Statistics in the Clinical Laboratory: Keys to Understanding
Download
Report
Transcript Statistics in the Clinical Laboratory: Keys to Understanding
Practical Applications of
Statistical Methods in
the Clinical Laboratory
Roger L. Bertholf, Ph.D., DABCC
Associate Professor of Pathology
Director of Clinical Chemistry & Toxicology
UF Health Science Center/Jacksonville
“[Statistics are] the only
tools by which an opening
can be cut through the
formidable thicket of
difficulties that bars the path
of those who pursue the
Science of Man.”
[Sir] Francis Galton (1822-1911)
“There are three kinds of
lies: Lies, damned lies,
and statistics”
Benjamin Disraeli (1804-1881)
What are statistics, and what
are they used for?
• Descriptive statistics are used to
characterize data
• Statistical analysis is used to distinguish
between random and meaningful
variations
• In the laboratory, we use statistics to
monitor and verify method performance,
and interpret the results of clinical
laboratory tests
“Do not worry about your
difficulties in mathematics, I
assure you that mine are
greater”
Albert Einstein (1879-1955)
“I don't believe in
mathematics”
Albert Einstein
Summation function
N
x x x
i 1
i
1
2
x3 x N
Product function
N
x x x
i
i 1
1
2
x3 x N
The Mean (average)
The mean is a measure of the centrality of
a set of data.
Mean (arithmetical)
N
1
x xi
N i 1
Mean (geometric)
N
x g N x1 x2 x3 x N N xi
i 1
Use of the Geometric mean:
The geometric mean is primarily used to
average ratios or rates of change.
Mean (harmonic)
N
N
xh
N
1 1 1
1
1
x1 x2 x3
x N i 1 xi
Example of the use of
Harmonic mean:
Suppose you spend $6 on pills costing 30
cents per dozen, and $6 on pills costing
20 cents per dozen. What was the
average price of the pills you bought?
Example of the use of
Harmonic mean:
You spent $12 on 50 dozen pills, so the
average cost is 12/50=0.24, or 24 cents.
This also happens to be the harmonic
mean of 20 and 30:
2
1 1
30 20
24
Root mean square (RMS)
xrms
x x x x
1
2
xi
N
N i 1
2
1
2
2
2
3
2
N
N
For the data set:
1, 2, 3, 4, 5, 6, 7, 8, 9, 10:
Arithmetic mean
5.50
Geometric mean
4.53
Harmonic mean
3.41
Root mean square
6.20
The Weighted Mean
N
xw
x w
i
i
i 1
N
w
i
i 1
Other measures of centrality
• Mode
The Mode
The mode is the value that occurs most
often
Other measures of centrality
• Mode
• Midrange
The Midrange
The midrange is the mean of the highest
and lowest values
Other measures of centrality
• Mode
• Midrange
• Median
The Median
The median is the value for which half of
the remaining values are above and half
are below it. I.e., in an ordered array of
15 values, the 8th value is the median.
If the array has 16 values, the median is
the mean of the 8th and 9th values.
Example of the use of median
vs. mean:
Suppose you’re thinking about building a
house in a certain neighborhood, and the
real estate agent tells you that the average
(mean) size house in that area is 2,500 sq.
ft. Astutely, you ask “What’s the median
size?” The agent replies “1,800 sq. ft.”
What does this tell you about the sizes of the
houses in the neighborhood?
Measuring variance
Two sets of data may have similar means,
but otherwise be very dissimilar. For
example, males and females have
similar baseline LH concentrations, but
there is much wider variation in females.
How do we express quantitatively the
amount of variation in a data set?
N
1
Mean difference (xi x )
N i 1
N
N
1
1
xi x
N i 1
N i 1
xx
0
The Variance
N
1
2
V ( xi x )
N i 1
The Variance
The variance is the mean of the squared
differences between individual data points
and the mean of the array.
Or, after simplifying, the mean of the squares
minus the squared mean.
The Variance
1
V
N
N
i 1
( xi x ) 2
1
N
1
x N
1
N
1
1 2
x N 2 x xi N x (1)
2
i
1
2 xi x N
2
i
x 2 2x 2 x 2
x2 x 2
2
x
The Variance
In what units is the variance?
Is that a problem?
The Standard Deviation
V
N
1
2
( xi x )
N i 1
The Standard Deviation
The standard deviation is the square root
of the variance. Standard deviation is
not the mean difference between
individual data points and the mean of
the array.
1
1
2
xx
(x x)
N
N
The Standard Deviation
In what units is the standard deviation?
Is that a problem?
The Coefficient of Variation*
CV 100
x
*Sometimes
called the Relative Standard
Deviation (RSD or %RSD)
Standard Deviation (or Error)
of the Mean
x
x
N
The standard deviation of an average
decreases by the reciprocal of the
square root of the number of data
points used to calculate the average.
Exercises
How many measurements must we
average to improve our precision by a
factor of 2?
Answer
To improve precision by a factor of 2:
1
1
05
.
2
N
1
N
2
05
.
2
N 2 4 (quadruplicate)
Exercises
• How many measurements must we
average to improve our precision by a
factor of 2?
• How many to improve our precision by a
factor of 10?
Answer
To improve precision by a factor of 10:
1
1
01
.
10
N
1
N
10
01
.
2
N 10 100 times!
Exercises
• How many measurements must we
average to improve our precision by a
factor of 2?
• How many to improve our precision by a
factor of 10?
• If an assay has a CV of 7%, and we
decide run samples in duplicate and
average the measurements, what
should the resulting CV be?
Answer
Improvement in CV by running duplicates:
CVdup
CV
7
4.9%
2 141
.
Population vs. Sample
standard deviation
• When we speak of a population, we’re
referring to the entire data set, which
will have a mean :
1
Population mean xi
N i
Population vs. Sample
standard deviation
• When we speak of a population, we’re
referring to the entire data set, which will
have a mean
• When we speak of a sample, we’re
referring to a subset of the population,
customarily designated “x-bar”
• Which is used to calculate the standard
deviation?
“Sir, I have found you an
argument. I am not
obliged to find you an
understanding.”
Samuel Johnson (1709-1784)
Population vs. Sample
standard deviation
1
2
(xi )
N i
s
1
2
(xi x )
N 1 i
Distributions
• Definition
Statistical (probability)
Distribution
• A statistical distribution is a
mathematically-derived probability
function that can be used to predict the
characteristics of certain applicable real
populations
• Statistical methods based on probability
distributions are parametric, since
certain assumptions are made about the
data
Distributions
• Definition
• Examples
Binomial distribution
The binomial distribution applies to events
that have two possible outcomes. The
probability of r successes in n attempts,
when the probability of success in any
individual attempt is p, is given by:
P(r; p, n) p (1 p)
r
n r
n!
r !(n r )!
Example
What is the probability that 10 of the 12
babies born one busy evening in your
hospital will be girls?
Solution
P(10;05
. ,12) 05
. (1 05
.)
10
12 10
0016
.
or 16%
.
12!
10!(12 10)!
Distributions
• Definition
• Examples
– Binomial
“God does arithmetic”
Karl Friedrich Gauss (1777-1855)
The Gaussian Distribution
What is the Gaussian distribution?
63
81
36
12
28
7
79
52
96
17
22
4
61
85 etc.
F
1
number
100
63
81
36
12
28
7
79
52
96
17
22
4
61
85
+
22
73
54
33
99
5
61
28
58
24
16
77
43
8
=
85
152
90
45
127
12
140
70
154
41
38
81
104
93
F
2
number
200
. . . etc.
x
Probability
The Gaussian Probability
Function
The probability of x in a Gaussian
distribution with mean and standard
deviation is given by:
1
( x ) 2 / 2 2
P ( x; , )
e
2
The Gaussian Distribution
• What is the Gaussian distribution?
• What types of data fit a Gaussian
distribution?
“Like the ski resort full of
girls hunting for husbands
and husbands hunting for
girls, the situation is
not as symmetrical as it
might seem.”
Alan Lindsay Mackay (1926- )
Are these Gaussian?
•
•
•
•
•
•
•
Human height
Outside temperature
Raindrop size
Blood glucose concentration
Serum CK activity
QC results
Proficiency results
The Gaussian Distribution
• What is the Gaussian distribution?
• What types of data fit a Gaussian
distribution?
• What is the advantage of using a
Gaussian distribution?
Probability
Gaussian probability
distribution
.67
.95
µ-3 µ-2 µ- µ µ+ µ+2 µ+3
What are the odds of an
observation . . .
• more than 1 from the mean (+/-)
• more than 2 greater than the mean
• more than 3 from the mean
Some useful Gaussian
probabilities
Range Probability Odds
68.3%
1 in 3
+/- 1.00
90.0%
1 in 10
+/- 1.64
95.0%
1 in 20
+/- 1.96
99.0%
1 in 100
+/- 2.58
That
Example
This
[On the Gaussian curve]
“Experimentalists think that
it is a mathematical theorem
while the mathematicians
believe it to be
an experimental fact.”
Gabriel Lippman (1845-1921)
Distributions
• Definition
• Examples
– Binomial
– Gaussian
"Life is good for only two
things, discovering
mathematics and
teaching mathematics"
Siméon Poisson (1781-1840)
The Poisson Distribution
The Poisson distribution predicts the
frequency of r events occurring
randomly in time, when the expected
frequency is
e
P ( r; )
r!
r
Examples of events described
by a Poisson distribution
?
• Lightning
• Accidents
• Laboratory?
A very useful property of the
Poisson distribution
V( r )
Using the Poisson distribution
How many counts must be collected in an
RIA in order to ensure an analytical CV
of 5% or less?
Answer
Since CV (100) (100)
x
and
0.05
400 counts
Distributions
• Definition
• Examples
– Binomial
– Gaussian
– Poisson
The Student’s t Distribution
When a small sample is selected from a
large population, we sometimes have to
make certain assumptions in order to
apply statistical methods
Questions about our sample
• Is the mean of our sample, x bar, the same as
the mean of the population, ?
• Is the standard deviation of our sample, s, the
same as the standard deviation for the
population, ?
• Unless we can answer both of these
questions affirmatively, we don’t know
whether our sample has the same distribution
as the population from which it was drawn.
Recall that the Gaussian distribution is
defined by the probability function:
1
( x ) 2 / 2 2
P ( x; , )
e
2
Note that the exponential factor contains both
and , both population parameters. The
factor is often simplified by making the
substitution:
z
(x )
The variable z in the equation:
z
(x )
is distributed according to a unit
gaussian, since it has a mean of
zero and a standard deviation of 1
Probability
Gaussian probability
distribution
.67
.95
-3
-2
-1
0
z
1
2
3
But if we use the sample mean and
standard deviation instead, we get:
(x x )
t
s
and we’ve defined a new quantity, t,
which is not distributed according
to the unit Gaussian. It is
distributed according to the
Student’s t distribution.
Important features of the
Student’s t distribution
• Use of the t statistic assumes that the
parent distribution is Gaussian
• The degree to which the t distribution
approximates a gaussian distribution
depends on N (the degrees of freedom)
• As N gets larger (above 30 or so), the
differences between t and z become
negligible
Application of Student’s t
distribution to a sample mean
The Student’s t statistic can also be
used to analyze differences
between the sample mean and the
population mean:
(x )
t
s
N
Comparison of Student’s t and
Gaussian distributions
Note that, for a sufficiently large N (>30), t
can be replaced with z, and a Gaussian
distribution can be assumed
Exercise
The mean age of the 20 participants in one
workshop is 27 years, with a standard
deviation of 4 years. Next door, another
workshop has 16 participants with a
mean age of 29 years and standard
deviation of 6 years.
Is the second workshop attracting older
technologists?
Preliminary analysis
• Is the population Gaussian?
• Can we use a Gaussian distribution for
our sample?
• What statistic should we calculate?
Solution
First, calculate the t statistic for the
two means:
t
( x1 x 2 )
s1 s 2
N1 N 2
(29 27)
2
2
6
4
20 16
1.19
( x1 x 2 )
2
1
2
2
s
s
N1 N 2
Solution, cont.
Next, determine the degrees of freedom:
N df N 1 N 2 2
16 20 2
34
Statistical Tables
df
t0.050
t0.025
t0.010
-
-
-
-
34
1.645
1.960
2.326
-
-
-
-
Conclusion
Since 1.16 is less than 1.64 (the t value
corresponding to 90% confidence limit),
the difference between the mean ages
for the participants in the two workshops
is not significant
The Paired t Test
Suppose we are comparing two sets of
data in which each value in one set has
a corresponding value in the other.
Instead of calculating the difference
between the means of the two sets, we
can calculate the mean difference
between data pairs.
(x1 x2 )
Instead of:
N
we use:
1
( x1 x2 ) ( x1i x2i )
N i 1
to calculate t:
(x1 x2 )
t
2
sd
N
Advantage of the Paired t
If the type of data permit paired analysis,
the paired t test is much more sensitive
than the unpaired t.
Why?
Applications of the Paired t
• Method correlation
• Comparison of therapies
Distributions
• Definition
• Examples
– Binomial
– Gaussian
– Poisson
– Student’s t
The 2 (Chi-square)
Distribution
There is a general formula that relates
actual measurements to their predicted
values
N
2
i 1
[ yi f (xi )]
2
2
i
The 2 (Chi-square)
Distribution
A special (and very useful) application of
the 2 distribution is to frequency data
( ni f i )
fi
i 1
N
2
2
Exercise
In your hospital, you have had 83 cases
of iatrogenic strep infection in your last
725 patients. St. Elsewhere, across
town, reports 35 cases of strep in their
last 416 patients.
Do you need to review your infection
control policies?
Analysis
If your infection control policy is roughly
as effective as St. Elsewhere’s, we
would expect that the rates of strep
infection for the two hospitals would be
similar. The expected frequency, then
would be the average
83 35
118
01034
.
725 416 1141
Calculating 2
First, calculate the expected frequencies at
your hospital (f1) and St. Elsewhere (f2)
f 1 725 01034
.
75 cases
f 2 416 01034
.
43 cases
Calculating 2
Next, we sum the squared differences
between actual and expected
frequencies
(ni f i )
fi
i
2
2
(83 75) 2 (35 43) 2
75
43
2.34
Degrees of freedom
In general, when comparing k sample
proportions, the degrees of freedom for
2 analysis are k - 1. Hence, for our
problem, there is 1 degree of freedom.
Conclusion
A table of 2 values lists 3.841 as the 2
corresponding to a probability of 0.05.
So the variation (2between strep
infection rates at the two hospitals is
within statistically-predicted limits, and
therefore is not significant.
Distributions
• Definition
• Examples
– Binomial
– Gaussian
– Poisson
– Student’s t
– 2
The F distribution
• The F distribution predicts the expected
differences between the variances of
two samples
• This distribution has also been called
Snedecor’s F distribution, Fisher
distribution, and variance ratio
distribution
The F distribution
The F statistic is simply the ratio of two
variances
V1
F
V2
(by convention, the larger V is the
numerator)
Applications of the F
distribution
There are several ways the F distribution
can be used. Applications of the F
statistic are part of a more general type
of statistical analysis called analysis of
variance (ANOVA). We’ll see more
about ANOVA later.
Example
You’re asked to do a “quick and dirty”
correlation between three whole blood
glucose analyzers. You prick your
finger and measure your blood glucose
four times on each of the analyzers.
Are the results equivalent?
Data
Analyzer 1
Analyzer 2
Analyzer 3
71
90
72
75
80
77
65
86
76
69
84
79
Analysis
The mean glucose concentrations for the
three analyzers are 70, 85, and 76.
If the three analyzers are equivalent, then
we can assume that all of the results are
drawn from a overall population with
mean and variance 2.
Analysis, cont.
Approximate by calculating the mean of
the means:
70 85 76
77
3
Analysis, cont.
Calculate the variance of the means:
(70 77) (85 77) (76 77)
Vx
3
38
2
2
2
Analysis, cont.
But what we really want is the variance of
the population. Recall that:
x
N
Analysis, cont.
Since we just calculated
Vx 38
2
x
we can solve for
Vx
N
N
2
2
2
x
N 4 38 152
2
2
x
Analysis, cont.
So we now have an estimate of the
population variance, which we’d like to
compare to the real variance to see
whether they differ. But what is the real
variance?
We don’t know, but we can calculate the
variance based on our individual
measurements.
Analysis, cont.
If all the data were drawn from a larger
population, we can assume that the
variances are the same, and we can simply
average the variances for the three data
sets.
V1 V2 V3
14.4
3
Analysis, cont.
Now calculate the F statistic:
152
F
10.6
14.4
Conclusion
A table of F values indicates that 4.26 is the
limit for the F statistic at a 95% confidence
level (when the appropriate degrees of
freedom are selected). Our value of 10.6
exceeds that, so we conclude that there is
significant variation between the
analyzers.
Distributions
• Definition
• Examples
– Binomial
– Gaussian
– Poisson
– Student’s t
– 2
–F
Unknown or irregular
distribution
• Transform
Probability
Probability
Log transform
x
log x
Unknown or irregular
distribution
• Transform
• Non-parametric methods
Non-parametric methods
• Non-parametric methods make no
assumptions about the distribution of
the data
• There are non-parametric methods for
characterizing data, as well as for
comparing data sets
• These methods are also called
distribution-free, robust, or sometimes
non-metric tests
Application to Reference
Ranges
The concentrations of most clinical
analytes are not usually distributed in a
Gaussian manner. Why?
How do we determine the reference range
(limits of expected values) for these
analytes?
Application to Reference
Ranges
• Reference ranges for normal, healthy
populations are customarily defined as the
“central 95%”.
• An entirely non-parametric way of expressing
this is to eliminate the upper and lower 2.5% of
data, and use the remaining upper and lower
values to define the range.
• NCCLS recommends 120 values, dropping the
two highest and two lowest.
Application to Reference
Ranges
What happens when we want to compare
one reference range with another? This
is precisely what CLIA ‘88 requires us to
do.
How do we do this?
“Everything should be made
as simple as possible, but
not simpler.”
Albert Einstein
Solution #1: Simple
comparison
Suppose we just do a small internal
reference range study, and compare our
results to the manufacturer’s range.
How do we compare them?
Is this a valid approach?
NCCLS recommendations
• Inspection Method: Verify reference
populations are equivalent
• Limited Validation: Collect 20 reference
specimens
– No more than 2 exceed range
– Repeat if failed
• Extended Validation: Collect 60
reference specimens; compare ranges.
Solution #2: Mann-Whitney*
Rank normal values (x1,x2,x3...xn) and the reference
population (y1,y2,y3...yn):
x1, y1, x2, x3, y2, y3 ... xn, yn
Count the number of y values that follow each x,
and call the sum Ux. Calculate Uy also.
*Also
called the U test, rank sum test, or Wilcoxen’s test.
Mann-Whitney, cont.
It should be obvious that: Ux + Uy = NxNy
If the two distributions are the same, then:
Ux = Uy = 1/2NxNy
Large differences between Ux and Uy indicate
that the distributions are not equivalent
“‘Obvious’ is the most
dangerous word in
mathematics.”
Eric Temple Bell (1883-1960)
Solution #3: Run test
In the run test, order the values in the two
distributions as before:
x1, y1, x2, x3, y2, y3 ... xn, yn
Add up the number of runs (consecutive values
from the same distribution). If the two data
sets are randomly selected from one
population, there will be few runs.
Solution #4: The Monte Carlo
method
Sometimes, when we don’t know anything
about a distribution, the best thing to do is
independently test its characteristics.
The Monte Carlo method
Asq xy x
2
x
Acir r
2
y
2
Acir
4
Asq
x
2
The Monte Carlo method
N
mean, SD
N
mean, SD
N
mean, SD
N
mean, SD
Reference population
The Monte Carlo method
With the Monte Carlo method, we have
simulated the test we wish to apply--that
is, we have randomly selected samples
from the parent distribution, and
determined whether our in-house data
are in agreement with the randomlyselected samples.
Analysis of paired data
• For certain types of laboratory studies,
the data we gather is paired
• We typically want to know how closely
the paired data agree
• We need quantitative measures of the
extent to which the data agree or
disagree
• Examples?
Examples of paired data
•
•
•
•
Method correlation data
Pharmacodynamic effects
Risk analysis
Pathophysiology
Correlation
50
45
40
35
30
25
20
15
10
5
0
0
5
10
15
20
25
30
35
40
45
50
Linear regression (least squares)
Linear regression analysis generates an
equation for a straight line
y = mx + b
where m is the slope of the line and b is the
value of y when x = 0 (the y-intercept).
The calculated equation minimizes the
differences between actual y values and the
linear regression line.
Correlation
50
45
40
y = 1.031x - 0.024
35
30
25
20
15
10
5
0
0
5
10
15
20
25
30
35
40
45
50
Covariance
Do x and y values vary in concert, or
randomly?
1
cov( x , y ) ( yi y )( xi x )
N i
1
cov( x , y ) ( yi y )( xi x )
N i
• What if y increases when x increases?
• What if y decreases when x increases?
• What if y and x vary independently?
Covariance
It is clear that the greater the covariance,
the stronger the relationship between x
and y.
But . . . what about units?
e.g., if you measure glucose in mg/dL,
and I measure it in mmol/L, who’s likely
to have the highest covariance?
The Correlation Coefficient
1
( yi y )( xi x )
cov( x , y ) N i
x y
y x
1 1
The Correlation Coefficient
• The correlation coefficient is a unitless
quantity that roughly indicates the
degree to which x and y vary in the
same direction.
• is useful for detecting relationships
between parameters, but it is not a very
sensitive measure of the spread.
Correlation
50
45
40
y = 1.031x - 0.024
= 0.9986
35
30
25
20
15
10
5
0
0
5
10
15
20
25
30
35
40
45
50
Correlation
50
45
40
y = 1.031x - 0.024
= 0.9894
35
30
25
20
15
10
5
0
0
5
10
15
20
25
30
35
40
45
50
Standard Error of the
Estimate
The linear regression equation gives us a
way to calculate an “estimated” y for
any given x value, given the symbol ŷ
(y-hat):
y mx b
Standard Error of the
Estimate
Now what we are interested in is the average
difference between the measured y and its
estimate, ŷ :
sy / x
1
2
( yi yi )
N i
Correlation
50
45
40
y = 1.031x - 0.024
= 0.9986
sy/x=1.83
35
30
25
20
15
10
5
0
0
5
10
15
20
25
30
35
40
45
50
Correlation
50
45
40
y = 1.031x - 0.024
= 0.9894
sy/x = 5.32
35
30
25
20
15
10
5
0
0
5
10
15
20
25
30
35
40
45
50
Standard Error of the
Estimate
If we assume that the errors in the y
measurements are Gaussian (is that a
safe assumption?), then the standard
error of the estimate gives us the
boundaries within which 67% of the y
values will fall.
2sy/x defines the 95% boundaries..
Limitations of linear
regression
• Assumes no error in x measurement
• Assumes that variance in y is constant
throughout concentration range
Alternative approaches
• Weighted linear regression analysis can
compensate for non-constant variance
among y measurements
• Deming regression analysis takes into
account variance in the x
measurements
• Weighted Deming regression analysis
allows for both
Evaluating method
performance
• Precision
Method Precision
• Within-run: 10 or 20 replicates
– What types of errors does within-run
precision reflect?
• Day-to-day: NCCLS recommends
evaluation over 20 days
– What types of errors does day-to-day
precision reflect?
Evaluating method
performance
• Precision
• Sensitivity
Method Sensitivity
• The analytical sensitivity of a method
refers to the lowest concentration of
analyte that can be reliably detected.
• The most common definition of
sensitivity is the analyte concentration
that will result in a signal two or three
standard deviations above background.
Signal
Signal/Noise threshold
time
Other measures of sensitivity
• Limit of Detection (LOD) is sometimes defined as
the concentration producing an S/N > 3.
– In drug testing, LOD is customarily defined as the lowest
concentration that meets all identification criteria.
• Limit of Quantitation (LOQ) is sometimes defined
as the concentration producing an S/N >5.
– In drug testing, LOQ is customarily defined as the lowest
concentration that can be measured within ±20%.
Question
At an S/N ratio of 5, what is the minimum
CV of the measurement?
If the S/N is 5, 20% of the measured
signal is noise, which is random.
Therefore, the CV must be at least 20%.
Evaluating method
performance
• Precision
• Sensitivity
• Linearity
Method Linearity
• A linear relationship between concentration
and signal is not absolutely necessary, but it
is highly desirable. Why?
• CLIA ‘88 requires that the linearity of
analytical methods is verified on a periodic
basis.
Ways to evaluate linearity
• Visual/linear regression
Signal
Concentration
Outliers
We can eliminate any point that differs
from the next highest value by more
than 0.765 (p=0.05) times the spread
between the highest and lowest values
(Dixon test).
Example: 4, 5, 6, 13
(13 - 4) x 0.765 = 6.89
Limitation of linear
regression method
If the analytical method has a high
variance (CV), it is likely that small
deviations from linearity will not be
detected due to the high standard error
of the estimate
Signal
Concentration
Ways to evaluate linearity
• Visual/linear regression
• Quadratic regression
Quadratic regression
Recall that, for linear data, the
relationship between x and y can be
expressed as
y = f(x) = a + bx
Quadratic regression
A curve is described by the quadratic
equation:
y = f(x) = a + bx + cx2
which is identical to the linear equation
except for the addition of the cx2 term.
Quadratic regression
It should be clear that the smaller the x2
coefficient, c, the closer the data are to
linear (since the equation reduces to the
linear form when c approaches 0).
What is the drawback to this approach?
Ways to evaluate linearity
• Visual/linear regression
• Quadratic regression
• Lack-of-fit analysis
Lack-of-fit analysis
• There are two components of the
variation from the regression line
– Intrinsic variability of the method
– Variability due to deviations from linearity
• The problem is to distinguish between
these two sources of variability
• What statistical test do you think is
appropriate?
Signal
Concentration
Lack-of-fit analysis
The ANOVA technique requires that method
variance is constant at all concentrations.
Cochran’s test is used to test whether this
is the case.
VL
0.5981 ( p 0.05)
Vi
i
Lack-of-fit method
calculations
• Total sum of the squares: the variance
calculated from all of the y values
• Linear regression sum of the squares:
the variance of y values from the
regression line
• Residual sum of the squares:
difference between TSS and LSS
• Lack of fit sum of the squares: the RSS
minus the pure error (sum of variances)
Lack-of-fit analysis
• The LOF is compared to the pure error to
give the “G” statistic (which is actually F)
• If the LOF is small compared to the pure
error, G is small and the method is linear
• If the LOF is large compared to the pure
error, G will be large, indicating significant
deviation from linearity
Significance limits for G
• 90% confidence = 2.49
• 95% confidence = 3.29
• 99% confidence = 5.42
“If your experiment needs
statistics, you ought to
have done a better
experiment.”
Ernest Rutherford (1871-1937)
Evaluating Clinical
Performance of laboratory
tests
• The clinical performance of a laboratory
test defines how well it predicts disease
• The sensitivity of a test indicates the
likelihood that it will be positive when
disease is present
Clinical Sensitivity
If TP as the number of “true positives”, and
FP is the number of “false positives”, the
sensitivity is defined as:
TP
Sensitivity
100
TP FN
Example
Of 25 admitted cocaine abusers, 23 tested
positive for urinary benzoylecgonine and
2 tested negative. What is the sensitivity
of the urine screen?
23
100 92%
23 2
Evaluating Clinical
Performance of laboratory
tests
• The clinical performance of a laboratory
test defines how well it predicts disease
• The sensitivity of a test indicates the
likelihood that it will be positive when
disease is present
• The specificity of a test indicates the
likelihood that it will be negative when
disease is absent
Clinical Specificity
If TN is the number of “true negative”
results, and FP is the number of falsely
positive results, the specificity is defined
as:
TN
Specificity
100
TN FP
Example
What would you guess is the specificity of
any particular clinical laboratory test?
(Choose any one you want)
Answer
Since reference ranges are customarily set
to include the central 95% of values in
healthy subjects, we expect 5% of values
from healthy people to be “abnormal”-this is the false positive rate.
Hence, the specificity of most clinical tests
is no better than 95%.
Sensitivity vs. Specificity
• Sensitivity and specificity are inversely
related.
Disease
+
Marker concentration
Sensitivity vs. Specificity
• Sensitivity and specificity are inversely
related.
• How do we determine the best
compromise between sensitivity and
specificity?
True positive rate
(sensitivity)
Receiver Operating
Characteristic
False positive rate
1-specificity
Evaluating Clinical
Performance of laboratory
tests
• The sensitivity of a test indicates the likelihood
that it will be positive when disease is present
• The specificity of a test indicates the likelihood
that it will be negative when disease is absent
• The predictive value of a test indicates the
probability that the test result correctly classifies
a patient
Predictive Value
The predictive value of a clinical laboratory
test takes into account the prevalence of a
certain disease, to quantify the probability
that a positive test is associated with the
disease in a randomly-selected individual,
or alternatively, that a negative test is
associated with health.
Illustration
• Suppose you have invented a new
screening test for Addison disease.
• The test correctly identified 98 of 100
patients with confirmed Addison disease
(What is the sensitivity?)
• The test was positive in only 2 of 1000
patients with no evidence of Addison
disease (What is the specificity?)
Test performance
• The sensitivity is 98.0%
• The specificity is 99.8%
• But Addison disease is a rare disorder-incidence = 1:10,000
• What happens if we screen 1 million
people?
Analysis
• In 1 million people, there will be 100 cases
of Addison disease.
• Our test will identify 98 of these cases (TP)
• Of the 999,900 non-Addison subjects, the
test will be positive in 0.2%, or about 2,000
(FP).
Predictive value of the
positive test
The predictive value is the % of all positives
that are true positives:
TP
PV
100
TP FP
98
100
98 2000
4.7%
What about the negative
predictive value?
• TN = 999,900 - 2000 = 997,900
• FN = 100 * 0.002 = 0 (or 1)
TN
PV
100
TN FN
997,900
100
997,900 1
100%
Summary of predictive value
Predictive value describes the usefulness
of a clinical laboratory test in the real
world.
Or does it?
Lessons about predictive
value
• Even when you have a very good test, it is
generally not cost effective to screen for
diseases which have low incidence in the
general population. Exception?
• The higher the clinical suspicion, the better
the predictive value of the test. Why?
Efficiency
We can combine the PV+ and PV- to give
a quantity called the efficiency:
TP TN
Efficiency
100
TP FP TN FN
The efficiency is the percentage of all
patients that are classified correctly by
the test result.
Efficiency of our Addison
screen
98 997,900
100 99.8%
98 2000 997,900 2
“To call in the statistician
after the experiment is done
may be no more than asking
him to perform
a postmortem examination:
he may be able to say what
the experiment died of.”
Ronald Aylmer Fisher (1890 - 1962)
Application of Statistics to
Quality Control
• We expect quality control to fit a Gaussian
distribution
• We can use Gaussian statistics to predict
the variability in quality control values
• What sort of tolerance will we allow for
variation in quality control values?
• Generally, we will question variations that
have a statistical probability of less than 5%
“He uses statistics as a
drunken man uses lamp
posts -- for support rather
than illumination.”
Andrew Lang (1844-1912)
Westgard’s rules
•
•
•
•
•
•
12s
13s
22s
R4s
41s
10x
1 in 20
1 in 300
1 in 400
1 in 800
1 in 600
1 in 1000
Some examples
+3sd
+2sd
+1sd
mean
-1sd
-2sd
-3sd
Some examples
+3sd
+2sd
+1sd
mean
-1sd
-2sd
-3sd
Some examples
+3sd
+2sd
+1sd
mean
-1sd
-2sd
-3sd
Some examples
+3sd
+2sd
+1sd
mean
-1sd
-2sd
-3sd
“In science one tries to
tell people, in such a way
as to be understood by
everyone, something that
no one ever knew before.
But in poetry, it's the
exact opposite.”
Paul Adrien Maurice Dirac (1902- 1984)