-
8/9/2019 Statistical Methods in Experimental Chemistry
1/103
Statistical Methods in Experimental Chemistry
Philip J. Grandinetti
January 20, 2000
-
8/9/2019 Statistical Methods in Experimental Chemistry
2/103
2
P. J. Grandinetti, January 20, 2000
-
8/9/2019 Statistical Methods in Experimental Chemistry
3/103
Contents
1 Statistical Description of Experimental Data 7
1.1 Random Error Distributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81.1.1 Univariate Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
The Mean . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
The Variance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
The Skewness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
The Kurtosis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
1.1.2 Bivariate Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
The Covariance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
1.1.3 Probability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
1.1.3.1 Permutations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161.1.3.2 Combinations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
1.1.3.3 Probability Distributions . . . . . . . . . . . . . . . . . . . . . . . . 17
Binomial Distribution . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
Poisson Distribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
Gaussian Distribution . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
Bi-dimensional Gaussian Distribution . . . . . . . . . . . . . . . . . . 24
Multi-dimensional Gaussian Distribution . . . . . . . . . . . . . . . . . 25
1.1.3.4 Central Limit Theorem . . . . . . . . . . . . . . . . . . . . . . . . . 26
1.2 The 2 test of a distribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
1.3 Systematic Errors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
Instrument Errors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
Personal Errors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
Method Errors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
1.4 Gross Errors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
1.5 Propagation of Errors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
3
-
8/9/2019 Statistical Methods in Experimental Chemistry
4/103
4 CONTENTS
1.6 Confidence Limits . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
1.6.1 Students t-distribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 381.7 The Two Sample Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
1.7.1 The t-test . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
1.7.1.1 Comparing a measured mean to a true value . . . . . . . . . . . . . 40
1.7.1.2 Comparing two measured means . . . . . . . . . . . . . . . . . . . . 41
1.7.2 Comparing Variances - The F-test . . . . . . . . . . . . . . . . . . . . . . . . 42
1.8 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
2 Modeling of Data 53
2.1 Maximum Likelihood . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
2.2 A Simple Example of Linear Least-Squares . . . . . . . . . . . . . . . . . . . . . . . 56
2.2.1 Finding Best-Fit Model Parameters . . . . . . . . . . . . . . . . . . . . . . . 56
2.2.2 Finding Uncertainty in Model Parameters . . . . . . . . . . . . . . . . . . . . 58
2.2.3 Finding Goodness of Fit . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
2.2.4 What to do if you dont know yi? . . . . . . . . . . . . . . . . . . . . . . . . 60
2.3 General Linear Least Squares . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
2.4 General Non-Linear Least Squares . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
2.4.1 Minimizing 2 without Derivative Information . . . . . . . . . . . . . . . . . 67
2.4.1.1 Downhill Simplex Method . . . . . . . . . . . . . . . . . . . . . . . . 67
2.4.2 Minimizing 2 with Derivative Information . . . . . . . . . . . . . . . . . . . 692.4.2.1 Directional Derivatives and Gradients . . . . . . . . . . . . . . . . . 72
2.4.2.2 Steepest Descent Method . . . . . . . . . . . . . . . . . . . . . . . . 74
2.4.2.3 Newton Raphson Method . . . . . . . . . . . . . . . . . . . . . . . . 75
2.4.2.4 Marquardt Method . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
2.5 Confidence Intervals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
2.5.1 Constant Chi-Squared Boundaries as Confidence Limits . . . . . . . . . . . . 79
2.5.2 Getting Parameter Errors from 2 with Non-Gaussian Errors . . . . . . . . . 81
2.5.2.1 Monte Carlo Simulations . . . . . . . . . . . . . . . . . . . . . . . . 81
2.5.2.2 Bootstrap Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . 822.6 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
A Computers 89
The Bit: . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
Integers: . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
Signed Integer Numbers . . . . . . . . . . . . . . . . . . . . . . . . . . 91
P. J. Grandinetti, January 20, 2000
-
8/9/2019 Statistical Methods in Experimental Chemistry
5/103
CONTENTS 5
Alphanumeric Characters: . . . . . . . . . . . . . . . . . . . . . . . . . 92
Floating point numbers: . . . . . . . . . . . . . . . . . . . . . . . . . . 93Roundoff Errors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
A.1 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
B Programming in C - A quick tutorial 95
Comments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
#include . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
main() . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
Another Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
To program or not to program . . . . . . . . . . . . . . . . . . . . . . 100
C.1 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
P. J. Grandinetti, January 20, 2000
-
8/9/2019 Statistical Methods in Experimental Chemistry
6/103
6 CONTENTS
P. J. Grandinetti, January 20, 2000
-
8/9/2019 Statistical Methods in Experimental Chemistry
7/103
Chapter 1
Statistical Description of
Experimental Data
A fundamental problem encountered by every scientist who performs a measurement is whether
her measured quantity is correct. Put another way, how can she know the difference between her
measured value and the true value?
Error in Measurement = Measured Value True ValueThe basic difficulty is that to answer this very important question she needs to know the true
value. Of course, if she knew the true value she wouldnt be making the measurement in the firstplace! So, how can she resolve this catch-22 situation?
The solution is for her to go back into the lab, perform the measurement many times on known
quantities, and study the errors. While she can never know the true value of her unknown quantity,
she can use her error measurements and the mathematics of probability theory to tell her the
probability that the true value for her unknown quantity lies within a given range.
Errors can be classified into two classes, random and systematic. Random errors are irrepro-
ducible, and usually are caused by a large number of uncorrelated sources. In contrast, systematic
errors are reproducible, and usually can be attributed to a single source. Random errors could also
be caused by a single source, particularly when the experimentalist cannot identify the source of her
errors. If she could identify the source she might be able to make the error behave systematically.
Often systematic errors can be eliminated by simply recalibrating the measurement device. You
may hear of a third class of errors called gross errors. These are described as errors that occur
occasionally and are large. We will discuss this particular type of error later on, however, these
errors technically do not define a new class of errors. All errors, regardless of the class, can be
studied and used to establish limits between which the true value lies.
7
-
8/9/2019 Statistical Methods in Experimental Chemistry
8/103
8 CHAPTER 1. STATISTICAL DESCRIPTION OF EXPERIMENTAL DATA
Measurement Number
31.4
31.3
31.2
31.1
31.0
30.9
30.8
N = 128
30.8 30.9 31.0 31.1 31.2 31.3 31.4
8
6
4
2
0
(A) (B)
Probability
Density
Temperature(C
)
Temperature (C)
Figure 1.1: (A) A set of 128 temperature measurements of a solution in a constant temperature
bath. (B) A normalized histogram of the measured values.
1.1 Random Error Distributions
1.1.1 Univariate Data
In this section we will assume that when we perform a measurement there are no systematic errors.
We are all familiar with random errors. For example, suppose you were measuring the temperature
of a solution in a constant temperature bath. You know that the temperature should be constant,
but on close inspection you find that is it not constant and appears to be randomly fluctuating
around the expected constant value (see Fig 1.1a). What do you report for the temperature of the
solution? On one extreme we could report only the average value, and on the other extreme wecould report all the values measured in the form of a histogram of measured values (see Fig 1.1b).
To use the mathematics of probability theory in helping us understand errors we make the
assumption that our histogram of measured values is governed by an underlying probability distri-
bution called the parent distribution. In the limit of an infinite number of measurements our
histogram or sample distribution becomes the parent distribution. This point is illustrated in
Fig. 1.2. Notice how the histogram of measured values more closely follows the values predicted
by the parent distribution as the number of measurements used to construct the histogram in-
creases. Thus, our first goal in understanding the errors in our measurements is to learn what is
the underlying parent distribution1
, p(x), that predicts the spread in our measured values. Oncewe know that distribution we can calculated the probability that the measured value lies within a
1We define our parent distribution so that it is normalized. That is, the area under the distribution is unity,
p(x)dx = 1. (1.1)
P. J. Grandinetti, January 20, 2000
-
8/9/2019 Statistical Methods in Experimental Chemistry
9/103
1.1. RANDOM ERROR DISTRIBUTIONS 9
N = 4096
30.8 30.9 31.0 31.2 31.3 31.4
8
6
4
2
0
N = 128
30.8 30.9 31.0 31.1 31.2 31.3 31.4
8
6
4
2
0
31.1
Measured Value
Probability
Density
N = 1024
30.8 30.9 31.0 31.1 31.2 31.3 31.4
8
6
4
2
0
Probability
Density
Probability
Density
N = 128
30.8 30.9 31.0 31.1 31.2 31.3 31.4
8
6
4
2
0
N = 1024
30.8 30.9 31.0 31.1 31.2 31.3 31.4
8
6
4
2
0
N = 4096
30.8 30.9 31.0 31.2 31.3 31.4
8
6
4
2
031.1
Measured Value
(a) (b)
N = 128
30.8 30.9 31.0 31.1 31.2 31.3 31.4
8
6
4
2
0
N = 1024
30.8 30.9 31.0 31.1 31.2 31.3 31.4
8
6
4
2
0
N = 4096
30.8 30.9 31.0 31.2 31.3 31.4
8
6
4
2
031.1
Measured Value
(c)
Figure 1.2: In (a), (b), and (c) are histograms constructed from three different sets of 4096 measure-
ments, all drawn from the same parent distribution. Shown from top to bottom are the histograms
constructed using only the first 128, the first 1024, and finally all 4096 measurements, respectively.
Notice how the histogram bin heights vary at low N and fall closer to the bin heights predicted by
the parent distribution for high N.
given range. That is
P(x, x+) =
x+x
p(x) dx, (1.2)
where p(x) is our parent distribution for the measured value x, and P(x, x+) is the probability
that the measured value lies between x and x+. In general, x and x+ are called the confidence
limits associated with a given probability P(x, x+). We will have further discussions on confidence
limits later in the text.
Note that when you report confidence limits you are not reporting any information concerning
the shape of the parent distribution. It is often useful to have such information. In the interests
of not having to plot a histogram for every measurement we report we ask the question: Can we
mathematically describe the parent distribution for our measured values by a finite (and perhaps
P. J. Grandinetti, January 20, 2000
-
8/9/2019 Statistical Methods in Experimental Chemistry
10/103
10 CHAPTER 1. STATISTICAL DESCRIPTION OF EXPERIMENTAL DATA
even small) number of parameters? While the answer to this question depends on the particular
parent distribution, there are a few parameters that by convention are often used to describe (inpart, or sometimes completely) the parent distribution. These are the mean, variance, skewness,
and kurtosis.
The Mean describes the average value of the distribution. It is also called the first moment
about the origin. It is defined as:
= limN
1
N
i
xi, (1.3)
where N corresponds to the number of measurements xi and is the mean of the parent distribution.
Experimentally we cannot make an infinite number of measurements so we define the experimental
mean according to:
x =1
N
i
xi, (1.4)
where x is the mean of the experimental distribution. The mean is usually approximated to be the
true value, particularly when the distribution is symmetric and assuming no systematic errors.
If the distribution is not symmetric the mean is often supplemented with two other parameters,
the median and the mode. The median cuts the area of the parent distribution in half. That is,xmedian
p(x)dx = 1/2. (1.5)
When the distribution is symmetric then the mean and median are the same. The mode is the
most probable value of the parent distribution. That is,
dp(xmode)
dx= 0, and
d 2p(xmode)
dx2< 0. (1.6)
The Variance characterizes the width of the distribution. It is also called the second moment
about the mean. It is defined as:
2 = limN
1
N
Ni
(xi )2, (1.7)
where 2 is the variance of the parent distribution. The experimental variance is defined as:
s2 =1
N 1Ni
(xi x)2, (1.8)
P. J. Grandinetti, January 20, 2000
-
8/9/2019 Statistical Methods in Experimental Chemistry
11/103
1.1. RANDOM ERROR DISTRIBUTIONS 11
mode
median
mean
Measured Value
probabilitydensity
Figure 1.3: Position of mean, median, and median for an asymmetric distribution. Note that in a
symmetric distribution the mean and median are the same.
where s2 is the variance of the experimental parent distribution. and s are defined as the
standard deviation of the parent distribution and experimental parent distribution, respectively.
The variance is approximated as the error dispersion of the set of measurements, and the standard
deviation is approximated as the uncertainty in the true value. Notice that in the expression
for s2 we use N 1 in the denominator in contrast to N for 2. Without going into detail, thedifference is because s2 is calculated using x instead of . In practice, if the difference between N
and N1 in the denominator affects the conclusions you make about your data, then you probablyneed to collect more data (i.e. increase N) and reanalyze your data.
The Skewness characterizes the degree of asymmetry of a distribution around its mean. It is
also called the third moment about the mean. It is defined as:
skewness =1
N
N
i=1xi
3
. (1.9)
Skewness is defined to have no dimensions. Positive skewness signifies a distribution with an asym-
metric tail extending out towards more positive x, while negative skewness signifies a distribution
whose tail extends out toward more negative x. A symmetric distribution should have zero skew-
ness (e.g. Gaussian). It is good practice to believe you have a statistically significant non-zero
skewness if it is larger than
15/N.
P. J. Grandinetti, January 20, 2000
-
8/9/2019 Statistical Methods in Experimental Chemistry
12/103
12 CHAPTER 1. STATISTICAL DESCRIPTION OF EXPERIMENTAL DATA
positive
skewness
negative
skewness
Measured Value
ProbabilityDensity
Figure 1.4: Positive skewness signifies a distribution with an asymmetric tail extending out towards
more positive x, while negative skewness signifies a distribution whose tail extends out toward more
negative x.
The Kurtosis measures the relative peakedness or flatness of a distribution relative to a normal
(i.e., Gaussian) distribution. It is also called the fourth moment about the mean. It is defined as:
kurtosis =
1
N
Ni=1
xi
4 3. (1.10)
Subtracting 3 makes the kurtosis zero for a normal distribution. A positive kurtosis is calledleptokurtic, a negative kurtosis is called platykurtic, and in between is called mesokurtic.
1.1.2 Bivariate Data
Lets consider a slightly different measurement situation. In this situation when we make a mea-
surement we obtain not a single number, but a pair of numbers. That is, each measurement yields
an (x, y) pair. Just as in the one parameter case, there are random errors that will result in a
distribution of (x, y) pairs around the true (x, y) pair value. Also, like the one parameter case we
can construct a histogram of the N measured (x, y) pairs, except in this case our histogram will be
a three dimensional plot of occurance versus x versus y.As before, in the limit of an infinite number of (x, y) pair measurements we would have the two
parameter parent distribution2 p(x, y). The mean (x, y) pair are calculated
x = limN
1
N
i
xi, and y = limN
1
N
i
yi. (1.12)
2We define our two parameter parent distribution so that it is normalized. That is, the area under the distribution
P. J. Grandinetti, January 20, 2000
-
8/9/2019 Statistical Methods in Experimental Chemistry
13/103
1.1. RANDOM ERROR DISTRIBUTIONS 13
negative
kurtosis
(platykurtic)
positive
kurtosis(leptokurtic)
Measured Value
ProbabilityDensit
y
Figure 1.5: The Kurtosis measures the relative peakedness or flatness of a distribution relative toa normal (i.e., Gaussian) distribution. A positive kurtosis is called leptokurtic, a negative kurtosis
is called platykurtic, and in between is called mesokurtic.
Similarly the variance is calculated
2x = limN
1
N
i
(xi x)2, and 2y = limN
1
N
i
(yi y)2. (1.13)
The Covariance In the two parameter case, however, we need to consider another type of
variance called the covariance. It is defined as
2xy = limN
1
N
i
(xi x)(yi y) (1.14)
If the errors in x and y are uncorrelated this term will be zero. What distinguishes these two
situations? Usually, a non-zero covariance implies that the two measured parameters are dependent
on each other, that is
y = F(x), or x = F(y),or both x and y are dependent on a common third parameter, that is
x =F
(z), or y =F
(z).
If x is the temperature of a solution in a water bath in New York City, and y is the temperature
of a solution in a water bath in San Francisco, then its unlikely that the errors in x and y will be
is unity,
p(x, y)dxdy = 1. (1.11)
P. J. Grandinetti, January 20, 2000
-
8/9/2019 Statistical Methods in Experimental Chemistry
14/103
14 CHAPTER 1. STATISTICAL DESCRIPTION OF EXPERIMENTAL DATA
x
y y
x
2xy = 0 2xy 0
(a) (b)
Figure 1.6: (a) Contour plot of a two parameter parent distribution having no correlation between
the two parameter errors. (b) Contour plot of a two parameter parent distribution having a strong
correlation between parameter errors.
correlated. This situation would have a distribution similar to the one in Fig. 1.6a. If, however,
x and y were the temperature of two different solutions in the same water bath, then it would be
reasonable if their errors were correlated as shown for the distribution in Fig. 1.6b.
Another parameter often used that is related to the covariance is the linear correlation coefficient
r. It is defined as
rxy =2xy
xy. (1.15)
An r2xy value of 1 implies a complete linear correlation between the measured x and y parameters,
while an r2xy value of zero implies no correlation whatsoever. It should be pointed out that even if
y is dependent on x in a non-linear fashion, the non-linear function may be fairly linear over the
small range of x and y associated with the random errors in their measurement.
A high covariance between two parameters means that any increase (or decrease) in the uncer-
tainty of one parameter will lead to a corresponding increase (or decrease) in the uncertainty of
the other parameter. This is illustrated in Fig. 1.7.
Finding the probability that a measured (x, y) pair will lie within a given range is a morecomplicated problem than in one-dimension. One could define a square region delimited by x and
x+ on the x-axis and y and y+ on the y-axis, such as
P(x, x+, y, y+) =
x+x
y+y
p(x, y)dxdy, (1.16)
as shown in Fig. 1.8. In case of high correlation (i.e. r2 close to 1), this approach will includes
P. J. Grandinetti, January 20, 2000
-
8/9/2019 Statistical Methods in Experimental Chemistry
15/103
1.1. RANDOM ERROR DISTRIBUTIONS 15
x
y y
x
2xy = 0 2xy 0
}y
1
y2 }
y1
y2
Distribution in x after
summing over all possible
values of y.
Distribution in x after
summing over all possible
values of y.
Distribution in x after
summing over values of y
between y1 and y2
Distribution in x after
summing over values of y
between y1 and y2
(a) (b)
Figure 1.7: When the distribution in two parameters are strongly correlated the width of the dis-
tribution in one of the parameters will be greatly reduced if the values of the other parameters arerestricted to a smaller set of possible values.
regions of low probability, and thus doesnt give an accurate representation of where the data is
likely to occur. Finding the probability that a measured (x, y) pair will lie within a given ellipse is
a better approach, which we will discuss later in the course.
1.1.3 Probability
Before we can go deeper with this statistical picture of random experimental errors, we need to
review the mathematics of probability theory. This will help us better understand the shapes ofparent distributions and their application to the study of random errors.
As you might have guessed, the mathematical theory of probability arose when a gambler
(Chevalier de Mere) wanted to adjust the odds so he could be certain of winning. Thus he enlisted
the help of the famous French mathematicians Blaise PASCAL and Pierre DE FERMAT.
x
y
x - x +
y -
y +
Figure 1.8: Cartesian confidence limits of two-dimensional probablity distribution.
P. J. Grandinetti, January 20, 2000
-
8/9/2019 Statistical Methods in Experimental Chemistry
16/103
16 CHAPTER 1. STATISTICAL DESCRIPTION OF EXPERIMENTAL DATA
In principle, calculating the probability of an event is straightforward:
Probability =Number of outcomes that are successful (winning)
Total number of outcomes (winning and losing)(1.17)
The difficulty lies in counting. In order to count the number of outcomes we appeal to combinatorics.
1.1.3.1 Permutations
A Permutationis an arrangement of outcomes in which the order is important. A common example
is the Election problem. Consider a club with 5 members, Joe, Cathy, Sally, Bob, and Pat. In
how many ways can we elect a president and a secretary? One solution is to make a tree, such as
the one below:
Joe
Joe, Cathy
Joe, Sally
Joe, Bob
Joe, Pat
Cathy
Cathy, Joe
Cathy, Sally
Cathy, Bob
Cathy, Pat
Pat
Pat, Joe
Pat, Cathy
Pat, Sally
Pat, Bob
Sally
Sally, Joe
Sally, Cathy
Sally, Bob
Sally, Pat
Bob
Bob, Joe
Bob, Cathy
Bob, Sally
Bob, Pat
Using such a tree diagram we can count that there are a total of 20 possible ways to elect a
president and a secretary in a club with 5 members. Another approach is to think of two boxes,
one for president, and one for secretary. If you pick the president first and secretary second then
youll have five choices for president 5 president, and four choices for secretary 4 secretary. The
total number of ways is the product of the two numbers
5 president 4 secretary = 20.
What if we wanted to elect a president, secretary, and treasurer? In this case a tree would be
alot of work. Using the boxes approach we would have 5 4 3 = 60 possibilities. In general, thenumber of ways r objects can be selected from n objects is
nPr = n (n 1) (n 2) (n r + 1), (1.18)
or more generally written as
nPr =n!
(n r)! . (1.19)
P. J. Grandinetti, January 20, 2000
-
8/9/2019 Statistical Methods in Experimental Chemistry
17/103
1.1. RANDOM ERROR DISTRIBUTIONS 17
For example, how many ways can a baseball coach arrange the batting order with 9 players?
Answer: 9P9 =9!
(9 9)! = 9! = 362, 880 ways.
1.1.3.2 Combinations
A Combination is an arrangement of outcomes in which the order is not important. A common
example is the Committee problem. Consider again our club with 5 members. In how many
ways can we form a three member committee? In this case the order is not important. That is,
{Joe, Cathy, Sally} = {Cathy, Joe, Sally} = {Cathy, Sally, Joe}. All arrangements of Cathy, Joe,and Sally are equivalent.
The total number of combinations of n objects taken r at a time is
nCr =nPr
r!, or
n!
r!(n r)! . (1.20)
For example, there are 5C3 =5!
3!2!= 10 possible combinations for a 3 member committee starting
with 5 members.
Heres another example: What is the number of ways a 5 card hand can be drawn from a 52
card deck?
52C5 =52!
5!47!= 2, 598, 960 ways.
nCr is called the binomial coefficient, and is also often written asn
r
. Both notations will be
used in this text.
Permutations and Combinations cannot be used to solve every counting problem, however, they
can be extremely helpful for certain types of counting problems.
1.1.3.3 Probability Distributions
Now that we know how to count, lets go back to the original problem of calculating the probability
of an event. Our starting point is the equation
Probability =Number of successful outcomes
Total number of outcomes. (1.21)
What if the names of everyone in our 5 member club are thrown in a hat, and two names are drawn,
with the first becoming president and the second becoming secretary. With this approach, what is
the probability that Pat is chosen as president and Cathy as secretary?
P. J. Grandinetti, January 20, 2000
-
8/9/2019 Statistical Methods in Experimental Chemistry
18/103
18 CHAPTER 1. STATISTICAL DESCRIPTION OF EXPERIMENTAL DATA
To calculate this probability we first have to count the outcomes. In this case there is only one
successful outcome, namely Pat as president and Cathy as secretary. The total number of outcomescan be calculated as the total number of permutations of drawing two names out of a group of five,
or
total number of outcomes = 5P2 = 20
Therefore, the probability is
Probability =1
20= 0.05.
That is, theres a 5% chance that Pat is chosen as president and Cathy as secretary.
Lets look at another example. What is the probability of being dealt a particular five card
hand from a 52 card deck?
number of successful outcomes = 1,
total number of outcomes = 52C5 = 2, 598, 960,
Probability =1
2, 598, 960= 3.847692 107.
Yet, another example. What is the probability of being dealt five cards that are all spades from
a 52 card deck?
number of successful outcomes = 13C5 =13!
5!8!
= 1, 287,
total number of outcomes = 52C5 = 2, 598, 960,
Probability =1, 287
2, 598, 960= 4.951980 104.
Binomial Distribution
Now lets look at the probabilities associated with independent events with the same probabil-
ities. For example, if I roll a die ten times, what is the probability that only 3 will come up sixes?
The first question we have to ask is what is the probability of rolling a particular sequence with
only 3 sixes. For example, one possibility is
X,X, 6,X,X,X, 6, 6, X , X
where X is a roll that was not 6. The probability of this particular sequence of independent events
is5
6 5
6 1
6 5
6 5
6 5
6 1
6 1
6 5
6 5
6=
5
6
7 16
3= 1.292044 103
P. J. Grandinetti, January 20, 2000
-
8/9/2019 Statistical Methods in Experimental Chemistry
19/103
1.1. RANDOM ERROR DISTRIBUTIONS 19
The second question we ask is how many different ways can we roll only 3 sixes. Clearly, the
total number of possibilities (i.e., combinations) is 10C3 or10
3. Assuming that all possible
combinations are equally probable, then to obtain the overall probability that I will roll only 3
sixes we simply multiply our calculated probability above by the number of combinations that give
only 3 sixes. That is,
P(3 sixes out of 10 rolls) = 10C3
5
6
7 16
3= 120 1.292044 103 = 0.15504536,
or roughly a 1 in 6.5 chance. Note that the solution to this problem would have been no different
if I had rolled ten different dice all at once and asked for the probability that only 3 came up sixes.
We can generalize this reasoning to the case where the probability of success is p (instead of
1/6), the probability of failure is (1 p) (instead of 5/6), the number of trials is n (instead of 10),and the number of successes is r (instead of 3). That is,
P(r,n,p) =
n
r
pr(1 p)nr (1.22)
This distribution of probabilities, for r = 0, 1, 2, . . . , n, is called the binomial distribution.
In general, the mean and variance of a discrete distribution3 is given by
r =rmaxr=0
rP(r), and 2r =rmaxr=0
(r r)2P(r). (1.24)
Thus, we find the mean of the binomial distribution to be
=n
r=0
r
n
r
pr(1 p)nr = np, (1.25)
and the variance of the binomial distribution to be
2 =n
r=0
(r )2
n
r
pr(1 p)nr
= np(1 p). (1.26)
For example, a coin is flipped 10 times. What is the probability of getting exactly r heads?
Since p = 1/2 and (1 p) = 1/2 then we have
P(r, 10, 1/2) =
10
r
(1/2)r(1/2)10r =
10
r
(1/2)10
3For a continuous distribution we have
x =
xp(x)dx, and 2x =
(x x)2p(x)dx. (1.23)
where x is a continuous variable.
P. J. Grandinetti, January 20, 2000
-
8/9/2019 Statistical Methods in Experimental Chemistry
20/103
20 CHAPTER 1. STATISTICAL DESCRIPTION OF EXPERIMENTAL DATA
r Prob.
0
1
2
3
4
5
6
7
8
9
10
9.77 X 10- 4
9.77 X 10- 3
0.117
0.205
0.246
0.205
0.117
0.044
9.77 X 10- 4
9.77 X 10- 3
0.044
0 2 4 6 8 10
0.00
0.05
0.10
0.15
0.20
0.25
Probability
r
= 5
= 1.58
(a)
0 2 4 6 8 10
0.00
0.05
0.10
0.15
0.20
0.25
0.30
0.35
r
= 1.67
= 1.18
Probability
r Prob.
0
12
3
4
5
6
7
8
9
10
0.162
0.323
0.00025
0.002
0.013
0.054
0.155
0.291
1.65 X10-80.0000008
0.00002
(b)
Figure 1.9: The binomial distribution. (a) A symmetric case where p = 1/2 and n = 10. (b) A
asymmetric case where p = 1/6 and n = 10.
In this case we get a symmetric distribution (see Fig. 1.9a) with a mean of = 5 (i.e., 5 out of 10
times we will get heads) and a variance of 2 = /2 = 2.5, and =
2.5.
Lets look at another example. A die is rolled 10 times. What is the probability of getting a
six r times? Since p = 1/6 and (1 p) = 5/6 then we have
P(r, 10, 1/6) =10!
r!(10 r)! (1/6)r(5/6)10r.
In this case we get an asymmetric distribution (see Fig. 1.9b) with a mean of = 10/6, 2 = 50/36,
and =
50/36.
Poisson Distribution
Lets look at a problem that more oriented towards physics and chemistry. If the probability
that an isolated excited state atom will emit a single photon in one second is 0.00050, what is the
probability that two photons would be emitted in one second from five identical non-interactingexcited state atoms?
Answer: P(2, 5, 0.0005) =5!
2!3!(0.00050)2(0.99950)3 = 2.5 106.
While this is a straightforward calculation, we run into difficulties in this type of calculation when
working with a larger number of atoms (i.e. higher n). For example, how would you calculate the
P. J. Grandinetti, January 20, 2000
-
8/9/2019 Statistical Methods in Experimental Chemistry
21/103
1.1. RANDOM ERROR DISTRIBUTIONS 21
0 5 10 15 20 25 300.00
0.05
0.10
0.15
0.20
0.25
0.30
0.35
0.40=1
=2
=4
=8
=16
r
P(r,)
Figure 1.10: The Poisson distribution.
probability that two photons would be emitted from 10,000 identical non-interacting excited stateatoms? My calculator cannot do factorial calculations with numbers greater than 69.
In this situation we can make an approximation for the binomial distribution in the limit that
n and p 0 such that np a finite number. This approximation is called the Poissondistribution and is given by
PPoisson(r,n,p) =(np)r
r!enp. (1.27)
This distribution most often describes the parent distribution when you are observing independent
random events that are occurring at a constant rate, such as photon counting experiments. The
mean of the Poisson distribution is
=r=0
r
(np)r
r!enp
= np. (1.28)
and the variance of the Poisson distribution is
2 =r=0
(r np)2 (np)
r
r!enp
= np. (1.29)
P. J. Grandinetti, January 20, 2000
-
8/9/2019 Statistical Methods in Experimental Chemistry
22/103
22 CHAPTER 1. STATISTICAL DESCRIPTION OF EXPERIMENTAL DATA
These two results are as expected for the binomial distribution in the limit that p 0. We define = np and rewrite the Poisson distribution as
PPoisson(r, ) =r
r!e. (1.30)
The Poisson distribution gives the probability for obtaining r events in a given time interval with
the expected number of events being .
Lets use the Poisson distribution to solve our problem above with 10,000 atoms. In this problem
n = 10, 000 and p = 0.00050 so = np = 5.0. Thus,
P(2, 5.0) =(5.0)2
2!
e5.0 = 0.084.
Theres an 8.4% chance that two photons would be emitted from 10,000 identical non-interacting
excited state atoms.
Gaussian Distribution In the limit of large n and p 0 the Poisson distribution serves as agood approximation for the binomial distribution. Is there an approximation that holds for the
limit of large n when p is not close to zero? Yes, under these conditions we can use the Gaussian
distribution as an approximation for the binomial. That is,
PGaussian(r,n,p) =
12np(1 p) exp1
2
(r
np)2
np(1 p) . (1.31)The mean of the Gaussian distribution is
=r=0
r
2np(1 p) exp
12
(r np)2np(1 p)
= np. (1.32)
and the variance of the Gaussian distribution is
2 =r=0
(r np)2
2np(1 p) exp
1
2
(r np)2np(1 p)
= np(1 p). (1.33)
Making the substitutions for = np and 2 = np(1 p) we can rewrite the Gaussian distributionin the form
PGaussian(r, ; ) =1
2exp
1
2
r
2. (1.34)
It turns out that the Gaussian distribution describes the distribution of random errors in many
kinds of measurements.
P. J. Grandinetti, January 20, 2000
-
8/9/2019 Statistical Methods in Experimental Chemistry
23/103
1.1. RANDOM ERROR DISTRIBUTIONS 23
In the same way that we make the transition from quantized to continuous observables in going
from quantum to classical mechanics we replace the integer r with a continuous parameter x andget
p Gaussian(x, ; ) =1
2exp
1
2
x
2. (1.35)
An important distinction in this transition is that p Gaussian now represents probability density
not probability. Recall from our earlier discussion that to get the probability that a continuous
measured quantity lies within a given range we need to integrate the probability density between
the limits that define the range. If your parent distribution is known to be Gaussian, and if you
choose your confidence limits to be equidistant from the mean, then the confidence limits for a
given probability can be written in the simple form
x = z, (1.36)
where z is a factor that is proportional to the percent confidence desired and is given in table 1.1.
For example, z = 0.67 for 50% confidence; that is, 50% of the total area under the Gaussian curve
lies between 0.67 from the mean . Conversely, we can say that for a given single measurement,x, the true mean will lie within the interval
= x z, (1.37)
-5 -4 -3 -2 - 0 2 3 4 5
0.40/
0.35/
0.30/
0.25/
0.20/
0.15/
0.10/
0.05/
0.00/
x
probabilitydensity
Figure 1.11: Gaussian Distribution.
P. J. Grandinetti, January 20, 2000
-
8/9/2019 Statistical Methods in Experimental Chemistry
24/103
24 CHAPTER 1. STATISTICAL DESCRIPTION OF EXPERIMENTAL DATA
with a percent confidence determined by z.
Bi-dimensional Gaussian Distribution So far weve only discussed univariate data. Lets
look at the case of a bivariate Gaussian distribution. Remember we have to take into account
the covariance (or linear correlation rxy) between the two variables. The functional form of this
distribution is
p Gaussian(x, y) =1
2xy
1 r2xy
exp
12(1 r2xy)
x xx
2+
y y
y
2 2rxy
x x
x
y y
y
. (1.38)
If you take a cross-section through the distribution to obtain, for example, the distribution in
x for a fixed value of y, then the mean of this cross-section distribution will be
x(y) = x + rxy(x/y)(y y), (1.39)
and the standard deviation will be
x(y) = x
1 r2xy. (1.40)
-5 -4 -3 -2 -1 0 1 2 3 4 5
z = (x ) /
Area between z = 0.67 is
50% of total area
+0.67-0.67
0.40/
0.35/
0.30/
0.25/
0.20/
0.15/
0.10/
0.05/
0.00/
probabilitydensity
Figure 1.12: 50% of the area under a Gaussian distribution lies between the limits 0.67 from themean. These limits define the 50% confidence limits for the true value.
P. J. Grandinetti, January 20, 2000
-
8/9/2019 Statistical Methods in Experimental Chemistry
25/103
1.1. RANDOM ERROR DISTRIBUTIONS 25
% confidence z
50 0.6760 0.84
75 1.15
90 1.64
95 1.96
97.5 2.24
99.0 2.58
99.5 2.81
99.95 3.50
Table 1.1: Percent Confidence (i.e., Area under Gaussian distribution) as a function of z values.
Multi-dimensional Gaussian Distribution In the more general case of multi-variate data we
have
p Gaussian(x) =1
(2)n/2|V| exp
1
2(x ) V1 (x )
. (1.41)
Here we use the symbol x to represent not one variable but m variables. For example, instead of the
variables (x,y ,z) we use a 3 dimensional vector x or (x1, x2, x3). In general x is a m-dimensional
vector whose elements (x1, x2, x3, . . . , xm) represent the m variables in our multi-variate data.Likewise is a m-dimensional vector whose elements (x1 , x2 , x3 , . . . , xm) represent the means
of the m variables in our multi-variate data.
To represent the variance, however, we use a m m dimension symmetric (i.e., Vi,j = Vj,i)matrix so that we can include all the covariances between variables. For example, in the bi-variate
case we have
V =
2x 2xy
2xy 2y
. (1.42)
Or in the three variable case we have
V =
21 212
213
212 22
223
213 223
23
. (1.43)
P. J. Grandinetti, January 20, 2000
-
8/9/2019 Statistical Methods in Experimental Chemistry
26/103
26 CHAPTER 1. STATISTICAL DESCRIPTION OF EXPERIMENTAL DATA
x
y
2xy 0
y2
(b)
Mean
Stand.
Dev.
x(y) = x + r(x/y)(y y),
x(y) = x/
1 r2
Figure 1.13: Bivariate Gaussian Distribution.
V1, the inverse of the covariance matrix, is called the curvature matrix, ,
V1 = . (1.44)
We will learn more about the curvature matrix later when we discuss modeling of data.
1.1.3.4 Central Limit Theorem
Our earlier assumption that the random errors in our measurements are governed by a parent dis-
tribution seems to be reasonable in terms of the binomial theorem and the probabilities associated
with quantum measurements. Wouldnt it be great if our measurement uncertainties were domi-
nated by just the uncertainty principle of quantum mechanics! But alas, we also have the workman
outside our building running a jackhammer that sends random vibrations into our instrument,
and/or the carelessness of Homer Simpson at the local power plant sending random voltage fluctu-
ations into our instrument, and/or a number of other seemingly random perturbations leading to
random errors in our measurements. Not being able to eliminate them we are still faced with the
task of determining the parent distribution governing our errors.
One helpful simplification is the Central Limit Theorem. This theorem says that if you
have different random error sources, each with their own parent distribution, then in the limit
that you have a infinite number of random error sources the final parent distribution for your
random errors will be Gaussian, even if none of the individual random error sources have Gaussian
P. J. Grandinetti, January 20, 2000
-
8/9/2019 Statistical Methods in Experimental Chemistry
27/103
1.2. THE 2 TEST OF A DISTRIBUTION 27
parent distributions. An important condition is that the component errors have the same order
of magnitude and that no single source of error dominates all the others. Since you really cantknow if you have a large number of random error sources you should be cautious about assuming
Gaussian errors. Nonetheless, the Gaussian distribution does seem to describe the errors in most
experiments, other than situation where the Poisson distribution should apply.
1.2 The 2 test of a distribution
If we suspect that a given set of observations comes from a particular parent distribution, how can
we test them for agreement with this distribution? For example, if I gave you two dice and ask you
to tell me if the dice are loaded4 what would you do?You could roll the dice a large number of times and make a histogram of the results. You dont
expect that you will roll 12 (i.e., both dice come up as 6) exactly 1/36 times the number of rolls,
but it should be close. In contrast, if the die came up 12 in over half the number of total rolls then
you might suspect that the dice are not obeying the statistics of the binomial distribution. How
much disagreement between the parent (binomial) distribution and our sample distribution can we
reasonably expect? To answer this question, we need a quantitative index of the difference between
the two distributions and a way to interpret this index.
Lets consider a variable x that weve measured N times. We can construct a histogram of
measured x values using a bin width of dx. If the errors in x are governed by a parent distributionp(x) then the expected number of x observations in the range dx is given by
h(xi) = N p(xi)dx.
How do we find the uncertainty in the bin heights? Pay close attention, this part is tricky.
For each bin height h(xi) theres a probability distribution governing the distribution of measured
heights. If we construct many histograms from different groups of measurements, then we would
expect some distribution of bin heights around the expected bin heights. Recall that counting
events are typically governed by a Poisson distribution. So while the spread in x can be governed
by a particular parent distribution, the spread in the frequency of occurance of a particular x valueis governed by the Poisson distribution. Recall that for a Poisson distribution the variance is equal
to the mean of the distribution. Thus we estimate the standard deviation of the bin heights as
i(h(xi)) =
h(xi). (1.45)
4One side of the die is weighted so that the probabilities for any given side coming up are not equal.
P. J. Grandinetti, January 20, 2000
-
8/9/2019 Statistical Methods in Experimental Chemistry
28/103
28 CHAPTER 1. STATISTICAL DESCRIPTION OF EXPERIMENTAL DATA
Now back to our original problem. The experimental histogram can be tested against a parent
distribution using the 2 test:
2 =Nbini
[h(xi) Np(xi)dx]2h(xi)
(1.46)
The 2 test characterizes the dispersion of the observed frequencies from the expected frequencies.
For good agreement between two distributions you might expect that each bin height would differ
from the expected heights by one standard deviation. So we might expect 2 to be equal to Nbin
for good agreement. More specifically, the expectation value for 2 is
2
= = N
bin nc, (1.47)
where is the degrees of freedom, Nbin is the number of sample frequencies, and nc is the number
of constraints or parameters calculated from the data to describe the potential parent distribution
p(x). Simply normalizing the parent distribution to the total number of events N is one constraint.
30.8 30.9 31.0 31.1 31.2 31.3 31.4
8
6
4
2
0
Measured Value
relative
occurance
0 2 4 6 8 10
0.00
0.05
0.10
0.15
0.20
0.25
0.30
0.35
Probability
relative occurance
Figure 1.14: The uncertainty in the number of occurances within a given bin is governed by the
Poisson distribution.
P. J. Grandinetti, January 20, 2000
-
8/9/2019 Statistical Methods in Experimental Chemistry
29/103
1.2. THE 2 TEST OF A DISTRIBUTION 29
If you obtained the mean and variance from your data and used them to describe your parent
distribution then you have used another two constraints.Often you will see the reduced 2 or
2 = 2/, (1.48)
which has an expectation value of2 = 1. Large values for 2 (i.e. 1) indicate poor agreement.Very small (i.e. 1) are also unreasonable, and imply something wrong somewhere.
We can also use the 2 test to compare two experimental distributions to decide whether or not
they were drawn from the same parent distribution. For example, to compare h(xi) and g(xi) you
would calculate
2 =
Nbini
[g(xi)
h(xi)]2
g(xi) + h(xi) . (1.49)
If a given bin contains no occurances for both histograms, then the sum over that bin is skipped
and the number of degrees of freedom (i.e., Nbin) is reduced by one. If the two experimental
distributions are constructed from different number of measurements then you need to scale the
individual distributions. In this case you would calculate
2 =Nbini
[
H/Gg(xi)
G/H h(xi)]2
g(xi) + h(xi)(1.50)
where H = i h(xi) and G = i g(xi). Having to scale the histograms for different number ofmeasurements is another constraint that will reduce the degrees of freedom by one.Interpreting the 2 test result can be simpler if the 2 probability function is used
p(2, ) =1
2/2(/2)(2)(2)/2e
2/2, (1.51)
where (x) is the Gamma function and is given by
(x) =
0euux1du with 0 x . (1.52)
p(2, ) is the distribution of 2 values as a function of the number of degrees of freedom, , and
has a mean of and a variance of 2. Q(2
|) is the probability that another distribution wouldgive a higher (i.e., worse) 2 than your value 20. It is calculated as follows:
Q(20|) =1
2/2(/2)
20
(2)(2)/2e2/2d(2). (1.53)
See table C.4 in Bevington for this function tabulated. If Q(2|) is reasonably close to 1, then theassumed distribution describes the spread of the data points well. If Q(2|) is small the assumed
P. J. Grandinetti, January 20, 2000
-
8/9/2019 Statistical Methods in Experimental Chemistry
30/103
30 CHAPTER 1. STATISTICAL DESCRIPTION OF EXPERIMENTAL DATA
0 5 10 15 20 25 30 350.0
0.1
0.2
0.3
0.4
2
p(2 ,)
= 10
= 1
= 2
= 4
= 8 = 6
Figure 1.15: The 2 distribution for = 1,2,4,6,8, and 10.
parent distribution is not a good estimate of the parent distribution. For example, for a sample
with 10 degrees of freedom ( = 10) and 2 = 2.56 the probability is 99% that another distribution
would give a higher (i.e., worse) 2 (i.e., the distributions agree well). Another example, if ( = 10)and 2 = 29.59 the probability is 0.1% that another distribution would give a higher (i.e., worse)
2 (i.e., the distributions dont agree well).
1.3 Systematic Errors
Systematic errors lead to bias in a measurement technique. Bias is the difference between your
measured mean and the true value
bias = B xt (1.54)
Systematic errors are reproducible, and usually can be attributed to a single source. The sources
of error can be divided into three subclasses:
Instrument Errors Arise from imperfections in the measuring device. For example, uncalibrated
glassware, bad power supply, faulty Pentium chip, ... If these errors are found they can be corrected
by calibration. Periodic calibration of equipment is always desirable because the response of most
P. J. Grandinetti, January 20, 2000
-
8/9/2019 Statistical Methods in Experimental Chemistry
31/103
1.4. GROSS ERRORS 31
30.8 30.9 31.0 31.2 31.3 31.4
8
6
4
2
031.1
Measured Value
Relative
Occurance
31.5
A B
A B
bias
True
Value
Figure 1.16: Method A has no bias so A = xtrue.
instruments changes with time as a result of wear, corrosion, or mistreatment.
Personal Errors Arise from carelessness, inattention, other limitations. For example, missedend points, poor record keeping, ... These errors can be minimized by care and self-discipline. It is
always a good habit to often check instrument readings, notebook entries, and calculations.
Method Errors Arise from non-ideal chemical or physical behavior of system under study. For
example, incomplete reactions, unknown decomposition, system nonlinearities, ... These errors are
the most difficult to detect. Some steps to take are analyses of standard samples, independent
analyses (i.e., another technique that confirms your result), blank determinations which can reveal
contaminant in the instrument.
1.4 Gross Errors
As we mentioned earlier these errors are not a third class of errors. They are either systematic or
random errors. It can be dangerous to reject gross errors (or outliers). There is no universal rule.
A common approach is the Q test. For more details see Rorabacher, Anal. Chem. 63, 139(1991).
P. J. Grandinetti, January 20, 2000
-
8/9/2019 Statistical Methods in Experimental Chemistry
32/103
32 CHAPTER 1. STATISTICAL DESCRIPTION OF EXPERIMENTAL DATA
X1 Xn Xq
d
w
Figure 1.17: Questionable measurements can be rejected if Qexp = d/w is greater than the value of
Qcrit given in table 1.2
If Qexp is greater than Qcrit than xq can be rejected.
Qexp =d
w=
|xq xn||x1 xq| (1.55)
where d is the difference between the questionable result and the nearest neighbor and w is the
spread in the entire data set. Or xq is the questionable result, xn is the nearest neighbor, and x1
is the furthest result from xq. Ideally, the only valid reason for rejecting data from a small set of
data is knowledge that a mistake was made in the measurement.
1.5 Propagation of Errors
If we calculate a result using experimentally measured variables how do we relate the uncertainty
in our experimental variable to the uncertainty in our calculated variable? Recall that the area
under the error distribution curve between the confidence limits is proportional to how confident
we are (i.e., the uncertainty) in our measurements. In general, propagating errors requires us to
Number of Qcrit
observations 90% confidence 95% confidence 99% confidence
3 0.941 0.970 0.994
4 0.765 0.829 0.926
5 0.642 0.710 0.821
6 0.560 0.625 0.7407 0.507 0.568 0.680
8 0.468 0.526 0.634
9 0.437 0.493 0.598
10 0.412 0.466 0.568
Table 1.2: Qcrit values for rejecting data from a small set of data.
P. J. Grandinetti, January 20, 2000
-
8/9/2019 Statistical Methods in Experimental Chemistry
33/103
-
8/9/2019 Statistical Methods in Experimental Chemistry
34/103
34 CHAPTER 1. STATISTICAL DESCRIPTION OF EXPERIMENTAL DATA
map our experimental error distribution curve through our calculation into an error distribution
curve for our calculated variable, and then find the interval in our calculated variable that gives usthe area that corresponds to our desired confidence. In Fig. 1.18 is an example where a Gaussian
experimental error distribution curve is mapped through a linear function to obtain a calculated
error distribution. In this example, the shape of the calculated error distribution is also Gaussian.
In Fig. 1.19 is another example where a Gaussian experimental error distribution curve is
mapped through a nonlinear function to obtain a calculated error distribution that is highly asym-
metric, and cannot be characterized by just a mean and variance. Clearly, it is important to realize
that the shape of the error distribution curve for the calculated variable depends on the functional
relationship between the experimentally measured variables and the calculated variables. This last
example would be one of the worst cases for error propagation. With this caveat, lets look at error
propagation and assume that the error distributions are Gaussian and map through the calculations
without too significant distortions in shape.
Previously we defined the uncertainty in a measured value in terms of its standard deviation.
We can do the same here,
2y(u , v , . . .) = limN
1
N
i
{y(ui, vi, . . .) y(u , v , . . .).}2
(1.56)
Here ui, vi, . . . are the experimental measurements. For example, we measured ui = 2.25, 2.05, 1.85, 1.79,
and 2.12, which have a mean of u = 2.01 and a variance of s2u = 0.03632. What is s2y(u) if
y(u) = 4u + 3? If we wanted to use Eq. 1.56 directly we would first calculate y(ui).
ui yi
2.25 12.00
2.05 11.20
1.85 10.40
1.79 10.16
2.12 11.48
Then we calculate the average y = 11.048 and a standard deviation s2y = 0.581.
Alternatively we could do a Taylor Series expansion of y(u , v , . . .) around y(u , v , . . .) and use it
in our expression for 2y .
y(u , v , . . .) = y(u , v , . . .) + (ui u) dy(u , v , . . .)du
+ (vi v) dy(u , v , . . .)dv
+ + higher-order terms
Assumingdy(u , v , . . .)
duand
dy(u , v , . . .)
dvare not near zero, as they were in our parabola example
earlier, then we can neglect higher-order terms, and write
P. J. Grandinetti, January 20, 2000
-
8/9/2019 Statistical Methods in Experimental Chemistry
35/103
1.5. PROPAGATION OF ERRORS 35
2y(u , v , . . .) = limN
1N
i
(ui u) dy(u , v , . . .)
du+ (vi v) dy(u , v , . . .)
dv+
2.
Expanding the squared term in brackets we obtain
2y(u,v , . . .) = lim
N
1
N
i
(ui u)
2
dy
du
2+ (vi v)
2
dy
dv
2+ + 2(ui u)(vi v)
dy
du
dy
dv
+ .
.
Breaking this into individual sums we have
2y(u , v , . . .) = limN
1
Ni(ui u)2
dy
du2
+ limN
1
Ni(vi v)2
dy
dv2
+
+ limN
1
N
i
2(ui u)(vi v)
dy
du
dy
dv
+ .
Finally we can write
2y(u , v , . . .) = 2u
dy
du
2+ 2v
dy
dv
2+ + 22uv
dy
du
dy
dv
+ . (1.57)
This is the error propagation equation. If ui, vi, . . . are all uncorrelated then 2uv = 0, and the
error propagation equation simplifies to
2y(u , v , . . .) = 2u
dy
du
2+ 2v
dy
dv
2+ . (1.58)
Thus, in our earlier example where y = 4u + 3, we can use the error propagation equation to
relate the variance in u to the variance in y. That is,
s2y = 16 s2u = 0.581.
Now lets look at some specific functions y(u , v , . . .) and see how the error propagation equation
is applied.
Simple Sums and Differences:
y = u a here a = constant.dy
du= 1
2y = 2u
P. J. Grandinetti, January 20, 2000
-
8/9/2019 Statistical Methods in Experimental Chemistry
36/103
36 CHAPTER 1. STATISTICAL DESCRIPTION OF EXPERIMENTAL DATA
Weighted Sums and Differences:
y = au bv here a, b = constants.
dy
du= a and
dy
dv= b
2y = a22u + b
22v 2ab2uv
Multiplication and Division:
y = auw
dydu
= aw and dydw
= au
2y = a2w22u + a
2u22w + 2a2uw2uw
If we divide both sides by y2 we get something that might be more familiar, that is
yy
2=
uu
2+
ww
2+ 2
uwuw
2
y/y is called the relative error in y, while y is the absolute error in y.
Powers:
y = aub
dy
du= abu(b1) = by
u
2y =2uu2
b2y2 or
yy
2= b2
2uu2
Exponentials:
y = aebu
dy
du= abebu = by
2y = 2u b2y2 or
yy
2= b22u
P. J. Grandinetti, January 20, 2000
-
8/9/2019 Statistical Methods in Experimental Chemistry
37/103
1.6. CONFIDENCE LIMITS 37
Logarithms:
y = a ln(bu)dy
du= a
bu(b) = a
u
2y = 2u
a2
u2= a2
2uu2
1.6 Confidence Limits
As we learned earlier, once you know the parent distribution governing the errors in your measure-
ments you can calculate the probability that a given measurement will lie within a certain intervalof possible values. That is,
P(x, x+) =
x+x
p(x) dx, (1.59)
where P(x, x+) is the probability that the measured value x lies between x and x+. The interval
between x and x+ is called the confidence interval and x and x+ are called confidence
limits.
Based on the Central Limit theorem we can often assume that our parent distribution is Gaus-
sian. In certain situations you may also know the variance of this Gaussian distribution but not the
mean. For example, most analytical balance manufacturers will print on the front of the balance
the standard deviation () for any measurement made with the balance. Assuming the distributionof errors are Gaussian, then given a single measurement, x, we can estimate that the true mean
to lie within the interval
= x z, (1.60)with a percent confidence determined by z, as given in Table 1.1.
If we performed the measurement twice we can use the average of the two measurements as a
better estimate for , but how do we determine the confidence limits in this case? We use the error
propagation equation. Given
x =x1 + x2
2, (1.61)
we can calculate the variance in x given the variances in x1 and x2,
2x =
1
2
221 +
1
2
222. (1.62)
Assuming 21 = 22 =
2 we obtain
2x = 2/2 or x = /
2. (1.63)
P. J. Grandinetti, January 20, 2000
-
8/9/2019 Statistical Methods in Experimental Chemistry
38/103
38 CHAPTER 1. STATISTICAL DESCRIPTION OF EXPERIMENTAL DATA
p(t ,)
= 10
= 1
= 2
= 4
= 3
= 6
-6 -5 -4 -3 -2 -1 0 1 2 3 4 5 60.00
0.05
0.10
0.15
0.20
0.25
0.30
0.35
0.40
0.45
t
Gaussian
Figure 1.20: Students t distribution for = 1,2,3,4,6, and 10. The unit Gaussian is also shown.
In general, after N measurements the standard deviation of the mean is given by
x = /
N. (1.64)
Therefore the confidence limits of the mean after N measurements is given by
= x z/
N. (1.65)
Note that this applies only if there are no systematic errors (or bias).
1.6.1 Students t-distribution
A problem with these equations is that they assume we already know , which requires an infinite
number of measurements. For a finite number of measurements we only have the experimental
variance s2. In the case where our distribution was Gaussian with variance 2, then we knew that
the distribution in z = ( x)/ had a zero mean and unit variance. If we know our distributionis Gaussian but only know s2, then we need to know the distribution in t = ( x)/s. Thisdistribution was derived by William Gossett and is called Students t distribution. It is
p(t, ) =((+ 1)/2)
(/2) 1
1 + (t2/))(+1)/2(1.66)
P. J. Grandinetti, January 20, 2000
-
8/9/2019 Statistical Methods in Experimental Chemistry
39/103
1.6. CONFIDENCE LIMITS 39
Degrees of Freedom
= N 1 t(95%) t(99%)1 12.7 63.7
2 4.30 9.92
3 3.18 5.84
4 2.78 4.60
5 2.57 4.03
6 2.45 3.71
7 2.36 3.50
10 2.23 3.17
30 2.04 2.7560 2.00 2.66
1.96 2.58
Table 1.3: Students t values as a function of for 95 and 99 Percent Confidence (i.e., 95 and 99
Percent of the Area under Students t distribution).
In contrast to the z-distribution this expression depends on t and , the number of degrees of
freedom. As expected, in the limit that N then t z. In table 1.3 are the values for t as afunction of the degrees of freedom (i.e., = N
1) for the 95% and 99% confidence limits. Using
Students t distribution we can use the average of N measurements and estimate the true mean
to lie within the interval
= x ts/
N, (1.67)
with a percent confidence determined by t and , as given in Table 1.3. From this table you will
notice that for a given confidence limit the t inflates the confidence limits to take into account a
finite number of measurements.
Lets look at an example. A chemist obtained the following data for the alcohol content of a
sample of blood: 0.084%, 0.089%, and 0.079%. How would he calculate the 95% confidence limits
for the mean assuming there is no additional knowledge about the precision of the method? First
he would calculate the mean
x = (0.084% + 0.089% + 0.079%)/3 = 0.084%,
and then the standard deviation
s =
(0.084% 0.084%)2 + (0.089% 0.084%)2 + (0.079% 0.084%)2
2= 0.0050%.
P. J. Grandinetti, January 20, 2000
-
8/9/2019 Statistical Methods in Experimental Chemistry
40/103
40 CHAPTER 1. STATISTICAL DESCRIPTION OF EXPERIMENTAL DATA
Then for 2 degrees of freedom he uses t = 4.30 to calculate the 95% confidence limits for the mean
as = 0.084% 0.012%.
How would the calculation change if on the basis of previous experience he knew that s =0.005% (i.e., he know the parent distribution variance)? In this case he would use z = 1.96 instead
of t = 4.30, and would calculate
= 0.084% 0.006%for the 95% confidence limits.
1.7 The Two Sample Problem
1.7.1 The t-test
1.7.1.1 Comparing a measured mean to a true value
The best way to detect a systematic error in an experimental method of analysis is to analyze
a standard reference material and compare the results with the known true value. If there is a
difference between your measured mean and the true value how would you know if its due to
random error or to some reproducible systematic error? Using the expression
= x
ts/
N (1.68)
we can rewrite it as
x = ts/
N Critical Value
. (1.69)
If | x| is less than or equal to ts/N then the difference is statistically insignificant at thepercent confidence determined by t. Otherwise the difference will be statistically significant.
A procedure for determination of sulfur in kerosenes was tested on a sample known from
preparation to contain 0.123% S. The results were %S= 0.112, 0.118, 0.115, and 0.119.
Is there a systematic error (or bias) in the method?
The first step is the calculate the mean.
x =0.112 + 0.118 + 0.115 + 0.119
4= 0.116%S
Next we calculate the difference between our mean and the true value
x = 0.116 0.123 = 0.007%S.
P. J. Grandinetti, January 20, 2000
-
8/9/2019 Statistical Methods in Experimental Chemistry
41/103
1.7. THE TWO SAMPLE PROBLEM 41
The experimental standard deviation is
s =
(0.112 0.116)2 + (0.118 0.116)2 + (0.115 0.116)2 + (0.119 0.116)23
= 0.0032.
At 95% confidence t has a value of 3.18 for N = 4. The critical value is
ts/
N = (3.18)(0.0032%)/
4 = 0.0051%S.
Since the difference between our mean and the true value (0.007% S) is greater than
the critical value (0.0051% S) we conclude that a systematic error is present with 95%
confidence.
1.7.1.2 Comparing two measured means
Often chemical analyses are used to determine whether two materials are identical. Then the
question arises . . . How can you know if the difference in the mean of two sets of (allegedy identical)
analyses is either real and constitutes evidence that the samples are different or simply due to
random errors in the two data sets.
To solve this problem we use the same approach as in the last section. The difference between
the two experimental means can be compared to a critical value to determine if the difference is
statistically significant.
x1
x2 =
tspooled
N1 + N2
N1N2 Critical Value
, (1.70)
where
spooled =
s21(n1 1) + s22(n2 1)
n1 + n2 2 . (1.71)
As before if |x1 x2| is less than or equal to the critical value then the difference is probablydue to random error, i.e., the sample are probably the same. If, however, |x1 x2| is greater thanthe critical value then the difference is probably real, i.e., the samples are not identical.
Two wine bottles were analyzed for alcohol content to determine whether they were
from difference sources. Six analyses of the first bottle gave a mean alcohol contentof 12.61% and s = 0.0721%. Four analyses of the second bottle gave a mean alcohol
content of 12.53% and s = 0.0782%.
First we calculate spooled.
spooled =
5 (0.0721)2 + 3 (0.0782)2
8= 0.0744%.
P. J. Grandinetti, January 20, 2000
-
8/9/2019 Statistical Methods in Experimental Chemistry
42/103
42 CHAPTER 1. STATISTICAL DESCRIPTION OF EXPERIMENTAL DATA
At 95% confidence t = 2.31 (for 8 degrees of freedom, i.e., (6 1 ) + ( 4 1) = 8). Then
tspooled
N1 + N2
N1N2= (2.31)(0.0744%)
6 + 4
6 4 = 0.111%.
Since x1 x2 = 12.61 12.53 = 0.08% we conclude that at the 95% confidence levelthere is no difference in the alcohol content of the two wine bottles.
1.7.2 Comparing Variances - The F-test
Sometimes you need to compare two measured variances. For example, is one method of analysis
is more precise than another? Or, are the differences between the variances for the two methods of
analysis statistically significant?To answer these types of questions we use the F-test. F is the ratio of the variance of sampling
A with A = NA 1 degrees of freedom and sampling B with B = NB 1 degrees of freedom,that is
Fexp =s2As2B
. (1.72)
In the limit of A and B then
F =2A2B
. (1.73)
p (F)
0 1 2 3 4 5 6 7 8 9 100.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
F
( = 3, = 5)
( = 10, = 12)
Figure 1.21: F distribution for A = 3 and B = 5.
P. J. Grandinetti, January 20, 2000
-
8/9/2019 Statistical Methods in Experimental Chemistry
43/103
1.7. THE TWO SAMPLE PROBLEM 43
An F value of 1 indicates that the two variances are the same, whereas F values differing from one
indicates that the two variances are different. When A and B are finite we will, of course, haverandom errors such that even if the variances are identical we would often expect to obtain Fexp
values different from 1. How likely is it that Fexp differs from one when A = B? That depends
on the shape of the distribution of F values. The parent distribution for F values is given by
p(F, A, B) =((A + B)/2)
(A/2)(B/2)
AA BB
F(A/2)1
(B + AF)(A+B)/2). (1.74)
A plot of this distribution is given in Fig. 1.21. As you can see, for A = 10 and B = 12 an F value
of 2 is not that unlikely. The basic approach, then, is to integrate the F distribution from infinity
down to some critical F value to define the range of F values that cover, say the top 5% worst
F-values. It is common to use tables that will give you the critical F value (limit) that correspondsto a given probability. In in Tables 1.5 and 1.4 are tables of critical F values that correspond to
2.5% and 5% significance.
Sometimes you will see two types of F tables called the one- and two-tailed F distribution
tables. If you are asking if the variance in A is greater than the variance in B? or if the variance
in B is greater than the variance in A? then you will use the one-tailed F distribution table. On
the other hand, if you are asking if the variance in A is not equal to the variance in B, then you
would use the two-tailed distribution table. In most reference texts, however, you will only find
the one-tailed F distribution table because the values for the two-tailed F distribution table can be
obtained from the one-tailed table. That is,
Pone-tailed = Ptwo-tailed/2. (1.75)
That is, the one tailed F table for 5% probability is the same as the two tailed F table for 10%
probability.
Lets look at an example where we ask if the variance in A is not equal to the variance in B.
Two chemists determined the purity of a crude sample of ethyl acetate. Chemist A made
6 measurements with a variance ofs2A = 0.00428 and Chemist B made 4 measurements
with a variance of s2B = 0.00983. Can we be 95% confident that these two chemists
differ in their measurement precision?In this case we are not asking if A is better than B, but only if A and B are different.
Therefore we use the two-tailed F distribution table. When calculating the experimental
ratio Fexp, the convention is to define s2A to be the large of the two variances s
2A and
s2B. With this definition Fexp will always be greater than 1. So we have
Fexp = 0.00983/0.00428 = 2.30.
P. J. Grandinetti, January 20, 2000
-
8/9/2019 Statistical Methods in Experimental Chemistry
44/103
44 CHAPTER 1. STATISTICAL DESCRIPTION OF EXPERIMENTAL DATA
From the two-tailed table with A = 3, B = 5, and P = 0.05 we find Fcrit = 7.764.
Since 2.30 < 7.764 we conclude that the F test shows no difference at the 95% confidencelevel.
Lets look at another example.
A water sample is analyzed for silicon content by two methods, one is supposed to
have improved precision. The original method gives xA = 143.6 and s2A = 66.8 after 5
measurments. The newer method gives xB = 149.0 and s2B = 8.50 after 5 measurments.
Is the new method an improvement over the original at 95% confidence (i.e., less than
5% chance that theyre the same.).
First we calculate the F value from the measurements.
Fexp = 66.8/8.5 = 7.86.
Then we use the one-tailed table and find Fcrit = 6.388 for A = 5 1 = 4 andB = 5 1 = 4. Since 7.86 > 6.388 we conclude that we are 95% confident that thenew method is better than the original.
1 =1 2 3 4 5 6 7 8 9 10 12 15 20
2
1 161.4 199.5 215.7 224.6 230.2 234.0 236.8 238.9 240.5 241.9 243.9 245.9 248.0 254.32 18.51 19.00 19.16 19.25 19.30 19.33 19.35 19.37 19.38 19.40 19.41 19.43 19.45 19.50
3 10.13 9.552 9.277 9.117 9.013 8.941 8.887 8.845 8.812 8.786 8.745 8.703 8.660 8.53
4 7.709 6.944 6.591 6.388 6.256 6.163 6.094 6.041 5.999 5.964 5.912 5.858 5.803 5.63
5 6.608 5.786 5.409 5.192 5.050 4.950 4.876 4.818 4.772 4.735 4.678 4.619 4.558 4.36
6 5.987 5.143 4.757 4.535 4.387 4.284 4.207 4.147 4.099 4.060 4.000 3.938 3.874 3.67
7 5.591 4.737 4.347 4.120 3.972 3.866 3.787 3.726 3.677 3.637 3.575 3.511 3.445 3.23
8 5.318 4.459 4.066 3.838 3.687 3.581 3.500 3.438 3.388 3.347 3.284 3.218 3.150 2.93
9 5.117 4.256 3.863 3.633 3.482 3.374 3.293 3.230 3.179 3.137 3.073 3.006 2.936 2.71
10 4.965 4.103 3.708 3.478 3.326 3.217 3.135 3.072 3.020 2.978 2.913 2.845 2.774 2.54
12 4.747 3.885 3.490 3.259 3.106 2.996 2.913 2.849 2.796 2.753 2.687 2.617 2.544 2.30
15 4.543 3.682 3.287 3.056 2.901 2.790 2.707 2.641 2.588 2.544 2.475 2.403 2.328 2.07
20 4.351 3.493 3.098 2.866 2.711 2.599 2.514 2.447 2.393 2.348 2.278 2.203 2.124 1.84
Table 1.4: Critical values of F for a one-tailed test P = 0.05 (or two-tailed test with P = 0.10). 1
is the number of degrees of freedom of the numerator and 2 is the number of degrees of freedom of
the denominator.
P. J. Grandinetti, January 20, 2000
-
8/9/2019 Statistical Methods in Experimental Chemistry
45/103
1.7. THE TWO SAMPLE PROBLEM 45
1 =1 2 3 4 5 6 7 8 9 10 12 15 20
2
1 647.8 799.5 864.2 899.6 921.8 937.1 948.2 956.7 963.3 968.6 976.7 984.9 993.1 1018
2 38.51 39.00 39.17 39.25 39.30 39.33 39.36 39.37 39.39 39.40 39.41 39.43 39.45 39.50
3 17.44 16.04 15.44 15.10 14.88 14.73 14.62 14.54 14.47 14.42 14.34 14.25 14.17 13.90
4 12.22 10.65 9.979 9.605 9.364 9.197 9.074 8.980 8.905 8.844 8.751 8.657 8.560 8.26
5 10.01 8.434 7.764 7.388 7.146 6.978 6.853 6.757 6.681 6.619 6.525 6.428 6.329 6.02
6 8.813 7.260 6.599 6.227 5.988 5.820 5.695 5.600 5.523 5.461 5.366 5.269 5.168 4.85
7 8.073 6.542 5.890 5.523 5.285 5.119 4.995 4.899 4.823 4.761 4.666 4.568 4.467 4.14
8 7.571 6.059 5.416 5.053 4.817 4.652 4.529 4.433 4.357 4.295 4.200 4.101 3.999 3.67
9 7.209 5.715 5.078 4.718 4.484 4.320 4.197 4.102 4.026 3.964 3.868 3.769 3.667 3.33
10 6.937 5.456 4.826 4.468 4.236 4.072 3.950 3.855 3.779 3.717 3.621 3.522 3.419 3.08
12 6.554 5.096 4.474 4.121 3.891 3.728 3.607 3.512 3.436 3.374 3.277 3.177 3.073 2.72
15 6.200 4.765 4.153 3.804 3.576 3.415 3.293 3.199 3.123 3.060 2.963 2.862 2.756 2.40
20 5.871 4.461 3.859 3.515 3.289 3.128 3.007 2.913 2.837 2.774 2.676 2.573 2.464 2.09
Table 1.5: Critical values of F for a one-tailed test P = 0.025 (or two-tailed test with P = 0.05).
1 is the number of degrees of freedom of the numerator and 2 is the number of degrees of freedom
of the denominator.
P. J. Grandinetti, January 20, 2000
-
8/9/2019 Statistical Methods in Experimental Chemistry
46/103
46 CHAPTER 1. STATISTICAL DESCRIPTION OF EXPERIMENTAL DATA
1.8 Problems
1. Why is it often assumed that experimental random errors follow a Gaussian parent distribu-
tion?
2. In what type of experiments would you expect random errors to follow a Poisson parent
distribution?
3. Indicate whether the Skewness and Kurtosis are positive, negative, or zero for the following
distribution.
RelativeFrequency
Measured Value
4. Describe the three classifications of systematic errors and what can be done to minimize eachone.
5. Describe the Q-test. When is it reasonable to reject data from a small set of data?
6. (a) Give a general explanation of confidence limits. (b) How would you find the 99% confidence
limits for a measured parameter that had the distribution in question 3 above?
7. The error propagation equation is
2
y(u , v , . . .) = 2
u yu2
+ 2
v yv2
+ 2
uv yuyv + Explain the assumptions involved in using this equation to relate the variances in u, v, . . . to
the variance in y.
8. The mean is defined as the first moment about the origin. Explain why the variance is defined
as the second moment about the mean and not the second moment about the origin.
P. J. Grandinetti, January 20, 2000
-
8/9/2019 Statistical Methods in Experimental Chemistry
47/103
1.8. PROBLEMS 47
9. An negatively charged oxygen is coordinated by 4 potassium cations. 50% of the potassium
cations are ion exchanged out and replaced by sodium cations. Assuming that the sodi-ums simply replace potassium cations randomly, what is the probability that the oxygen is
coordinated by four potassiums in the ion exchanged material?
10. We discussed two equations for obtaining the confidence limits for the mean after N mea-
surements of a variable. They were
= x zN
,
and
= x tsN
.
If you had measured a concentration in triplicate, which equation would you use and why?
Describe a situation where you could correctly use the other equation for your triplicate
measurements.
11. In the 2 test of a distribution explain what value is used for the variance of the bin heights
and why.
12. In your own words describe the Central Limit theorem and one experimental situation where
the experimental distribution of noise is not described by this theorem.
13. Given the mean and standard deviation from the parent distribution in question 3, would
it be correct if you calculated the 50% confidence limits for a single measurement using the
equation = x .67? Explain why or why not.
14. Draw a distribution that has a positive skewness.
15. Draw a distribution that is leptokurtic.
16. If the probability of getting an A in this course is 38%, what is the probability that 15 people
will receive an A if the class has 30 students enrolled?
17. (a) Explain the conditions under which the Poisson distribution is a good approximation for
the binomial distribution. (b) Explain the conditions under which the Gaussian distribution
is a good approximation for the binomial distribution.
18. If you had to guess which parent distribution governs the errors in your experiment before
you made a single measurement, which distribution would you guess and why?
P. J. Grandinetti, January 20, 2000
-
8/9/2019 Statistical Methods in Experimental Chemistry
48/103
48 CHAPTER 1. STATISTICAL DESCRIPTION OF EXPERIMENTAL DATA
19. When using the chi-squared test to compare distributions A and B below, how many degrees
of freedom would you use to calculate 2 and why?
8
6
4
2
0
(A) (B)
Probability
Density
30.8 30.9 31.0 31.1 31.2 31.3 31.4
8
6
4
2
0
Probability
Density
30.8 30.9 31.0 31.1 31.2 31.3 31.4
20. Given the following measurements 2.12, 2.20, 2.12, 2.39, and 2.25, use the Q-test and deter-
mine which, if any, of these numbers can be rejected at the 90% confidence level.
21. Seven measurements of the pH of a buffer solution gave the following results:
5.12, 5.20, 5.15, 5.17, 5.16, 5.19, 5.15
Calculate (i) the 95% and (ii) 99% confidence limits for the true pH. (Assume that there are
no systematic errors.)
22. The solubility product of barium sulfate is 1.31010, with a standard deviation of 0.11010.Calculate the standard deviation of the calculated solubility of barium sulfate in water.
23. A 0.1 M solution of acid was used to titrate 10 mL of 0.1 M solution of alkali and the following
volumes of acid were recorded:
9.88, 10.18, 10.23, 10.39, 10.25 mL
Calculate the 95% confidence limits of the mean and use them to decide if there is any evidence
of systematic error.
24. Seven measurements of the concentration of thiol in a blood lysate gave:
1.84, 1.92, 1.94, 1.92, 1.85, 1.91, 2.07
Verify that 2.07 is not an outlier.
P. J. Grandinetti, January 20, 2000
-
8/9/2019 Statistical Methods in Experimental Chemistry
49/103
1.8. PROBLEMS 49
25. The following data give the recovery of bromide from spiked samples of vegetable matter,
measured by using a gas-liquid chromatographic method. The same amount of bromide wasadded to each specimen.
Tomato: 777 790 759 790 770 758 764 g/g
Cucumber: 782 773 778 765 789 797 782 g/g
(a) Test whether the recoveries from the two vegetables have variances which differ signifi-
cantly.
(b) Test whether the mean recovery rates differ significantly.
26. Calculate the probability of a rolling a one exactly 120 times in 720 tosses of a die.
27. Explain the circumstances under which the Students t distribution is used, and its relation-
ship to the Gaussian distribution.
28. (a) Give a general explanation of confidence limits. (b) How would you find the 99% confidence
limits for a measured parameter that had the distribution in question 3?
29. Give a real life example of a bivariate data set expected to have a linear correlation coefficient
rxy = 2xyxy near zero and an example o