Sie sind auf Seite 1von 107

Chapter 4:

Summary Statistics

Note: In these notes, I will be using


M instead of X-bar to symbolize
sample mean.

Summarizing Data
Mean

N will refer to the


number of scores
N = 28
M = arithmetic mean
= sum; add up

2127
M
75.96
28

Summarizing Data
Ordered data
32
34
40
62
63
65
68
68
70
76
76
77
78
79

80
80
81
81
87
88
88
88
89
90
93
95
97
102

Summarizing Data
Median
32
34
40
62
63
65
68
68
70
76
76
77
78
79

80
80
81
81
87
88
88
88
89
90
93
95
97
102

Place scores in order from


lowest to highest
Can also list from highest to
lowest
Median (Mdn) = middle score
Half of scores are above the
median, half are below

Summarizing Data
Median
32
34
40
62
63
65
68
68
70
76
76
77
78
79

80
80
81
81
87
88
88
88
89
90
93
95
97
102

Since we have an even number


of scores (N = 28), there is no
one score that is in the middle
Here we have 2 middle scores:
79 and 80 (13 scores above, 13
scores below)
Take the mean of the 2 middle
scores

159
Mdn
79.5
2

Summarizing Data
Score

Frequency

32
34
40
62
63
65
68
70
76
77
78
79
80
81
87
88
89
90
93
95
97
102

1
1
1
1
1
1
2
1
2
1
1
1
2
2
1
3
1
1
1
1
1
1

Mode

Data are still ordered


from lowest to highest
Frequency column
shows how often each
score appears
Mode (Mo) is the score
which appears most
frequently
Mo = 88

Measures of Central Tendency


Mean (M) = X = 2127 = 75.95
N
28
Median (Mdn) = 79 + 80 = 159
= 79.5
2
2
Mode (Mo) = 88
All are accurately calculated, but give a slightly
different impression of where the middle is.
In a normal distribution, these three measures will
be the same (or very close)

Displaying Data -- Histogram


Graph of frequencies

3.0

Frequency

2.0

1.0

0.0

SCORE

Displaying Data -- Histogram


Graph of frequencies

(8 categories)

10

Frequency

0
35.0

45.0

55.0

65.0

75.0

SCORE

85.0

95.0

105.0

Displaying Data -- Histogram


Graph of frequencies

(16 categories)

Frequency

0
32.5

37.5

42.5

47.5

52.5

57.5

62.5

67.5

72.5

SCORE

77.5

82.5

87.5

92.5

97.5

102.5 107.5

Histogram
Graph of frequencies

(16 categories with normal curve)

Frequency

0
32.5

37.5

42.5

47.5

52.5

57.5

62.5

67.5

72.5

SCORE

77.5

82.5

87.5

92.5

97.5

102.5 107.5

Whats a Skewed Distribution?


Salaries of 3 companies
Company A

Company B

C ompany C

$10,000
$10,000
$12,000
$12,000
$12,000
$12,000
$12,000
$12,000
$14,000
$14,000

$10,000
$10,000
$12,000
$12,000
$12,000
$12,000
$12,000
$12,000
$14,000
$380,000

$10,000
$10,000
$12,000
$12,000
$12,000
$12,000
$12,000
$12,000
$15,000
$380,000

Note the dramatic effect of one salary on the mean, especially comparing A to B

Whats a Skewed Distribution?


Salaries of 3 companies
Company A

$10,000
$10,000
$12,000
$12,000
$12,000
$12,000
$12,000
$12,000
$14,000
$14,000

M
Mdn
Mo

$12,000
$12,000
$12,000

Company B

C ompany C

$10,000
$10,000
$10,000
$10,000
$12,000
$12,000
$12,000
$12,000
Means$12,000
for Company B$12,000
and C are
$12,000
pulled
in the direction$12,000
of the one
$12,000
$12,000
unusual score
$12,000
$12,000
$14,000
$15,000
$380,000
$380,000

$48,600
$12,000
$12,000

$48,700
$12,000
$12,000

Whats a Skewed Distribution?


Salaries of 3 companies
Company A

Company B

C ompany C

$10,000
$10,000
$10,000
$10,000
$10,000
$10,000
$12,000
$12,000
$12,000
$12,000
$12,000
$12,000
$12,000
$12,000
$12,000
$12,000 and Modes
$12,000
$12,000 the
But the Medians
remain unaffected
$12,000
$12,000
$12,000
impact of
a few unusually
high or low scores
is seen
$12,000
most clearly$12,000
on the MEAN $12,000
$14,000
$14,000
$15,000
$14,000
$380,000
$380,000

M
Mdn
Mo

$12,000
$12,000
$12,000

$48,600
$12,000
$12,000

Not skewed

Skewed

$48,700
$12,000
$12,000

Skewed

Examples of Shapes of Skewed


Distributions
Positively Skewed (aka skewed to the
right)
14

12

Frequency

10

2
0
95.0

105.0
100.0

115.0
110.0

125.0
120.0

135.0
130.0

145.0
140.0

155.0
150.0

IQ

Mode 105 (& 121)


121.1

Median 120.5

Mean

Positively Skewed Distribution


14

12

10

Std. Dev = 3.98

Mean = 7.3
N = 50.00

0
2.0

4.0

murder rate

6.0

8.0

10.0

12.0

14.0

16.0

18.0

20.0

Positively Skewed Distribution


30

20

10

Std. Dev = 1.58


Mean = 1.56
N = 124.00

0
0.00

1.00
.50

2.00

1.50

2.50

3.00

4.00

3.50

average daily alcohol use

5.00

4.50

6.00
5.50

7.00

6.50

7.50

8.00

Positively Skewed Distribution


300

Frequency

200

100

Mean gets
dragged to
the right

0
15.0

25.0
20.0

35.0
30.0

45.0
40.0

55.0
50.0

65.0
60.0

75.0
70.0

85.0
80.0

90.0

AGE

Mode 21
31.34

Median

25

Mean

Examples of Shapes of Skewed


Distributions
Negatively Skewed (aka skewed to the
left)
3.5

3.0

Frequency

2.5

2.0

1.5

1.0

.5
0.0
70.0

72.5

75.0

77.5

80.0

82.5

85.0

87.5

90.0

Score on exam

Mean 82.48
Mode 88.8

Median 84

Negatively Skewed Distribution


7

Frequency

1
0
2.81

3.06

3.31

2.94

3.19

3.56
3.44

3.81
3.69

3.94

Median

3.55

OVERALL COLLEGE GPA

Mean 3.51
Mode 3.71

Negatively Skewed Distribution


14

12

10

Mean gets
dragged
to the left

Frequency

2
0
3.19

3.31
3.25

3.44
3.38

3.56
3.50

3.69
3.63

3.81
3.75

3.94
3.88

4.00

MAJOR AREA GPA

Mean 3.78
Mode 3.98

Median

3.82

Skewed Distributions
Relative positions of measures of central
tendency

Salaries in 2 Companies

Means and Medians suggest that the employees at these 2 companies have similar pay

M = $60,000
Mdn = $60,000

M = $60,000
Mdn = $60,000

Salaries in 2 Companies
But looking at the distribution of pay, the shapes are
different
6.0

5.5
5.0

4.5
4

Frequency

Frequency

4.0
3.5
3.0
2.5

2.0
1.5

1.0
.5

0.0

M = $60,000
Mdn = $60,000

M = $60,000
Mdn = $60,000

Range
6.0

A rough measure of dispersion


(variability)
6

5.5
5.0

4.5
4

3.5

Frequency

Frequency

4.0

3.0
2.5
2.0

1.5
1.0

.5
0.0

R = Maximum Minimum
R = $110,000 $10,000
$40,000
R = $ 100,000

R = $ 80,000
R = $ 40,000

Comparing the ranges of these 2 companies suggests that


there is some difference between them.

Standard Deviation
A more precise measure of dispersion
Comparable to the mean in that it takes all of the
data points into account (not just highest and
lowest)
More commonly used than Range
Is basis for many later measures and tests well be
using
Standard = average
Deviation = away from normal

Standard Deviation
Scores (X)
3
4
5
6
7

We are going to
calculate a standard
deviation (SD) for
this set of 5 scores.

N
25

5
5

The mean is 5, so
how far away (on
average) do these 5
scores deviate from
the mean?

Standard Deviation
Scores (X)
3
4
5

X M (deviation score)

6
7

N
25

5
5

35

= -2

45

= -1

55

= 0

65

= 1

75

= 2

Subtracting
each score
from the
mean to get
the deviation
score

Standard Deviation
Scores (X)
3
4
5

X M (deviation score)

6
7

N
25

5
5

35

= -2

45

= -1

55

= 0

65

= 1

75

= 2

(X- M)
N

0
5

then taking
the average of
these 5
numbers -- !
Always results
in 0. The
positive &
negative
numbers
cancel each
other out to

Standard Deviation
Scores (X)
3

This method
wont work
positives &
negatives
always cancel
each other
out to zero!

4
5

X M (deviation score)

6
7

N
25

5
5

35

= -2

45

= -1

55

= 0

65

= 1

75

= 2

(X- M)
N

0
5

Need another
method to get
the information
we want.

Standard Deviation
Deviation method
(X M)2

Scores (X)
3
4

XM

(deviation score)

-2

-1

4
1
0
1
4

We square EACH deviation score. Why?


Because in the X M column, the
positives and negatives cancel each
other out (sum to zero).
What happens when we square these
deviation scores?

Standard Deviation
Deviation method
(X M)2

Scores (X)
3
4

XM

(deviation score)

-2

-1

4
1
0
1
4

We still get a measure of distance from


the mean (a squared version), and also
lose the pesky negatives ( since negative
x negative = positive).

Standard Deviation
Deviation method
(X M)2

Scores (X)
3
4

XM

(deviation score)

-2

-1

4
1
0
1
4

s2

Add up these 5 numbers (SS is the sum


of squares)
and then
divide by (n 1)

SS
n- 1
10

Standard Deviation
Deviation method
(X M)2

Scores (X)
3
4

XM

(deviation score)

-2

-1

4
1
0
1
4

s2

NOW when we find the (near)average of


THESE 5 scores, we get a meaningful
number an expected squared distance
from the mean.

SS
n- 1
10

4
2.5

Standard Deviation
Deviation method
(X M)2

Scores (X)
3
4

XM

(deviation score)

-2

-1

4
1
0
1
4

s2 2.5
Taking the square root of the
variance gets this measure back
into the original scale.

2.5 1.58

Standard Deviation
Deviation method
Scores (X)
3
4
5
6
7

The standard deviation for this set of


numbers is 1.58
It is a measure of how much variation
(departure from the mean) these numbers
have
A larger number tells us that the
numbers are generally spread out (i.e., far
away from each other)
A smaller number tells us that many of
the numbers are quite close to the mean.

Standard Deviation
Computational method for SS
The deviation method (the last few pages) will give you
the standard deviation, and it is easier to see where that
end number comes from (a measure of how far from the
mean the numbers are).
However, the deviation method can be a bit tedious in
hand calculations, especially with certain types of raw
data.
We have another more direct method (computational
2 hand
method) of getting to the SS when using
X
2
calculations.
SS X

Standard Deviation
Computational method
Scores (X)

SS

3
4

X2
9

16

25

49

N
25

5
5

SS 10
s

SS
n 1

10
1.58
4

X2 = 135

NOTICE: this is NOT 252

25
SS 135

36

Standard Deviation
Second example deviation method

Scores (X)
10
12
13
15
18
20
22
25

Another dataset,
showing you how to
calculate standard
deviation using both
methods.

Standard Deviation
Second example deviation method

Scores (X)
10
12
13
15
18
20
22
25

M = 135 / 8
= 16.875

XM
-6.875
-4.875
-3.875
-1.875
1.125
3.125
5.125
8.125

(X
M)2
47.27
23.77
15.02
3.52
1.27
SS
s2 9.77
n- 1
192.91
26.27

7
66.02
27.56

Standard Deviation
Second example deviation method

Scores (X)
10
12
13
15
18
20
22
25

M = 135 / 8
= 16.875

XM
-6.875
-4.875
-3.875
-1.875
1.125
3.125
5.125
8.125

(X
M)2
47.27
23.77
15.02
3.52
1.27
9.77
s2 27.56
26.27

s 27.56 5.25
66.02

Standard Deviation
Second example computation method

Scores (X)
10
12
13
15
18
20
22
25

M = 135 / 8
= 16.875

X2
100
144
169
225
324
400
434
625

SS

135

SS 2471

SS 192.875
s

SS
n 1

X2 = 2471

192.875
5.249
7

Standard Deviation
Third example deviation method
Scores
(X)
6
2
8
5
4
4
7/7
M = 36
= 5.143

XM
.857
-3.14
2.86
-.143
-1.143
-1.143
1.857

(X
M)2
.735
9.878
8.16
.020
1.306
1.306
3.449
SS
2
s
n- 1
24.857

6
4.143

Standard Deviation
Third example deviation method
Scores
(X)
6
2
8
5
4
4
7/7
M = 36
= 5.143

XM
.857
-3.14
2.86
-.143
-1.143
-1.143
1.857

(X
M)2
.735
9.878
8.16
.020
1.306
1.306
3.449
s2 4.143

s 4.143 2.035

Standard Deviation
Third example computation method
Scores
(X)
6
2
8
5
4
4
7
M = 36 / 7
= 5.143

X2
36
4
64
25
16
16
49

SS
SS

36
210

SS 24.857
s

X2 = 210

SS
n 1

24.857
2.035
6

Kurtosis
Another check on normality
40

Frequency

30

20

10

0
16.0

18.0

20.0

22.0

24.0

26.0

28.0

30.0

32.0

Distance (mm) from center of pituitary to pteryo-maxillary fissure


Does the distribution
have the right shape to be
normal?
This one fits pretty well under the normal curve
(Mesokurtic)

Kurtosis
10

Frequency

0
1.5

2.5

3.5

4.5

5.5

6.5

7.5

8.5

9.5

Score

This one is flatter in the center, and doesnt fit


under the corners
This one would be considered platykurtic

Kurtosis
55
50
45
40

Frequency

35
30
25
20
15
10
5
0
23.00

23.50

24.00

24.50

25.00

Score

This one is higher in the center, and doesnt fill out


the shoulders
This one would be considered leptokurtic

A quick check on kurtosis


In a normal curve, the standard deviation is
approximately 1/6 of the size of the range
So in order to check if a distribution fits this ratio, you
take the SD and divide by the range
IF the distribution is normal, you should get a result of
approximately .17 (somewhere between .15 and .19)
Normal:

s
.17
R
Anything LARGER than that (such as .38 or .47) would
indicate a platykurtic distribution one that has MORE
variability than normal
Anything SMALLER than that (such as .11 or .09) would
suggest that the data are more closely packed or

Assessing Kurtosis
In a normal curve,
the standard
deviation is
approximately 1/6
of the range

40

Frequency

30

s
.17
R

20

10

In this example:
0
16.0

18.0

20.0

22.0

24.0

26.0

28.0

30.0

32.0

Distance (mm) from center of pituitary to pteryo-maxillary fissure

s
2.93

.19
R
15
Verdict: Mesokurtic

Assessing Kurtosis
Normal:

10

s
.17
R

Frequency

Inthisexample
:
s
2.42

.30
R
8

0
1.5

2.5

3.5

4.5

5.5

Score

6.5

7.5

8.5

9.5

Verdict: Platykurtic

Assessing Kurtosis
Normal:

50

s
.17
R

Frequency

40

30

20

Inthisexample
:

10

s
.226

.11
R
2

0
22.95

23.55

24.15

24.75

25.35

Score

Verdict:
Leptokurtic

Is this distribution normal?


Skewness & Kurtosis

40
45
50
60
60
65
70
74
75
75

77
80
80
80
85
85
85
90
90
90

93
93
93
93
94
94
95
100
100
105

105
105
105
110
120
127
130
137
140
180

X = 3675
X2 = 365971

Is this distribution normal?


Skewness
18

M = X /N
15

=
3675/40

12

=
91.875

Mdn = 91.5

0
40.0

60.0

80.0

100.0

120.0

140.0

160.0

180.0

Mo = 93 &
105

Conclusion: Not (very) skewed

Is this distribution normal?


Skewness
2

3675
SS 365971

18

40

15

SS 28330.375
12

0
40.0

60.0

80.0

100.0

120.0

140.0

160.0

180.0

28330.375
26.952
39
s
26.952

.193
R
140

Conclusion: Mesokurtic

Chapter 5:
Probability Concepts
and Screening

Definitions
Random variable: a numerical quantity that takes on
different values depending on chance
Population: the set of all possible values for a
random variable
Event: an outcome or set of outcomes
Probability: the proportion of times an event is
expected to occur in the population
Ideas about probability are founded on relative
frequencies (proportions) in populations.

Types of random variables


Discrete random variables: a countable set of possible
outcome (e.g., the number of cases in an SRS from the
population)
Binomial: a family of discrete random variables
Binomial random variable: the random number of
successes in n independent Bernoulli trials (a
Bernoulli trial has two possible outcomes: success
or failure)
Binomials random variables have two parameters
n number of trials
p probability of success of each trial
Continuous random variable: an unbroken continuum of
possible outcome (e.g., the average weight of newborns

Classical Model
a priori, or before the fact predictions, ideas about
what will happen using reason alone
(Based on conditions that all possible outcomes are
equally likely and that only 1 can occur)
probability = # of specific events out of total
possible events:
p = s/t
This will give us a number between 0 and 1
0 ------------------.50------------------1.00
no chance
certainty
of occurring
of occurring

Classical/Theoretical Model
Tossing a coin or rolling dice are good examples
In flipping a coin:
Head is just as likely as tails
Can only have 1 outcome (cant be both H & T)
In rolling a die:
Getting a 3 is just as likely as getting a 5
Can only have 1 outcome

Classical/Theoretical Model
What is the probability of flipping a coin and getting
a Head as the outcome?
Number of specific outcomes that satisfy this
condition
1 (getting a Head)
How many total possible outcomes are there?
2 (H or T)
p = s / t = 1 / 2 = .50
Classical model predicts a 50% chance of getting a
Head

Classical/Theoretical Model
What is the probability of rolling a die and getting a 3?
How many specific outcomes satisfy this condition?
1 (i.e., tossing a 3)
How many total possible outcomes are there?
6
p=

s/t

= 1/6

.167

Classical model predicts a 17% chance of getting a 3


on any given toss of a die

Classical Model
What is the probability of tossing a die and getting
an even number as an outcome?
How many specific outcomes satisfy this condition?
3 (i.e., getting a 2, 4 or 6 mutually exclusive
outcomes)
How many total possible outcomes are there?
6
p =

s/t

= 3/6

.50

Classical model predicts a 50% chance of getting an


even number on any given toss of a die

Long-run/Empirical model
a posteriori viewpoint, after the fact, data have
already been collected
Rather than looking at what SHOULD occur, this is
what HAS actually occurred over a large number of
events
Can compare this number to the expected outcomes
as predicted by the classical model
Long-run because the more data you have, the
more accurate the results are going to be

Long-run/Empirical model
Coin flip game: a coin will be flipped
Heads = Ill give you $5
Tails = you give me $5
So youd be hoping for heads
With multiple tosses, you might generally expect to
break even, or maybe one person would be $5 ahead
of the other

Long-run/Empirical model
On 10 flips, the outcome is:
5 tails (sounds good)
6 tails (I got a bit lucky)
7 tails (more lucky)
8 tails (hm)
9 tails (???)
At what point do you (should you) become
suspicious?

Long-run/Empirical model
Same scenario:
Heads = Ill give you $5
Tails = you give me $5
BUT instead of there being 10 tosses, there will
be 100

Long-run/Empirical model
On 100 flips, the outcome is:
50 tails
60 tails
70 tails
80 tails
90 tails
Percentages are the same as with 10 tosses, but now
its based on a larger number
60% based on 100 is a lot more suspicious than
60% based on 10

Connection to real data


Are outcomes different?
Fair/biased coin = control/experimental groups
E.g., Are the results we got with an
experimental treatment that much different
from the control group?
Different enough to be suspicious?

Connection to the normal curve


Weve talked a lot about areas and percentages
under different parts of the normal curve
These ideas can be directly translated into
statements of probability
What is the probability that any individual will
have an IQ score greater than 100?
If 50% of people have scores in this range, then
stating this as a probability would be:
p = .50

Connection to the normal curve


Converting percentages to probability
statements is simple
Since a percentage is a proportion out of 100,
simply divide the percentage by 100
50 %

= > 50 / 100

=> .50

The z table already gives the information in the


form of probability

Rules and Properties of Probability


Conditional Probability
Conditional Probability
Let Pr(B|A) represent the conditional probability
of B given A.
This denotes the probability of B given that A is
evident. By definition,

Pr(AandB)
Pr(B| A)
Pr(A)
(as long as Pr(A) > 0)

Rules and Properties of Probability


Multiplication
General Rule for Multiplication
Start with the definition of conditional probability:
Pr(AandB)
Pr(B| A)
Pr(A)

then rearrange the formula as follows:

Pr A andB Pr A Pr B| A

Rules and Properties of


Probability
Independent Events
Events A and B are independent if and only if
Pr(A and B) = Pr(A) Pr(B)
General Rule of Addition:
Pr(A or B) = Pr(A) + Pr(B) - Pr(A and B)

Combining Probabilities

1 outcome
Finding the probability that ONE of several
acceptable outcomes may occur
For example, what is the probability that you will
get a Head OR a Tail when you flip a coin?
In this type of situation, you add the
probabilities together:
p = Head OR Tail
p = .50
+
.50
= 1.00

Combining Probabilities
1 outcome
p = Head
p = .50

OR
+

Tail
.50

1.00

In this case, weve expanded the range of


acceptable outcomes from just Head Head
OR Tail
This makes it more likely that we will have an
outcome that fits our criteria
Rule: Add probabilities for Or statements (if
independent and mutually exclusive)

Combining Probabilities
1 outcome
Finding the probability that one of several
acceptable outcomes may occur
Outcomes must be all mutually exclusive
What is the probability of rolling a die and getting
an even number OR a 3?

ADD-OR rule
p = even number OR
=
.50
= .667

.167

Combining Probabilities
Multiple outcomes
Finding the probability that multiple events will
occur
For example, what is the probability that you will
flip 2 coins and get 2 Heads?
In this type of situation, you multiply the
probabilities:
p = Head AND Head
p = .50
x
.50
= .25

Combining Probabilities
Multiple outcomes
p = Head
p = .50

AND
x

Head
.50

.25

In this case, weve made it more difficult to meet the


criteria than if we had just specified 1 Head on 1 coin
flip
This makes it less likely that we will have an
outcome that fits our criteria
Rule: Multiply probabilities for And statements
applied to multiple outcomes (independent and
mutually exclusive)

Combining Probabilities
Multiple outcomes
You will draw one card from a deck, replace that card
and draw a second card from the deck (the
replacement will make the two draws independent)
What is the probability that you will get an Ace on
the first draw AND and an Ace on the second draw?
MULT-AND rule:
p = (Ace) AND
= (4/52)
X
= .077
X

(Ace)
(4/52)
.077 =

.006

Combining probabilities
When you have one outcome desired, and more
than one way to achieve that outcome, it makes it
easier to achieve (higher probability)
When you have multiple outcomes that need to
occur together, it makes it more difficult to achieve
than any one of the outcomes alone

Applying probability
Vital statistics rates
We can combine these rules and properties to
determine certain information of interest
Mortality rate: standard way to compare death rate
across different circumstances
Numerator: # of people who died during a given
period of time
Denominator: # of people who were at risk of
dying during the same period
Denominator may be difficult to calculate, so the
number of people alive in the population halfway
through the time period is sometimes used as an
estimate

Applying probability
Vital statistics rates
Morbidity rate: calculated like mortality, but
conveys the rate disease of disease in relation to
the population
Prevalence: number of individuals with the
disease at one point in time divided by the
population at risk at that time
Incidence: number of new cases during a given
time span divided by the population at risk (at
the beginning of the interval)

Applying probability
Vital statistics rates
NOTE:
(Because prevalence does not involve a period of
time, it is actually a proportion, but is often
mistakenly termed a rate.)
The term incidence is sometimes used
erroneously when the term prevalence is meant.
One way to distinguish between them is to look for
units: An incidence rate should always be
expressed in terms of a unit of time.

Applying probability
Vital statistics rates
Prevalence and incidence can be compared to crosssectional and longitudinal studies
Prevalence is like a snapshot (like a cross-sectional
study); may see cross-sectional studies referred to as
prevalence studies
Incidence needs a period of time to pass, like cohort
studies (begin at a given time and continue to
examine outcomes over the specific span of the
study)

Applying probability
Screening
Screening used to distinguish those who are apparently
well from those who have a decently high probability of
having the disease or condition under study (with the goal
of further testing and/or tx)
Generally employed when:
Target disease is serious enough to warrant
Test is proven and acceptable to detect the disease
early enough for intervention
There is tx available
Two probabilities are used to measure the ability of a test
to distinguish between those who have the disease and
those who do not
Compare the screening results to definitive diagnosis
results

Applying probability
Screening
Sensitivity: does the test return a positive result
on those who actually have the disease?
(Missing out on finding people who have it would
reduce the sensitivity)
Sensitivit
y

#ofpeoplewhotested
()atscreening
x 100
total#ofpeoplescreened
whohavethedisease

Specificity: does the test return a negative result


on those who are free of the disease? (False
negatives would reduce the specificity)
Specificit
y

#ofpeoplewithoutthedisease
whoscreened
(-)
X 100
total#ofpeoplescreened
whoarewithoutthedisease

Applying probability
Screening

Screening
Higher cutoff (27)

Screening
Higher cutoff

Hits: true positive screenings

Screening
Higher cutoff

True negatives

Screening
Higher cutoff
Some false negatives (misses) and
some correct negatives

Screening
Higher cutoff

Misses: false negatives

Screening
Lower cutoff (22)

Screening
Lower cutoff (22)

Hits: true positive screenings

Screening
Lower cutoff

True negatives

Screening
Lower cutoff
Some hits and
some false positives

Screening
Lower cutoff
False positives

Applying probability
Screening
False positives with low cutoff

False negatives with high cutoff

Applying probability
Screening

To detect all glaucomatous eyes (i.e., 100% sensitivity), cutoff must be


at 22 mm. But this will also result in nonglaucomatous eyes at the high
end to be included (specificity drops below 100%). As long as there is
overlap in (+) and (-) populations with regard to the screening
characteristic, both sensitivity AND specificity can never be 100%

Applying probability
Screening
Which is worse? False positive or false negative?
Where should we err in putting the cutoff?
Depends:
Is the disease rare? Then high sensitivity is valuable
Is the disease silent for a while before symptoms?
Is the disease lethal?
Is there effective treatment available?
Will see a parallel between these questions, and
those relating to Type I/II errors and power (related to
hypothesis/ significance testing)

Applying probability
Predictive value (positive)
Predictive value of a test: the probability of
disease given a positive result, the chance that a
patient with a positive test has the disease
This is a conditional probability in which the
event of the disease being present is dependent
(i.e., conditional) on having a positive test result

P(T
|
D
)
P(D
)

P(D | T )
P(T | D ) P(D ) P(T | D ) P(D )

Applying probability
Predictive value (positive)
The probability that a test is
positive, given that the disease is
present (i.e., the sensitivity of
the test)

The best guess (or prior


probability) that the
patient has the disease to
begin with (i.e., the
prevalence of the disease)

P(T
|
D
)
P(D
)

P(D | T )

P(T | D ) P(D ) P(T | D ) P(D )

The denominator. The probability that a positive test occurs at all,


P(T +),
which can occur in one of two ways:
a positive test when the disease is present, and
a positive test when the disease is not present
(each weighted by the prior probability of that outcome)

Applying probability
Predictive value (negative)
Predictive value of a test: the probability of
absence of disease given a negative result: the
chance that a patient with a negative test is
free of the disease
This is a conditional probability in which the
event of the disease being present is dependent
(i.e., conditional) on having a negative test

result
P(T
|
D
)
P(D
)

P(D | T )

P(T | D ) P(D ) P(T | D ) P(D )

Applying probability
Predictive value (negative)

The probability that a test is


negative, given that the disease
is not present (i.e., the specificity
of the test)

The best guess (or prior


probability) that the
patient does not have the
disease to begin with (i.e.,
the non-prevalence of the
disease)

P(T | D ) P(D )
P(D | T )

P(T | D ) P(D ) P(T | D ) P(D )

The denominator. The probability that a negative test result


occurs at all, P(T -),
which can occur in one of two ways:
a negative test when the disease is not present, and
a negative test when the disease is present
(each weighted by the prior probability of that outcome)

Applying probability
Screening exercise
A newly developed test produced positive results in 138 of 150
known diabetics and in 24 of 150 persons known not to have
diabetes.
a.

What is the sensitivity of the new test?

b.

What is the specificity of the new test?

c.

What is the false-positive rate of the new test?

d.

Suppose a fasting blood sugar is obtained with known


sensitivity and specificity of 0.80 and 0.96, respectively. If this
test is applied to the same group that the new test used (150
persons with diabetes and 150 persons without diabetes),
what is the predictive validity of a positive result on the blood
sugar screening?

Das könnte Ihnen auch gefallen