1 0 1 2 5 10 15 20
gesell$age norm quantiles Index
5. Observations 18 and 19 stand out with relatively high Cook’s D. The

cutoff line is only a rough guideline. Those two were flagged with high
influence and standardized residual, respectively, also. Be sure to examine
the scatter plot carefully to see why 18 and 19 stand out.
6. Consider doing two additional analyses: Analyze the data after omitting
case 18 only and analyze the data after omitting case 19 only. Refit the
regression model for each of these two scenarios. Provide a summary
table such as the following, giving the relevant summary statistics for
the three analyses. Discuss the impact that observations 18 and 19 have
individually on the fit of the model.
When observation 18 is omitted, the estimated slope is not significantly
different from zero (p-value = 0.1489), indicating that age is not an impor-
tant predictor of Gesell score. This suggests that the significance of age
as a predictor in the original analysis was due solely to the presence of ob-
servation 18. Note the dramatic decrease in R2 after deleting observation
18.
The fit of the model appears to improve when observation 19 is omitted.
For example, R2 increases noticeably and the p-value for testing the sig-
nificance of the slope decreases dramatically (in a relative sense). These
tendencies would be expected based on the original plot. However, this
improvement is misleading. Once observation 19 is omitted, observation
18 is much more influential. Again the significance of the slope is due to
the presence of observation 18.
Feature Full data Omit 18 Omit 19

b0 109.87 105.63 109.30
b1 -1.13 -0.78 -1.19
SE(b0) 5.07 7.16 3.97
SE(b1) 0.31 0.52 0.24
R2 0.41 0.11 0.57
p-val for H0 : β1 = 0 0.002 0.149 0.000
Can you think of any reasons to justify doing the analysis without observation
18?
If you include observation 18 in the analysis, you are assuming that the mean
Gesell score is linearly related to age over the entire range of observed ages.
Observation 18 is far from the other observations on age (age for observation
18 is 42; the second highest age is 26; the lowest age is 7). There are no
children with ages between 27 and 41, so we have no information on whether
the relationship is roughly linear over a significant portion of the range of ages. I
am comfortable deleting observation 18 from the analysis because it’s inclusion
forces me to make an assumption that I can not check using these data. I am
only willing to make predictions of Gesell score for children with ages roughly
between 7 and 26. However, once this point is omitted, age does not appear to
be an important predictor.
A more complete analysis would delete observation 18 and 19 together.
What would you expect to see if you did this?
8.10 Weighted Least Squares

Earlier I indicated that nonconstant error variance can be addressed (some-
times) with weighted least squares. The scedastic function is the conditional
variance of Y given X. Y is said to be heteroscedastic if it’s variance
depends on the value of X (variance changes), and homoscedastic if it’s
variance does not depend on X (constant variance).
Recall the LS (OLS or ordinary LS) line chooses the values of β0 and β1
that minimize n
X
{Yi − (β0 + β1Xi)}2
i=1
over all possible choices of β0 and β1. The weighted LS (WLS) line chooses the
values of β0 and β1 that minimize
n
X
wi{Yi − (β0 + β1Xi)}2
i=1
over all possible choices of β0 and β1. If σY |X depends up X, then the correct
choice of weights is inversely proportional to variance, wi ∝ σY2 |X .
Consider the following data and plot of y vs. x and standardized OLS resid-
uals vs x. It is very clear that variability increases with x.
#### Weighted Least Squares
# R code to generate data
set.seed(7)
n <- 100
# 1s, Xs uniform 0 to 100
X <- matrix(c(rep(1,n),runif(n,0,100)), ncol=2)
# intercept and slope (5, 5)
beta <- matrix(c(5,5),ncol=1)
8.10: Weighted Least Squares 331
# errors are X*norm(0,1), so variance increases with X
e <- X[,2]*rnorm(n,0,1)
# response variables
y <- X %*% beta + e
# put data into data.frame

wlsdat <- data.frame(y, x = X[,2])
# fit regression
lm.y.x <- lm(y ~ x, data = wlsdat)
# put residuals in data.frame

wlsdat$res <- lm.y.x$residuals
# ggplot: Plot the data with linear regression fit

library(ggplot2)
p <- ggplot(wlsdat, aes(x = x, y = y))
print(p)
# ggplot: Plot the residuals

library(ggplot2)
p <- ggplot(wlsdat, aes(x = x, y = res))
print(p)
● ●
●
600
● 100 ●
● ●
● ● ● ●
●
●
●
● ●
● ●
● ● ● ●
●
● ●
● ● ● ●
● ● ● ● ●
● ●●
● ●
●
● ● ●
● ● ●
● ● ● ● ● ●
400 ● ● ● ● ● ● ●
● ● ● ● ●
● ● ● ● ●
● ● ● ●
● ● ● ● ●
● ●● ● ●
●●
● 0 ●
●
●●●
●
●
●
●
● ●
● ● ●● ●
● ● ●
● ●
res
● ●
y
● ● ●
● ●
● ● ●
● ● ● ●
● ●
●
● ●
● ● ●
● ● ●
●● ● ● ●
●
● ● ●
●
● ●●
● ●
● ● ● ●
● ● ●
● ●
200 ●
● ●
● ● ●
●
●
●
●
●
● ● ● −100
● ●
● ● ●
●
●
●● ●
●
●●●● ●
● ●
●
●●●
●●
● ●
●
● ●
0
0 25 50 75 100 0 25 50 75 100
x x
In order to use WLS to solve this problem, we need some form for σY2 |X .
Finding that form is a real problem with WLS. It can be useful to plot the
absolute value of the standardized residual vs. x to see if the top boundary
seems to follow a general pattern.
# ggplot: Plot the absolute value of the residuals
library(ggplot2)
p <- ggplot(wlsdat, aes(x = x, y = abs(res)))
print(p)
150
100 ●
●
abs(res)
● ●
● ●
●
●
●
●
● ●
●
●
●
● ●
● ●
● ●
50 ● ●
● ●
●
●
●
● ● ● ● ●
● ●
● ● ● ●
● ●● ● ●
● ● ●
● ● ● ● ● ●
●
● ● ● ● ●
● ● ●
● ● ●
● ●
●● ● ●
●
●● ● ●
●
● ● ● ●
● ● ● ●
● ● ● ●● ●
0 ●● ● ●
0 25 50 75 100
x
It is plausible the upper boundary is linear, so let us try wi = x12 . Stan-

dardized residuals from this WLS fit look very good. Note that raw (non-
standardized) residuals will still have the same pattern — it is essential to use
standardized residuals here.
# fit regression
lm.y.x.wt <- lm(y ~ x, data = wlsdat, weights = x^(-2))
# put residuals in data.frame

wlsdat$reswt <- lm.y.x.wt$residuals
wlsdat$wt <- lm.y.x.wt$weights^(1/2)
# ggplot: Plot the absolute value of the residuals

library(ggplot2)
p <- ggplot(wlsdat, aes(x = x, y = reswt*wt))
print(p)
8.10: Weighted Least Squares 333
●
●
● ●
●
●
●
1 ●
●
● ●
● ● ● ● ●
●
● ● ● ●
●
● ● ● ● ●
● ● ●
●
reswt * wt
●
● ●
● ● ● ●
● ● ●
● ● ● ●
● ● ● ●
● ●
●
● ● ●
0 ●
●
●
● ● ●
● ● ●
● ●
●
● ●
●
●●
●
● ● ●
● ●
● ●
●
−1 ● ● ●
● ●
●
●
●
●
●
● ●
● ● ●
● ●
●
−2
0 25 50 75 100
x
Compare also the OLS fitted equation:

summary(lm.y.x)
##
## Call:
## lm(formula = y ~ x, data = wlsdat)
##
## Residuals:
## -168.18 -24.94 2.54 24.97 125.37
##
## Coefficients:
## (Intercept) 7.631 10.426 0.73 0.47
## x 5.035 0.179 28.12 <2e-16 ***
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## F-statistic: 790 on 1 and 98 DF, p-value: <2e-16
to the WLS fitted equation:
summary(lm.y.x.wt)
##
## Call:
## lm(formula = y ~ x, data = wlsdat, weights = x^(-2))
##
## Weighted Residuals:
## -1.894 -0.671 0.178 0.596 2.313
##
## Coefficients:

## (Intercept) 5.0931 0.5493 9.27 4.6e-15 ***
## x 5.1069 0.0939 54.40 < 2e-16 ***
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## F-statistic: 2.96e+03 on 1 and 98 DF, p-value: <2e-16
Clearly the weighted fit looks better, although note that everything is based
on the weighted SS. In practice it can be pretty difficult to determine the correct
set of weights, but WLS works much better than OLS if appropriate. I actually
simulated this data set using β0 = β1 = 5. Which fit actually did better?
Chapter 9
Introduction to the
Bootstrap
Learning objectives
explain the bootstrap principle for hypothesis tests and inference.
decide (for simple problems) how to construct a bootstrap procedure.
9.1 Introduction
Statistical theory attempts to answer three basic questions:
1. How should I collect my data?
2. How should I analyze and summarize the data that I’ve collected?
3. How accurate are my data summaries?
Question 3 consitutes part of the process known as statistical inference. The
bootstrap makes certain kinds of statistical inference1. Let’s look at an example.
1
Efron (1979), “Bootstrap methods: another look at the jackknife.” Ann. Statist. 7, 1–26
336 Ch 9: Introduction to the Bootstrap
Example: Aspirin and heart attacks, large-sample theory Does
aspirin prevent heart attacks in healthy middle-aged men? A controlled, ran-
domized, double-blind study was conducted and gathered the following data.
(fatal plus non-fatal)
heart attacks subjects
aspirin group: 104 11037
placebo group: 189 11034
A good experimental design, such as this one, simplifies the results! The ratio
of the two rates (the risk ratio) is
104/11037
θ̂ = = 0.55.
189/11034
Because of the solid experimental design, we can believe that the aspirin-takers
only have 55% as many heart attacks as the placebo-takers.
We are not really interested in the estimated ratio θ̂, but the true ratio, θ.
That is the ratio if we could treat all possible subjects, not just a sample of
them. Large sample theory tells us that the log risk ratio has an approximate
Normal distribution. The standard error of the log risk ratio is estimated simply
by the square root of the sum of the reciprocals of the four frequencies:
r
1 1 1 1
SE(log(RR)) = + + + = 0.1228
104 189 11037 11034
The 95% CI for log(θ) is
log(θ̂) ± 1.96 × SE(log(RR)), (−0.839, −0.357),
and expontiating gives the CI on the ratio scale,
exp{log(θ̂) ± 1.96 × SE(log(RR))}, (0.432, 0.700).
The same data that allowed us to estimate the ratio θ with θ̂ = 0.55 also
allowed us to get an idea of the estimate’s accuracy.
Example: Aspirin and strokes, large-sample theory The aspirin

study tracked strokes as well as heart attacks.
9.2: Bootstrap 337
strokes subjects
aspirin group: 119 11037
placebo group: 98 11034
The ratio of the two rates (the risk ratio) is
119/11037
θ̂ = = 1.21.
98/11034
It looks like aspirin is actually harmful, now, however the 95% interval for the
true stroke ratio θ is (0.925, 1.583). This includes the neutral value θ = 1, at
which aspirin would be no better or worse than placebo for strokes.
9.2 Bootstrap
The bootstrap is a data-based simulation method for statistical inference, which
can be used to produce inferences like those in the previous slides. The term
“bootstrap” comes from literature. In “The Adventures of Baron Munchausen”,
by Rudolph Erich Raspe, the Baron had fallen to the bottom of a deep lake,
and he thought to get out by pulling himself up by his own bootstraps.
9.2.1 Ideal versus Bootstrap world, sampling distri-

butions
Ideal world
1. Population of interest
2. Obtain many simple random samples (SRSs) of size n
3. For each SRS, calculate statistic of interest (θ)
4. Sampling distribution is the distribution of the calculated statistic
Bootstrap world
1. Population of interest; One empirical distribution based on a sample of
size n
2. Obtain many bootstrap resamples of size n
3. For each resample, calculate statistic of interest (θ∗)
4. Bootstrap distribution is the distribution of the calculated statistic
5. Bootstrap distribution estimates the sampling distribution centered at the
statistic (not the parameter).
Example: Aspirin and strokes, bootstrap Here’s how the bootstrap

works in the stroke example. We create two populations:
the first consisting of 119 ones and 11037 − 119 = 10918 zeros,
the second consisting of 98 ones and 11034 − 98 = 10936 zeros.
We draw with replacement a sample of 11037 items from the first population,
and a sample of 11034 items from the second population. Each is called a
bootstrap sample. From these we derive the bootstrap replicate of θ̂:
Proportion of ones in bootstrap sample 1
θ̂∗ = .
Proportion of ones in bootstrap sample 2
Repeat this process a large number of times, say 10000 times, and obtain 10000
bootstrap replicates θ̂∗. The summaries are in the code, followed by a histogram
of bootstrap replicates, θ̂∗.
#### Example: Aspirin and strokes, bootstrap
# sample size (n) and successes (s) for sample 1 (aspirin) and 2 (placebo)
n <- c(11037, 11034)
s <- c( 119, 98)
# data for samples 1 and 2, where 1 = success (stroke), 0 = failure (no stroke)
dat1 <- c(rep(1, s[1]), rep(0, n[1] - s[1]))
dat2 <- c(rep(1, s[2]), rep(0, n[2] - s[2]))
# draw R bootstrap replicates
R <- 10000
# init location for bootstrap samples
bs1 <- rep(NA, R)
bs2 <- rep(NA, R)
# draw R bootstrap resamples of proportions
for (i in 1:R) {
# proportion of successes in bootstrap samples 1 and 2
# (as individual steps for group 1:)
resam1 <- sample(dat1, n[1], replace = TRUE)
success1 <- sum(resam1)
bs1[i] <- success1 / n[1]
# (as one line for group 2:)
bs2[i] <- sum(sample(dat2, n[2], replace = TRUE)) / n[2]
}
# bootstrap replicates of ratio estimates
rat <- bs1 / bs2
# sort the ratio estimates to obtain bootstrap CI
rat.sorted <- sort(rat)
# 0.025th and 0.975th quantile gives equal-tail bootstrap CI
9.2: Bootstrap 339
CI.bs <- c(rat.sorted[round(0.025*R)], rat.sorted[round(0.975*R+1)])

CI.bs
## [1] 0.9399 1.5878
## Plot the bootstrap distribution with CI
# First put data in data.frame for ggplot()
dat.rat <- data.frame(rat)
library(ggplot2)
p <- ggplot(dat.rat, aes(x = rat))
, binwidth=0.02
# vertical line at 1 and CI
p <- p + geom_vline(xintercept=1, colour="#BB0000", linetype="dashed")
p <- p + geom_vline(xintercept=CI.bs[1], colour="#00AA00", linetype="longdash")
p <- p + labs(title = "Bootstrap distribution of relative risk ratio, strokes")
p <- p + xlab("ratio (red = 1, green = bootstrap CI)")
print(p)
## Warning: position stack requires constant width: output may be incorrect
Bootstrap distribution of relative risk ratio, strokes
2
density
1.0 1.5 2.0

ratio (red = 1, green = bootstrap CI)
In this simple case, the confidence interval derived from the bootstrap
(0.94, 1.588) agrees very closely with the one derived from statistical theory
(0.925, 1.583). Bootstrap methods are intended to simplify the calculation of
inferences like those using large-sample theory, producing them in an automatic
way even in situations much more complicated than the risk ratio in the aspirin
example.
9.2.2 The accuracy of the sample mean

For sample means, and essentially only for sample means, an accuracy formula
(for the standard error of the parameter) is easy to obtain (using the delta
method). We’ll see how to use the bootstrap for the sample mean, then for the
more complicated situation of assessing the accuracy of the median.
Bootstrap Principle The plug-in principle is used when the underly-

ing distribution is unknown and you substitute your best guess for what that
distribution is. What to substitute?
Empirical distribution ordinary bootstrap
Smoothed distribution (kernel) smoothed bootstrap
Parametric distribution parametric bootstrap
Satisfy assumptions such as the null hypothesis
This substitution works in many cases, but not always. Keep in mind that the
bootstrap distribution is centered at the statistic, not the parameter. Imple-
mention is done by Monte Carlo sampling.
The bootstrap is commonly implemented in one of two ways, nonparamet-
rically or parametrically. An exact nonparametric bootstrap requires nn
samples! That’s one for every possible combination of each of n observation po-
sitions taking the value of each of n observations. This is sensibly approximated
by using the Monte Carlo strategy of drawing a large number (1000 or 10000)
of random resamples. On the other hand, a parametric bootstrap first
assumes a distribution for the population (such as a normal distribution) and
estimates the distributional parameters (such as the mean and variance) from
the observed sample. Then, the Monte Carlo strategy is used to draw a large
number (1000 or 10000) of samples from the estimated parametric distribution.
Example: Mouse survival, two-sample t-test, mean Sixteen mice

were randomly assigned to a treatment group or a control group. Shown are
their survival times, in days, following a test surgery. Did the treatment prolong
survival?
9.2: Bootstrap 341
Group Data n Mean SE
Control: 52, 104, 146, 10, 9 56.22 14.14
51, 30, 40, 27, 46
Treatment: 94, 197, 16, 38, 7 86.86 25.24
99, 141, 23
Difference: 30.63 28.93
Numerical and graphical summaries of the data are below. There seems to
be a slight difference in variability between the two treatment groups.
#### Example: Mouse survival, two-sample t-test, mean
treatment <- c(94, 197, 16, 38, 99, 141, 23)
control <- c(52, 104, 146, 10, 51, 30, 40, 27, 46)
survive <- c(treatment, control)
group <- c(rep("Treatment", length(treatment)), rep("Control", length(control)))
mice <- data.frame(survive, group)
library(plyr)
# ddply "dd" means the input and output are both data.frames
mice.summary <- ddply(mice,
"group",
function(X) {
data.frame( m = mean(X$survive),
s = sd(X$survive),
n = length(X$survive)
)
}
)
# standard errors
mice.summary$se <- mice.summary$s/sqrt(mice.summary$n)
# individual confidence limits
mice.summary$ci.l <- mice.summary$m - qt(1-.05/2, df=mice.summary$n-1) * mice.summary$se
mice.summary$ci.u <- mice.summary$m + qt(1-.05/2, df=mice.summary$n-1) * mice.summary$se
mice.summary
## group m s n se ci.l ci.u
## 1 Control 56.22 42.48 9 14.16 23.57 88.87
## 2 Treatment 86.86 66.77 7 25.24 25.11 148.61
diff(mice.summary$m) #£
## [1] 30.63
p <- ggplot(mice, aes(x = survive))
p <- p + facet_grid(group ~ .)
p <- p + labs(title = "Mouse survival following a test surgery") + xlab("Survival (days)")
print(p)
Mouse survival following a test surgery
4
Control
2
count
0
4
Treatment
2
0
0 50 100 150 200
Survival (days)
√
The standard error for the difference is 28.93 = 25.242 + 14.142, so the
observed difference of 30.63 is only 30.63/28.93=1.05 estimated standard errors
greater than zero, an insignificant result.
The two-sample t-test of the difference in means confirms the lack of sta-
tistically significant difference between these two treatment groups with a p-
value=0.3155.
t.test(survive ~ group, data = mice)
##
##
## data: survive by group
## t = -1.059, df = 9.654, p-value = 0.3155
## -95.42 34.15
## mean in group Control mean in group Treatment
## 56.22 86.86
But these are small samples, and the control sample does not look normal.
We could do a nonparametric two-sample test of difference of medians. Or, we
could use the bootstrap to make our inference.
Example: Mouse survival, two-sample bootstrap, mean Here’s

how the bootstrap works in the two-sample mouse example. We draw with
replacement from each sample, calculate the mean for each sample, then take
9.2: Bootstrap 343
the difference in means. Each is called a bootstrap sample of the difference in
means. From these we derive the bootstrap replicate of µ̂:
µ̂∗ = x̄∗ − ȳ ∗.
Repeat this process a large number of times, say 10000 times, and obtain 10000
bootstrap replicates µ̂∗. The summaries are in the code, followed by a histogram
of bootstrap replicates, µ̂∗.
#### Example: Mouse survival, two-sample bootstrap, mean
R <- 10000
bs1 <- rep(NA, R)
bs2 <- rep(NA, R)
# draw R bootstrap resamples of means
for (i in 1:R) {
bs2[i] <- mean(sample(control, replace = TRUE))
bs1[i] <- mean(sample(treatment, replace = TRUE))
}
# bootstrap replicates of difference estimates
bs.diff <- bs1 - bs2
sd(bs.diff)
## [1] 27
# sort the difference estimates to obtain bootstrap CI
diff.sorted <- sort(bs.diff)
CI.bs <- c(diff.sorted[round(0.025*R)], diff.sorted[round(0.975*R+1)])
CI.bs
## [1] -21.97 83.10
dat.diff <- data.frame(bs.diff)
library(ggplot2)
p <- ggplot(dat.diff, aes(x = bs.diff))
, binwidth=5
p <- p + labs(title = "Bootstrap distribution of difference in survival time, median")
print(p)
Bootstrap distribution of difference in survival time, median
0.015
0.010
density
0.005
0.000
−50 0 50 100 150

Example: Mouse survival, two-sample bootstrap, median For

most statistics (such as the median) we don’t have a formula for the limiting
value of the standard error, but in fact no formula is needed. Instead, we use
the numerical output of the bootstrap program. The summaries are in the code,
followed by a histogram of bootstrap replicates, η̂ ∗.
Group Data (n) Median est. SE
Control: 52, 104, 146, 10, (9) 46 ?
51, 30, 40, 27, 46
Treatment: 94, 197, 16, 38, (7) 94 ?
99, 141, 23
Difference: 48 ?
#### Example: Mouse survival, two-sample bootstrap, median
sort(control)
## [1] 10 27 30 40 46 51 52 104 146
sort(treatment)
## [1] 16 23 38 94 99 141 197
R <- 10000
bs1 <- rep(NA, R)
bs2 <- rep(NA, R)
# draw R bootstrap resamples of medians
for (i in 1:R) {
9.2: Bootstrap 345
bs2[i] <- median(sample(control, replace = TRUE))

bs1[i] <- median(sample(treatment, replace = TRUE))
}
# bootstrap replicates of difference estimates
bs.diff <- bs1 - bs2
sd(bs.diff)
## [1] 40.43
diff.sorted <- sort(bs.diff)
CI.bs <- c(diff.sorted[round(0.025*R)], diff.sorted[round(0.975*R+1)])
CI.bs
## [1] -29 111
dat.diff <- data.frame(bs.diff)
library(ggplot2)
p <- ggplot(dat.diff, aes(x = bs.diff))
, binwidth=5
p <- p + labs(title = "Bootstrap distribution of difference in survival time, median")
print(p)
Bootstrap distribution of difference in survival time, median
0.03
0.02
density
0.01
0.00
−100 0 100
9.2.3 Comparing bootstrap sampling distribution from

population and sample
Example: Law School, correlation of (LSAT, GPA) The popula-

tion of average student measurements of (LSAT, GPA) for the universe of 82
law schools are in the table below. Imagine that we don’t have all 82 schools
worth of data. Consider taking a random sample of 15 schools, indicated by
the +’s.
9.2: Bootstrap 347
School LSAT GPA School LSAT GPA School LSAT GPA
1 622 3.23 28 632 3.29 56 641 3.28
2 542 2.83 29 587 3.16 57 512 3.01
3 579 3.24 30 581 3.17 58 631 3.21
4+ 653 3.12 31+ 605 3.13 59 597 3.32
5 606 3.09 32 704 3.36 60 621 3.24
6+ 576 3.39 33 477 2.57 61 617 3.03
7 620 3.10 34 591 3.02 62 637 3.33
8 615 3.40 35+ 578 3.03 62 572 3.08
9 553 2.97 36+ 572 2.88 64 610 3.13
10 607 2.91 37 615 3.37 65 562 3.01
11 558 3.11 38 606 3.20 66 635 3.30
12 596 3.24 39 603 3.23 67 614 3.15
13+ 635 3.30 40 535 2.98 68 546 2.82
14 581 3.22 41 595 3.11 69 598 3.20
15+ 661 3.43 42 575 2.92 70+ 666 3.44
16 547 2.91 43 573 2.85 71 570 3.01
17 599 3.23 44 644 3.38 72 570 2.92
18 646 3.47 45+ 545 2.76 73 605 3.45
19 622 3.15 46 645 3.27 74 565 3.15
20 611 3.33 47+ 651 3.36 75 686 3.50
21 546 2.99 48 562 3.19 76 608 3.16
22 614 3.19 49 609 3.17 77 595 3.19
23 628 3.03 50+ 555 3.00 78 590 3.15
24 575 3.01 51 586 3.11 79+ 558 2.81
25 662 3.39 52+ 580 3.07 80 611 3.16
26 627 3.41 53+ 594 2.96 81 564 3.02
27 608 3.04 54 594 3.05 82+ 575 2.74
55 560 2.93
#### Example: Law School, correlation of (LSAT, GPA)

School <- 1:82
LSAT <- c(622, 542, 579, 653, 606, 576, 620, 615, 553, 607, 558, 596, 635,
581, 661, 547, 599, 646, 622, 611, 546, 614, 628, 575, 662, 627,
608, 632, 587, 581, 605, 704, 477, 591, 578, 572, 615, 606, 603,
535, 595, 575, 573, 644, 545, 645, 651, 562, 609, 555, 586, 580,
594, 594, 560, 641, 512, 631, 597, 621, 617, 637, 572, 610, 562,
635, 614, 546, 598, 666, 570, 570, 605, 565, 686, 608, 595, 590,
558, 611, 564, 575)
GPA <- c(3.23, 2.83, 3.24, 3.12, 3.09, 3.39, 3.10, 3.40, 2.97, 2.91, 3.11,
3.24, 3.30, 3.22, 3.43, 2.91, 3.23, 3.47, 3.15, 3.33, 2.99, 3.19,
3.03, 3.01, 3.39, 3.41, 3.04, 3.29, 3.16, 3.17, 3.13, 3.36, 2.57,
3.02, 3.03, 2.88, 3.37, 3.20, 3.23, 2.98, 3.11, 2.92, 2.85, 3.38,
2.76, 3.27, 3.36, 3.19, 3.17, 3.00, 3.11, 3.07, 2.96, 3.05, 2.93,
3.28, 3.01, 3.21, 3.32, 3.24, 3.03, 3.33, 3.08, 3.13, 3.01, 3.30,
3.15, 2.82, 3.20, 3.44, 3.01, 2.92, 3.45, 3.15, 3.50, 3.16, 3.19,
3.15, 2.81, 3.16, 3.02, 2.74)
Sampled <- c(0, 0, 0, 1, 0, 1, 0, 0, 0, 0, 0, 0, 1, 0, 1, 0, 0, 0, 0, 0, 0, 0,

0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0,
1, 0, 1, 0, 0, 1, 0, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 1)
# law = population
law <- data.frame(School, LSAT, GPA, Sampled)
law$Sampled <- factor(law$Sampled)
# law.sam = sample
law.sam <- subset(law, Sampled == 1)
library(ggplot2)
p <- ggplot(law, aes(x = LSAT, y = GPA))
p <- p + geom_point(aes(colour = Sampled, shape = Sampled), alpha = 0.5, size = 2)
p <- p + labs(title = "Law School average scores of LSAT and GPA")
print(p)
Law School average scores of LSAT and GPA

3.50
3.25
Sampled
GPA
0
3.00 1
2.75
500 550 600 650 700

LSAT
Let’s bootstrap the sample of 15 observations to get the bootstrap sampling

distribution of correlation (for sampling 15 from the population). From the
bootstrap sampling distribution we’ll calculate a bootstrap confidence interval
for the true population correlation, as well as a bootstrap standard deviation for
the correlation. But how well does this work? Let’s compare it against the true
sampling distribution by drawing 15 random schools from the population of 82
9.2: Bootstrap 349
schools and calculating the correlation. If the bootstrap works well (from our
hopefully representative sample of 15), then the bootstrap sampling distribution
from the 15 schools will be close to the true sampling distribution.
The code below does that, followed by two histograms. In this case, the
histograms are noticeably non-normal, having a long tail toward the left. In-
ferences based on the normal curve are suspect when the bootstrap histogram
is markedly non-normal. The histogram on the left is the nonparametric boot-
strap sampling distribution using only the n = 15 sampled schools with 10000
bootstrap replicates of cdorr(x∗). The histogram on the right is the true sam-
orr(x∗) from the population of law
pling distribution using 10000 replicates of cd
school data, repeatedly drawing n = 15 without replacement from the N = 82
points. Impressively, the bootstrap histogram on the left strongly resembles
the population histogram on the right. Remember, in a real problem we would
only have the information on the left, from which we would be trying to infer
the situation on the right.
R <- 10000
bs.pop <- rep(NA, R)
bs.sam <- rep(NA, R)
# draw R bootstrap resamples of medians
for (i in 1:R) {
# sample() draws indicies then bootstrap correlation of LSAT and GPA
# population
bs.pop[i] = cor(law [sample(seq(1,nrow(law )), nrow(law.sam)
, replace = TRUE), 2:3])[1, 2]
# sample
bs.sam[i] = cor(law.sam[sample(seq(1,nrow(law.sam)), nrow(law.sam)
, replace = TRUE), 2:3])[1, 2]
}

diff.sorted <- sort(bs.pop)
CI.bs.pop <- c(diff.sorted[round(0.025*R)], diff.sorted[round(0.975*R+1)])
# population correlation
cor(law [, c(2,3)])[1,2]
## [1] 0.76
CI.bs.pop
## [1] 0.4297 0.9271
sd(bs.pop)
## [1] 0.1295

diff.sorted <- sort(bs.sam)
CI.bs.sam <- c(diff.sorted[round(0.025*R)], diff.sorted[round(0.975*R+1)])
# sample correlation
cor(law.sam[, c(2,3)])[1,2]
## [1] 0.7764
CI.bs.sam
## [1] 0.4638 0.9638
sd(bs.sam)
## [1] 0.1335
law.bs.df <- data.frame(corr = c(bs.pop, bs.sam), group = c(rep("Pop",R),rep("Sam",R)))
library(ggplot2)
p <- ggplot(law.bs.df, aes(x = corr, fill=group))
p <- p + geom_histogram(binwidth = .01, alpha = 0.5, position="identity")
p <- p + labs(title = "Sampling distribution of 15 observation from 82 (Pop) vs 15 (Sam, BS)")
xlab("Correlation")
print(p)
Sampling distribution of 15 observation from 82 (Pop) vs 15 (Sam, BS)
300
group
200
count
Pop
Sam
100
0.0 0.5 1.0

Correlation
Chapter 10
Power and Sample size

Learning objectives
assess the power of a test or
determin the required sample size for a study.
7. Distinguish between statistical significance and scientific relevance.
10. Identify and explain the statistical methods, assumptions, and limita-
tions.
12. Make evidence-based decisions by constructing and deciding between
testable hypotheses using appropriate data and methods.
10.1 Power Analysis

The meaning of statistical power Power is the probability (1 − β)
of detecting an effect, given that the effect is really there. In other words,
it is the probability of correctly rejecting the null hypothesis when it is in fact
false. For example, let’s say that we have a simple study with drug A and a
placebo group, and that the drug truly is effective; the power is the probability
of finding a difference between the two groups. So, imagine that we had a power
of 1 − β = 0.8 and that this simple study was conducted many times. Having
power of 0.8 means that 80% of the time, we would get a statistically significant
difference between the drug A and placebo groups. This also means that 20%
352 Ch 10: Power and Sample size
of the times that we run this experiment, we will not obtain a statistically
significant effect between the two groups, even though there really is an effect
in reality. That is, the probability of a Type-II error is β = 0.2.
One-sample power figure Consider the plot below for a one-sample one-
tailed greater-than t-test. If the null hypothesis, H0 : µ = µ0, is true, then
the test statistic t is follows the null distribution indicated by the hashed area.
Under a specific alternative hypothesis, H1 : µ = µ1, the test statistic t follows
the distribution indicated by the solid area. If α is the probability of making
a Type-I error (rejecting H0 when it is true), then “crit. val.” indicates the
location of the tcrit value associated with H0 on the scale of the data. The
rejection region is the area under H0 that is at least as far as “crit. val.” is from
µ0. The power (1 − β) of the test is the green area, the area under H1 in the
rejection region. A Type-II error is made when H1 is true, but we fail to reject
H0 in the red region. (Note, for a two-tailed test the rejection region for both
tails under the H1 curve contribute to the power.)
#### One-sample power
# Power plot with two normal distributions
# http://stats.stackexchange.com/questions/14140/how-to-best-display-graphically-type-ii-bet
x <- seq(-4, 4, length=1000)

hx <- dnorm(x, mean=0, sd=1)
plot(x, hx, type="n", xlim=c(-4, 8), ylim=c(0, 0.5),

ylab = "",
xlab = "",
main= expression(paste("Type-II Error (", beta, ") and Power (", 1-beta, ")")), axes=FALSE)
#shift = qnorm(1-0.025, mean=0, sd=1)*1.7

shift = qnorm(1-0.05, mean=0, sd=1)*1.7 # one-tailed
xfit2 <- x + shift
yfit2 <- dnorm(xfit2, mean=shift, sd=1)
#axis(1, at = c(-qnorm(.025), 0, shift, -4),

# labels = expression("p-value", 0, mu, -infinity ))
#axis(1, at = c(-qnorm(.025), 0, shift),
# labels = expression((t[alpha/2]), mu[0], mu[1]))
axis(1, at = c(-qnorm(.05), 0, shift),
labels = expression("crit. val.", mu[0], mu[1]))
axis(1, at = c(-4, 4+shift),
labels = expression(-infinity, infinity ), lwd=1, lwd.tick=FALSE)
## Print null hypothesis area

10.1: Power Analysis 353
#col_null = "#DDDDDD"
#polygon(c(min(x), x,max(x)), c(0,hx,0), col=col_null)
#lines(x, hx, lwd=2)
col_null = "#AAAAAA"
polygon(c(min(x), x,max(x)), c(0,hx,0), col=col_null, lwd=2, density=c(10, 40), angle=-45, bor
lines(x, hx, lwd=2, lty="dashed", col=col_null)
# The alternative hypothesis area
## The red - underpowered area

lb <- min(xfit2)
#ub <- round(qnorm(.975),2)
ub <- round(qnorm(.95),2)
col1 = "#CC2222"
i <- xfit2 >= lb & xfit2 <= ub

polygon(c(lb,xfit2[i],ub), c(0,yfit2[i],0), col=col1)
## The green area where the power is

col2 = "#22CC22"
i <- xfit2 >= ub
polygon(c(ub,xfit2[i],max(xfit2)), c(0,yfit2[i],0), col=col2)
# Outline the alternative hypothesis

lines(xfit2, yfit2, lwd=2)
axis(1, at = (c(ub, max(xfit2))), labels=c("", expression(infinity)),

col=col2, lwd=1, lwd.tick=FALSE)
#legend("topright", inset=.05, title="Color",

# c("Null hypoteses","Type II error", "True"), fill=c(col_null, col1, col2), horiz=FALSE)
legend("topright", inset=.015, title="Color",
c("Null hypothesis","Type-II error", "Power"), fill=c(col_null, col1, col2),
angle=-45,
density=c(20, 1000, 1000), horiz=FALSE)
abline(v=ub, lwd=2, col="#000088", lty="dashed")
arrows(ub, 0.45, ub+1, 0.45, lwd=3, col="#008800")

arrows(ub, 0.45, ub-1, 0.45, lwd=3, col="#880000")
Type−II Error (β) and Power (1 − β)
Color
Null hypothesis
Type−II error
Power
−∞ µ0 crit. val. µ1 ∞
Example: IQ drug Imagine that we are evaluating the effect of a puta-

tive memory enhancing drug. We have randomly sampled 25 people from a
population known to be normally distributed with a µ of 100 and a σ of 15.
We administer the drug, wait a reasonable time for it to take effect, and then
test our subjects’ IQ. Assume that we were so confident in our belief that the
drug would either increase IQ or have no effect that we entertained one-sided
(directional) hypotheses. Our null hypothesis is that after administering the
drug µ ≤ 100 and our alternative hypothesis is µ > 100.
These hypotheses must first be converted to exact hypotheses. Converting
the null is easy: it becomes µ = 100. The alternative is more troublesome.
If we knew that the effect of the drug was to increase IQ by 15 points, our
exact alternative hypothesis would be µ = 115, and we could compute power,
the probability of correctly rejecting the false null hypothesis given that µ is
really equal to 115 after drug treatment, not 100 (normal IQ). But if we already
10.1: Power Analysis 355
knew how large the effect of the drug was, we would not need to do inferential
statistics. . .
One solution is to decide on a minimum nontrivial effect size. What
is the smallest effect that you would consider to be nontrivial? Suppose that
you decide that if the drug increases µIQ by 2 or more points, then that is a
nontrivial effect, but if the mean increase is less than 2 then the effect is trivial.
Now we can test the null of µ = 100 versus the alternative of µ = 102. Con-
sider the previous plot. Let the left curve represent the distribution of sample
means if the null hypothesis were√true, µ = 100. This sampling distribution
has a µ = 100 and a σȲ = 15/ 25 = 3. Let the right curve represent the
sampling distribution if the exact alternative hypothesis is true, µ = 102. Its
µ is 102 and, assuming the drug has no effect on the variance in IQ scores, also
has σȲ = 3.
The green area in the upper tail of the null distribution (gray hatched curve)
is α. Assume we are using a one-tailed α of 0.05. How large would a sample
mean need be for us to reject the null? Since the upper 5% of a normal distribu-
tion extends from 1.645σ above the µ up to positive infinity, the sample mean
IQ would need be 100 + 1.645(3) = 104.935 or more to reject the null. What
are the chances of getting a sample mean of 104.935 or more if the alternative
hypothesis is correct, if the drug increases IQ by 2 points? The area under the
alternative curve from 104.935 up to positive infinity represents that probabil-
ity, which is power. Assuming the alternative hypothesis is true, that µ = 102,
the probability of rejecting the null hypothesis is the probability of getting a
sample mean of 104.935 or more in a normal distribution with µ = 102, σ = 3.
Z = (104.935 − 102)/3 = 0.98, and P (Z > 0.98) = 0.1635. That is, power is
about 16%. If the drug really does increase IQ by an average of 2 points, we
have a 16% chance of rejecting the null. If its effect is even larger, we have a
greater than 16% chance.
Suppose we consider 5 (rather than 2) the minimum nontrivial effect size.
This will separate the null and alternative distributions more, decreasing their
overlap and increasing power. Now, Z = (104.935 − 105)/3 = −0.02, P (Z >
−0.02) = 0.5080 or about 51%. It is easier to detect large effects than
small effects.
Suppose we conduct a 2-tailed test, since the drug could actually decrease
IQ; α is now split into both tails of the null distribution, 0.025 in each tail. We
shall reject the null if the sample mean is 1.96 or more standard errors away from
the µ of the null distribution. That is, if the mean is 100 + 1.96(3) = 105.88 or
more (or if it is 100−1.96(3) = 94.12 or less) we reject the null. The probability
of that happening if the alternative is correct (µ = 105) is: Z = (105.88 −
105)/3 = 0.29, P (Z > 0.29) = 0.3859, and P (Z < (94.12 − 105)/3) = P (Z <
−3.63) = 0.00014, for a total power = (1 − β) = 0.3859 + 0.00014, or about
39%. Note that our power is less than it was with a one-tailed test. If you
can correctly predict the direction of effect, a one-tailed test is
more powerful than a two-tailed test.
Consider what
√ would happen if you increased sample size to 100. Now
the σȲ = 15/ 100 = 1.5. With the null and alternative distributions are
narrower, and should overlap less, increasing power. With σȲ = 1.5 the sample
mean will need be 100 + (1.96)(1.5) = 102.94 (rather than 105.88 from before)
or more to reject the null. If the drug increases IQ by 5 points, power is:
Z = (102.94 − 105)/1.5 = −1.37, P (Z > −1.37) = 0.9147, or between 91 and
92%. Anything that decreases the standard error will increase
power. This may be achieved by increasing the sample size N or
by reducing the σ of the dependent variable. The σ of the dependent
variable may be reduced by reducing the influence of extraneous variables upon
the dependent variable (eliminating “noise” in the dependent variable makes it
easier to detect the signal).
Now consider what happens if you change the significance level, α. Let us
reduce α to 0.01. Now the sample mean must be 2.58 or more standard errors
from the null µ before we reject the null. That is, 100 + 2.58(1.5) = 103.87
(rather than 102.94 with α = 0.05). Under the alternative, Z = (103.87 −
105)/1.5 = −0.75, P (Z > −0.75) = 0.7734 or about 77%, less than it was
with α = 0.05. Reducing α reduces power.
Please note that all of the above analyses have assumed that we have used a
normally distributed test statistic, as Z = (Ȳ −µ0)/σȲ will be if the dependent
variable is normally distributed in the population or if sample size is large
enough to invoke the central limit theorem (CLT). Remember that using Z
10.2: Effect size 357
also requires that you know the population σ rather than estimating it from
the sample data. We more often estimate the population σ, using Student’s t
as the test statistic. If N is fairly large, Student’s t is nearly normal, so this is
no problem. For example, with a two-tailed α = 0.05 and N = 25, we went
out ±1.96 standard errors to mark off the rejection region. With Student’s
t on N − 1 = 24 df we should have gone out ±2.064 standard errors. But
1.96 versus 2.06 is a relatively trivial difference, so we should feel comfortable
with the normal approximation. If, however, we had N = 5, df = 4, critical
t = ±2.776, then the normal approximation would not do. A more complex
analysis would be needed.
10.2 Effect size

For the one-sample test, the effect size in σ units is d = (µ1 − µ0)/σ. For our
IQ problem with minimum nontrivial effect size at 5 IQ points, d = (105 −
100)/15 = 1/3. Cohen’s1 conventions for small, medium, and large effects for
a two-sample difference test between two means is in the table below.
One- or two-sample difference of means
Size of effect d % variance
small 0.2 1
medium 0.5 6
large 0.8 16
Cohen has conventions for other tests (correlation, contingency tables, etc.),
but they should be used with caution.
What is a small or even trivial effect in one context may be a large effect
in another context. For example, Rosnow and Rosenthal (1989) discussed a
1988 biomedical research study on the effects of taking a small, daily dose of
aspirin. Each participant was instructed to take one pill a day. For about
half of the participants the pill was aspirin, for the others it was a placebo.
The dependent variable was whether or not the participant had a heart attack
during the study. In terms of a correlation coefficient, the size of the observed
1
Cohen, J. (1988). Statistical power analysis for the behavior sciences. (2nd ed.). Hillsdale, NJ:
Erlbaum.
effect was r = 0.034. In terms of percentage of variance explained, that is
0.12%. In other contexts this might be considered a trivial effect, but it this
context it was so large an effect that the researchers decided it was unethical
to continue the study and the contacted all of the participants who were taking
the placebo and told them to start taking aspirin every day.
10.3 Sample size

Before you can answer the question ”how many subjects do I need,” you will
have to answer several other questions, such as:
How much power do I want?
What is the likely size (in the population) of the effect I am trying to
detect, or, what is smallest effect size that I would consider of importance?
What criterion of statistical significance will I employ?
What test statistic will I employ?
What is the standard deviation (in the population) of the criterion vari-
able?
For correlated samples designs, what is the correlation (in the population)
between groups?
If one considers Type I and Type II errors equally serious, then one should have
enough power to make α = β. If employing the traditional 0.05 criterion of
statistical significance, that would mean you should have 95% power. However,
getting 95% power usually involves expenses too great – that is, too many
samples.
A common convention is to try to get at least enough data to have 80%
power. So, how do you figure out how many subjects you need to have the
desired amount of power. There are several methods, including:
You could buy an expensive, professional-quality software package to do
the power analysis.
You could buy an expensive, professional-quality book on power analysis
and learn to do the calculations yourself and/or to use power tables and
figures to estimate power.
You could try to find an interactive web page on the Internet that will do
10.3: Sample size 359
the power analysis for you. This is probably fine, but be cautious.
You could download and use the G Power program, which is free, not too
difficult to use, and generally reliable (this is not to say that it is error
free).
You could use the simple guidelines provided in Jacob Cohen’s “A Power
Primer” (Psychological Bulletin, 1992, 112, 155-159).
The plots below indicate the amount of power for a given effect size and
sample size for a one-sample t-test and ANOVA test. This graph makes clear
the diminishing returns you get for adding more and more subjects if you already
have moderate to high power. For example, let’s say we’re doing a one-sample
test and we an effect size of 0.2 and have only 10 subjects. We can see that we
have a power of about 0.15, which is really, really low. Going to 25 subjects
increases our power to about 0.25, and to 100 subjects increases our power to
about 0.6. But if we had a large effect size of 0.8, 10 subjects would already
give us a power of about 0.8, and using 25 or 100 subjects would both give a
power at least 0.98. So each additional subject gives you less additional power.
This curve also illustrates the “cost” of increasing your desired power from 0.8
to 0.98.
# Power curve plot for one-sample t-test with range of sample sizes
# http://stackoverflow.com/questions/4680163/power-vs-effect-size-plot/4680786#4680786
P <- 3 # number of groups for ANOVA

fVals <- seq(0, 1.2, length.out=100) # effect sizes f for ANOVA
dVals <- seq(0, 3, length.out=100) # effect sizes d for t-Test
#nn <- seq(10, 25, by=5) # group sizes
nn <- c(5,10,25,100) # group sizes
alpha <- 0.05 # test for level alpha
# function to calculate one-way ANOVA power for given group size

getFPow <- function(n) {
critF <- qf(1-alpha, P-1, P*n - P) # critical F-value
# probabilities of exceeding this F-value given the effect sizes f

# P*n*fVals^2 is the non-centrality parameter
1-pf(critF, P-1, P*n - P, P*n * fVals^2)
}
# function to calculate one-sample t-Test power for given group size

getTPow <- function(n) {
critT <- qt(1-alpha, n-1) # critical t-value
# probabilities of exceeding this t-value given the effect sizes d

# sqrt(n)*d is the non-centrality parameter

1-pt(critT, n-1, sqrt(n)*dVals)
}
powsF <- sapply(nn, getFPow) # ANOVA power for for all group sizes
powsT <- sapply(nn, getTPow) # t-Test power for for all group sizes
#dev.new(width=10, fig.height=5)
par(mfrow=c(1, 2))
matplot(dVals, powsT, type="l", lty=1, lwd=2, xlab="effect size d",
ylab="Power", main="Power one-sample t-test", xaxs="i",
xlim=c(-0.05, 1.1), col=c("blue", "red", "darkgreen", "green"))
#legend(x="bottomright", legend=paste("N =", c(5,10,25,100)), lwd=2,
# col=c("blue", "red", "darkgreen", "green"))
legend(x="bottomright", legend=paste("N =", nn), lwd=2,
col=c("blue", "red", "darkgreen", "green"))
#matplot(fVals, powsF, type="l", lty=1, lwd=2, xlab="effect size f",
# ylab="Power", main=paste("Power one-way ANOVA, ", P, " groups", sep=""), xaxs="i",
# xlim=c(-0.05, 1.1), col=c("blue", "red", "darkgreen", "green"))
##legend(x="bottomright", legend=paste("Nj =", c(10, 15, 20, 25)), lwd=2,
## col=c("blue", "red", "darkgreen", "green"))
#legend(x="bottomright", legend=paste("Nj =", nn), lwd=2,
# col=c("blue", "red", "darkgreen", "green"))
library(pwr)
pwrt2 <- pwr.t.test(d=.2,n=seq(2,100,1),
sig.level=.05,type="one.sample", alternative="two.sided")
#plot(pwrt£n, pwrt£power, type="b", xlab="sample size", ylab="power")
matplot(matrix(c(pwrt2$n,pwrt3$n,pwrt5$n,pwrt8$n),ncol=4),
matrix(c(pwrt2$power,pwrt3$power,pwrt5$power,pwrt8$power),ncol=4),
type="l", lty=1, lwd=2, xlab="sample size",
ylab="Power", main="Power one-sample t-test", xaxs="i",
xlim=c(0, 100), ylim=c(0,1), col=c("blue", "red", "darkgreen", "green"))
legend(x="bottomright", legend=paste("d =", c(0.2, 0.3, 0.5, 0.8)), lwd=2,
col=c("blue", "red", "darkgreen", "green"))
10.3: Sample size 361
Power one−sample t−test Power one−sample t−test
1.0
1.0
0.8
0.8
0.6
0.6
Power
Power
0.4
0.4
0.2
0.2
N=5 d = 0.2
N = 10 d = 0.3
N = 25 d = 0.5
0.0
N = 100 d = 0.8
0.0 0.2 0.4 0.6 0.8 1.0 0 20 40 60 80 100
effect size d sample size
Reasons to do a power analysis There are several of reasons why one

might do a power analysis. (1) Perhaps the most common use is to determine
the necessary number of subjects needed to detect an effect of a given size. Note
that trying to find the absolute, bare minimum number of subjects needed in the
study is often not a good idea. (2) Additionally, power analysis can be used to
determine power, given an effect size and the number of subjects available. You
might do this when you know, for example, that only 75 subjects are available
(or that you only have the budget for 75 subjects), and you want to know if you
will have enough power to justify actually doing the study. In most cases, there
is really no point to conducting a study that is seriously underpowered. Besides
the issue of the number of necessary subjects, there are other good reasons for
doing a power analysis. (3) For example, a power analysis is often required as
part of a grant proposal. (4) And finally, doing a power analysis is often just
part of doing good research. A power analysis is a good way of making sure
that you have thought through every aspect of the study and the statistical
analysis before you start collecting data.
Limitations Despite these advantages of power analyses, there are some
limitations. (1) One limitation is that power analyses do not typically generalize
very well. If you change the methodology used to collect the data or change
the statistical procedure used to analyze the data, you will most likely have to
redo the power analysis. (2) In some cases, a power analysis might suggest a
number of subjects that is inadequate for the statistical procedure. For example
(beyond the scope of this class), a power analysis might suggest that you need
30 subjects for your logistic regression, but logistic regression, like all maximum
likelihood procedures, require much larger sample sizes. (3) Perhaps the most
important limitation is that a standard power analysis gives you a “best case
scenario” estimate of the necessary number of subjects needed to detect the
effect. In most cases, this “best case scenario” is based on assumptions and
educated guesses. If any of these assumptions or guesses are incorrect, you may
have less power than you need to detect the effect. (4) Finally, because power
analyses are based on assumptions and educated guesses, you often get a range
of the number of subjects needed, not a precise number. For example, if you do
not know what the standard deviation of your outcome measure will be, you
guess at this value, run the power analysis and get X number of subjects. Then
you guess a slightly larger value, rerun the power analysis and get a slightly
larger number of necessary subjects. You repeat this process over the plausible
range of values of the standard deviation, which gives you a range of the number
of subjects that you will need.
Other considerations After all of this discussion of power analyses and

the necessary number of subjects, we need to stress that power is not the only
consideration when determining the necessary sample size. For example, differ-
ent researchers might have different reasons for conducting a regression analysis.
(1) One might want to see if the regression coefficient is different from zero, (2)
while the other wants to get a very precise estimate of the regression coefficient
with a very small confidence interval around it. This second purpose requires a
larger sample size than does merely seeing if the regression coefficient is different
from zero. (3) Another consideration when determining the necessary sample
size is the assumptions of the statistical procedure that is going to be used (e.g.,
10.4: Power calculation via simulation 363
parametric vs nonparametric procedure). (4) The number of statistical tests
that you intend to conduct will also influence your necessary sample size: the
more tests that you want to run, the more subjects that you will need (mul-
tiple comparisons). (5) You will also want to consider the representativeness
of the sample, which, of course, influences the generalizability of the results.
Unless you have a really sophisticated sampling plan, the greater the desired
generalizability, the larger the necessary sample size.
10.4 Power calculation via simulation

Using the principles of the bootstrap (to be covered later) we can estimate
statistical power through simulation.
Example: IQ drug, revisited Recall that we sample N = 25 people

from a population known to be normally distributed with a µ of 100 and a σ
of 15. Consider the first one-sided alternative H0 : µ = 100 and H1 : µ > 100.
Assume the minimum nontrivial effect size was that the drug increases µIQ
by 2 or more points, so that the specific alternative to consider is H1 : µ = 102.
What is the power of this test?
We already saw how to calculate this analytically. To solve this computa-
tionally, we need to simulate samples of N = 25 from the alternative distribu-
tion (µ = 102 and σ = 15) and see what proportion of the time we correctly
reject H0.
#### Example: IQ drug, revisited
# R code to simulate one-sample one-sided power
# Strategy:
# Do this R times:
# draw a sample of size N from the distribution specified by the alternative hypothesis
# That is, 25 subjects from a normal distribution with mean 102 and sigma 15
# Calculate the mean of our sample
# Calculate the associated z-statistic
# See whether that z-statistic has a p-value < 0.05 under H0: mu=100
# If we reject H0, then set reject = 1, else reject = 0.
# Finally, the proportion of rejects we observe is the approximate power
n <- 25; # sample size of 25

mu0 <- 100; # null hypothesis mean of 100
mu1 <- 102; # alternative mean of 102

#mu1 <- 105; # alternative mean of 105
sigma <- 15; # standard deviation of normal population
alpha <- 0.05; # significance level
R <- 10000; # Repetitions to draw sample and see whether we reject H0

# The proportion of these that reject H0 is the power
reject <- rep(NA, R); # allocate a vector of length R with missing values (NA)
# to fill with 0 (fail to reject H0) or 1 (reject H0)
for (i in 1:R) {
sam <- rnorm(n, mean=mu1, sd=sigma); # sam is a vector with 25 values
ybar <- mean(sam); # Calculate the mean of our sample sam
z <- (ybar - mu0) / (sigma / sqrt(n)); # z-statistic (assumes we know sigma)

# we could also have calculated the t-statistic, here
pval <- 1-pnorm(z); # one-sided right-tail p-value

# pnorm gives the area to the left of z
# therefore, the right-tail area is 1-pnorm(z)
if (pval < 0.05) {

reject[i] <- 1; # 1 for correctly rejecting H0
} else {
reject[i] <- 0; # 0 for incorrectly fail to reject H0
}
power <- mean(reject); # the average reject (proportion of rejects) is the power
power
## [1] 0.166
# 0.1655 for mu1=102
# 0.5082 for mu1=105
Our simulation (this time) with µ1 = 102 gave a power of 0.166 (exact
answer is P (Z > 0.98) = 0.1635). Rerunning with µ1 = 105 gave a power
of 0.5082 (exact answer is P (Z > −0.02) = 0.5080). Our simulation well-
approximates the true value, and the power can be made more precise by in-
creasing the number of repetitions calculated. However, two to three decimal
precision is quite sufficient.
Example: Head breadth Recall the head breadth example in Chapter 3

comparing maximum head breadths (in millimeters) of modern day Englishmen
with ancient Celts. The data are summarized below.
Descriptive Statistics: ENGLISH, CELTS
Variable N Mean SE Mean StDev Minimum Q1 Median Q3 Maximum
ENGLISH 18 146.50 1.50 6.38 132.00 141.75 147.50 150.00 158.00
CELTS 16 130.75 1.36 5.43 120.00 126.25 131.50 135.50 138.00
Imagine that we don’t have the information above. Imagine we have been
invited to a UK university to take skull measurements for 18 modern day En-
glishmen, and 16 ancient Celts. We have some information about modern day
skulls to use as prior information for measurement mean and standard devia-
tion. What is the power to observe a difference between the populations? Let’s
make some reasonable assumptions that allows us to be a bit conservative. Let’s
assume the sampled skulls from each of our populations is a random sample
with common standard deviation 7mm, and let’s assume we can’t get the full
sample but can only measure 15 skulls from each population. At a significance
level of α = 0.05, what is the power for detecting a difference of 5, 10, 15, 20,
or 25 mm?
The theoretical two-sample power result is not too hard to derive (and is
available in text books), but let’s simply compare the power calculated exactly
and by simulation.
For the exact result we use R library pwr. Below is the function call as well
as the result. Note that we specified multiple effect sizes (diff/SD) in one call
of the function.
# R code to compute exact two-sample two-sided power
library(pwr) # load the power calculation library
pwr.t.test(n = 15,
d = c(5,10,15,20,25)/7,
sig.level = 0.05,
power = NULL,
type = "two.sample",
alternative = "two.sided")
##
## Two-sample t test power calculation
##
## n = 15
## d = 0.7143, 1.4286, 2.1429, 2.8571, 3.5714
## sig.level = 0.05
## power = 0.4717, 0.9652, 0.9999, 1.0000, 1.0000
## alternative = two.sided
##
## NOTE: n is number in *each* group
To simulate the power under the same circumstances, we follow a similar
strategy as in the one-sample example.
# R code to simulate two-sample two-sided power
# Strategy:
# Do this R times:
# draw a sample of size N from the two hypothesized distributions
# That is, 15 subjects from a normal distribution with specified means and sigma=7
# Calculate the mean of the two samples
# Calculate the associated z-statistic
# See whether that z-statistic has a p-value < 0.05 under H0: mu_diff=0
# If we reject H0, then set reject = 1, else reject = 0.
# Finally, the proportion of rejects we observe is the approximate power
n <- 15; # sample size of 25

mu1 <- 147; # null hypothesis English mean
mu2 <- c(142, 137, 132, 127, 122); # Celt means
sigma <- 7; # standard deviation of normal population
alpha <- 0.05; # significance level
R <- 2e4; # Repetitions to draw sample and see whether we reject H0

# The proportion of these that reject H0 is the power
power <- rep(NA,length(mu2)); # allocate a vector to store the calculated power in
for (j in 1:length(mu2)) { # do for each value of mu2
reject <- rep(NA, R); # allocate a vector of length R with missing values (NA)
# to fill with 0 (fail to reject H0) or 1 (reject H0)
for (i in 1:R) {
sam1 <- rnorm(n, mean=mu1 , sd=sigma); # English sample
sam2 <- rnorm(n, mean=mu2[j], sd=sigma); # Celt sample
ybar1 <- mean(sam1); # Calculate the mean of our sample sam

ybar2 <- mean(sam2); # Calculate the mean of our sample sam
# z-statistic (assumes we know sigma)

# we could also have calculated the t-statistic, here
z <- (ybar2 - ybar1) / (sigma * sqrt(1/n+1/n));
pval.Left <- pnorm(z); # area under left tail

pval.Right <- 1-pnorm(z); # area under right tail
# p-value is twice the smaller tail area
pval <- 2 * min(pval.Left, pval.Right);
if (pval < 0.05) {

reject[i] <- 1; # 1 for correctly rejecting H0
} else {
reject[i] <- 0; # 0 for incorrectly fail to reject H0
# the average reject (proportion of rejects) is the power

power[j] <- mean(reject);
}
power
## [1] 0.4928 0.9765 1.0000 1.0000 1.0000
Note the similarity between power calculated using both the exact and sim-
ulation methods. If there is a power calculator for your specific problem, it is
best to use that because it is faster and there is no programming. However,
using the simulation method is better if we wanted to entertain different sam-
ple sizes with different standard deviations, etc. There may not be a standard
calculator for our specific problem, so knowing how to simulate the power can
be valuable.
Mean Sample size Power
µE µC diff SD nE nC exact simulated
147 142 5 7 15 15 0.4717 0.4928
147 137 10 7 15 15 0.9652 0.9765
147 132 15 7 15 15 0.9999 1
147 127 20 7 15 15 1.0000 1
147 122 25 7 15 15 1.0000 1
Appendix A
Custom R functions
A.1 Ch 2. Estimation in One-Sample Problems
# a function to compare the bootstrap sampling distribution with a
# normal distribution with mean and SEM estimated from the data
bs.one.samp.dist <- function(dat, N = 10000) {
n <- length(dat)
sam <- matrix(sample(dat, size = N * n, replace = TRUE), ncol = N)
par(mfrow = c(2, 1), mar = c(3, 2, 2, 1), oma = c(1, 1, 1, 1))
hist(dat, freq = FALSE, breaks = 6, main = "Plot of data with smoothed density curve")
rug(dat)
hist(sam.mean, freq = FALSE, breaks = 25, main = "Bootstrap sampling distribution of the m
xlab = paste("Data: n =", n, ", mean =", signif(mean(dat), digits = 5),
", se =", signif(sd(dat)/sqrt(n)), digits = 5))
x <- seq(min(sam.mean), max(sam.mean), length = 1000)
points(x, dnorm(x, mean = mean(dat), sd = sd(dat)/sqrt(n)), type = "l",
lwd = 2, col = "red")
rug(sam.mean)
par(old.par)
}
A.2: Ch 3. Two-Sample Inferences 369
# Function ot plot t-distribution with shaded p-value

t.dist.pval <- function(t.summary) {
par(mfrow = c(1, 1))
lim.extreme <- max(4, abs(t.summary$statistic) + 0.5)
lim.lower <- -lim.extreme
lim.upper <- lim.extreme
x.curve <- seq(lim.lower, lim.upper, length = 200)
y.curve <- dt(x.curve, df = t.summary$parameter)
plot(x.curve, y.curve, type = "n", ylab = paste("t-dist( df =", signif(t.summary$parameter
3), ")"), xlab = paste("t-stat =", signif(t.summary$statistic,
5), ", Shaded area is p-value =", signif(t.summary$p.value, 5)))
if ((t.summary$alternative == "less") | (t.summary$alternative == "two.sided")) {
x.pval.l <- seq(lim.lower, -abs(t.summary$statistic), length = 200)
y.pval.l <- dt(x.pval.l, df = t.summary$parameter)
polygon(c(lim.lower, x.pval.l, -abs(t.summary$statistic)), c(0,
y.pval.l, 0), col = "gray")
}
if ((t.summary$alternative == "greater") | (t.summary$alternative ==
"two.sided")) {
x.pval.u <- seq(abs(t.summary$statistic), lim.upper, length = 200)
y.pval.u <- dt(x.pval.u, df = t.summary$parameter)
polygon(c(abs(t.summary$statistic), x.pval.u, lim.upper), c(0,
y.pval.u, 0), col = "gray")
}
points(x.curve, y.curve, type = "l", lwd = 2, col = "blue")
}
A.2 Ch 3. Two-Sample Inferences

# a function to compare the bootstrap sampling distribution of the
# difference of means from two samples with a normal distribution with
# mean and SEM estimated from the data
bs.two.samp.diff.dist <- function(dat1, dat2, N = 10000) {
n1 <- length(dat1)
n2 <- length(dat2)
sam1 <- matrix(sample(dat1, size = N * n1, replace = TRUE), ncol = N)
sam2 <- matrix(sample(dat2, size = N * n2, replace = TRUE), ncol = N)
# calculate the means and take difference between populations
sam1.mean <- colMeans(sam1)
sam2.mean <- colMeans(sam2)
diff.mean <- sam1.mean - sam2.mean
par(mfrow = c(3, 1), mar = c(3, 2, 2, 1), oma = c(1, 1, 1, 1))
hist(dat1, freq = FALSE, breaks = 6, main = paste("Sample 1", "\n",
"n =", n1, ", mean =", signif(mean(dat1), digits = 5), ", sd =",
370 Ch A: Custom R functions
signif(sd(dat1), digits = 5)), xlim = range(c(dat1, dat2)))

rug(dat1)
hist(dat2, freq = FALSE, breaks = 6, main = paste("Sample 2", "\n",

"n =", n2, ", mean =", signif(mean(dat2), digits = 5), ", sd =",
signif(sd(dat2), digits = 5)), xlim = range(c(dat1, dat2)))
rug(dat2)
hist(diff.mean, freq = FALSE, breaks = 25, main = paste("Bootstrap sampling distribution o

"\n", "mean =", signif(mean(diff.mean), digits = 5), ", se =",
signif(sd(diff.mean), digits = 5)))
points(density(diff.mean), type = "l")
x <- seq(min(diff.mean), max(diff.mean), length = 1000)
points(x, dnorm(x, mean = mean(diff.mean), sd = sd(diff.mean)), type = "l",
lwd = 2, col = "red")
rug(diff.mean)
par(old.par)
}

ADA1 Notes F14

Uploaded by

Copyright:

Available Formats

ADA1 Notes F14

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

ADA1 Notes F14

Uploaded by

Copyright:

Available Formats

Lecture notes for

Advanced Data Analysis 1 (ADA1)

Erik B. Erhardt Edward J. Bedrick Ronald M. Schrader

1 Summarizing and Displaying Data 27

2 Estimation in One-Sample Problems 59

4 Checking Assumptions 130

5 One-Way Analysis of Variance 150

6 Nonparametric Methods 181

7 Categorical Data Analysis 232

8 Correlation and Regression 281

9 Introduction to the Bootstrap 335

10 Power and Sample size 351

A Custom R functions 368

Syllabus Website: http://statacumen.com/teaching/ada1

Let’s see what’s there:

Learning outcomes, syllabus

ClickerQ s — Ch 0, Learning outcomes, General

Statistics can be challenging because we operate at the higher levels of

Setup Install R and Rstudio on your computer, as outlined in Step 0 on the

1. Upper-left: Program editor window, where you’ll write code.

 Ctrl-Enter (submit current line or selection in editor window to console)

Workflow is how you get things done. Here’s my recommendation:

R as calculator In the following sections, look at the commands and output

Vectors A vector is a set of numbers like a column of a spreadsheet. Below

Assign variables Assigning objects to varibles.

Extracting subsets It will be an extremely important skill to understand

#### Extracting subsets

Boolean: True/False Subsets can be selected based on which elements

Boolean operators compare TRUE/FALSE values and return TRUE/-

ClickerQ s — Ch 0, R building blocks, Subset

ClickerQ s — Ch 0, R building blocks, T/F selection 2

How’d you do?

R command review This is a summary of the commands we’ve seen so

0.4 Plotting with ggplot2

# ggplot2 includes a dataset "mpg"

# ? gives help on a function or dataset

Geom: is the “type” of plot

Help is available. Try searching for examples, too.

When certain aesthetics are defined, an appropriate legend is chosen and

I encourage you to experiment with aesthetics!

1. Assign variables to aesthetics colour, size, and shape.

The behavior of the aesthetics is predictable and customizable.

Aesthetic Discrete Continuous

Let’s see a couple examples.

 Typically, small multiples will display different subsets of the data.

Experiment with faceting of different types. What relationships would you

# examples of subsetting the scatterplot in facets

# plot all four in one arrangement

4 5 6 8 2seater compact midsize

0.4.1 Improving plots

How can this plot be improved?

How can this plot be improved?

2seater compact midsize minivan pickup subcompact suv

Problem: The classes are in alphabetical order, which is somewhat arbitrary.

pickup suv minivan 2seater midsize subcompact compact

pickup suv minivan 2seater midsize subcompact compact

. . . or replace with boxplots

pickup suv minivan 2seater midsize subcompact compact

. . . or jitter those points with reduced-opacity boxplots on top

. . . even better with the boxplots with jittered points on top

Ctrl-Enter (submit current line or selection in editor window to console)

Typically, small multiples will display different subsets of the data.