-
1
Biostatistics
SGU
July 2014
2
Data is Everywhere
Research LiteratureHypothesis: Surgeon-directed institutional
peer review, associated withpositive physician feedback, can
decrease the morbidity and mortalityrates associated with carotid
endarterectomy. Results: Stroke ratedecreased from 3.8% (1993-1994)
to 0%(1997-1998). The mortality ratedecreased from 2.8% (1993-1994)
to 0% (1997-1998). (average) Lengthof stay decreased from 4.7 days
(1993-1994) to 2.6 days (1997-1998).The (average) total cost
decreased from $13344 (1993-1994) to $9548(1997-1998).
Archives of Surgery, August 2000
Popular Press
For the first time, an influential doctors group is recommending
thatsome children as young as 8 be given cholesterol-fighting drugs
to wardoff future heart problems... With one-third of U.S. children
overweightand about 17 percent obese, the new recommendations are
important,said Dr. Jennifer Li, a Duke University childrens heart
specialist.
cnn.com, July 8, 2008
3
Data provides Information
Good Data Can Be Analyzed and Summarized toProvide Useful
Information
Bad Data Can Be Analyzed and Summarized toProvide
Incorrect/Harmful/Non-informative
Information
4
Steps in Research Project
Planning Design Data Collection Data Analysis Presentation
Interpretation
-
5
Biostatistics
Design of Studies
Sample size Selection of study participants Role of
randomization
Data Collection Variability
Important patterns in data are obscured by
variability.Distinguish real patterns from random variation.
Inference
Draw general conclusions from limited data e.g. survey
Summarize
What summary measures will best convey the results How to convey
uncertainty in results
Interpretation
What do the results mean in terms of practice, the program,the
population etc.
6
1954 Salk Polio Vaccine Trial
School Children
Vaccinatedn = 200, 745
Placebon = 201, 229
Polio Cases
Vaccine 82Placebo 162
Reference: Meier P, The Biggest Public Health Experiment Ever:
The 1954 Field Trial of theSalk Poliomyelitis Vaccine, In:
Statistics: A Guide to the Unknown, 1972.
7
Design: Features of the Polio Trial
Comparison Group Randomized Placebo Controls Double Blind
Objective: The groups should be equivalentexcept for the factor
(vaccine) being investigated.
Question: Could the results be due to chance?
8
There were almost twice as many polio casesin the placebo
compared to vaccine group.
COULD WE GET SUCH GREAT IMBALANCE BY CHANCE?
Polio Cases
Vaccine 82 out of 200,745Placebo 162 out of 201,229
p-value=?
Statistical methods tell us how to make these
probabilitycalculations.
-
9
Types of Data
1 Binary (dichotomous) data
Polio: Yes/No Cure: Yes/No Gender: Male/Female
2 Categorical data
Race/ethnicity nominalno ordering Country of birth nominalno
ordering Degree of agreement ordinalordering
3 Continuous data (finer measurements) Blood pressure Weight
Height Age
4 Time to Event data
Time in remission
10
There are Different Statistical Methods for Different Types of
Data
Binary Data To compare the number of polio cases in the
2treatment arms of the Salk Polio vaccine, you coulduse
Fishers Exact Test Chi-Square Test
Continuous Data To compare blood pressure in a clinical
trialevaluating 2 blood pressure lowering medications, youcould
use
2-sample t-Test Wilcoxon Rank Sum (nonparametric) Test
11
Sample Mean (X )
Add up data, then divide by sample size (n)The sample size nis
the number of
observations(pieces of data)
Example: n = 5 Systolic Blood Pressures (mmHg)
X1 = 120
X2 = 80
X3 = 90
X4 = 110
X5 = 95
X =120 + 80 + 90 + 110 + 95
5= 99 mmHg
12
Notes on Sample Mean (X )
1 Formula
X =
ni=1 Xin
2 Also called sample average or arithmetic mean
3 Sensitive to extreme valuesOne data point could make a great
change in sample mean
4 Why is it called the sample mean?To distinguish it from
population mean
-
13
Population Versus Sample
Population The entire group about which you want information
Blood pressures of all 18-year-old male collegestudents in the
U.S.
Sample A part of the population from which we actuallycollect
information; used to draw conclusions aboutthe whole population
Sample of blood pressures from n = 518-year-old male college
students in the U.S.
14
Population Versus Sample
The sample mean X is not the population mean
PopulationPopulation mean
SampleSample mean X
We dont know the population mean but we would like to We draw a
sample from the population We calculate the sample mean X How close
is X to ? Statistical theory will tell us how close X is to
15
STATISTICAL INFERENCE IS THE PROCESS OFTRYING TO DRAW
CONCLUSIONS ABOUT THE
POPULATION FROM THE SAMPLE
We will return to this later
16
Sample Median
The median is the middle number
80 90 95 110 120
Median
1 Not sensitive to extreme values.If 120 became 200
Median no change Mean big change (becomes 115)
-
17
Sample Median
2 If the sample size is an even number,Average the two middle
numbers
80 90 95 110 120 125
Median
95 + 110
2= 102.5 mmHg
18
19
Describing Variability
How Can We Describe the Spreadof the Distribution?
Minimum and Maximum Range=Max-Min SAMPLE STANDARD DEVIATION (S
or SD)
m
20
Describing Variability
The sample variance is the average of thesquare of the
deviations about the sample
mean
s2 =
ni=1(Xi X )2
n 1
Sample variance (s2) Sample standard deviation (s or SD) is the
square root of s2
Why n 1?Stay tuned
-
21
Calculating s
Example: n = 5Systolic Blood Pressures(mmHg)
X1 = 120
X2 = 80
X3 = 90
X4 = 110
X5 = 95
Sample Mean X = 99 (mmHg)Sample Variance s2 = 255Sample Standard
Deviation (SD) s = 15.97 (mmHg)
22
Notes on s
1 The bigger s is, the more variability there is
2 s measures the spread about the mean
3 s can equal 0 only if there is no spread all n observations
have the same value
4 The units of s are the same as the units of the data(e.g.
mmHg)
5 Often abbreviated SD
6 s is the best estimate of the population standard
deviation
Interpretation
Most of the population will be within about 2 standarddeviations
(s) of the mean X
For a normally (Gaussian) distributed population, most isabout
95%
23
More Notes about SD:Why do we divide by n 1 instead of n?
We want to replace X with in the formula for s2
s2 =
(Xi X )2
n 1
Because we dont know , we use X
But (Xi X )2 tends to be smaller than (Xi )2So, to compensate we
divide by a smaller number:n 1 instead of n
n 1 is called the degrees of freedom of the variance.Why?
The sum of the deviations is zero The last deviation can be
found once we know the other n 1 Only n 1 of the squared deviations
can vary freely
The term degrees of freedom arises in other statistics It is not
always n 1, but it is in this case
24
-
25
Other Measures of Variation
Standard deviation (SD or S) Minimum and maximum observation
Range=Max-Min
What Happens To These as Sample Size Increases?
. Tend to increase?
. Tend to decrease?
. Remain about the same?
26
Continuous Variables: Histograms
Means and medians do not tell the whole story Differences in
spread (variability) Differences in shape of the distribution
Histograms are a way of displaying thedistribution of a set of
data by charting the
number (or percentage) of observations whosevalues fall within
pre-defined numerical ranges
27
How to Make a Histogram
Table 20: Resident Population by Age and State (2000)State
Percent State Percent State Percent
Alabama 13.0 Louisiana 11.6 Ohio 13.3Alaska 5.7 Maine 14.4
Oklahoma 13.2Arizona 13.0 Maryland 11.3 Oregon 12.8Arkansas 14.0
Massachusetts 13.5 Pennsylvania 15.6California 10.6 Michigan 12.3
Rhode Island 14.5Colorado 9.7 Minnesota 12.1 South Carolina
12.1Connecticut 13.8 Mississippi 12.1 South Dakota 14.3Delaware
13.0 Missouri 13.5 Tennessee 12.4Florida 17.6 Montana 13.4 Texas
9.9Georgia 9.6 Nebraska 13.6 Utah 8.5Hawaii 13.3 Nevada 11.0
Vermont 12.7Idaho 11.3 New Hampshire 12.0 Virginia 11.2Illinois
12.1 New Jersey 13.2 Washington 11.2Indiana 12.4 New Mexico 11.7
West Virginia 15.3Iowa 14.9 New York 12.9 Wisconsin 13.1Kansas 13.3
North Carolina 12.0 Wyoming 11.7Kentucky 12.5 North Dakota 14.7
Source: Statistical Abstract of the United States,
2001.www.census.gov/prod/2002pubs/01statab/stat-ab01.html 28
Divide into intervals (equal) Count number in each
Count the observations in each class.Here are the counts:
Class Count Class Count Class Count
4.1 to 5.0 0 9.1 to 10.0 3 14.1 to 15.0 55.1 to 6.0 1 10.1 to
11.0 2 15.1 to 16.0 26.1 to 7.0 0 11.1 to 12.0 9 16.1 to 17.0 07.1
to 8.0 0 12.1 to 13.0 14 17.1 to 18.0 18.1 to 9.0 1 13.1 to 14.0 12
18.1 to 19.0 0
www.census.gov/prod/2002pubs/01statab/stat-ab01.html
-
29
How to Make a Simple Histogram
Divide range of data into intervals (bins) of equal width Count
number of observations in each class Draw the histogram Label
scales
Percent of residents over 65
Num
ber
of s
tate
s
4 6 8 10 12 14 16 18 20
02
46
810
1214
30
Pictures of Data: Histograms
80 100 120 140 160 180
05
1015
Systolic Blood Pressure (mm Hg)
Num
ber
of M
en in
Pop
ulat
ion
80 100 120 140 160 180
010
2030
4050
Systolic Blood Pressure (mm Hg)
Num
ber
of M
en in
Pop
ulat
ion
80 100 120 140 160 180
01
23
45
Systolic Blood Pressure (mm Hg)
Bin width:5 mm Hg
Bin width:20 mm Hg
Bin width:1 mm Hg
31
How many intervals (bins) should you have in a histogram?
There is no perfect answer to thisDepends on sample size nRough
Guideline: # Intervals
n
n Number of Intervals
10 about 350 about 7
100 about 10
Histogram applet
athttp://www.stat.sc.edu/~west/javahtml/Histogram.html
32
Other Types of Histograms
0 1 2 3 4
05
1015
2025
3035
IgM concentrations (g/l) in 324 childrenF
requ
ency
0 1 2 3 4
0.0
0.2
0.4
0.6
0.8
1.0
IgM concentrations (g/l)
Rel
ativ
e F
requ
ency
(%
)
0 1 2 3 4
0.0
0.2
0.4
0.6
0.8
1.0
IgM concentrations (g/l)R
elat
ive
Fre
quen
cy (
%)
FrequencyHistogram
RelativeFrequencyHistogram
RelativeFrequencyPolygon
http://www.stat.sc.edu/~west/javahtml/Histogram.html
-
33
Stem and Leaf Plot
9 | 79
10 | 1166788999
11 | 0112333444555667777889
12 | 00111111233445555566777777889
13 | 0011123333456667788999
14 | 0111224446
15 | 003
16 | 05
34
Boxplots
100
110
120
130
140
150
160
Sample Median
25th Percentile
75th Percentile
Largest Non-Outlier
Smallest Non-Outlier
Outlier
35
Shapes of the Distribution
1988
1976
Source: Sibergeld, Annual Rev. Public Health, 1997.
Many Distributions are Not Symmetric
36
Shapes of the Distribution
Symmetrical andbell-shaped
Positively skewed orskewed to the right
Negatively skewed orskewed to the left
Bimodal Reverse J-shaped Uniform
-
37
Distribution Characteristics
Mode Median Mean
38
Note on Shapes of Distributions
Right Skewed (positively skewed)Long right tailMean >
Mediane.g. hospital stays
Left Skewed (negatively skewed)Long left tailMean <
Mediane.g. humidity (cant get over 100%)
Symmetric Right and left sides are mirror imagesLeft tail looks
like right tailMean Median Mode
Outlier An individual observation that falls outside theoverall
pattern of the graph.
39
Mean=Balancing Point
40
The Histogram and the Probability Density
80 100 120 140 160 180
01
23
45
Systolic Blood Pressure (mm Hg)N
umbe
r of
Men
in P
opul
atio
n
80 100 120 140 160 180
010
2030
Systolic Blood Pressure (mm Hg)
Num
ber
of M
en in
Pop
ulat
ion
80 100 120 140 160 180
0.00
00.
010
0.02
0
Systolic Blood Pressure (mm Hg)P
roba
bilit
y D
ensi
ty
MediumSample
LargeSample
EntirePopulation
-
41
The Probability Density
The probability density is a smooth idealized curve that
showsthe shape of the distribution in the population
This is generally a theoretical distribution that we can
neversee: we can only estimate it from the distribution presented
bya representative (random) sample from the population
Areas in an interval under the curve represent the percent ofthe
population in the interval
42
What is the most well-known Distribution?
The Normal (Gaussian) Distribution
25 30 35 40 45 50
0.00
0.02
0.04
0.06
0.08
Serum Albumin (g/l)
Fre
quen
cy
43
The Normal (Gaussian) Distribution
Symmetric Bell-shaped Mean Median Mode
Applet at
http://stat-www.berkeley.edu/~stark/Java/Html/StandardNormal.htm
44
The Normal Distribution
There are lots of normal distributions:
You can tell which normal distribution you have by knowing
themean and standard deviation:
Mean () is the center Standard deviation () measures the spread
(variability)
http://stat-www.berkeley.edu/~stark/Java/Html/StandardNormal.htm
-
45
The Normal Distribution
Areas under a normal curve represent the proportion of total
valuesdescribed by the curve that fall in that range:
3 2 1 0 1 2 3
The shaded area isapproximately 29% of thetotal area under the
curve
46
The 68-95-99.7 Rule
In any normal distribution, approximately:
. 68% of the observations fall within one standard deviation
ofthe mean.
. 95% of the observations fall within two standard deviationsof
the mean.
. 99.7% of the observations fall within three standarddeviations
of the mean.
*more precisely, 1.96
47
Distributions of Heights in Females Age 18-24
Approximately normal Mean 65 inches Standard deviation 2.5
inches
57.5 60 62.5 65 67.5 70 72.5
68%
57.5 60 62.5 65 67.5 70 72.5
68%
95%
57.5 60 62.5 65 67.5 70 72.5
68%
95%
99.7%
The rule says that if a
population is normally
distributed then approximately
68% of the population will be
within 1 SD of X
It doesnt guarantee that
exactly 68% of your sample of
data will fall within 1 SD of X .
Why? The rule works better if
the sample size is big.
48
-
49
Standard Normal Distribution = 0 and = 1
=0 =2
50
Standard Normal ScoresZ -Scores
How many standard deviations from the population mean are
you?
Standard Score(Z ) =observation population mean
standard deviation
A standard score of
Z = 1 = observation lies one SD above the mean Z = 2 =
observation lies two SD above the mean Z = 1 = observation lies one
SD below the mean Z = 2 = observation lies two SD below the
mean
51
Z -Scores
Example: Female Heights, mean= 65, s = 2.5 inches
1 Height = 72.5 inches
Z =72.5 65
2.5= +3.0
2 Height = 60 inches
Z =60 65
2.5= 2.0
52
Whats the usefulness of standard normal scores?
1 It tells you how many SD(s) an observation is from the
mean.
2 Thus, it is a way of quickly assessing how unusual
anobservation is.
Suppose the mean height is 65 inches, and s = 2.5
Is 72.5 inches unusually tall? If we know Z = 3.0, does that
help us?
-
53
Assuming the population has a normal distribution:
Fraction of Population that isWithin Z More than Z More than Z
More than Z
SDs of SDs above SDs below SDs abovethe mean the mean the mean
or below
the mean
Z0.5 38.29 % 30.85% 30.85% 61.71%1.0 68.27 % 15.87% 15.87%
31.73%1.5 86.64 % 6.68% 6.68% 13.36%2.0 95.45 % 2.28% 2.28%
4.55%2.5 98.76 % 0.62% 0.62% 1.24%3.0 99.73 % 0.13% 0.13% 0.27%3.5
99.95 % 0.02% 0.02% 0.05%
54
Normal Probability Applet
atwww-stat.stanford.edu/~naras/jsm/FindProbability.html
55
Problems
Suppose the population is normally distributed:
1 If you have a standard score of Z = 2, what % of thepopulation
would have scores greater than you?
2 If you have a standard score of Z = 2, what % of thepopulation
would have scores less than you?
56
3 If you have a standard score of Z = 3, what % of thepopulation
would have scores greater than you?
4 If you have a standard score of Z = 1.5, what % of
thepopulation would have scores less than you?
www-stat.stanford.edu/~naras/jsm/FindProbability.html
-
57
5 Suppose we callunusualobservations those that are either
atleast 2 SD above the mean or about 2 SD below the mean.What % are
unusual?
In other words, what % of the observations will have astandard
score either Z > +2.0 or Z < 2.0?What % would have |Z | >
2?
6 What % of the observations would have |Z | > 1.0 (i.e.,
morethan 1 SD away from the mean)?
58
7 What % of the observations would have |Z | > 3.0?
8 What percent of observations would have |Z | > 1.15?
The above results will turn out to be very important later in
ourdiscussion of p-values.
59
Normal Distribution
z P z P z P z P z P0.00 1.0000 0.30 0.7642 0.60 0.5485 0.90
0.3681 1.20 0.23010.01 0.9920 0.31 0.7566 0.61 0.5419 0.91 0.3628
1.21 0.22630.02 0.9840 0.32 0.7490 0.62 0.5353 0.92 0.3576 1.22
0.22250.03 0.9761 0.33 0.7414 0.63 0.5287 0.93 0.3524 1.23
0.21870.04 0.9681 0.34 0.7339 0.64 0.5222 0.94 0.3472 1.24
0.21500.05 0.9601 0.35 0.7263 0.65 0.5157 0.95 0.3421 1.25
0.21130.06 0.9522 0.36 0.7188 0.66 0.5093 0.96 0.3371 1.26
0.20770.07 0.9442 0.37 0.7114 0.67 0.5029 0.97 0.3320 1.27
0.20410.08 0.9362 0.38 0.7039 0.68 0.4965 0.98 0.3271 1.28
0.20050.09 0.9283 0.39 0.6965 0.69 0.4902 0.99 0.3222 1.29
0.19710.10 0.9203 0.40 0.6892 0.70 0.4839 1.00 0.3173 1.30
0.19360.11 0.9124 0.41 0.6818 0.71 0.4777 1.01 0.3125 1.31
0.19020.12 0.9045 0.42 0.6745 0.72 0.4715 1.02 0.3077 1.32
0.18680.13 0.8966 0.43 0.6672 0.73 0.4654 1.03 0.3030 1.33
0.18350.14 0.8887 0.44 0.6599 0.74 0.4593 1.04 0.2983 1.34
0.18020.15 0.8808 0.45 0.6527 0.75 0.4533 1.05 0.2937 1.35
0.17700.16 0.8729 0.46 0.6455 0.76 0.4473 1.06 0.2891 1.36
0.17380.17 0.8650 0.47 0.6384 0.77 0.4413 1.07 0.2846 1.37
0.17070.18 0.8572 0.48 0.6312 0.78 0.4354 1.08 0.2801 1.38
0.16760.19 0.8493 0.49 0.6241 0.79 0.4295 1.09 0.2757 1.39
0.16450.20 0.8415 0.50 0.6171 0.80 0.4237 1.10 0.2713 1.40
0.16150.21 0.8337 0.51 0.6101 0.81 0.4179 1.11 0.2670 1.41
0.15850.22 0.8259 0.52 0.6031 0.82 0.4122 1.12 0.2627 1.42
0.15560.23 0.8181 0.53 0.5961 0.83 0.4065 1.13 0.2585 1.43
0.15270.24 0.8103 0.54 0.5892 0.84 0.4009 1.14 0.2543 1.44
0.14990.25 0.8026 0.55 0.5823 0.85 0.3953 1.15 0.2501 1.45
0.14710.26 0.7949 0.56 0.5755 0.86 0.3898 1.16 0.2460 1.46
0.14430.27 0.7872 0.57 0.5687 0.87 0.3843 1.17 0.2420 1.47
0.14160.28 0.7795 0.58 0.5619 0.88 0.3789 1.18 0.2380 1.48
0.13890.29 0.7718 0.59 0.5552 0.89 0.3735 1.19 0.2340 1.49
0.1362
Tabulated values are the proportion of thestandard Normal
distribution outside therange z, where z is a standard
Normaldeviatealso called two-sided p-values.
60
Normal Distribution
z P z P z P z P z P1.50 0.1336 1.80 0.0719 2.10 0.0357 2.40
0.0164 2.70 0.00691.51 0.1310 1.81 0.0703 2.11 0.0349 2.41 0.0160
2.71 0.00671.52 0.1285 1.82 0.0688 2.12 0.0340 2.42 0.0155 2.72
0.00651.53 0.1260 1.83 0.0672 2.13 0.0332 2.43 0.0151 2.73
0.00631.54 0.1236 1.84 0.0658 2.14 0.0324 2.44 0.0147 2.74
0.00611.55 0.1211 1.85 0.0643 2.15 0.0316 2.45 0.0143 2.75
0.00601.56 0.1188 1.86 0.0629 2.16 0.0308 2.46 0.0139 2.76
0.00581.57 0.1164 1.87 0.0615 2.17 0.0300 2.47 0.0135 2.77
0.00561.58 0.1141 1.88 0.0601 2.18 0.0293 2.48 0.0131 2.78
0.00541.59 0.1118 1.89 0.0588 2.19 0.0285 2.49 0.0128 2.79
0.00531.60 0.1096 1.90 0.0574 2.20 0.0278 2.50 0.0124 2.80
0.00511.61 0.1074 1.91 0.0561 2.21 0.0271 2.51 0.0121 2.81
0.00501.62 0.1052 1.92 0.0549 2.22 0.0264 2.52 0.0117 2.82
0.00481.63 0.1031 1.93 0.0536 2.23 0.0257 2.53 0.0114 2.83
0.00471.64 0.1010 1.94 0.0524 2.24 0.0251 2.54 0.0111 2.84
0.00451.65 0.0989 1.95 0.0512 2.25 0.0244 2.55 0.0108 2.85
0.00441.66 0.0969 1.96 0.0500 2.26 0.0238 2.56 0.0105 2.86
0.00421.67 0.0949 1.97 0.0488 2.27 0.0232 2.57 0.0102 2.87
0.00411.68 0.0930 1.98 0.0477 2.28 0.0226 2.58 0.0099 2.88
0.00401.69 0.0910 1.99 0.0466 2.29 0.0220 2.59 0.0096 2.89
0.00391.70 0.0891 2.00 0.0455 2.30 0.0214 2.60 0.0093 2.90
0.00371.71 0.0873 2.01 0.0444 2.31 0.0209 2.61 0.0091 2.91
0.00361.72 0.0854 2.02 0.0434 2.32 0.0203 2.62 0.0088 2.92
0.00351.73 0.0836 2.03 0.0424 2.33 0.0198 2.63 0.0085 2.93
0.00341.74 0.0819 2.04 0.0414 2.34 0.0193 2.64 0.0083 2.94
0.00331.75 0.0801 2.05 0.0404 2.35 0.0188 2.65 0.0080 2.95
0.00321.76 0.0784 2.06 0.0394 2.36 0.0183 2.66 0.0078 2.96
0.00311.77 0.0767 2.07 0.0385 2.37 0.0178 2.67 0.0076 2.97
0.00301.78 0.0751 2.08 0.0375 2.38 0.0173 2.68 0.0074 2.98
0.00291.79 0.0735 2.09 0.0366 2.39 0.0168 2.69 0.0071 2.99
0.0028
-
61
Is every variable normally distributed?
Absolutely not.
Then why do we spend so much timestudying the normal
distribution?
1 Some variables are normally distributed.
2 A bigger reason is the Central Limit Theorem(next lecture)
62
Population versus Sample
The population of interest could be
All women between ages 30 and 40 All patients with a particular
disease
The sample is a small number of individuals from the
population.The sample is a subset of the population.
63
Population versus Sample
Sample mean X versus population mean ()
e.g. mean blood pressure We know the sample X (e.g., X = 99
mmHg) We dont know the population mean but we would like to
Sample proportion versus population proportion
e.g. proportion of individuals with health insurance We know the
sample proportion (e.g. 80%) We dont know the population
proportion
Key QuestionHow close is the sample mean (or proportion) to
the population mean (or proportion)?
64
Population versus Sample
A parameter A number that describes the population.A parameter
is a fixed number, but in practice we do not know itsvalue.
Example:population meanpopulation proportion
A statistic A number that describes a sample of data.A statistic
can be calculated. We often use a statistic to estimatean unknown
parameter.
Example:sample meansample proportion
-
65
Sources of Error
Errors from Biased SamplingThe study systematically favors
certain outcomes
. Voluntary response
. Non-response
. Convenience sampling
Solution:Random sampling
Errors from (Random) Sampling
. Caused by chance occurrence
. Get a bad sample because of bad luck(by bad I mean not
representative)
. Can be controlled by taking a larger sample
Using mathematical statistics, we can figure out how
muchpotential error there is from random sampling (standard
error)
66
Some Examples of Potentially Biased Sampling
Example Blood pressure study of women age 30-40 Volunteers
Non-random; selection bias Family members
Non-random; not independent Telephone survey; random digit
dial
Random or non-random sample?
Example Clinic Population 100 consecutive patients
Random or non-random sample? Convenience samples are sometimes
assumed to
be random.
67
Example: Literary Digest poll of 1936 presidential election
Election result: 62% voted for RooseveltDigest prediction: 43%
voted for Roosevelt
Problem: Sampling Bias
Selection Bias
Mail questionnaire to 10 million people Sources: telephone
books, clubs Poor people are unlikely to have telephone
(only 25% had telephones)
Non Response Bias
Only about 20% responded (2.4 million) Responders different than
non-responders
68
Bottom Line
When a selection procedure is biased, taking a larger sampledoes
not help
. This just repeats the mistake on a larger scale
Non-respondents can be very different from respondents. When
there is a high non-response rate, look out for
non-response bias
-
69
Random Sample
When a sample is randomly selected from a population, it
iscalled a random sample
In a simple random sample each individual in the populationhas
an equal chance of being chosen for the sample
Random sampling helps control systematic bias But even with
random sampling, there is still sampling
variability or error
70
Sampling Variability
If we repeatedly choose samples from the same population,
astatistic will take different values in different samples
IDEAIf the statistic does not changemuch if you repeated the
study
(you get the same answer each time),then it is fairly
reliable(not a lot of variability)
71
Example
Estimate the proportion of persons in a population who
havehealth insurance
Choose a sample of size n = 1373.
Sample 1
n = 1373 p =1100
1373= .8012
Is the sample proportion reliable? If we took another sample of
another 1373 persons,
would the answer bounce around a lot?
72
Sample 1
p =1100
1373= .8012
Sample 2
p =1090
1373= .7939
Sample 3p = .8347
Sample 4p = .7786
and so on
-
73
The Sampling Distribution
Sample Proportion with Health Insurance
0.76 0.78 0.80 0.82 0.84
010
2030
Samples of size n= 1373
Histogramof 1000Sample
Proportions
74
The spread of the sampling distribution depends on the sample
size
Sample Proportion with Health Insurance
0.70 0.75 0.80 0.85 0.90
05
1015
Samples of size n= 300Sample Proportion with Health
Insurance
0.70 0.75 0.80 0.85 0.90
05
1015
2025
30
Samples of size n= 1000
Proportions based on
sample size n = 300
Proportions based on
sample size n = 1000
75
Lets explore this...
76
Population distribution of Health Insurance
No Health Insurance Health Insurance
Per
cent
age
0.0
0.2
0.4
0.6
0.8
1.0
p = .80
-
77
Lets do an experiment...
Take 500 separate random samples from this population
ofpatients, each with n = 20 patients
For each of the 500 samples, we will plot the health insurance
status record the sample proportion
Ready, set,go...
78
Sample 1
No Health Insurance Health Insurance
Per
cent
age
0.0
0.2
0.4
0.6
0.8
1.0
p = 0.9
Sample 2
No Health Insurance Health Insurance
Per
cent
age
0.0
0.2
0.4
0.6
0.8
1.0
p = 0.6
79
So we did this 500 times...Lets look at a histogram of the 500
sample meansEach based on a sample of size 20
0.0 0.2 0.4 0.6 0.8 1.0
0.0
0.5
1.0
1.5
2.0
2.5
3.0
p = 0.8s = 0.11
80
Lets do ANOTHER experiment...
Take 500 separate random samples from this population
ofpatients, each with n = 50 patients
For each of the 500 samples, we will plot the health insurance
status record the sample proportion
Ready, set,go...
-
81
Sample 1
No Health Insurance Health Insurance
Per
cent
age
0.0
0.2
0.4
0.6
0.8
1.0
p = 0.8
Sample 2
No Health Insurance Health InsuranceP
erce
ntag
e
0.0
0.2
0.4
0.6
0.8
1.0
p = 0.7
82
So we did this 500 times...Lets look at a histogram of the 500
sample meansEach based on a sample of size 50
0.0 0.2 0.4 0.6 0.8 1.0
01
23
45
67
p = 0.8s = 0.06
83
Lets do ANOTHER experiment...
Take 500 separate random samples from this population
ofpatients, each with n = 100 patients
For each of the 500 samples, we will plot the health insurance
status record the sample proportion
Ready, set,go...
84
Sample 1
No Health Insurance Health Insurance
Per
cent
age
0.0
0.2
0.4
0.6
0.8
1.0
p = 0.76
Sample 2
No Health Insurance Health Insurance
Per
cent
age
0.0
0.2
0.4
0.6
0.8
1.0
p = 0.83
-
85
So we did this 500 times...Lets look at a histogram of the 500
sample meansEach based on a sample of size 100
0.0 0.2 0.4 0.6 0.8 1.0
02
46
8
p = 0.8s = 0.04
86
Lets Review
No Health Insurance Health Insurance
0.0 0.2 0.4 0.6 0.8 1.0
0.0 0.2 0.4 0.6 0.8 1.0
0.0 0.2 0.4 0.6 0.8 1.0
Population
n = 20
n = 50
n = 100
p = .8
p = 0.799
p = 0.803
p = 0.798
sp = 0.11
sp = 0.06
sp = 0.04
87
Population distribution of blood pressures
Systolic Blood Pressure (mm Hg)
Per
cent
age
of M
en in
Pop
ulat
ion
80 100 120 140 160
0.00
00.
005
0.01
00.
015
0.02
00.
025
0.03
0
= 125 mm Hg = 14 mm Hg
88
Lets do an experiment...
Take 500 separate random samples from this population ofmen,
each with n = 20 subjects
For each of the 500 samples, we will plot a histogram of the
sample BP values record the sample mean sample standard
deviation
Ready, set,go...
-
89
Sample 1
80 100 120 140 160 180
0.00
0.01
0.02
0.03
0.04
X = 125.17
s = 12.36
Sample 2
80 100 120 140 160 180
0.00
0.01
0.02
0.03
0.04
X = 124.3
s = 11.65
90
So we did this 500 times...Lets look at a histogram of the 500
sample meanseach based on a sample of size 20
80 100 120 140 160 180
0.00
0.02
0.04
0.06
0.08
0.10
0.12
X = 125
sX = 3.07
91
Lets do ANOTHER experiment...
Take 500 separate random samples from this population ofmen,
each with n = 50 subjects
For each of the 500 samples, we will plot a histogram of the
sample BP values record the sample mean sample standard
deviation
Ready, set,go...
92
Sample 1
80 100 120 140 160 180
0.00
0.01
0.02
0.03
0.04
X = 124.98
s = 14.05
Sample 2
80 100 120 140 160 180
0.00
0.01
0.02
0.03
0.04
X = 126.72
s = 13.64
-
93
So we did this 500 times...Lets look at a histogram of the 500
sample meansEach based on a sample of size 50
80 100 120 140 160 180
0.00
0.05
0.10
0.15
0.20
X = 125.01
sX = 1.93
94
Lets do ANOTHER experiment...
Take 500 separate random samples from this population ofmen,
each with n = 100 subjects
For each of the 500 samples, we will plot a histogram of the
sample BP values record the sample mean sample standard
deviation
Ready, set,go...
95
Sample 1
80 100 120 140 160 180
0.00
0.01
0.02
0.03
0.04
X = 127.32
s = 14.93
Sample 2
80 100 120 140 160 180
0.00
0.01
0.02
0.03
0.04
X = 125.06
s = 13.15
96
So we did this 500 times...Lets look at a histogram of the 500
sample meansEach based on a sample of size 100
80 100 120 140 160 180
0.00
0.05
0.10
0.15
0.20
0.25
X = 124.93
sX = 1.41
-
97
Lets Review
80 100 120 140 16090 100 110 120 130 140 150 160
80 100 120 140 160
80 100 120 140 160
80 100 120 140 160
Population
n = 20
n = 50
n = 100
= 125
X = 124.997
X = 125.015
X = 124.934
= 14
sX = 3.07
sX = 1.93
sX = 1.41
98
Population distribution of hospital length of stay
Length of Stay (in days)
Per
cent
age
0 5 10 15 20 25 30
0.00
0.05
0.10
0.15
0.20
= 4 days = 3 days
99
Lets do an experiment...
Take 500 separate random samples from this population ofhospital
admissions, each with n = 16 patients
For each of the 500 samples, we will plot a histogram of the
sample LOS values record the sample mean sample standard
deviation
Ready, set,go...
100
Sample 1
0 5 10 15 20 25
0.00
0.05
0.10
0.15
0.20
0.25
0.30
X = 4.7
s = 2.88
Sample 2
0 5 10 15 20 25
0.00
0.05
0.10
0.15
0.20
0.25
0.30
X = 5.01
s = 2.73
-
101
So we did this 500 times...Lets look at a histogram of the 500
sample meansEach based on a sample of size 16
0 5 10 15 20 25
0.0
0.1
0.2
0.3
0.4
0.5
X = 4.08
sX = 0.74
102
Lets do ANOTHER experiment...
Take 500 separate random samples from this population ofmen,
each with n = 64 subjects
For each of the 500 samples, we will plot a histogram of the
sample LOS values record the sample mean sample standard
deviation
Ready, set,go...
103
Sample 1
0 5 10 15 20 25
0.00
0.05
0.10
0.15
0.20
0.25
0.30
X = 4.26
s = 2.72
Sample 2
0 5 10 15 20 25
0.00
0.05
0.10
0.15
0.20
0.25
0.30
X = 4.08
s = 2.45
104
So we did this 500 times...Lets look at a histogram of the 500
sample meansEach based on a sample of size 64
0 5 10 15 20 25
0.0
0.2
0.4
0.6
0.8
X = 4.1
sX = 0.37
-
105
Lets do ANOTHER experiment...
Take 500 separate random samples from this population ofmen,
each with n = 256 subjects
For each of the 500 samples, we will plot a histogram of the
sample LOS values record the sample mean sample standard
deviation
Ready, set,go...
106
Sample 1
0 5 10 15 20 25
0.00
0.05
0.10
0.15
0.20
0.25
0.30
X = 4.48
s = 3.32
Sample 2
0 5 10 15 20 25
0.00
0.05
0.10
0.15
0.20
0.25
0.30
X = 4.29
s = 2.76
107
So we did this 500 times...Lets look at a histogram of the 500
sample meansEach based on a sample of size 256
0 5 10 15 20 25
0.0
0.5
1.0
1.5
2.0
X = 4.1
sX = 0.19
108
Lets Review
0 5 10 15 20 250 10 20
0 5 10 15 20 25
0 5 10 15 20 25
0 5 10 15 20 25
Population
n = 16
n = 64
n = 256
= 4
X = 4.081
X = 4.104
X = 4.1
s = 3
sX = 0.74
sX = 0.37
sX = 0.19
-
109
Variation in sample mean values tied to size of each sampleNOT
the number of samples
500 5000 500 5000 500 5000
23
45
67
8
Simulations Simulations Simulations
n=16 n=64 n=256
110
The Sampling Distribution
The sampling distribution of a sample statistic refers to what
thedistribution of the statistic would look like if we chose a
largenumber of samples from the same
populationwww.ruf.rice.edu/~lane/stat_sim/sampling_dist/index.html
111
Sampling Distribution of a Sample Mean
The sampling distribution of a sample mean is a
theoreticalprobability distributionIt describes the distribution
of
all sample means from all possible random samples of the same
size taken from a population
112
In real research it is impossible to estimate the
samplingdistribution of a sample mean by actually taking
multiplerandom samples from the same population
no research would ever happen if a study needed to berepeated
multiple times to understand this sampling behavior
Simulations are useful to illustrate a concept, but not
tohighlight a practical approach!
Luckily, there is some mathematical machinery thatgeneralizes
some of the patterns we saw in the simulationresults
www.ruf.rice.edu/~lane/stat_sim/sampling_dist/index.html
-
113
Amazing Result
Mathematical statisticians have figured out how to predict
whatthe sampling distribution will look like without actually
repeatingthe study numerous times and having to choose a sample
each time
Often, the sampling distribution willlook normal
Sample Proportion with Health Insurance
0.76 0.78 0.80 0.82 0.84
010
2030
Samples of size n= 1373 114
The Big Idea
Its not practical to keep repeating a study to evaluate
samplingvariability and to determine the sampling
distribution.Mathematical statisticians have figured out how to
calculate itwithout doing multiple studies.
The sampling distribution of a statistic is often
normallydistributed.
This mathematical result comes from the CENTRAL LIMITTHEOREM.
For the theorem to work, it requires the samplesize (n) to be large
(usually n > 60 suffices).
Statisticians have derived formulas to calculate the
standarddeviation of the sampling distribution and its called
thestandard error of the statistic.
115
Central Limit Theorem
If the sample size is large, the distribution of sample
meansapproximates a normal distribution:
Mean Value
Num
ber
of O
ccur
renc
es
1 2 3 4 5 6
Mean Value
1 2 3 4 5 6
Mean Value
1 2 3 4 5 6
One Die Two Dice Five Dice
116
Illustration of the Central Limit Theorem
0 2 4 6 8 10
0 2 4 6 8 10
0 2 4 6 8 10
0 2 4 6 8 10
Population
Means based
on n = 16
Means based
on n = 32
Means based
on n = 64
-
117
Why is the normal distribution so important in the study of
statistics?
Its not because things in nature are always normally
distributed(although sometimes they are)
Its because of the Central Limit Theorem:The sampling
distribution of statistics (like a sample mean) oftenfollows a
normal distribution if the sample sizes are large
118
Why is the sampling distribution so important?
If a sampling distribution has a lot of variability (i.e. has a
bigstandard error), then if you took another sample its likely
youwould get a very different result
About 95% of the time the sample mean (or proportion) willbe
within 2 standard errors of the population mean (orproportion)
This tells us how close the sample statistic should be to
thepopulation parameter
119
Standard Errors (SE)
Measures the precision of your sample statistic A small SE means
it is more precise The SE is the standard deviation of the sampling
distribution
of the statistic
Mathematical statisticians have come up with formulas for
thestandard error. There are different formulae for:
Standard error of the mean (SEM) Standard error of a
proportion
These formulae always involve the sample size n. As thesample
size gets bigger, the standard error gets smaller.
120
The standard deviationIS NOT
The standard error of a statistic
Standard deviation measures the variability among
individualobservations.
Standard error measures the precision of a statistic such as
thesample mean or proportion that is calculated from anumber (n) of
different observations. The samplemean and sample proportion are
trying to estimatethe population mean or population proportion.
-
121
Standard Error of the MeanSEM
This is a measure of the precision of the sample mean:
SEM =sn
Example
Measure systolic blood pressure on random sample of 100
studentsSample size n = 100Sample mean X = 123.4 mmHgSample SD s =
14.0 mmHg
SEM =14100
= 1.4mmHg
122
Notes on SEM
1 The smaller SEM is, the more precise X is.
2 SEM depends on n and s.
3 SEM gets smaller if s gets smaller n gets bigger
123
Question:
How close to population mean () is sample mean (X)?
ANSWER
The standard error of the sample mean tells us 95% of the
timethe population mean will lie within about 2 standard errors of
thesample mean
X 2SEM
123.4 2 1.4
123.4 2.8
Why is this true? Because of the Central Limit Theorem
INTERPRETATION
We are 95% confident that the sample mean is within 2.8 mmHgof
the population mean. The 95% error bound is 2.8.
124
95% Confidence Interval for Population Mean
X 2SEM
(More accurately: X 1.96SEM)
The CI gives the range of plausible values for Example: Blood
pressure n = 100, X = 123.4 mmHg, s = 1495% CI is
123.4 2 1.4123.4 2.8
Ways to write a confidence interval:
120.6 to 126.2 (120.6, 126.2) (120.6126.2) We are highly
confident that the population mean falls in the
range 120.6 to 126.2.
-
125
Notes on Confidence Intervals
1 Interpretation:Plausible values for the population mean with
highconfidence
2 Are all CIs 95%? No.
It is the most commonly used A 99% CI is wider A 90% CI is
narrower To be more confident you need a bigger interval For a 99%
CI
you need 2.576 SEM95% CI you need 2 SEM (actually its 1.96
SEM)90% CI you need 1.645 SEMWhere do these come from?
126
Notes on Confidence Intervals
3 The length of CI decreases when n increases s decreases Level
of confidence decreases (e.g. 90%, 80% vs 95%)
4 Confidence interval is only accounting for random
samplingerror not other systematic sources of error of bias
Examples
BP measurement is always +5 too highOnly those with high BP
agree to participate(non response bias)
127
Notes on Confidence Intervals
5 Technical InterpretationThe CI works (includes ) 95% of the
time
6 Confidence Interval
Appletwww.stat.sc.edu/~west/javahtml/ConfidenceInterval.html
128
Underlying Assumptions for a 95% CI for the Population Mean
X 2SEM
X 2 sn
Random sample of populationImportant
Sample size n is at least 60 to use 2SEMCentral Limit Theorem
requires large n
www.stat.sc.edu/~west/javahtml/ConfidenceInterval.html
-
129
What if the sample size is smaller than 60?
There needs to be a small correction in the formulaX 2SEM needs
to be slightly bigger.
How much bigger 2 needs to be depends on the sample
size.Computers or statistical tables refer to the degrees offreedom
= n 1. One looks up the correct number in at-table or
t-distribution with n 1 degrees of freedom. You can think of
degrees of freedom like a corrected sample
size. In this case its n 1 because we had to estimate
oneparameter by X . But its not always n 1.
X t SEM
X t sn
130
Value of t.95 used for 95% Confidence Interval for Mean
df t df t
1 12.706 12 2.1792 4.303 13 2.1603 3.182 14 2.1454 2.776 15
2.1315 2.571 20 2.0866 2.447 25 2.0607 2.365 30 2.0428 2.306 40
2.0219 2.262 60 2.000
10 2.228 120 1.98011 2.201 1.960
Notes
Most people use t = 2 once n gets above 60 or so Sometimes
people use 1.96 when n gets bigger (> 120) Value of t depends on
the level of confidence and sample size
131
Students t-Distribution
df 0.2 0.1 0.05 0.02 0.01 0.0011 3.078 6.314 12.706 31.821
63.657 636.6192 1.886 2.920 4.303 6.965 9.925 31.5993 1.638 2.353
3.182 4.541 5.841 12.9244 1.533 2.132 2.776 3.747 4.604 8.6105
1.476 2.015 2.571 3.365 4.032 6.8696 1.440 1.943 2.447 3.143 3.707
5.9597 1.415 1.895 2.365 2.998 3.499 5.4088 1.397 1.860 2.306 2.896
3.355 5.0419 1.383 1.833 2.262 2.821 3.250 4.781
10 1.372 1.812 2.228 2.764 3.169 4.58711 1.363 1.796 2.201 2.718
3.106 4.43712 1.356 1.782 2.179 2.681 3.055 4.31813 1.350 1.771
2.160 2.650 3.012 4.22114 1.345 1.761 2.145 2.624 2.977 4.14015
1.341 1.753 2.131 2.602 2.947 4.07316 1.337 1.746 2.120 2.583 2.921
4.01517 1.333 1.740 2.110 2.567 2.898 3.96518 1.330 1.734 2.101
2.552 2.878 3.92219 1.328 1.729 2.093 2.539 2.861 3.88320 1.325
1.725 2.086 2.528 2.845 3.85021 1.323 1.721 2.080 2.518 2.831
3.81922 1.321 1.717 2.074 2.508 2.819 3.79223 1.319 1.714 2.069
2.500 2.807 3.76824 1.318 1.711 2.064 2.492 2.797 3.74525 1.316
1.708 2.060 2.485 2.787 3.72526 1.315 1.706 2.056 2.479 2.779
3.70727 1.314 1.703 2.052 2.473 2.771 3.69028 1.313 1.701 2.048
2.467 2.763 3.67429 1.311 1.699 2.045 2.462 2.756 3.65930 1.310
1.697 2.042 2.457 2.750 3.646
Tabulated values correspond toa given two-tailed p-value
fordifferent degrees of freedom.
132
Students t-Distribution
df 0.2 0.1 0.05 0.02 0.01 0.00131 1.309 1.696 2.040 2.453 2.744
3.63332 1.309 1.694 2.037 2.449 2.738 3.62233 1.308 1.692 2.035
2.445 2.733 3.61134 1.307 1.691 2.032 2.441 2.728 3.60135 1.306
1.690 2.030 2.438 2.724 3.59136 1.306 1.688 2.028 2.434 2.719
3.58237 1.305 1.687 2.026 2.431 2.715 3.57438 1.304 1.686 2.024
2.429 2.712 3.56639 1.304 1.685 2.023 2.426 2.708 3.55840 1.303
1.684 2.021 2.423 2.704 3.55141 1.303 1.683 2.020 2.421 2.701
3.54442 1.302 1.682 2.018 2.418 2.698 3.53843 1.302 1.681 2.017
2.416 2.695 3.53244 1.301 1.680 2.015 2.414 2.692 3.52645 1.301
1.679 2.014 2.412 2.690 3.52046 1.300 1.679 2.013 2.410 2.687
3.51547 1.300 1.678 2.012 2.408 2.685 3.51048 1.299 1.677 2.011
2.407 2.682 3.50549 1.299 1.677 2.010 2.405 2.680 3.50050 1.299
1.676 2.009 2.403 2.678 3.49651 1.298 1.675 2.008 2.402 2.676
3.49252 1.298 1.675 2.007 2.400 2.674 3.48853 1.298 1.674 2.006
2.399 2.672 3.48454 1.297 1.674 2.005 2.397 2.670 3.48055 1.297
1.673 2.004 2.396 2.668 3.47656 1.297 1.673 2.003 2.395 2.667
3.47357 1.297 1.672 2.002 2.394 2.665 3.47058 1.296 1.672 2.002
2.392 2.663 3.46659 1.296 1.671 2.001 2.391 2.662 3.46360 1.296
1.671 2.000 2.390 2.660 3.460 1.282 1.645 1.960 2.326 2.576
3.291
-
133
t-distribution Applets
http://www.stat.sc.edu/~west/applets/tdemo.html
http:
//www.econtools.com/jevons/java/Graphics2D/tDist.html
134
Example: Blood pressure
n = 5 X = 99 mmHg s = 15.97
95% CI is X 2.776 SEM
99 2.776 7.142
99 19.83
The 95% CI for mean blood pressure is
(79.17, 118.83)
(79.17 118.83)
Rounding off is okay too: (79, 119)
135
Confusion between SD and SEM
Standard deviation (s) - measures spread in the data
Standard error (s/n) - measures the precision of the sample
mean
The standard error of the sample mean depends on the sample
size.
Does the standard deviation depend on the sample size too?
136
PROPORTIONS (p)
Proportion of individuals with health insurance Proportion of
patients who became infected Proportion of patients who are cured
Proportion of individuals who are hypertensive Proportion of
individuals positive on a blood test Proportion of adverse drug
reactions Proportion of premature infants who survive
On each individual in the study, we record a binary
outcome(Yes/No; Success/Failure) rather than a continuous
measurement
http://www.stat.sc.edu/~west/applets/tdemo.htmlhttp://www.econtools.com/jevons/java/Graphics2D/tDist.htmlhttp://www.econtools.com/jevons/java/Graphics2D/tDist.html
-
137
Proportions
How accurate of an estimate is the sample proportion of
thepopulation proportion?
What is the standard error of a proportion?
138
Example
n = 200 patientsX = 90 adverse drug reactionThe estimated
proportion who experience an adverse drug reactionis
p = 90/200 = .45
or 45%
NOTES
There is uncertainty about this rate because it involved onlyn =
200 patients
If we had studied a much larger number of patients, would wehave
gotten a much different answer?
The sample proportion is p = .45 But it is not the true rate of
adverse drug reactions in the
population
139
The Sampling Distribution of a Proportion
Sample Proportion
Num
ber
of S
ampl
es
0.35 0.40 0.45 0.50 0.55
02
46
810
12
The standard error of asample proportion is
SE (p) =
p (1 p)
n
140
95% CI for a Proportion
p 1.96SE (p)
p 1.96
p (1 p)n
p is the sample proportionn is the sample size
Example
n = 200 patientsX = 90 adverse drug reactions
p = 90/200 = .45
.45 1.96.45 .55
200.45 1.96 0.035
.45 0.07The 95% confidence interval is (.38 .52).
-
141
Interpreting a 95% CI for a Proportion
Plausible range of values for population proportion Highly
confident that population proportion is in the interval The method
works 95% of the time
p
142
Notes on 95% CI for Proportions
1 Random (or representative) sampleSuppose the 200 patients were
sicker?Suppose the 200 patients were consecutive?
2 The confidence interval does not address your definition
ofdrug reaction and whether thats a good or bad definition.
Itaccounts only for sampling variation.
3 Can also have CI with different levels of confidence
143
4 Sometimes 1.96SE (p) is called
95% Error BoundMargin of Error
5 The formula for a 95% CI is ONLY APPROXIMATE. It workswell if
the number of failures (drug reactions) and
successes(non-reactions) are both at least 5.Otherwise, you need to
use a computer to perform somethingcalled exact binomial
calculations.You do NOT use the t-correction for small sample sizes
likewe did for sample means. We use exact binomialcalculations.
144
Example
Study of survival of premature infants. All premature babies
born at Johns Hopkins during a 3 year
period (Allen, et al. NEJM, 1993)
n = 39 infants born at 25 weeks gestation 31 survived 6
months
p =31
39= 0.79
95% CI .63 .91(based on exact binomial calculations)
Source: Motulsky, Intuitive Biostatistics
-
145
Are confidence intervals needed even though all infants
werestudied?
Are the 39 infants a sample? Seems like its the whole
population.
It makes sense to calculate a CI when the sample
isrepresentative of a larger population about which you
wish to make inferences. It is reasonable to thinkthat these
data from several years at one hospital are
representative of data from other years at otherhospitals, at
least at big-city university hospitals in
the United States.
146
Comparison of 2 Groups
Are the Population Means Different(Continuous Data)
Two Situations
1 Paired Design
Before-after data Twin data
2 Two Independent Sample Design
School Children
Treatment A
Treatment B
147
Paired Design
BeforeAfter
Why Pairing?
Controls extraneous noise Everyone acts as own control
148
Example: Blood pressure and Oral Contraceptive Use
Subjects Ten non-pregnant, pre-menopausal women 16-49years old
who were beginning a regimen of oralcontraceptive (OC)
Methods Measure blood pressure prior to starting OC use,
andthree-months after consistent OC use
Goal Identify any changes in average blood pressureassociated
with OC use in such women
Rosner, Fundamentals of Biostatistics, (2005).
-
149
Example: Blood pressure and Oral Contraceptive UseDifference
BP Before OC BP After OC After-Before1. 115 1282. 112 1153. 107
1064. 119 1285. 115 1226. 138 1457. 126 1328. 105 1099. 104 102
10. 115 117
sample mean 115.6 120.4
The sample average of the differences is 4.8. The sample
standard deviation (s) of the differences iss = 4.57.
150
Calculate a 95% CI for the Expected Change in Blood Pressure
95% CI for population mean BP change
X t.95,df =9 SEM
4.8 2.262 4.5710
4.8 2.262 1.4451.53 mm Hg to 8.07 mm Hg
Notes
1 Where does 2.262 come from?See the t-distribution with 9
degrees of freedom
2 The BP change could be due to factors other than
oralcontraceptives. A control group of comparable women whowere not
taking oral contraceptives would strengthen thisstudy.
151
3 The number 0 is NOT in confidence interval (1.53 8.07)
0 1.53 8.07
Because 0 is not in the interval, suggests there is a
significantchange in BP over time
There is a significant increase in blood pressure
152
Hypothesis TestingSignificance Testingand p-values
Want to draw a conclusion about a population parameter:
In a population of women who use oralcontraceptives, is the
average (expected) change in
blood pressure (After-Before) 0 or not?
Sometimes statisticians use the term expected for thepopulation
average
is the expected (population) mean change in blood pressure
Choose between two competing possibilities for using asingle
imperfect (paired) sample
Null hypothesis H0: = 0Alternative hypothesis H1: 6= 0
We reject H0 if the sample mean is far away from 0.
-
153
The Hypotheses
We set up mutually exclusive, exhaustive possibilities for the
truth:
The null hypothesis H0Typically represents the hypothesis that
there is noeffect or difference.It represents current beliefs or
state of knowledge.For example, there is no effect of oral
contraceptiveson blood pressure:
H0 : = 0
The alternative hypothesis H1Typically represents what you are
trying to prove.For example, oral contraceptives affect
bloodpressure:
H1 : 6= 0
154
Do we have sufficient evidence to reject H0 and claim H1 is
true?
If X is close to zero, it is consistent with H0 If X is far from
zero, is it consistent with H1
How do we decide if X = 4.8 ismore consistent with H0 or H1?
155
The p-value
What is the probability of observing an extreme sample
meanlike4.8 mm Hgif the null hypothesis (H0: = 0 ) were true?
The answer is called the p-value
If that probability (p-value) is small, it suggests the
observedresult was unlikely if H0 is true.
This would provide evidence against H0 If that probability
(p-value) is large, it suggests the observed
result quite probably if H0 is true. This would provide evidence
for H0
156http://xkcd.com/892/
-
157
How are p-values calculated?
1 First, measure the distance between the sample mean andwhat
you would expect the sample mean to be if H0: = 0were true:
t =sample mean 0
SEM
t =4.8
4.57/
10=
4.8
1.45= 3.31
The value t = 3.31 is called the test statistic We observed a
sample mean that was 3.31 standard
deviations of the mean (SEM) away from what we would
haveexpected the mean to be if OC has no effect (i.e., d = 0)
158
The t-statistic is analogous to the Z -score on pages 37-44
Z =observation mean
SDt =
X 0SEM
Z t
observation sample meanstandard deviation standard error
mean 0 because we are calculating p-values under the scenario
thatH0: = 0
159
How are p-values calculated?
2 Next, calculate the probability of getting a test statistic as
ormore extreme than what you observed (t=3.31) if H0 wastrue:
This p-value comes from the normal distribution. How unusual is
it to get a standard normal score as extreme as
3.31? Not likely at all (p < .01)
3.313.31
160
How are p-values calculated?
If the sample size is small (n < 60), a small t-correction
mustbe made
Instead of a normal distribution, t-distribution is used withn 1
degrees of freedom
The p-values gets a little larger
Use the t table. This procedure is called a paired t-test with n
1 degrees of
freedom. In the oral contraceptive example, we performed apaired
t-test with 9 degrees of freedom.
-
161
Interpreting the p-value
The p-value in the blood pressure/OC example is .0089
Interpretation If the true before OC/after OC blood
pressuredifference is 0 amongst all women taking OCs, thenthe
chance of seeing a mean difference asextreme/more extreme as 4.8 in
a sample of 10women is .0089
162
Using the p-value to make a decision
1 p-values are probabilities (numbers between 0 and 1).Small
p-values are measures of evidence against H0 in favor ofH1.
2
The p-value is the probability of obtaining a resultas/or more
extreme than you did by chance alone
assuming the null hypothesis H0 is true.
3 If the p-value is small either
(a) A very rare event occurred and H0 is trueOR
(b) H0 is false
163
Using the p-value to make a decision
The p-value in the blood pressure/OC example is .0089
. This p-value is small
. So there is a small probability of observing our data
(orsomething more extreme) if H0 is true
. We reject H0
164
Using the p-value to make a decision
4 p-value is a continuum of evidenceGuidelines?
p = .10: suggestive p = .05: magical cutoff p = .01: strong
evidence
5 How precise should p-values be?
2 decimal places suffice (p = .07) Sometimes 3 decimal places if
p < .01
. p = .007
If the p-value is really small, p < .001 is fine If the
p-value is really big, p > .20 is fine
-
165
Blood PressureOC ExampleSummary
Methods The changes in blood pressures after oral contraceptive
usewere calculated for 10 women. A paired t-test was used
todetermine if there was a significant change in blood pressureand
a 95% confidence was calculated for the mean bloodpressure change
(after-before).
Result Blood pressure measurements increased on average 4.8 mm
Hgwith standard deviation 4.57. The 95% confidence interval forthe
mean change was 1.5 8.1. There was evidence that bloodpressure
measurements after oral contraceptive use weresignificantly higher
than before oral contraceptive use(p = .0089).
Discussion A limitation of this study is that there was no
comparisongroup of women who did not use oral contraceptives. We
donot know if blood pressures may have risen even without
oralcontraceptive usage.
166
SummaryPaired t-test
1 Designate null and alternative hypotheses
2 Collect data Compute change for each paired set of
observations Compute Xd , the sample mean of the paired differences
Compute sd , the sample standard deviation of the differences
3 Calculate the test statistic
t =Xd 0SEM
=Xd 0sd/n
4 Compare t to a t-distribution to get a p-value If p is small,
Reject H0 If p is large, Fail to Reject H0
167
Two Types of Errors
Type I error: Claim H1 is true when in fact H0 is true
Type II error: Do not claim H1 is true when in fact H1 is
true
The probability of making a Type I error is called the -levelThe
probability of making a Type II error is called the -levelThe
probability of NOT making a Type II error is called the power
168
The p-value and the -level
Some people will only call a p-value significant if it is less
thansome cutoff (e.g., .05). This cutoff is called the -level
The -level is the probability of a type I error. It is the
probabilityof falsely rejecting H0.
Statistically significantThe p-value is less than a preset
threshold value, .
Do Not Say
The result is statistically significantThe result is
statistically significant at = .05The result is significant (p <
.05)
Instead Give the p-value and Interpret
The result is statistically significant (p = .009)
-
169
One-Sided versus Two-Sided p-values
Two-sided p-value: (p = .009) Probability of a result as or
moreextreme than observed (either X < 4.8 or X > 48)
One-sided p-value: (p = .0045) Probability of a more
extremepositive result than observed (X > 4.8)
You never know what direction study results will go...In this
course, we will use two-sided p-values exclusively.This is what is
typically done in the scientific/medical literature.
170
Connection between CIs and HTs
The CI gives plausible values for the population parameter data
take me to the truth
Hypothesis testing postulates two choice for the
populationparameter
here are two possibilities for the truth, data help me choose
one
171
Connection between CIs and HTs
If 0 is not in the 95% CI, then we reject H0 that = 0 atlevel =
.05 (the p-value < .05)
0 1.53 8.07
Why? CI starts at Xd and captures 2 standard errors in
either
direction If 0 is not in the 95% CI, the d is more than 2
standard errors
from 0 (either above or below) So the distance (t) will be >
2 or < 2
and the resulting p-value< .05
172
Connection between CIs and HTs
In this BP-OC example, the 95% confidence interval tells usthat
the p-value is less than .05, but it doesnt tell us that itis p =
.009
The confidence interval and the p-value are complementary.You
cant get an exact p-value from just looking at aconfidence
interval
I like to report both
-
173
More on the p-value
STATISTICAL SIGNIFICANCE DOESNOT IMPLY CAUSATION
Blood Pressure Example
There could be other factors that could explain the change
inblood pressure.
A significant p-value is only ruling out random sampling asthe
explanation.
Need a Comparison Group
Self-selected (may be okay)Randomized (better)
174
More on the p-value
STATISTICAL SIGNIFICANCEIS NOT THE SAME AS
SCIENTIFIC SIGNIFICANCE
Example: Blood Pressure and Oral Contraceptives
n = 100, 000
X = .03 mmHg
s = 4.57
p-value = .04
Big n can sometimes produce a small p-value even though
themagnitude of the effect is very small (not
scientificallysignificant)
Supplement with a CI: 95% CI is .002 .058 mmHg
175
The Language of Hypothesis (Significance) Testing
Suppose the p-value is p = .40How might this result be
described?
Not statistically significant Do not reject H0
Can we also say?
Accept H0 Claim H0 is true
Statisticians much prefer the double negativeDo not reject
H0
176
More on the p-value
NOT REJECTING H0 IS NOT THE SAMEAS ACCEPTING H0
Example: Blood Pressure and Oral Contraceptives
n = 5
X = 5.0 mmHg
s = 4.57
p-value = .07
We cannot reject H0 at significance level = .05.Are we convinced
there is no effect of OC on BP?
Maybe we should have taken a bigger sample. Interesting trend,
but not proven beyond a reasonable doubt
Look at the confidence interval 95% CI (-.67, 10.7)
Innocent until proven guilty
-
177
Comparing Two Independent Groups
Controlled Trial in Peru of Bismuth Subsalicylate (Pepto
Bismol)in Infants with Diarrheal Disease
Infants
Controls n = 84
Treatment n = 85
Control Tx
n 84 85Mean stool output ml/kg 260 182Standard deviation(s) 254
197
Scientific Question: Is there a treatment effect?
178
Note
The data are not paired. There are different infants in each
group.
2 independent groups How do we calculate
Confidence interval for difference p-value to determine if the
difference in two groups is
significant 2-sample (unpaired) t-test
179
95% CI for the Difference in Means of Two Independent (Unpaired)
Groups
Generic CI formula:
estimate 1.96 SE
(X1 X2) 1.96 SE (X1 X2)
SE (X1 X2) = standard error of thedifference of 2 sample
means
The standard error of the difference for two independentsamples
is calculated differently than we did for paired designs.
Statisticians have developed formulae for the standard error
ofthe difference. These formulae depend on sample size in
bothgroups and standard deviations in both groups.
180
The SE of the Difference in Sample Means
Principle: Variation from independent sources can be added
Variance(X1 X2) = (SE (X1))2 + (SE (X2))2
SE (X1 X2) =
(SE (X1))2 + (SE (X2))2
Formula depends on n1, n2, s1, s2 There are other slightly
different equations for SE (X1 X2)
But they all give similar answers
-
181
The SE of the Difference in Sample Mean
Control Tx
n 84 85Mean stool output ml/kg 260 182Standard deviation(s) 254
197
SE (X1 X2) =(
254/
84)2
+(
197/
85)2
=
27.712 + 21.372
= 34.94
182
Example: Pepto Bismol RCT
95% CI for Difference in Mean
78 1.96 SE (X1 X2)
78 1.96 34.94
78 68.48
9 to 147
Note
The confidence interval does not include 0. Thus, p < .05
183
Hypothesis Test to Compare Two Independent Groups
Two-Sample (Unpaired) t-test
Are the expected stool outputs equal in the two groups?
H0 : 1 = 2
H1 : 1 6= 2
t =difference in means 0SE of the difference
t =260 182
34.94=
78
34.94= 2.23
184
Notes on the 2-sample t-test
1 This is a 2-sample (unpaired) t-test
2 The value t = 2.23 is the test statistic
3 We calculate a p-value which is the probability of obtaining
atest statistic as extreme as we did if H0 was true.
2.232.23
4 How is the probability computed? If sample sizes are
large(both greater than 60) a normal distribution is used.
5 If sample sizes are small, a small t correction is required (a
tdistribution is used with n1 + n2 2 degrees of freedom; thatis the
degrees of freedom is the total sample size from bothgroups minus
2).An assumption that is also required is that both populationsare
approximately normally distributed. (Results can be
highlyinfluenced by wild observations or outliers.)
-
185
DiarrheaPepto BismolSummary
Question Is there a difference in mean stool output between the
twotreatment groups?
Methods The stool output was calculated for 84 infants
randomized toplacebo and 85 infants randomized to Pepto Bismol. A
95%confidence interval was calculated for the difference in
meanstool output between the two groups and a two-sample t-testwas
used to determine if there was a significant differentbetween the
two groups.
Result The mean stool outputs in the treated and control groups
were182 and 260 respectively. The control group stool output
wassignificantly higher than the treated group (p = .03).
Thecontrol group was 78 ml/kg higher than the treated group(95%
confidence interval 9 147 ml/kg).
186
Nonparametric Alternative to the 2-sample t-test
Mann-Whitney-Wilcoxon Rank Sum Test
Objective Assess if the two populations are different?Advantages
Does not assume populations are normally
distributed. The two-sample t-test requires thatassumption with
small sample sizes
Uses only ranksdo not need precise numericaloutcomes
Not sensitive to outliersDisadvantage of the Nonparametric
Test
Nonparametric methods are often less sensitive(powerful) for
finding true differences becausethey throw away information (they
use onlyranks)
Need full data set, not just summary statistics
187
Example: Health Intervention Study
Evaluate an intervention to educate high school students
abouthealth and lifestyle
Y = Post Pretest Score
Randomize
Intervention (I) 5 0 7 2 19
Control (C) 6 -5 -6 1 4
Only 5 individuals in eachsample
With such a small sample, we need to be sure scoreimprovements
are normally distributed if we want to use t-testBIG assumption
Alternative: Mann- Whitney-Wilcoxon non-parametric test!
188
Rank the pooled data:
Order: Rank: Group:
Find the average rank in the 2 groups:
Intervention group average rank: 3+5+7+9+105 = 6.8
Control group average rank: 1+2+4+6+85 = 4.2
p-value calculations:
Statisticians have developed formulae and tables to determine
theprobability of observing such an extreme discrepancy (6.8 vs
4.2)by chance alone. Thats the p-value.In the example, p = .22.
-
189
Health Intervention StudySummary
Question Is there a difference in test score change between
theintervention and control groups?
Design 10 high school students were randomized to either receive
atwo-month health and lifestyle education program (or noprogram).
Each student was administered a test regardinghealth and lifestyle
issues prior to randomization and after thetwo-month period.
Statistics Differences in the two test scores (after-before)
were computedfor each student. Mean and median test score changes
werecomputed for each study group. A Mann-Whitney rank sumtest was
used to determine if there was a statistically
significantdifference in test score change between the intervention
andcontrol groups at the end of the two-month study period.
Result The median score change was four points higher in
theintervention group than in the control group. The difference
intest score improvements between the intervention and
controlgroups was not statistically significant (p = .17)
190
Note
In the health insurance study, the p-value was .22.. No
significant difference in test scores between the intervention
and control group (p = .22)
The two-sample t-test would give a different answer (p = .11).
Different statistical procedures can give different p-values. If
the largest observation 19 was changed to 100, the p-value
based on the Mann-Whitney test would not change but
thetwo-sample t-test would.
191
The t-test or the nonparametric test?
Statisticians will not always agree, but there are some
guidelines:
Use nonparametric test if sample size is small and
distributionlooks skewed. You might also do a t-test, too, and
compare.
Only ranks availableOtherwise, use t-test
192
Example: Exposure of Young Infants to Environmental Tobacco
Smoke
Objective This study examined the degree to which breast-feeding
andcigarette smoking by mothers and smoking by other
householdmembers contributed to the exposure of infants to
theproducts of tobacco smoker (urinary cotinine level).
Method We report median values and interquartile ranges for
eachgroup. Comparisons between groups are made with theWilcoxon
rank sum test because the distributions of urinecotinine values are
positively skewed.
Source: Mascola et al., AJPH, 1998, 88:893-895.
-
193
Extension of the 2-sample t-testAnalysis of VarianceOne-Way
ANOVA
The t-test compares two populations Analysis of variance is a
generalization of the two-samplet-test to compare three or more
populations
The test statistic from ANOVA calculations is called theF
-test
A p-value is then calculated Are there any differences among the
populations?
An alternative strategy is to perform lots of two-samplet-tests
(pairwise)
That could be a lot of statistical testing! Instead, perform an
ANOVA
No significant differences Stop. No further analysis
necessarySignificant differences Do two-sample t-tests to find
them
194
Example: Pulmonary Disease
Goal: Does passive smoking have a measurable effect onpulmonary
health?
Methods: Measure mid-expiratory flow (FEF) in liters persecond
(amount of air expelled per second) in sixsmoking groups.
1 Nonsmokers (NS)
2 Passive Smokers (PS)
3 Noninhaling Smokers (NI)
4 Light Smokers (LS)
5 Moderate Smokers (MS)
6 Heavy Smokers (HS)
White and Froeb. Small-Airways Dysfunction in Non-Smokers
Chronically Exposedto Tobacco Smoke, NEJM 302: 13 (1980)
195
One strategy is to perform lots of two-sample t-tests...
Group Group Mean FEF sd FEFnumber name (L/s) (L/s) n
1 NS 3.78 0.79 2002 PS 3.30 0.77 2003 NI 3.32 0.86 504 LS 3.23
0.78 2005 MS 2.73 0.81 2006 HS 2.59 0.82 200
In this example, there would be 15 comparisons... It would be
nice to have one catch-all test which would tell
you whether there were any differences amongst the six
groups
196
Mean FEF 2 SE
Group
FE
F (
L/s)
NS PS NI LS MS HS
2.5
2.7
2.9
3.1
3.3
3.5
3.7
Based on a one-way analysis of variance, there are
significantdifferences in pulmonary function among these groups(p
< .001).
Pairwise two-sample t-tests show very significant
differencesbetween nonsmokers and all other groups.
There were no significant differences between passive
smokers,noninhalers and light smokers; and between moderate
andheavy smokers.
-
197
Smoking and FEFSummary
Subjects A sample of over 3,000 persons was classified into one
of sixsmoking categorizations based on responses to smoking
relatedquestions
Methods 200 men were randomly selected from each of five
smokingclassification groups (non-smoker, passive smokers,
lightsmokers, moderate smokers, and heavy smokers), as well as
50men classified as non-inhaling smokers for a study designed
toanalyze the relationship between smoking and
respiratoryfunction
198
Smoking and FEFSummary
Statistics ANOVA was used to test for any differences in FEF
levelsamongst the six groups of men
Individual group comparisons were performed with aseries of two
sample t-tests, and 95% confidence intervalswere constructed for
the mean difference in FEF betweeneach combination of groups
Results Analysis of variance showed statistically significant(p
< .001) differences in FEF between the six groups ofsmokers.
Non-smokers had the highest mean FEF value, 3.78 L/s,and this
was statistically significantly larger than the fiveother
smoking-classification groups
The mean FEF value for non-smokers was 1.19 L/shigher than the
mean FEF for heavy smokers (95% CI1.031.35 L/s), the largest mean
difference between anytwo smoking groups
199
Whats the rationale behind analysis of variance?
H0 : 1 = 2 = = k
H1 : at least one mean is different
The variation in the sample means between groups is compared
tothe variation within a group.
Group
FE
F (
L/s)
NS PS NI LS MS HS
2.5
2.7
2.9
3.1
3.3
3.5
3.7
If the between group variation is a lot bigger than the within
groupvariation, that suggests there are some differences among
thepopulations.
http://www.ruf.rice.edu/~lane/stat_sim/one_way/index.html
200
Overuse of Hypothesis TestsBad Statistics!!
SampleAge n Mean
< 20 97 17.8 20 88 24.6
http://www.ruf.rice.edu/~lane/stat_sim/one_way/index.html
-
201
Comparing Two Proportions
Study: Clinical trial of AZT to prevent
maternal-infanttransmission of HIV.
Randomize
AZTn = 121
9 infectedinfants
Placebon = 127
31 infectedinfants
Conner et al. New England J. of Medicine 331:1173-1190
(1994)
202
Notes on Design
Random assignment of TxHelps insure 2 groups are
comparablePatient & physician could not request particular
Tx
Double blindPatient & physician did not know Tx
assignment
Definition of infectionTwo positive cultures (infant > 32
weeks)
203
HIV Transmission Rates
AZT 9/121 = .074 (7.4%)Placebo 31/127 = .244 (24.4%)
Note
These are NOT the true population parameters for thetransmission
rates.There is sampling variability
204
HIV Transmission Rates
95% confidence intervals
AZT 95% CI .03 .14Placebo 95% CI .17 .32
1 Is the difference significant, or can it be explained by
chance?
2 As CIs do not overlap, suggests significance. But whats
thep-value?
3 Note: if the CIs did overlap, it would still be possible to
get ap < .05.
Want a direct method for testing 2 independent proportions
-
205
Display the Data in a 2 2 Table(2 rows and 2 columns)
AZT Placebo
HIV transmission(infected)
Yes 9 31 40
No 112 96 208
121 127 248
206
Hypothesis Testing
H0 : p1 = p2
H1 : p1 6= p2
p1 = Proportion infected on AZTp2 = Proportion infected on
placebo
1 Fishers Exact Test
2 (Pearsons) Chi-Square Test (2)
207
Fishers Exact Test
As with all hypothesis tests, start by assuming H0 is true:AZT
is not effective
Imagine putting 40 red balls (the infected) and 208 blue
balls(non-infected) in a jar. Shake it up.
Now choose 121 ballsthats AZT group. The remaining balls are the
placebo group.
We can calculate the probability you get 9 or fewer red
ballsamong the 121. That is the one-sided p-value.
The two-sided p-value is just about (but not exactly) twicethe
one-sided p-value. It accounts for the probability ofgetting either
extremely few red balls or a lot of red balls inthe AZT group.
The p-value is the probability of obtaining a result as or
moreextreme (more imbalance) than you did by chance alone.
208
Notes on Fishers Exact Test
Calculations are difficult Always appropriate to test equality
of two proportions Computers are usually used Exact p-value (no
approximations)
no minimum sample size requirements
-
209
HIV-AZTSummary
Study We conducted a randomized, double-blind,
placebo-controlledtrial of the efficacy and safety of zidovudine
(AZT) in reducingthe risk of maternal-infant HIV transmission
Methods HIV transmission rates for both the placebo and AZT
groupswere calculated as the ratio of HIV infected infants (based
oncultures at 32 weeks) divided by the total number of infantsand
95% confidence intervals were calculated. The transmissionrates for
the two groups were compared by Fishers Exact Test.
Results The maternal infant HIV transmission rate for the
AZTgroup was 7.4%(95% CI 3.5% 13.7%)
The maternal infant HIV transmission rate for theplacebo group
was 24.4%(95% CI 17.2% 32.8%)
AZT significantly reduced the rate of HIV transmissioncompared
to placebo (p < .001)
210
The Chi-Square Approximate Method
Works for big sample sizes If all 4 numbers in the 2 2 table are
5 or more it is okay
The only advantage of this method over Fishers Exact Test isyou
dont need a computer to do it.
211
The Chi-Square Approximate Method
Looks at discrepancies between observed and expected.
O = observed
E = expected =row total column total
grand total
Expected refers to the values for the cell counts that would
beexpected if the null hypothesis is true
212
The Chi-Square Approximate Method
1 Calculate expected counts assuming H0 is true
2 Calculate a test statistic to measure the difference
betweenwhat we observe and what we expect
Test Statistic 2 =
4 cells
(O E )2
E
3 Use a chi-square table with 1 degree of freedom to get
ap-value
How likely is it to get such a big discrepancy between
theobserved and expected?
-
213
2 Distribution with 1 Degree of Freedom
0 1 5
3.84
214
Performing the 2 Test for a 22 Table
AZT Placebo
HIV transmission(infected)
Yes 9 31 40
No 112 96 208
121 127 248
Observed = 9
Expected = 121 40248
= 19.52
215
Performing the 2 Test for a 22 Table
AZT Placebo
HIVYes 9 31 40
No 112 96 208
121 127 248
AZT Placebo
Yes 40
No 208
121 127 248
Observed Expected
216
Performing the 2 Test for a 22 TableAZT Placebo
HIVYes 9 31 40No 112 96 208
121 127 248
AZT Placebo
HIVYes 19.52 20.48 40No 101.48 106.52 208
121 127 248
(9 19.52)2
19.52+
(112 101.48)2
101.48+
(31 20.48)2
20.48+
(96 106.52)2
106.52
= 13.19 2
13.19
The p-value is about p = .0003 It is NOT a coincidence that
the
square of Z on page 141 is almost the2. One is nearly the square
of theother:
13.19 3.63
-
217
2 Distribution with 1 Degree of Freedom
This table assumes that you have one degree of freedomthe case
when analyzing a22 table:
2 P 2 P 2 P 2 P 2 P 2 P0.0 1.0000 2.5 0.1138 5.0 0.0253 7.5
0.0062 10.0 0.0016 12.5 0.00040.1 0.7518 2.6 0.1069 5.1 0.0239 7.6
0.0058 10.1 0.0015 12.6 0.00040.2 0.6547 2.7 0.1003 5.2 0.0226 7.7
0.0055 10.2 0.0014 12.7 0.00040.3 0.5839 2.8 0.0943 5.3 0.0213 7.8
0.0052 10.3 0.0013 12.8 0.00030.4 0.5271 2.9 0.0886 5.4 0.0201 7.9
0.0049 10.4 0.0013 12.9 0.00030.5 0.4795 3.0 0.0833 5.5 0.0190 8.0
0.0047 10.5 0.0012 13.0 0.00030.6 0.4386 3.1 0.0783 5.6 0.0180 8.1
0.0044 10.6 0.0011 13.1 0.00030.7 0.4028 3.2 0.0736 5.7 0.0170 8.2
0.0042 10.7 0.0011 13.2 0.00030.8 0.3711 3.3 0.0693 5.8 0.0160 8.3
0.0040 10.8 0.0010 13.3 0.00030.9 0.3428 3.4 0.0652 5.9 0.0151 8.4
0.0038 10.9 0.0010 13.4 0.00031.0 0.3173 3.5 0.0614 6.0 0.0143 8.5
0.0036 11.0 0.0009 13.5 0.00021.1 0.2943 3.6 0.0578 6.1 0.0135 8.6
0.0034 11.1 0.0009 13.6 0.00021.2 0.2733 3.7 0.0544 6.2 0.0128 8.7
0.0032 11.2 0.0008 13.7 0.00021.3 0.2542 3.8 0.0513 6.3 0.0121 8.8
0.0030 11.3 0.0008 13.8 0.00021.4 0.2367 3.9 0.0483 6.4 0.0114 8.9
0.0029 11.4 0.0007 13.9 0.00021.5 0.2207 4.0 0.0455 6.5 0.0108 9.0
0.0027 11.5 0.0007 14.0 0.00021.6 0.2059 4.1 0.0429 6.6 0.0102 9.1
0.0026 11.6 0.0007 14.1 0.00021.7 0.1923 4.2 0.0404 6.7 0.0096 9.2
0.0024 11.7 0.0006 14.2 0.00021.8 0.1797 4.3 0.0381 6.8 0.0091 9.3
0.0023 11.8 0.0006 14.3 0.00021.9 0.1681 4.4 0.0359 6.9 0.0086 9.4
0.0022 11.9 0.0006 14.4 0.00012.0 0.1573 4.5 0.0339 7.0 0.0082 9.5
0.0021 12.0 0.0005 14.5 0.00012.1 0.1473 4.6 0.0320 7.1 0.0077 9.6
0.0019 12.1 0.0005 14.6 0.00012.2 0.1380 4.7 0.0302 7.2 0.0073 9.7
0.0018 12.2 0.0005 14.7 0.00012.3 0.1294 4.8 0.0285 7.3 0.0069 9.8
0.0017 12.3 0.0005 14.8 0.00012.4 0.1213 4.9 0.0269 7.4 0.0065 9.9
0.0017 12.4 0.0004 14.9 0.0001
218
Chi-Square for Associations in r c Tables
219
Summary of Methods for Comparing Proportions
Fishers Exact Test Always works with large or small sample
sizeHighly computational; need a computer
2-Test Works with larger sample sizeCalculations easy to doOne
of the most popular statistical methods inscientific
literatureExtends to larger tables
220
Note on p-Values and Sample Size
Will the p-value change if we have smaller sample sizes
butproportions remain about the same?Suppose our sample size were
about 1/4 the original:
AZT Placebo
HIVtransmission
Yes 2 8 10
No 28 24 52
30 32 62
AZT 2/30 = 6.7%Placebo 8/32 = 25%
p = .083
-
221
Note
The p-value depends not only on the observed differencebetween
the proportions, but also on the sample sizes
If the sample sizes that two proportions were based on
werebigger, the p-value would get smaller
222
Relative Risk
Ratio of Proportions:
Relative risk = p1/p2
AZT Example
1 The risk of HIV with AZT relative to placebo:Relative risk=
p1/p2 = .074/.244 = .30
2 The risk of HIV with placebo relative to AZT:Relative risk=
p2/p1 = 3.29The risk of HIV transmission with placebo is more than
3times higher compared to AZT
223
3 You can testH0: Relative Risk= 1H1: Relative Risk6= 1Using any
of the methods for comparing proportions(2, Fishers)
224
The Relative Risk versus the p-Value
The relative risk tells you the magnitude of the
disease-exposureassociation.
The p-value (calculated using either Fishers exact test or the
2
statistic) tells you if the observed result can be explained
bychance.
A big relative risk does not necessarily mean that the p-value
issmall.
The p-value depends both on the magnitude of the relative risk
aswell as the sample size.
-
225
Describing the Association Between Two Continuous Variables
1 Scatter plot
2 Correlation coefficient
3 Simple linear regression
226
Association between body weight (X ) and plasma volume (Y )
Body Weight (kg) Plasma Volume (l)
1 58.0 2.752 70.0 2.863 74.0 3.374 63.5 2.765 62.0 2.626 70.5
3.497 71.0 3.058 66.0 3.12
227
The Scatter Plot
60 65 70
2.6
2.8
3.0
3.2
3.4
Body Weight (kg)
Pla
sma
Vol
ume
(l)
Scatter diagram of plasma volume andbody weight showing the
linear regression
line228
The Correlation Coefficient
Measures the direction and magnitude ofthe linear association
between X and Y
The correlation coefficient is between -1 and +1
r = 1 r = 0 r = 1
r = .7 r = .7
-
229
Examples of the Correlation Coefficient
Perfect Positive Uncorrelated
Weak Positive Weak Negative
230
Properties of Correlation Coefficient
1 Corr(X ,Y ) = r
2 1 r 1
r = 1: Perfect positive association r > 0: Positive
association r = 0: No association r < 0: Negative association r
= 1: Perfect negative association
3 Closer to 1 and 1: stronger relationship4 Sign: direction of
association
5 r = 0: no linear association
231
Correlation Slider:noppa5.pc.helsinki.fi/koe/corr/cor7.html
Correlation Guessing
Game:http://istics.net/stat/Correlations/
232
Correlation measures linear association
r = 0
A strong relationship along a curve for which r = 0
noppa5.pc.helsinki.fi/koe/corr/cor7.htmlhttp://istics.net/stat/Correlations/
-
233
Four Scatterplotsall have r = .7
Anscombes Data
234
NOTES AND CAVEATS ON CORRELATION COEFFICIENT
Measuring only linear relationships Other kinds of relationships
are also important Look at and graph the data Sensitive to outliers
X values are measured not controlled by the experimental
design. That is, X and Y are random
Example where r is appropriate
X = height Y = weight
Example where r is not appropriate
Clinical study at different dosesX = dose of drug Y =
Response
235
Body WeightBlood Plasma
60 65 70
2.6
2.8
3.0
3.2
3.4
Body Weight (kg)
Pla
sma
Vol
ume
(l)
r = .76
236
How Close Do the Points Fall to the Line?
This is measured by the correlation coefficient But what line?
This is measured by regression
-
237
Simple Linear Regression
Y is the dependent variable X is the independent variable
Predictor Regressor Covariate
We try to predict Y from X Called simple because there is only
one independent variableX
If there are several independent variables, its called
multiplelinear regression
238
Simple Linear Regression
Fit a straight line to the data
55 60 65 70 75
2.2
2.4
2.6
2.8