Chapter 3: Analysis of

Experimental Data
 3.1 Introduction

• Our discussion in this chapter will consider the

analysis of data to determine errors, precision, and
general validity of experimental measurements.
• The experimentalist should always know the validity
of data.
• The automobile test engineer must know the
accuracy of the speedometer and gas gage in order
to express the fuel-economy performance with
confidence.
• A nuclear engineer must know the accuracy and
precision of many instruments just to make some
simple radioactivity measurements with confidence.
• In order to specify the performance of an amplifier,
an electrical engineer must know the accuracy with
which the appropriate measurements of voltage,
distortion, and so forth, have been conducted.
 3.1 Introduction

• Errors will creep into all experiments regardless of

the care exerted.
• Some of these errors are of a random nature, and
some will be due to gross blunders on the part of
the experimenter.
• Bad data due to obvious blunders may be discarded
immediately.
• But what of the data points that just “look” bad?
• We cannot throw out data because they do not
conform with our hopes and expectations unless we
see something obviously wrong.
• If such “bad” points fall outside the range of
normally expected random deviations, they may be
discarded on the basis of some consistent statistical
data analysis.
• The keyword here is “consistent.” The elimination of
data points must be consistent
 3.2 Causes and Types of Experimental Errors

• First, let us distinguish between single-sample and

multisample data.

• Single-sample data are those in which some

uncertainties may not be discovered by repetition.

• Multisample data are obtained in those instances

where enough experiments are performed so that
the reliability of the results can be assured by
statistics.

• Frequently, cost will prohibit the collection of

multisample data, and the experimenter must be
content with single-sample data and prepared to
extract as much information as possible from such
experiments.
 3.2 Causes and Types of Experimental Errors

• If one measures pressure with a pressure gage and

a single instrument is the only one used for the
entire set of observations,
• then some of the error that is present in the
measurement will be sampled only once no matter
how many times the reading is repeated.
• Consequently, such an experiment is a single-
sample experiment.
• On the other hand, if more than one pressure gage
is used for the same total set of observations,
• then we might say that a multisample experiment
has been performed.
• The number of observations will then determine the
success of this multisample experiment in
accordance with accepted statistical principles.
 3.2 Causes and Types of Experimental Errors

• types of errors that may cause uncertainty in an

experimental measurement.
• gross blunders in apparatus or instrument construction
which may invalidate the data.
• fixed errors (or systematic errors) which will cause
repeated readings to be in error by roughly the same
amount but for some unknown reason.
• random errors, which may be caused by personal
fluctuations, random electronic fluctuations in the
apparatus or instruments, various influences of friction,
and so forth.
• Random errors usually follow a certain statistical
distribution, but not always.
• In many instances it is very difficult to distinguish
between fixed errors and random errors.
 3.3 Error Analysis on a Commonsense Basis

• Suppose that we have satisfied ourselves with the

uncertainty in some basic experimental
measurements, taking into consideration:
• instrument accuracy,
• competence of the people using the instruments etc.

• Eventually, the primary measurements must be

combined to calculate a particular result that is
desired.

• We shall be interested in knowing the uncertainty in

the final result due to the uncertainties in the
primary measurements.

• This may be done by a commonsense analysis of

the data which may take many forms.
 3.3 Error Analysis on a Commonsense Basis

• One rule of thumb that could be used is that the error in

the result is equal to the maximum error in any
parameter used to calculate the result.
• Another commonsense analysis would combine all the
errors in the most detrimental way in order to determine
the maximum error in the final result
• Consider the calculation of electric power fromP = EI
• where E and I are measured as
• E = 100 V ± 2 V
• I = 10 A ± 0.2 A
• The nominal value of the power is 100×10 = 1000W.
• By taking the worst possible variations in voltage and
current, we could calculate
• Pmax = (100 + 2)(10 + 0.2) = 1040.4 W
• Pmin = (100 − 2)(10 − 0.2) = 960.4 W
 3.3 Error Analysis on a Commonsense Basis

• Thus, using this method of calculation, the

uncertainty in the power is +4.04 percent, −3.96
percent.

• It is quite unlikely that the power would be in error

by these amounts because the voltmeter variations
would probably not correspond with the ammeter
variations.

• When the voltmeter reads an extreme “high,” there

is no reason that the ammeter must also read an
extreme “high” at that particular instant;

• indeed, this combination is most unlikely.

 3.3 Error Analysis on a Commonsense Basis

• The simple calculation applied to the electric-power

equation above is a useful way of inspecting
experimental data to determine what errors could
result in a final calculation;

• however, the test is too severe and should be used

only for rough inspections of data.

• It is significant to note, however, that if the results of

the experiments appear to be in error by more than
the amounts indicated by the above calculation,
then the experimenter had better examine the data
more closely.
 3.3 Error Analysis on a Commonsense Basis

• As another example we might conduct an

experiment where heat is added to a container of
water.

• If our temperature instrumentation should indicate a

drop in temperature of the water, our good sense
would tell us that something is wrong

• And the data point(s) should be thrown out.

• No sophisticated analysis procedures are necessary

to discover this kind of error.
 3.4 Uncertainty Analysis and Propagation
of Uncertainty
• A more precise method

• a certain pressure reading might be expressed as

• p = 100 kPa ± 1 kPa

• When the plus or minus notation is used to

designate the uncertainty, the person making this
designation is stating the degree of accuracy with
which he or she believes the measurement has
been made.
 3.4 Uncertainty Analysis and Propagation
of Uncertainty
• To add a further specification of the uncertainty of a
particular measurement, the experimenter specify
certain odds for the uncertainty.

• The above equation for pressure might thus be

written

• p = 100 kPa ± 1 kPa (20 to 1)

• In other words, the experimenter is willing to bet

with 20 to 1 odds that the pressure measurement is
within ±1 kPa.

• It is important to note that the specification of such

odds can only be made by the experimenter based
on the total laboratory experience.
 3.4 Uncertainty Analysis and Propagation
of Uncertainty
• Suppose a set of measurements is made and the
uncertainty in each measurement may be
expressed with the same odds.

• These measurements are then used to calculate

some desired result of the experiments.

• We wish to estimate the uncertainty in the

calculated result on the basis of the uncertainties in
the primary measurements.

• The result R is a given function of the independent

variables x1, x2, x3, . . . , xn. Thus,
 3.4 Uncertainty Analysis and Propagation
of Uncertainty
• Let wR be the uncertainty in the result and w1,w2, . . . ,
wn be the uncertainties in the independent variables.

• If the uncertainties in the independent variables are all

given with the same odds, then the uncertainty in the
result having these odds is
 3.4 Uncertainty Analysis and Propagation
of Uncertainty
• In many cases the result function of above Eq. takes the
form of a product of the respective primary variables raised
to exponents and expressed as

• When the partial differentiations are performed, we obtain

• Dividing by R

• Inserting this relation in Eq given in previous slide gives

 3.4 Uncertainty Analysis and Propagation
of Uncertainty
• When the result function has an additive form, R will be
expressed as

• When the partial differentiations are performed, we obtain

• The uncertainty in the result may then be expressed as

 3.4 Uncertainty Analysis and Propagation
of Uncertainty
• uncertainty propagation in the result wR depends on the
squares of the uncertainties in the independent variables
wn.

• if the uncertainty in one variable is significantly larger

than the uncertainties in the other variables, then it is the
largest uncertainty that predominates

• and the others may probably be neglected.

• Ex: suppose there are three variables with a product of

sensitivity and uncertainty [(∂R/∂x)wx] of magnitude 1,
and one variable with a magnitude of 5.

• (52 + 12 + 12 + 12)1/2 =√28 = 5.29

 3.6 Statistical Analysis of Experimental Data

• it is important to define some pertinent terms

• When a set of readings of an instrument is taken, the

individual readings will vary somewhat from each other,

• The experimenter may be concerned with the mean of all

the readings.

• If each reading is denoted by xi and there are n readings,

the arithmetic mean is given by

• The deviation di for each reading is defined by

 3.6 Statistical Analysis of Experimental Data

• We may note that the average of the deviations of all the

readings is zero since:

• The average of the absolute values of the deviations is

given by
• Note that this quantity is
not necessarily zero.
• The standard deviation or root-mean-square deviation is

• σ2 is called the variance

• This is sometimes called the population or biased
standard deviation because it strictly applies only when a
large number of samples is taken to describe the
population.
 3.6 Statistical Analysis of Experimental Data

• In many circumstances the engineer will not be able to

collect as many data points as necessary to describe the
underlying population.

• Generally speaking, it is desired to have at least 20

measurements in order to obtain reliable estimates of
standard deviation.

• For small sets of data an unbiased or sample standard

deviation is defined by
 3.6 Statistical Analysis of Experimental Data

• There are other kinds of mean values of interest from

time to time in statistical analysis.

• The median is the value that divides the data points in

half.

• For example, if measurements made on five production

resistors give 10, 12, 13, 14, and 15 kΩ,

• The median value would be 13 kΩ.

• The arithmetic mean, however, would be

• Rm = (10 + 12 + 13 + 14 + 15)/5= 12.8 kΩ

 3.6 Statistical Analysis of Experimental Data

• Sometimes it is appropriate to use a geometric mean

when studying phenomena which grow in proportion to
their size.

• This would apply to certain biological processes and to

growth rates in financial resources.

• The geometric mean is defined by

 3.6 Statistical Analysis of Experimental Data

• As an example of the geometric mean, consider the 5-

year record of a mutual fund investment:

• The average growth rate is therefore

• Average growth = [(0.89)(1.1124)(1.1111)(1.1364)]1/4
• = 1.0574
• To see that this is indeed a valid average growth rate, we
can observe that
• (1000)(1.0574)4 = 1250
 3.7 Probability Distribution

• Probability
• The probability that one will get a head when flipping an
coin is ½
• The probability that a tail will occur is also ½
• The probability that either a head or a tail will occur is ½+ ½
or unity
 3.7 Probability Distribution

• Suppose two dice are thrown and we wish to know the

probability that both will display a 6.
• The probability for a 6 on a single die is 1/6
• By a short listing of the possible arrangements it can be
seen that there can be 36 possibilities
• two 6s represents only one of these possibilities1/36

• For a throw of 7 there are 6 possible ways

• the probability of getting a 7 is 6/36 or 1/6

• There are only 2 ways of getting an 11;

• thus, the probability is 2/36 or 1/18 .

• The probability of getting either a 7 or an 11 is 6/36+ 2/36

 3.7 Probability Distribution

• If several independent events occur at the same time such

that each event has a probability pi, the probability that all
events will occur is given as the product of the probabilities
of the individual events.

• This rule could be applied to the problem of determining the

probability of double 6,
• This reasoning could not be applied to the problem of
obtaining a 7 on the two dice
• because the number on each die is not independent of the number
on the other die,
 3.7 Probability Distribution

• Suppose we toss a horseshoe some distance x

• we would not always meet with success to toss it to the
same distance
• On the first toss the horseshoe might travel a distance x1,
on the second toss a distance of x2
• The curve shown in figure is
called a probability distribution.
• It shows how the probability of
success in a certain event is
distributed over the distance x.
 3.7 Probability Distribution

• Each value of the ordinate p(x) gives the probability that

the horseshoe will land between x and x+Δx
• We might consider the deviation from xm as the error in
the throw.
• The area under the curve is unity since it is certain that
• the horseshoe will land somewhere.
• We should note that more
than one variable may be
present in a probability
distribution.

• n: number of success
• N: number of events
• p: probability of success
 3.7 Probability Distribution

• When a limited number of observations is made and the

raw data are plotted, we call the plot a histogram.
• For example, the following distribution of throws might be
observed for a horseshoes player:

• Histogram with • Histogram with

Δx = 10 cm. Δx = 20 cm.
 3.7 Probability Distribution

• A cumulative frequency diagram could be employed for

these data, as shown in below figure.
• If this figure had been constructed on the basis of a very
large number of throws,
• Then we could appropriately refer to the ordinate as the
probability that the horseshoe will land within a distance x
of the target.

• Cumulative frequency diagram.

 3.8 The Gaussian or Normal Error Distribution

• Suppose an experimental observation is made and some

particular result is recorded.
• We know that the observation has been subjected to many
random errors.
• These random errors may make the final reading either too
large or too small
• Assuming that
• there are many small errors that contribute to the final error
• each small error is of equal magnitude
• each small error is of equally likely to be positive or negative,
• then gaussian or normal error distribution may be derived.
• If the measurement is designated by x, the gaussian
distribution gives the probability that the measurement will
lie between x and x + dx and is written as:
• xm is the mean reading
• and σ is the standard deviation
 3.8 The Gaussian or Normal Error Distribution

• A plot of Eq
3.8 The Gaussian or Normal Error Distribution

• the good horseshoes player’s

throws will be bunched around the
target.
• The better the player is at the
game, the more closely the throws
will be grouped around the mean
3.8 The Gaussian or Normal Error Distribution

• We may determine the precision of

a set of experimental
measurements through an
application of the normal error
distribution.

• for sets of data where a large

number of measurements is taken,
experiments indicate that the
measurements do indeed follow a
distribution like that shown in the
below figure

• Thus, as a matter of experimental

verification, the gaussian
distribution is believed to represent
the random errors
 3.8 The Gaussian or Normal Error Distribution

• we see that the maximum probability occurs at x = xm, and

the value of this probability is

• It is seen from above eq that smaller values of the standard

deviation produce larger values of the maximum probability

• P(xm) is sometimes called a measure of precision of the

data because it has a larger value for smaller values of the
standard deviation.
 3.9 Comparison of Data with Normal Distribution
• We should determine if the data are following a normal
distribution before making too many conclusions about the
mean value, variances, and so forth.
• Specially constructed probability graph is available for this
purpose
• The graph uses the coordinate system shown below
• ordinate has the percent of readings at or below the value of
the abscissa, and the abscissa is the value of a particular
reading
 3.9 Comparison of Data with Normal Distribution
• Thus, to determine if a set of data points is distributed
normally, we plot the data on probability graph and see how
well they match with the theoretical straight line.

• It is to be noted that the largest reading cannot be plotted on

the graph because the ordinate does not extend to 100
percent

• In assessing the validity of the data, we should not place as

much reliance on the points near the upper and lower ends
of the curve

• since they are closer to the “tails” of the probability

distribution and are thus less likely to be valid.
 3.11 Method of Least Squares
• Suppose we have a set of observations x1, x2, . . . , xn.
• The sum of the squares of their deviations from some mean
value is

• Now, suppose we wish to minimize S with respect to the

mean value xm.We set

• where n is the number of observations.We find that

• or the mean value which minimizes the sum of the squares

of the deviations is the arithmetic mean
 3.11 Method of Least Squares
• Suppose that the two variables x and y are measured over a
range of values.

• Suppose further that we wish to obtain a simple analytical

expression for y as a function of x.

• The simplest type of function is a linear one; hence, we

might try to establish y as a linear function of x.

• The problem is one of finding the best linear function, for the
data may scatter a considerable amount.

• the method of least squares gives a reliable way to obtain a

functional relationship
 3.11 Method of Least Squares
• We seek an equation of the form

• We therefore wish to minimize the quantity

• This is accomplished by setting the derivatives with

respect to a and b equal to zero.
• Performing these operations, there results

• Solving above Eqs. simultaneously gives

 3.11 Method of Least Squares
• Designating the computed value of y as ˆy, we have

• and the standard error of estimate of y for the data is

 3.11 Method of Least Squares
• The method of least squares may also be used for
determining higher-order polynomials for fitting data.
• One only needs to perform additional differentiations to
determine additional constants.
• For example, if it were desired to obtain a least-squares fit
according to the quadratic function

• the quantity

• would be minimized by setting the following derivatives

equal to zero:
 3.11 Method of Least Squares
• Expanding and collecting terms, we have

• These equations may then be solved for the constants a, b,

and c.
 3.14 Standard Deviation of the Mean
• We have taken the arithmetic mean value as the best
estimate of the true value of a set of experimental
measurements.
• How good (or precise) is this arithmetic mean value which is
taken as the best estimate of the true value of a set of
readings?
• To obtain an experimental answer to this question it would
be necessary to repeat the set of measurements and to find
a new arithmetic mean.
• In general, we would find that this new arithmetic mean
would differ from the previous value, and thus we would not
be able to resolve the problem until a large number of sets
of data was collected.
• We would then know how well the mean of a single set
approximated the mean which would be obtained with a
large number of sets.
• Consequently, we wish to know the standard deviation of
the mean of a single set of data from this true value.
 3.14 Standard Deviation of the Mean
• It turns out that the problem may be resolved with a
statistical analysis.

• The result is 

• Where
• σm = standard deviation of the mean value
• σ = standard deviation of the set of measurements
• n = number of measurements in the set