BMTRY 701

Biostatistical Methods II

Elizabeth Garrett-Mayer, PhD

Associate Professor
Director of Biostatistics, Hollings Cancer Center
[email protected]
Biostatistical Methods II
Description: 
 This is a one-semester course intended for graduate students pursuing
degrees in biostatistics and related fields such as epidemiology and
bioinformatics. 
 Topics covered will include linear, logistic, poisson, and Cox
regression. 
 Advanced topics will be included, such as ridge regression or
hierarchical linear regression if time permits.
 Estimation, interpretation, and diagnostic approaches will be
discussed. 
 Software instruction will be provided in class in R. 
 Students will be evaluated via homeworks (55%), two exams
(35%) and class participation (10%). 
 This is a four credit course.
Biostatistical Methods II
Textbooks: 
(1) Introduction to Linear Regression Analysis (4th Edition). 
Montgomery, Peck and Vining.    Wiley; New York, 2006.
(2) Regression with Modeling Strategies: With Applications to Linear
Models, Logistic Regression, and Survival Analysis.  Frank E.
Harrell, Jr.  Springer; New York, 2001.

Prerequisites:  Biometry 700

Course Objectives:  Upon successful completion of the course, the

student will be able to
• Apply, interpret and diagnose linear regression models
• Apply, interpret and diagnose logistic, poisson and Cox
regresssion models
 Biostatistical Methods II
Instructor: Elizabeth Garrett-Mayer
Website: http://people.musc.edu/~elg26/teaching/methods2.2010/methods2.2010.htm

Contact Info:  Hollings Cancer Center, Rm 118G

  [email protected] (preferred mode of contact is email)

  792-7764
Time: Mondays and Wednesdays, 1:30-3:30
Location: Cannon 301, Room 305V

Office Hours: Tuesdays 2 :00 – 3:30pm

Biostatistical Methods II
 Lecture schedule is on the website
 Second time teaching this class
 New textbooks this year
• syllabus is a ‘work in progress’
• timing of topics subject to change
• lectures may appear on website last-minute
 Computing
• R
• integrated into lecture time
 Homeworks, articles, datasets will also be posted to website
• some/most problems will be from textbook
• some datasets will be from R library
 If you want printed versions of lectures:
• download and print prior to lecture; OR
• work interactively on your laptop during class
 We will take a break about halfway through each lecture
Expectations
 Academic
• Participate in class discussions
• Invest resources in YOUR education
• Complete homework assignments on time
• The results of the homework should be communicated so that a person
knowledgeable in the methodology could reproduce your results.
• Create your own study groups
 Challenge one another
 everyone needs to contribute
 you may do homeworks together, but everyone must turn in his/her own
homework.
 written sections of homework should be ‘independently’ developed
 General
• Be on time to class
• Be discrete with interruptions (pages, phones, etc.)
• Do NOT turn in raw computer output
Other Expectations

 Knowledge of Methods I!
 You should be very familiar with
• confidence intervals
• hypothesis testing
 t-tests
 Z-tests
• graphical displays of data
• exploratory data analysis
 estimating means, medians, quantiles of data
 estimating variances, standard deviations
About the instructor
 B.A. from Bowdoin College, 1994
• Double Major in Mathematics and Economics
• Minor in Classics
 Ph.D. in Biostatistics from Johns Hopkins, 2000
• Dissertation research in latent class models, Adviser Scott Zeger
 Assistant Professor in Oncology and Biostatistics at JHU, 2000-
2007
 Taught course in Statistics for Psychosocial Research for 8 years
 Applied Research Areas:
• oncology
 Biostats Research Areas:
• latent variable modeling
• class discovery in microarray data
• methodology for early phase oncology clinical trials
 Came to MUSC in Feb 2007
Computing

 Who knows what?

 Who WANTS to know what?

 Who will bring a laptop to class?

 What software do you have and/or prefer?

Regression
 Purposes of Regresssion
1. Describe association between Y and X’s
2. Make predictions:
 Interpolation: making prediction within a range of X’s
 Extrapolation: making prediction outside a range of X’s
3. To “adjust” or “control for” confounding variables
 What is “Y”?
• an outcome variable
• ‘dependent’ variable
• ‘response’
 Type of regression depends on type of Y
• continuous (linear regression)
• binary (logistic regression)
• time-to-event (Cox regression)
• rare event or rate (poisson regression)
Some motivating examples

 Example 1: Suppose we are interested in

studying the relationship between fasting blood
glucose (FBG) levels and the number hours per
day of aerobic exercise. Let Y denote the fasting
blood glucose level
• Let X denote the number of hours of exercise
• One may be interested in studying the relationship of
Y and X
 Simple linear regression can be used to quantify
this relationship
Some motivating examples
 Example 2: Consider expanding example 1 to include
other factors that could be related FBG.
 Let X1 denote hours of exercise
 Let X2 denote BMI
 Let X3 indicate if the person has diabetes
 . . . (other covariates possible)
 One may be interested in studying the relationship of all
X′s on Y and identifying the “best” combination of factors
 Note: Some of the X′s may correlated (e.g., exercise and
bmi)
 Multiple (or multivariable, not multivariate) linear
regression can be used to quantify this relationship
Some motivating examples

 Example 3: Myocardial infarction (MI, heart attack) is

often a life-altering event
 Let Y denote the occurrence (Y = 1) of an MI after
treatment, let Y = 0 denote no MI
 Let X1 denote the dosage of aspirin taken
 Let X2 denote the age of the person
 . . . (other covariates possible)
 One may be interested in studying the relationship of all
X′s on Y and identifying the “best” combination of factors
 Multiple LOGISTIC regression can be used to quantify
this relationship
More motivating examples
 Example 4: This is an extension of Ex 3: Myocardial
infarction. Let the interest be now on when the first
 MI occurs instead of “if” one occurs.
 Let Y denote the occurrence (Y = 1) of an MI after
treatment, let Y = 0 denote no MI observed
 Let Time denote the length of time the individual is
observed
 Let X1 denote the dosage of aspirin taken
 . . . (other covariates possible)
 Survival Analysis (which, in some cases, is a regression
model) can be used to quantify this relationship of aspirin
on MI
More motivating examples
 Example 5: Number of cancer cases in a city
 Let Y denote the count (non-negative integer value) of
cases of a cancer in a particular region of interest
 Let X1 denote the region size in terms of “at risk”
individuals
 Let X2 denote the region
 . . . (other covariates possible)
 One may be interested in studying the relationship of the
region on Y while adjusting for the population at risk
sizes
 POISSON regression can be used to quantify this
relationship
Brief Outline

 Linear regression: half semester (through spring

break)
 Logistic regression
 Cox regression (survival)
 Poisson regression
 Hierarchical regression or ridge regression?
Linear Regression

 Outcome is a CONTINUOUS variable

 Assumes association between Y and X is a ‘straight line’
 Assumes relationship is ‘statistical’ and not ‘functional’
• relationship is not perfect
• there is ‘error’ or ‘noise’ or ‘unexplained variation’
 Aside:
• I LOVE graphical displays of data
• This is why regression is especially fun
• there are lots of neat ways to show your data
• prepare yourself for a LOT of scatterplots this semester
Graphical Displays
 Scatterplots: show associations between two variables
(usually)
 Also need to understand each variable by itself
 Univariate data displays are important
 Before performing a regresssion, we should
• identify any potential skewness
• outliers
• discreteness
• multimodality
 Top choices for univariate displays
• boxplot
• histogram
• density plot
• dot plot
Linear regression example
 The authors conducted a pilot study to assess the use of
toenail arsenic concentrations as an indicator of
ingestion of arsenic-containing water. Twenty-one
participants were interviewed regarding use of their
private (unregulated) wells for drinking and cooking, and
each provided a sample of water and toenail clippings.
Trace concentrations of arsenic were detected in 15 of
the 21 well-water samples and in all toenail clipping
samples.
 Karagas MR, Morris JS, Weiss JE, Spate V, Baskett C,
Greenberg ER. Toenail Samples as an Indicator of
Drinking Water Arsenic Exposure. Cancer Epidemiology,
Biomarkers and Prevention 1996;5:849-852.
Purposes of Regression
1. Describe association
• hypothesis: as arsenic in well water increases, level of arsenic in
nails also increases.
• linear regression can tell us
 how much increase in nail level we see on average for a 1 unit
increase in well water level of arsenic
2. Predict
• linear regresssion can tell us
 what level of arsenic we would expect in nails for a given level in
well water.
 how precise our estimate of arsenic is for a given level of well water
3. Adjust
• linear regression can tell us
 what the association between well water arsenic and nail arsenic is
adjusting for other factors such as age, gender, amount of use of
water for cooking, amount of use of water for drinking.
Boxplot
 Arsenic in ppm

0.0 0.5 1.0 1.5 2.0

Nails
Graphical Displays

Arsenic in ppm

Water 0.00 0.04 0.08 0.12

Histogram

 Bins the data

 x-axis represents variable values
 y-axis is either
• frequency of occurrence
• percentage of occurence
 Visual impression can depend on bin width
 often difficult to see details of highly skewed
data
Histogram
10 15
Frequency

5
0

0.0 0.5 1.0 1.5 2.0 2.5

Arsenic in Nails, ppm

10 15
Frequency

5
0

0.00 0.02 0.04 0.06 0.08 0.10 0.12 0.14

Arsenic in Water, ppm

Histogram
Frequency

8
4
0

0.0 0.5 1.0 1.5 2.0

Arsenic in Nails, ppm

Frequency

15
5
0

0.00 0.05 0.10 0.15

Arsenic in Water, ppm

Density Plot

 Smoothed density based on kernel density

estimates
 Can create similar issues as histogram
• smoothing parameter selection
• can affect inferences
 Can be problematic for ‘ceiling’ or ‘floor’ effects
Density Plot

0.0 1.0 2.0

Density

0.0 0.5 1.0 1.5 2.0 2.5

Arsenic in Nails, ppm

40
Density

20
0

0.00 0.05 0.10 0.15

Arsenic in Water, ppm

Dot plot
 My favorite for
• small datasets
• when displaying data by groups
2.0
1.5
Arsenic, ppm

1.0
0.5
0.0

Male Female
And…the scatterplot
2.0
Level of Arsenic in Nails (ppm)

1.5
1.0
0.5
0.0

0.00 0.02 0.04 0.06 0.08 0.10 0.12 0.14

Level of Arsenic in Well Water (ppm)

Measuring the association between X and Y

 Y is on the vertical
 X predicts Y
 Terminology:
• “Regress Y on X”
• Y: dependent variable, response, outcome
• X: independent variable, covariate, regressor,
predictor, confounder
 Linear regression  a straight line
• important!
• this is key to linear regression
Simple vs. Multiple linear regresssion

 Why ‘simple’?
• only one “x”
• we’ll talk about multiple linear regression later…
 Multiple regression
• more than one “X”
• more to think about: selection of covariates
 Not linear?
• need to think about transformations
• sometimes linear will do reasonably well
Association versus Causation

 Be careful!
 Association ≠ Causation
 Statistical relationship does not mean X causes
Y
 Could be
• X causes Y
• Y causes X
• something else causes both X and Y
• X and Y are spuriously associated in your sample of data
 Example: vision and number of gray hairs
Basic Regression Model

Yi   0  1 X i   i

 Yi is the value of the response variable in the ith

individual
 β0 and β1 are parameters
 Xi is a known constant; the value of the covariate in the
ith individual
 εi is the random error term
 Linear in the parameters
 Linear in the predictor
Basic Regression Model

 NOT linear in the parameters:

Yi   0  X i1   i
Yi   0  1 2 X i   i
 NOT linear in the predictor:

log(Yi )   0  1 X i   i
Yi 2   0  1 X i   i
Model Features
 Yi is the sum of a constant piece and a random piece:
• β0 + β1Xi is constant piece (recall: x is treated as constant)
• εi is the random piece
 Attributes of error term
• mean of residuals is 0: E(εi) = 0
• constant variance of residuals : σ2(εi ) = σ2 for all i
• residuals are uncorrelated: cov(εi, εj) = 0 for all i, j; i ≠ j
 Consequences
• Expected value of response
 E(Yi) = β0 + β1Xi
 E(Y) = β0 + β1X
• Variance of Yi |Xi= σ2
• Yi and Yj are uncorrelated
Probability Distribution of Y

 For each level of X, there is a probability

distribution of Y
 The means of the probability distributions vary
systematically with X.
Parameters

 β0 and β1 are referred to as “regression

coefficients”
 Remember y = mx+b?
 β1 is the slope of the regression line
• the expected increase in Y for a 1 unit increase in X
• the expected difference in Y comparing two
individuals with X’s that differ by 1 unit
• Expected? Why?
Parameters

 β0 is the intercept of the regression line

• The expected value of Y when X = 0
• Meaningful?
 when the range of X includes 0, yes
 when the range of X excluded 0, no
• Example:
 Y = baby’s weight in kg
 X = baby’s height in cm
 β0 is the expected weight of a baby whose height is 0 cm.
SENIC Data
 Will be used as a recurring example
 SENIC = Study on the Efficacy of Nosocomial Infection
Control
 The primary objective of the SENIC Project was to
determine whether infection surveillance and control
programs have reduced the rates of nosocomial
(hospital-acquired) infection in the United States
hospitals.
 This data set consists of a random sample of 113
hospitals selected from the original 338 hospitals
surveyed.
 Each line of the data set has an ID number and provides
information on 11 other variables for a single hospital.
 The data used here are for the 1975-76 study period.
SENIC Data
SENIC Simple Linear Regression Example

 Hypothesis: The number of beds in a given

hospital is associated with the average length of
stay. . scatter los beds
 Y=?
 X= ? 20

 Scatterplot
15
length of stay
10
5

0 200 400 600 800

number of beds
 20
Stata
Regression

15
Results

10
5
0 200 400 600 800
number of beds

length of stay Fitted values

. regress los beds

Source | SS df MS Number of obs = 113

-------------+------------------------------ F( 1, 111) = 22.33
Model | 68.5419355 1 68.5419355 Prob > F = 0.0000

̂
Residual | 340.668443 ̂
111 3.06908508
0
-------------+------------------------------
R-squared
Adj R-squared
=
=
0.1675
0.1600
1
Total | 409.210379 112 3.6536641 Root MSE = 1.7519

------------------------------------------------------------------------------
los | Coef. Std. Err. t P>|t| ˆ
[95% Conf. Interval]
-------------+----------------------------------------------------------------
beds | .0040566 .0008584 4.73 0.000 .0023556 .0057576
_cons | 8.625364 .2720589 31.70 0.000 8.086262 9.164467
------------------------------------------------------------------------------
Another Example: Famous data

 Father and sons heights data from Karl Pearson

(over 100 years ago in England)
 1078 pairs of fathers and sons
 Excerpted 200 pairs for demonstration
 Hypotheses:
• there will be a positive association between heights of
fathers and their sons
• very tall fathers will tend to have sons that are shorter
than they are
• very short fathers will tend to have sons that are taller
than they are
Scatterplot of 200 records of father son data
plot(father, son, xlab="Father's Height, Inches", ylab="Son's Height, Inches",
xaxt="n",yaxt="n",ylim=c(58,78), xlim=c(58,78))
axis(1, at=seq(58,78,2))
axis(2, at=seq(58,78,2))

78
74
Son's Height, Inches

70
66
62
58

58 60 62 64 66 68 70 72 74 76 78

Father's Height, Inches

Regression Results
> reg <- lm(son~father)
> summary(reg)

Call:
lm(formula = son ~ father)

Residuals:
Min 1Q ̂
Median 0 3Q Max
-7.72874 -1.39750 -0.04029 1.51871 7.66058

Coefficients:
Estimate Std. Error t value Pr(>|t|)
(Intercept) 39.47177 3.96188 9.963 < 2e-16 ***
father 0.43099 0.05848 7.369 4.55e-12 ***
---

̂1 ˆ
Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

Residual standard error: 2.233 on 198 degrees of freedom

Multiple R-squared: 0.2152, Adjusted R-squared: 0.2113
F-statistic: 54.31 on 1 and 198 DF, p-value: 4.549e-12
 This is where the term “regression” came from
78
74

x  67.7
Son's Height, Inches

y  68.6
70
66
62

x=y
regression line
58

58 60 62 64 66 68 70 72 74 76 78

Father's Height, Inches

Aside: Design of Studies
 Does it matter if the study is randomized?
observational?
 Yes and no
 Regression modeling can be used regardless
 The model building will often depend on the nature of the
study
 Observational studies:
• adjustments for confounding
• often have many covariates as a result
 Randomized studies:
• adjustments may not be needed due to randomization
• subgroup analyses are popular and can be done via regression
Estimation of the Model

 The Method of Least Squares

 Intuition: we would like to minimize the residuals:

 i  Yi   0  1 X i

 Minimize/maximize: how to do that?

 Can we minimize the sum of the residuals?
Least Squares

 Minimize the distance between the fitted line and

the observed data
 Take absolute values?
 Simpler? Square the errors.
 LS estimation:
• Minimize Q:
N
Q   (Yi   0  1 X i ) 2
i 1
Least Squares

 Derivation
 Two initial steps: reduce the following
N

 i
( X
i 1
 X ) 2

(X
i 1
i  X )(Yi  Y )
Least Squares
N
Q   (Yi   0  1 X i ) 2
i 1
Least Squares