B D HALL AND D R WHITE

AN INTRODUCTION TO
MEASUREMENT UNCERTAINTY

MEASUREMENT STANDARDS LABORATORY OF NEW ZEALAND

B D HALL AND D R WHITE

AN INTRODUCTION
TO MEASUREMENT
U N C E R TA I N T Y

M E A S U R E M E N T S TA N D A R D S L A B O R AT O R Y O F N E W Z E A L A N D
The Measurement Standards Laboratory (msl) is New Zealand’s
National Metrology Institute, operating as a business group of
Callaghan Innovation.

Visit msl on-line: www.measurement.govt.nz

Published by
Measurement Standards Laboratory of New Zealand,
Callaghan Innovation,
69 Gracefield Road,
Lower Hutt 5010,
New Zealand.

ISBN: 978-0-473-40582-3 (digital)

ISBN: 978-0-473-40581-6 (print)

The digital edition, published in June 2020, is an unabridged republication

of the original print edition published in September 2017.

DOI: https://doi.org/10.5281/zenodo.3872590

B. D. Hall: https://orcid.org/0000-0002-4249-6863
D. R. White: https://orcid.org/0000-0002-8156-7943

The cover illustration is from the 1800 French print Usage des Nouvelles Mesures, by L. F. Labrousse. The
upper panel depicts the litre, the gram, and the metre; the lower panel shows the are (100 square
metres), the franc (a decimal currency), and the stère (1 cubic metre of wood). These decimal units
became legal in France on 4 November 1800, five years after the metric system was introduced (source:
Bibliotèque nationale de France).

This work is licensed under a Creative Commons Attribution-No Derivatives 3.0 Licence
(CC BY-ND 3.0 NZ, see http://creativecommons.org/licenses/by-nd/3.0/nz/).

This licence allows for redistribution, commercial and non-commercial, as long as

the work is passed along unchanged and in whole, with appropriate credit.
Contents

Preface 6

Introduction 7

Measurement errors and models 9

Describing unpredictability 10

Observations of variability 14

Measurement uncertainty 18

Confidence and decisions 20

Measurement models 22

Calculating uncertainty 25

Uncertainty analysis 30

Using a ruler 34

Calibrating a thermometer 41

Final words 49
Preface

This introduction to measurement uncertainty is intended for

metrology professionals working in calibration laboratories and
metrology institutes, as well as students in tertiary-level science and
engineering programmes.

The subject matter is presented with an emphasis on developing

models of the physical measurement process. Since the first publi-
cation of the Guide to the Expression of Uncertainty in Measurement in
1993,1 there has been a need to develop appropriate measurement
models for uncertainty analysis. Yet guidance on how to do so is
1
The current version of this guide was
released in 2008:
lacking. This booklet addresses that problem.
BIPM, IEC, IFCC, ISO, IUPAC, IUPAP
and OIML, 2008, “Evaluation of Mea-
The level of mathematics and statistics used here is basic and is surement Data – Guide to the Expression
typically covered by high school studies. However, modelling is a of Uncertainty in Measurement JCGM
process that uses mathematical language to describe an abstraction 100:2008 (GUM 1995 with minor correc-
tions)”, 1st edition (Sèvres: BIPM Joint
of a measurement process. This may be unfamiliar to some readers. Committee for Guides in Metrology).

Many people have contributed to this booklet. We are grateful to

our colleagues at the Measurement Standards Laboratory of New
Zealand, with whom we have discussed, at great length and at
times with some vigour, the subtleties and nuances of measurement
uncertainty that arise in our work. In particular, we would like to
thank Eleanor Howick, Jeremy Lovell-Smith, Murray Early and
Peter Saunders for their comments and contributions.

B D Hall
D R White

July, 2017.
1 Introduction

Measurements provide us with information about quantities of

interest. Measurements often provide answers to questions like:
what time is it, how fast is the car travelling, what is the alcohol
concentration of the blood sample? Measurement results are usu-
ally needed because there is a decision to be made: will I be late for
the train, is the car travelling too fast, is the driver’s blood-alcohol
concentration above the legal limit? Unfortunately, no measurement
result is exact. So poor decision outcomes can result: I might miss
the train, fail to identify a speeding car or decide to prosecute a
sober driver.

In modern society, measurements inform important decisions about

public health, safety, the environment, trade, industry, commerce,
etc. So the accuracy of information obtained by measurement has to
be considered as part of sound decision making.

The difference between a measurement result and the actual value

of a quantity of interest is called measurement error. It is the presence
of measurement errors, and the fact that they are unpredictable,
that leads to uncertainty about the correctness of decisions.

The investigation of measurement errors is called uncertainty anal-

ysis. This identifies and assesses factors that can contribute to
measurement error. An uncertainty analysis produces a model that
describes the inherently unpredictable behaviour of a measurement
and provides statistical information about the accuracy of results
that can be used when making decisions.

This booklet introduces the most important ideas involved in uncer-

tainty analysis: the notion of an unpredictable measurement error,
which leads to uncertainty in the interpretation of measurement
results; the notion of probability, which can be used to describe
unpredictable events; and the idea that a physical model can relate
the outcome of a measurement to those factors that influence the
measurement system. We explain how to formulate measurement
models and how to use them to evaluate measurement uncertainty.
We then apply these ideas to case studies of measurement and
8 an introduction to measurement uncertainty

calibration.

1.1 Mathematical language and notation

Mathematical expressions and symbols are used in this booklet

to give clear, concise, representations of the quantities involved
in a measurement process. The symbols are, however, just tokens
for things that can be described in other ways – the mathemati-
cal language is not strictly necessary. For readers who find the
mathematical expressions difficult to follow, try to think about the
meaning of each symbol using words. For instance, equation (2.1),
on page 9,
Y = y − Ey ,

could be written as1

1
The quantity intended to be mea-
measurand = measured value − measurement error, sured is called the measurand.

or even

the value of the measurand can be calculated by taking the measured

value and subtracting the value of the measurement error.

Notation: A simple notation is used to distinguish between known

and unknown quantities: upper-case letters represent unknown
quantities, while lower-case letters represent things that have a
known value. We also add subscripts to symbols to help label the
quantities.

For example, the power consumed by a household might be rep-

resented by the upper-case letter P (the exact amount of power
consumed is not known), while a reading of the household power
meter could be represented by pmeter , a lower-case letter (the meter
reading is a known number).

In some situations we need to indicate that there are quantities that

vary while others remain fixed. To do this, an ‘index’ is appended
to the symbol for the quantity that varies. For example, the ith
power meter reading could be written as pmeter [i ].
2 Measurement errors and models

We can never determine the value of a quantity exactly, because

nuisance factors always influence the outcome of a measurement.
For this reason, a measured value is only ever approximately equal
to the quantity of interest.1
1
We can also say that a measurement
estimates the quantity of interest.

The relationship between the quantity intended to be measured, Y,

and the measured value, y, can be expressed as

Y = y − Ey , (2.1)

where the term Ey is called the measurement error and is due to the
2
The capital letters used for the mea-
surand Y and the measurement error
combined effects of many influence factors on the outcome of the Ey remind us that these quantities are
measurement.2 Equation (2.1) tells us that we could correct the not precisely known, whereas y, the
measurement result, is.
measurement result y, and obtain Y, if the error Ey were known.

Ey

y Y

This equation is an example of a measurement model. Models de-

scribe a relationship between the quantity of interest and other
quantities that can influence the measurement outcome. They are
central to uncertainty analysis, because they can be used to evaluate
the quantities intended to be measured as well as to evaluate mea-
surement uncertainty. In equation (2.1), it is uncertainty about the
value of the error Ey that leads to uncertainty in the value y as an
estimate of Y.

In later chapters, we explain how to develop models and how

to use them to obtain information about the unpredictability of
measurement results. Models need not be complicated; but, the
striking thing is, they refer to quantities that are not precisely
known.
3 Describing unpredictability

When something is unpredictable we can say that its behaviour

is random. Measurement outcomes are random because many
small factors influence the results in ways that cannot be foreseen.
Nonetheless, the quantities that we intend to measure are not
considered random (we would like to know, for example, the time,
the speed, the alcohol concentration, etc).

When carrying out an uncertainty analysis, we identify the factors

that can perturb a measurement result. We then try to describe the
behaviour of each factor using notions of probability.

This chapter introduces the notions of probability that are needed

to describe influence factors and measurement errors.

3.1 Probability

Probability statements describe the behaviour of unpredictable

outcomes. For example, when rolling a die we cannot predict the
outcome, but we might say that there is a 1/6 probability of rolling
a ‘two’.
1
A relative frequency is the number
of times that something in particular
happens divided by the total number
Our assessment of 1/6 is made on the basis of experience and other of outcomes. For example, if a coin is
tossed four times, resulting in three
information available to us; we did not measure it. Nevertheless, ‘heads’, then the relative frequency
the assumption that each of the six outcomes is equally likely could of ‘heads’ is 3/4 and the relative
be checked by measurement. We could roll the die many times to frequency of ‘tails’ is 1/4.

check that the relative frequency of sixes does not differ from 1/6 by
more than can reasonably be expected, considering the number of
2
For this reason, we avoid interpreta-
tions of the word ‘probability’ when
throws.1 the meaning cannot be understood
in terms of the relative frequency of
outcomes. For example, the winner of
Real measurements don’t involve dice, but the possibility of check- an election, or that it will rain tomor-
ing assumptions is important. It should always be possible, at least row, are unique events that cannot be
in principle, to look more closely at the validity of any probability repeated. The meaning of ‘probability’
when applied to such cases is different.
statement that describes unpredictable behaviour.2
 describing unpredictability 11

3.2 Probability distributions

When there is a discrete number of possible outcomes, such as

when rolling a die, a value of probability can be attributed to each.
However, when a measurement result can take on any value, it is
the probability of an outcome occurring within a range of values
that is of interest. Probability distributions are used in such cases.

A probability distribution can be depicted as a curve over a hori-

3
We are using the term ‘probability
distribution’ loosely here. More
zontal axis representing the possible outcomes, as in Figure 3.1.3 correctly, the curve in Figure 3.1 is
The higher the curve above the axis, the greater the probability of called a probability density function.
observing the values under that part of the curve.4
4
The probability of observing a value
within a particular range is equal to
the area under the curve over that
range. So, for example, the area of
the shaded region in figure 3.1 is
3.3 Different distributions about 0.68, which is the probability of
observing a value within one standard
deviation of the mean.

In uncertainty analysis, probability distributions are used to de-

scribe the unpredictability of influence factors and measurement
errors. The behaviour of the measurement error in a result is almost
always described by a Gaussian distribution,5 but the behaviour
of influence factors and other errors may be better described with
5
Here, we are following the Guide to
the Expression of Uncertainty in Mea-
different types of distribution. We briefly mention the uniform and surement. There will be exceptions, of
triangular distributions below. course, but they are rare in the context
of test and calibration measurements.

In order to define a distribution, its type must be specified together

with two other parameters: one that fixes its location, the other
that sets the scale of possible outcomes. In general, the distribution
mean fixes the location, and the distribution standard deviation sets
the scale. However, other ways of specifying scale and location are
sometimes more convenient.

3.3.1 The Gaussian distribution

The Gaussian distribution (Figure 3.1) is of particular importance Figure 3.1: A Gaussian distribution.
in test and measurement, because it is generally presumed to de- The centre of the distribution is the
mean; the half-width of the grey band
scribe the unpredictable error in a measurement result. A Gaussian is the standard deviation.
distribution is also often used to describe the error that arises when
estimates of influence factors are obtained by statistical methods
(e.g., taking the mean of a sample of observations, as in chapter 4).

When random behaviour follows a Gaussian distribution, values

are more likely to be observed near the centre of the distribution
(near the mean). Nevertheless, the range is theoretically unlimited,
so occasionally values far from the mean will be observed.
12 an introduction to measurement uncertainty

3.3.2 The uniform distribution

The uniform (rectangular) distribution (Figure 3.2) attributes the

same probability to all values over a range; it is used to describe
situations where all outcomes are considered equally likely.

Often, it will be convenient to specify the scale of a uniform dis- Figure 3.2: A uniform distribution. The
centre of the distribution is the mean;
tribution in terms of its width. However, when calculating the the half-width of the grey band is the
uncertainty, this information must always be converted into a stan- standard deviation.
dard deviation (see §3.3.4).

Rounding error

A uniform distribution is often used to describe the error generated when a number
is displayed using a fixed number of digits. That is, some quantity X is measured and
the result is rounded to
x = X + Ex , (3.1)
where Ex is the rounding error (see, for example, Fig. 10.3 on page 36).

Neither X or Ex are known, but we can say that x is an estimate of X (x is approxi-

mately equal to X).

If many different X values are rounded, the corresponding distribution of Ex values

will behave like a uniform distribution with a mean of zero and a range equal to the
magnitude of the least significant digit.

For example, a digital instrument with a resolution of 0.01 will generate a roughly
uniform distribution of round-off errors in the range from −0.005 to 0.005.

3.3.3 The triangular distribution

A triangular distribution, shown in Figure 3.3, attributes the highest

probability to values close to the mean. However, unlike a Gaussian
distribution, the range is limited.

As in the case of the uniform distribution, the scale of the distribu- Figure 3.3: A triangular distribution.
tion may be conveniently specified in terms of the width. The centre of the distribution is the
mean; the half-width of the grey band
is the standard deviation.
The triangular distribution may be used when values are more
likely to occur near the centre of a range than near the limits. This
is the case, for example, for the difference between a pair of uni-
formly distributed quantities when the uniform distributions have
the same width (see the example on page 38).
 describing unpredictability 13

3.3.4 Relationship between distribution width and standard deviation

The standard deviation of a distribution is always needed when

calculating measurement uncertainty. Table 3.1 gives formulae to
convert the range a of a distribution into a standard deviation. For
example, a triangular distribution with a width of 2 has a standard
√
deviation of 2/ 24 = 0.41.

distribution standard deviation

√ Table 3.1: Formulae to obtain the
uniform a/ 12 standard deviation in terms of the
√ range of the distribution a.
triangular a/ 24
4 Observations of variability

Sometimes the effect of an influence factor can be observed directly

as variability in data. In such cases, the standard deviation of a
distribution that describes the behaviour can be calculated from a
sample of observations.

The type of distribution that best describes the variability observed

in test and measurement situations is often Gaussian. In this chap-
ter, observations are assumed to vary because of the influence of a
random error that follows a Gaussian distribution.

4.1 Samples and estimates

Suppose that indications in a measurement system vary, although

the quantity of interest Y is considered fixed.

We express the ith indication as

w[i ] = Y + Erandom [i ] , (4.1)

where the values of Erandom [i ] – that give rise to variability – have a

Gaussian distribution with a mean of zero.

We shall consider two ways of measuring Y: we can take just one in-
dication, or we can take a sample of values and calculate the mean.
The outcome in either case is affected by Erandom ; there will always
be an unpredictable error in the value y obtained as an estimate of
Y. This situation is described, in general, by equation (2.1), repeated
here,
Y = y − Ey .
The behaviour of the error Ey is best described by a Gaussian dis-
tribution with a mean of zero in both cases. However, the standard
deviation that scales the distribution width will depend on how y is
obtained. A sample of n observations

w [1], w [2], · · · , w [ n ]

is needed to evaluate this parameter.

 observations of variability 15

4.2 A single observation and the sample standard deviation

When a single indication is taken as an estimate of Y (the index i is

no longer needed),1 we can write simply that
1
It may appear odd to make a mea-
y=w. surement of Y with just a single indi-
cation when a sample is still needed to
work out the standard deviation of the
The standard deviation of the distribution describing Ey in this case distribution describing Ey .
is s
n However, sometimes the variability
1
n − 1 i∑
sn = (w − w[i ])2 , of Ey can be assessed at some other
=1 convenient time. Thereafter, measure-
ments may be made that involve just a
which involves a calculation of the sample mean single reading, while the information
obtained earlier about the distribution
n
1 of Ey can still be used. An example of
w=
n ∑ w [i ] . this is given in chapter 11.
i =1

The statistic sn is called the sample standard deviation.

Example: variable voltage readings

The following 10 readings were taken from a digital multimeter configured to mea-
sure a DC voltage.

9.998 mV
10.002 mV
10.000 mV
10.002 mV
10.000 mV
10.001 mV
10.003 mV
10.002 mV
10.000 mV
9.999 mV

The sample statistics are:

sample mean = 10.0007 mV
sample standard deviation = 0.0016 mV
16 an introduction to measurement uncertainty

4.3 Sample mean and the standard deviation of the mean

The sample mean provides a better estimate of Y than just one

indication2
1 n 2
The variability of the means of
y = w = ∑ w [i ] . different samples of observations
n i =1
will be less than the variability of
When a sample mean is taken as a measurement of Y, the standard individual observations in those
samples.
deviation of the distribution describing Ey is
s
n
1
SEw =
n ( n − 1) ∑ (w − w[i])2 ,
i =1

which is related to the sample standard deviation

sn
SEw = √ .
n

The statistic SEw is called the standard error of the mean.

Example (continued from page 15): variable voltage readings

The sample means of 10 samples, each of 10 observations, are shown below.

10.000 70 mV
10.000 02 mV
10.000 26 mV
10.000 00 mV
9.999 78 mV
9.999 67 mV
9.999 53 mV
9.999 12 mV
9.999 76 mV
9.999 70 mV

The first value in this list, 10.000 70 mV, was calculated from the 10 data points shown
on page 15. The standard error of the mean for that sample can be calculated from the
sample standard deviation obtained earlier

0.0016
√ mV = 0.000 51 mV .
10
This is, as expected, close to the standard deviation of the 10 values above, 0.000 43.
 observations of variability 17

4.4 Variability and a constant offset

Often, the indications of a simple measuring instrument can be

modelled by a pair of influence factors: one, Erandom , changes with
each indication while the other, Eoffset , remains constant. In this
case, the ith indication can be expressed as3
3
Equations like (4.1) and (4.2) are
w[i ] = Y + Eoffset + Erandom [i ] . (4.2) called observation equations. They
| {z } describe an indication (observation) in
constant terms of the quantity of interest and
the other influence factors.
When this situation occurs, the methods of measuring Y described
in sections 4.2 and 4.3 are no longer appropriate: they would esti-
mate the value of Y + Eoffset , not Y.4 A different way of measuring
the offset must be found. One common approach is to compare
4
The sample standard deviation
and standard error of the mean are
the indications of the instrument with indications from a more statistics that only depend on the
accurate one. In chapter 11, for example, the fixed offset of a ther- variability of data; the value of these
statistics would not change if all
mometer being calibrated is measured by comparing readings of observations were to increase, or
the thermometer with readings from a more accurate reference decrease, by the same amount.
thermometer.

4.5 Sampling

When collecting observations, it is important to ensure that a rep-

resentative sample of the variability of an influence quantity is ob-
tained. Careful planning may be needed to ensure that a sufficient
range of conditions is covered when observations are collected.

For example, in chapter 11 the effect of mechanical imperfections in

the bore of a thermometer is an influence factor. A representative
sample of data affected by this factor can be obtained by recording
readings over a suitable range of temperature, so that the changing
height of the liquid column in the bore samples the irregularities
that affect measurements.
5 Measurement uncertainty

In the context of test and calibration work, the word uncertainty has
two slightly different meanings.

On one hand, when discussing the measurement error in a result

or the unpredictable behaviour of an influence quantity, the word
uncertainty is associated with the scale of the variability.

On the other hand, when reporting results, an uncertainty state-

ment is made about a range of values considered likely to include
the quantity of interest. This is an inference about the actual value
of a quantity, not a description of random behaviour.

These different meanings of ‘uncertainty’ correspond to two specific

terms: standard uncertainty and expanded uncertainty, respectively.

5.1 Standard uncertainty

A standard uncertainty is the standard deviation of a distribution

that describes the random behaviour of an influence quantity or
measurement error.

If y is a measurement of Y, the usual notation for the standard

uncertainty is u(y) – that is, u(y) is the standard deviation of the
probability distribution chosen to describe how measurements of Y
vary.

5.2 Expanded uncertainty

Having carried out a measurement procedure and processed the

data, including a calculation of the standard uncertainty in the
measured value (described in chapter 7), the result may be reported
as an interval that is considered likely to include the quantity of
interest. The extent of this interval represents the uncertainty in y
 measurement uncertainty 19

– the measured value – as an estimate of Y. The half-width of the

interval is called the expanded uncertainty.

Figure 5.1: The expanded uncertainty

U is the half-width of the uncertainty
y−U y+U interval.

5.2.1 Evaluating expanded uncertainty

An expanded uncertainty is calculated by multiplying the standard

1
Typically, k ≈ 2 for a level of con-
fidence close to 95%. However, k
uncertainty of a result u(y) by a factor k, called the coverage factor. will become larger when a measure-
This coverage factor has an associated level of confidence (or coverage ment depends on a small number of
observations.
probability).1
U = k u(y) .
Typically, a result may be reported as a measured value and an ex-
panded uncertainty at a particular level of confidence. For example,

Y = 50.000 12 mm ± 0.000 32 mm , at a level of confidence of 95% .

Note that it is conventional to use an upper-case ‘U’ for expanded

uncertainty. However, the value of U is known, so this is an excep-
tion to the convention used in this booklet of representing unknown
quantities by upper case letters.

5.2.2 Level of confidence

The unpredictable behaviour of the measurement error associated

Y
with a result is presumed to follow a Gaussian distribution. This
assumption gives meaning to the term level of confidence.

The level of confidence refers to the performance of the procedure

used to carry out the measurement: it is not a statement about a
single measurement result. If a measurement procedure that reports
results to a 95% level of confidence were to be used over and over
again (independent measurements), 95% of the intervals produced
are expected to cover the measurand (Figure 5.2).

Unfortunately there is no way of knowing whether Y is covered in a

particular case, because measurement errors are unpredictable.

Figure 5.2: The width and position

of intervals depends on the measure-
ment error. Solid intervals cover the
measurand, dotted intervals do not.
6 Confidence and decisions

Measurements provide information about quantities – the time

is 12:04 pm, our speed is 55 km/h, etc. However, that information
will always be considered in a particular context when decisions
are to be made. Usually a measured value will be compared with
a value of the same kind of quantity – are we late for the midday
train, are we exceeding the 50 km/h speed limit, etc? In this chapter,
we consider what the notions of probability and measurement
uncertainty allow us to say about the quality of decisions based on
such comparisons.

In chapter 5, we saw that a result can be reported as an interval

likely to cover the actual value of the quantity of interest. The
width of this interval accounts for the typical size of measurement
errors, but we can never be sure that a particular interval covers the
quantity of interest. To look at the implications of this, we consider
the problem of testing an instrument to determine whether it is
operating correctly. A laboratory will measure certain instrument
parameters and compare them with manufacturer specifications.

Suppose that one such parameter is an offset voltage that is speci-

fied to lie in the range from −15.0 µV to 15.0 µV. The laboratory can
measure this with an expanded uncertainty of ±5 µV at a 95 % level
of confidence.

Figure 6.1 shows four possible test outcomes in which the mea-
sured value comes close to the upper specification limit. In the case
labelled ‘A‘, the measured voltage is slightly higher than 20 µV, so
the interval does not include the upper limit 15 µV. In the cases
labelled ‘B‘ and ‘C‘, the results are within ±5 µV of the upper limit;
in measurement ‘B’ the value is above 15 µV, and in ‘C’ it is below.
In the case labelled ‘D’, the measured voltage is a little less than
10 µV, so, again, the interval does not include 15 µV.

How should these different outcomes be interpreted by the labora-

tory? What kind of statement can be made about the instrument
meeting manufacturer specifications?
 confidence and decisions 21

15 µV
Figure 6.1: Four different test out-
A comes. In A we can state, with at least
95 % level of confidence, that the test
parameter exceeds the limit of 15 µV.
B In B and C we cannot state that the test
parameter lies either above or below
the limit with 95 % confidence. In D
C we can state, with at least 95 % level of
confidence, that the test parameter is
under the limit.
D

In chapter 5, we saw that 95 % of intervals obtained from indepen-

dent measurements are expected to cover the quantity of interest.
So, we must now step back and consider the different possible
outcomes of many measurements. We cannot be sure about a par-
ticular result, but we can state our confidence that the instrument
meets, or misses, specifications in some cases.

For instance, when a result like ‘A’ occurs, the laboratory can
state that the parameter is above the specified range (i.e., out of
specification) at a level of confidence of at least 95 %. That is, on
95 % of occasions when a result like ‘A’ is obtained, the actual value
is covered by the interval and hence out of specification. Similarly,
when a result like ‘D’ occurs, the laboratory can state that the
parameter is inside the specified range (i.e., meets the specification)
at a level of confidence of at least 95 %. However, when results like
‘B’ or ‘C’ arise the interval straddles the limit, so we know that
the actual offset voltage could be above, or below, the specified
level more frequently. In such cases, a statement about meeting,
or missing, the specification cannot be made with a 95 % level of
confidence.

It is also worth noting that, by allowing for measurement uncer-

tainty, we have effectively narrowed the range of acceptable mea-
sured values by 10 µV in this example. This shrinkage is necessary
to reduce the risk of incorrectly accepting a measured value as
being in-specification, when it is not, to no more than 5%.

acceptable values
Figure 6.2: The range of measured
values that are acceptable is reduced
5 µV 5 µV by the expanded measurement uncer-
tainty.

30 µV
7 Measurement models

A measurement model describes how the various quantities in-

volved in a measurement relate to the quantity of interest. A model
is expressed as an equation, or as a set of equations, that can be
thought of as a mathematical recipe for calculating the measur-
and if exact values for the influence quantities and errors were
known. A measurement model can also be used to calculate the
measurement uncertainty associated with a result.

In this chapter we describe how to develop measurement mod-

els. We begin by looking for ways to obtain a simple model, as a
starting point, and then show how the model can be extended to
include as many influence factors as needed.

7.1 Getting started

One way to begin modelling is to think about how measured values

would be used to calculate a result. This calculation will include
the values of important quantities, but it usually leaves out values
of residual errors and influence quantities of lesser importance. An
initial measurement model can be obtained by writing down this
calculation as an equation with the measured values replaced by
symbols for the corresponding quantities.
 measurement models 23

Example: a length measurement

To model a measurement of the distance between two points (chapter 10), we can
start by thinking about how to calculate this length when given ruler readings at
each point. We would, of course, subtract one reading from the other. So, an initial
model of the measurement can be written down as the difference between the point
locations (i.e., by subtracting the symbols representing each location rather than the
ruler readings). If one location is X A , and the other XB , the model can be written as

L = XB − X A .

Another good way to start a model can be to start with an obser-

vation equation that describes the measurement process (see chap-
ter 4). For example, in §4.4 we saw that equation (4.2), repeated
here
w[i ] = Y + Eoffset + Erandom [i ] ,
describes how observations of a quantity Y are affected by a ran-
dom factor Erandom and a fixed offset Eoffset . To obtain a model for
a measurement based on one observation, we can rearrange this
equation so that the measurand Y appears on the left

Y = w[i ] − Eoffset − Erandom [i ] .

7.2 Extending a model

Once we have an initial model, other influence factors and error

terms can be included, until we are satisfied that the model ade-
quately describes the measurement.

A model can be extended in two ways, which can be applied repeat-

edly and in any order:

1. New terms representing additional influence factors, or errors,

can be incorporated in the model equation.

2. Terms already in the model can be replaced by sub-models

(other equations) representing measurements of the correspond-
ing quantities.
24 an introduction to measurement uncertainty

Example: temperature measurement

Chapter 11 describes a situation where a reference thermometer is used to estimate

the temperature T in the vicinity of a thermometer being calibrated. A reading from
this thermometer is taken to be a measurement of T, although the reference ther-
mometer actually senses the temperature of its own environment, Tprt . So, the initial
model for the measurement of temperature near the thermometer being calibrated is

T = Tprt .

The two thermometers are some distance apart, so there may be a small temperature
difference between Tprt and T. We need to account for this in the model. As an exam-
ple of extending a model using step 1 above, a term for the temperature difference
between the thermometer locations, Ebath , can be included

T = Tprt − Ebath .

Example: a length measurement (continued)

Ruler readings approximate the actual locations X A and XB . The readings are subject
to errors that need to be included in the model.

As an example of extending a model using step 2 above, we can develop sub-models

for the individual measurements of X A and XB .

At each location, the scale marking on the ruler closest to the point of interest is
taken as the reading. This rounds the actual location to the value of the nearest scale
division.

To model a measurement of X A , we introduce a rounding error E A (see §3.3.2).

X A = x A − EA .

This expresses the fact that the reading x A is only approximately equal to X A . A
similar sub-model describes the measurement of XB .

These sub-models can be incorporated to give us a set of model equations describing

the length measurement

X A = x A − EA
X B = x B − EB
L = XB − X A ,

or the equations can be combined algebraically into one

L = ( x B − EB ) − ( x A − E A )
= x B − x A − EB + E A .
8 Calculating uncertainty

A measurement result can be calculated once all quantities have

been measured, or estimated in some other way. The uncertainty
in this result, as an estimate of the quantity of interest, can be
calculated from the model.
1
We can say that two estimates are
independent when knowing the error in
In this chapter, we describe how to calculate the uncertainty when one would not help us determine the
all quantity estimates are independent.1 error in the other.

8.1 Models involving only addition and subtraction

When a model only adds or subtracts terms, the uncertainty can

be calculated by taking the square root of the sum of the squared
standard uncertainties. That is, when the model has the form

Y = w ⊕ X1 ⊕ X2 ⊕ · · · ,

where w is a number,2 X1 , X2 , · · · are quantities, and ‘⊕’ means

either ‘+’ or ‘−’, then the standard uncertainty of the result
2
The number w could, for instance, be
an instrument reading.
q
u ( y ) = u ( x1 )2 + u ( x2 )2 + · · ·

is evaluated from the standard uncertainties of the estimates of

X1 , X2 , · · · . Note, the number w makes no contribution to the
uncertainty, because it has a known value.

Example: a fixed offset measurement

In chapter 11, a measurement model for the fixed offset of a thermometer is given by
(equation 11.8)
Dglass = d − Ed·ran + Eprt·cal + Ebath .
This equation only adds and subtracts terms.

A value of d = 0.0052 °C has been obtained from a sample of readings. Estimates of

the other quantities, with uncertainties, are
26 an introduction to measurement uncertainty

Influence Estimate Uncertainty

Ed·ran ed·ran = 0.0000 °C u(ed·ran ) = 0.0019 °C
Eprt·cal eprt·cal = 0.0000 °C u(eprt·cal ) = 0.0078 °C
Ebath ebath = 0.0000 °C u(ebath ) = 0.0029 °C

Using this information, an estimate of Dglass can be calculated directly from the model

dglass = d − ed·ran + eprt·cal + ebath

= 0.0052 °C − 0.0000 °C + 0.0000 °C + 0.0000 °C
= 0.0052 °C .

The combined standard uncertainty of this estimate is calculated by using the addition
and subtraction rule
q
u(dglass ) = u(ed·ran )2 + u(eprt·cal )2 + u(ebath )2 (8.1)
q
= (0.0019)2 + (0.0078)2 + (0.0029)2
= 0.0085 °C .

8.2 Models involving only multiplication and division

When a model only multiplies or divides terms, the relative standard

uncertainty of the result can be evaluated by taking the square root
of the sum of the squared relative standard uncertainties. That is,
when a model has the form

Y = w ⊗ X1 ⊗ X2 ⊗ · · · ,

where ⊗ can be either × or ÷, the relative uncertainty in y, as an

estimate of Y,
s
u(y) u ( x1 ) 2 u ( x2 ) 2
   
= + +··· ,
y x1 x2
is calculated from the estimates of X1 , X2 , · · · and the associated
standard uncertainties. Note, again, that there is no uncertainty
associated with the value w, because it has a known value..

Example: the area of a rectangular sheet

The area of a rectangular sheet can be found as the product of the length L along one
side with the width W along the other side

A = L·W .

In chapter 10 we determine a standard measurement uncertainty of 0.41 mm when

using a particular calibrated ruler.
 calculating uncertainty 27

This ruler is used to measure the length and width of a sheet, obtaining measured
values l = 263 mm and w = 95 mm. So, the measured value of the area is

a = l·w
= 263 mm · 95 mm
= 24 985 mm2 .

The uncertainty in this estimate of the area is found by first calculating the relative
standard uncertainty using the rule for multiplication and division
s
u( a) u(l ) 2 u(w) 2
   
= +
a l w
s
0.41 2 0.41 2
   
= +
263 95
q
= 2.430 × 10−6 + 18.626 × 10−6
= 4.589 × 10−3 .

Then the standard uncertainty is

u( a) = 4.589 × 10−3 · 24 985 mm2

= 115 mm2 .

8.3 Mixed models

When a model adds or subtracts some terms and multiplies or

divides others, the uncertainty can still be evaluated using the
rules above. However, the model must be decomposed first into
a number of intermediate equations, each of which that can be
handled by one rule or the other.

Example: a linear model

A simple linear observation equation involving both multiplicative and additive errors
can often be used to describe the indications of simple measuring instruments.

We consider a meter measuring DC voltage, which indicates a value w in response to

an applied voltage V. An observation equation is

w = Again V + Eoffset .

The meter is, of course, designed to measure the applied voltage, so it is reasonable
to assume that Again ≈ 1.0 and Eoffset ≈ 0.0 V. These values correspond to a per-
fect meter, but we can use them as estimates of the actual values, and include some
uncertainty to allow for the imperfect behaviour of a real meter.
28 an introduction to measurement uncertainty

The quantity of interest is the applied voltage V, so we can re-arrange the observation
equation to obtain a measurement model

V = (w − Eoffset )/Again .

This equation involves both subtraction and division, so it cannot be handled by just
one of the rules above: it must be decomposed.

As a first step, the difference between the indication and the offset

V1 = w − Eoffset

provides us with an equation that can be handled by the addition-subtraction rule.

This intermediate step could be thought of as a model for an ‘offset-corrected-
indication’.

Next, dividing the ‘offset-corrected-indication’ by the gain factor gives an equation

that can be handled by the multiplication-division rule

V2 = V1 /Again .

This is as far as we need to go. The result V2 happens to be the quantity of interest
(V = V2 ). So, the decomposition here has led to two simpler models.

Suppose now we have a reading

w = 10.015 V
and the estimates
again = 1.000 and eoffset = 0.000 V ,
with a relative standard uncertainty in the gain of

u( again )/again = 0.005

and a standard uncertainty in the offset of

u(eoffset ) = 0.05 V .

The measured value of applied voltage can be calculated by working through these
equations. First, calculate v1 , the estimate of V1 ,

v1 = w − eoffset = 10.015 − 0.000 = 10.015 V .

The standard uncertainty of this value as an estimate of V1 is

q q
u(v1 ) = u(eoffset )2 = (0.05)2 = 0.05 V .

Second, calculate v2 , the estimate of V2 ,

v2 = v1 /again = 10.015/1.000 = 10.015 V .

 calculating uncertainty 29

The relative standard uncertainty of this as an estimate of V2 is

v
u u(v ) 2  u( again ) 2
u
u ( v2 ) 1
=t +
v2 v1 again
s
0.05 2 0.005 2
   
= +
10.015 1.000
= 0.007 07 .

Note that results obtained in the first step are needed in the second step.

The final result is

v = v2 = 10.015 V ,

with a standard uncertainty of

u(v) = 10.015 × 0.007 07

= 0.071 V .
9 Uncertainty analysis

The preceding chapters have looked at distinct aspects of uncer-

tainty analysis: the notion of unpredictable measurement errors and
influence factors; the use of probability and statistics to describe
such unpredictability; the construction of measurement models to
describe the unpredictable behaviour of measurements; and the use
of these models to calculate results and measurement uncertainty.
We now put these different aspects together and show how to carry
out full uncertainty analyses.

This chapter describes the process of uncertainty analysis in general

terms. It offers guidance on how to identify influence factors and
sources of error in a measurement system, and it discusses how the
design of a measurement procedure should aim to take care of as
many factors as possible, so that they do not contribute to the final
measurement error. The next two chapters present case studies in
which uncertainty analysis is applied to simple measurements.

9.1 Uncertainty analysis

An uncertainty analysis will follow the steps below. In practice,

we go back and forth over these steps as our understanding of the
measurement improves.

1. Identify the measurand – the quantity of interest.

2. Identify factors that can influence the outcome of a measure-

ment.1
1
These may be called influence quanti-
ties, or just influences.
3. Consider each influence factor and identify those that can be
eliminated by careful experimental design.

4. Obtain an estimate of each remaining influence and determine

the type of probability distribution that best describes its unpre-
dictable behaviour and the standard deviation of that distribu-
tion.
 uncertainty analysis 31

5. Develop a model that describes how the measurement result

depends on the influence factors.

6. Use the measurement model and data collected during the

measurement to evaluate a measured value and measurement
uncertainty.

Each step, except 6, requires detailed knowledge of the measure-

ment process, the equipment involved and the environment. Step 6
is a routine data-processing task.2
2
A common misconception about
uncertainty analysis is that it involves
a lot of maths. That’s not true. The
most important steps, 1 to 5, require
9.2 Identify the measurand measurement expertise.

Sometimes it is easy to identify the quantity of interest (the mea-

surand). For instance, to decide whether a car is travelling above
the legal speed limit, its speed must be measured so that it can be
compared to the limit.

In test and calibration work, however, more thought is often needed

to identify what is being measured and which influence factors
need to be considered.

For instance, to test an instrument that measures blood-alcohol

concentration (BAC), a reference sample might be examined under
strictly controlled conditions. To pass the test, the instrument will
have to indicate a result that falls within a pre-determined range of
BAC values.

The measurand, in this case, is the BAC of the test sample presented to
the instrument. The measurement error is the difference between the
nominal value (attributed to the sample) and the actual BAC (the
measurand). The error depends on things like sample preparation,
sample stability and environmental factors, but not the instrument’s
performance.3 A decision about whether an instrument is operating
correctly will depend on how accurately the BAC of the test sample
3
So, things like instrument noise,
digits in the digital display, etc, do
is known. The pre-determined range of acceptable values will take not contribute to the measurement
account of the variability in instrument performance. uncertainty in the BAC of the test
sample.

On the other hand, when the instrument is used to measure blood-

alcohol concentration, the measurand is the BAC of a sample taken
from an individual at a particular time. The performance of the
instrument is now likely to be paramount, but there will be other
influences too: physiological factors, the nature and stability of the
sample, etc. The measurement error will be the difference between
the instrument indication and the BAC of sample taken at some
particular time. It is uncertainty about the size of this error that
32 an introduction to measurement uncertainty

must be considered when deciding whether someone’s BAC has

exceeded a limit.

9.3 Identify influence factors

Identifying things that can influence the outcome of a measurement

is a difficult part of any uncertainty analysis. Some factors are
subtle and easily overlooked.4
4
For example, mass balances can be
sensitive to altitude and geographical
Broadly speaking, look for influences from the following areas: location because of variations in the
local gravitational acceleration!

• imperfect behaviour of measuring instruments, including:

noise, fixed digit resolution, drift, calibration and adjustment
errors;

• environmental factors, including: chemical (humidity, contami-

nation, corrosion, electrochemical effects, etc); optical (absorption,
reflection, the atmosphere, the sun, etc); magnetic (electrical
transformers, high-current cables, magnets, etc); electric (static
electricity, high-voltage cables, electrical earthing and shielding,
etc); thermal (air conditioning, people, heat from instruments,
drafts, etc); mechanical (vibration, stress, weight, etc), and so on;

• operator factors and the measurement process, including: skill

in assembling and operating the measurement system and the
sample, or component, to be measured; sampling errors (are
the measurements representative of the quantity of interest, was
data collected over a wide enough range to ensure that influence
factors were adequately sampled);

• sample behaviour, including: stability (drift, contamination,

degradation, etc); ease of handling, etc.

Sources of information about potential influence factors are many

and varied. Datasheets, manuals and application notes for instru-
ments may be useful, including similar instruments from different
manufacturers. Also, documentary standard test-methods are some-
times designed to address the more troublesome influences in order
to produce more reproducible results.

Above all, it is important to keep a detailed record of everything

considered. The assessment of influences has to be objective. The
choices and assumptions made must be available for review, so that
the measurement model can be validated and possibly refined later.
 uncertainty analysis 33

9.4 Classify the influences

Influence factors can be sorted into three broad categories:

Prerequisites: conditions that must be met, otherwise the measure-

ment would fail out-right, such as broken connections, battery
failure, etc;

Conditions of measurement: influences that can be eliminated by a

suitable procedure; for instance: allowing sufficient time for
a system to settle before taking a reading, using careful align-
ment to avoid reading errors such as parallax, using a 4-wire
measurement technique to eliminate lead resistance effects, etc;

Components of measurement error: influence factors that will give rise

to measurement errors: terms representing these influences must
be included in the measurement model.

The influences that cannot be eliminated need to be assessed and

included in the measurement model. A probability distribution that
describes the unpredictable behaviour of each influence must be
selected (see chapter 3) and the standard deviation (scale) must be
5
In more general treatments of uncer-
tainty another parameter, called the
specified.5 degrees of freedom, is also associated
with each influence. The degrees of
freedom is an indication of how well
the scale parameter for a distribution is
9.4.1 Trivial influence factor estimates known. The number of degrees of free-
dom is needed to calculate a coverage
factor for expanded uncertainty (see
chapter 5).
It is not uncommon for some estimates of influence factors to have
trivial values that contribute nothing to the result. This makes
them easy to ignore.6 Nonetheless, it is important to consider every
influence and the accuracy of each estimate. This information is
6
By design, measurement procedures
try to reduce errors and bias. Hence,
needed to calculate the combined measurement uncertainty. An the best estimate of a residual error is
example is given on page 27, where the gain of a simple voltmeter often zero.
is estimated as unity and the offset error is estimated as zero.
10 Using a ruler

This chapter looks at using a calibrated steel ruler to make a simple

length measurement. Uncertainty analysis is applied by following
the guidelines presented in the previous chapter.

10.1 Overview

An object, referred to as the ‘work piece’, has a feature delimited

by well-defined edges. The purpose of the measurement is to
determine the length of this feature by placing a ruler beside the
work piece, close to these edges.

The feature is roughly 700 mm long.

A 1 m long calibrated steel ruler marked with 1 mm divisions is

used.

10.2 The measurand

The measurand is the length L of the feature at ambient tempera-

ture.

10.3 Identify influence quantities

In this case, there are well-defined edges that are easy to compare
with ruler markings. However, in general, difficulty in locating end
points can contribute to measurement error.

Curvature of the ruler introduces errors.

Scale markings on the ruler vary slightly in width and are not
located at exactly the right positions. These imperfections introduce
 using a ruler 35

errors.

The ruler and work piece must be at a uniform temperature, be-

cause the dimensions of the ruler and work piece may respond
differently to temperature.

The ruler must be parallel to the feature of interest to avoid a

geometrical error (called cosine error), which causes the reading to
be greater than the actual length.

Figure 10.1: A cosine error will make

the length appear larger. In the
figure, the reading l = L/ cos θ is
L larger than the actual length L due to
misalignment by an angle θ.

The observer should be positioned perpendicular to the ruler scale

and to the feature edge when readings are made. If not, a reading
error (called parallax error) will occur. For example, if the work
piece is a different thickness than the ruler, and both are sitting on a
flat surface, then a parallax error will be introduced if the observer
is not looking directly down towards the scale.

Figure 10.2: When the angle of obser-

vation is not perpendicular, the edge of
the object appears to correspond to a
different location on the ruler. This is
called parallax error.

parallax error

A ruler reading is taken to be the value of the scale marking clos-

est to the feature edge. So, a measured value is rounded to the
resolution of the ruler scale, which introduces an error.

If one end of the ruler is used to locate one edge, an error may be
introduced due to the difficulty in aligning the end of the ruler with
the feature, and possibly also because of mechanical wear at the
end of the ruler.
36 an introduction to measurement uncertainty

Figure 10.3: Reading the scale to the

nearest marking introduces a rounding
error, Eres . The size of this error can
be no larger than half the width of the
smallest scale division.

Eres

10.4 Assess influence quantities

10.4.1 Considerations for the measurement procedure

The ruler must be kept flat; it should be supported and can be

placed on edge to avoid bending.

Calibrating a ruler characterises the accuracy of scale markings.

Here we assume that there is no need to interpolate between mark-
ings and instead read the location of a feature to the nearest ruler
graduation. If greater precision is required, we would consider
using a finer scale.

The ruler and work piece should be allowed enough time to reach
a uniform temperature. Avoid handling the objects during the
measurement, as this can affect their temperature. The ambient
temperature should be measured.

The ruler and work piece should be aligned carefully. They should
be parallel and as close together as possible. Avoid using the end of
the ruler to make a measurement, because the rule may not be cut
accurately through the last scale marking, or the end may be worn.
Instead, take readings from a position that is perpendicular to the
work-piece feature and along the ruler scale.

10.4.2 Estimate influence quantities

Rough estimates of the various influences, tabulated below, suggest

that only errors due to the resolution of the scale need be included
in a simple measurement model.

Resolution error Eres : The number on the scale marking closest to

a feature end-point is taken as the reading. This rounds the
 using a ruler 37

Factor Typical magnitude

Table 10.1: Rough estimates of the
Cosine error 0.01 mm typical sizes of the errors involved in a
Parallax error 0.08 mm length measurement.
Resolution error 0.29 mm
Ruler scale error 0.05 mm
Thermal expansion 0.04 mm

location to the nearest scale division. As explained in §3.3.2, a

uniform distribution is often used to describe rounding errors
Eres .

The distribution is centred on zero, because rounding errors are

equally likely to be positive or negative: our best estimate of Eres
is eres = 0.00 mm.

The width of the smallest division, 1.0 mm, determines the range
of possible error. Using Table 3.1, this range can be converted to
a standard deviation
1.0
u(eres ) = √ = 0.29 mm .
12

10.5 The measurement model

A model of this measurement is described on page 23, in chapter 7.

The measurand is the distance between the locations of the feature
end-points:
L = XB − X A .

Following step 2 on page 23, the measurement of each location can

be described using a sub-model that accounts for the rounding er-
ror when reading the ruler. For instance, the location X A is related
to the ruler reading x A and the rounding error E A by

X A = x A − EA .

Substituting these sub-models for X A and XB in the equation for L

above gives a more detailed model,

L = ( x B − EB ) − ( x A − E A ) ,

which can be re-arranged to group the readings and the errors,

L = x B − x A − ( EB − E A ) . (10.1)

This model includes the most significant component of error identi-

fied in Table 10.1 and is sufficiently detailed for our purposes.
38 an introduction to measurement uncertainty

10.6 Evaluate the result and measurement uncertainty

Suppose now that readings of the feature end-points are:

x A = 123 mm
x B = 825 mm .

The associated scale reading errors, E A and EB , are independent

and our best estimates of each are 0.00 mm, so the measured value
of the feature length

l = 825 − 123 − 0 + 0
= 702 mm .

To calculate the measurement uncertainty, we note that the model

(10.1) has only two independent additive terms E A and EB , so
from §8.1 the uncertainty is the square root of the sum of squared
standard uncertainties
q
u ( l ) = u ( e B )2 + u ( e A )2
p
= 0.292 + 0.292
= 0.41 mm .

Note: the triangular distribution

When a pair of rounded readings are added, or subtracted, the error in the sum, or
difference, can be described by a triangular distribution (see §3.3.3).

The ruler measurement is an example of this.

The range of the reading error is 1.0 mm (half a division either side of the reading).
After subtracting one reading from the other, the range of possible error in the dif-
ference can be as much as 2.0 mm (one division either side of the difference between
readings).

The standard deviation of a triangular distribution with a range of 2.0 mm can be

calculated using the formula from Table 3.1:

2.0
√ = 0.41 mm .
24

This is the same number that we obtained for u(l ) above. In fact, the pair of round-
ing errors can be treated individually, or as a single error described by a triangular
distribution.
 using a ruler 39

10.7 Repeat measurements

A single measurement of L does not provide us with a way to inde-

pendently check the assumption that resolution error dominates the
measurement error. A more robust measurement procedure would
involve several measurements.

Among the various influences considered in §10.4.2, all but thermal

expansion are expected to vary when measurements are repeated
and the ruler is re-aligned with the work piece each time.1
1
Repeat measurements should also
vary the lateral position of the ruler
A sample of repeat measurement results can be used to investigate relative to the feature (i.e., the readings
whether there are other significant influences. If the rounding error of the edge locations will change).
is dominant, as assumed, values in the sample should not vary by
much more than one digit in the smallest scale division.

Example: repeated measurements

From the rough estimates in §10.4.2, we expect repeat measurements to vary by not
much more than 1 mm. A set of results like

702, 702, 701, 702, 701

is consistent with this assumption, and so is

702, 702, 702, 702, 702

However, much more variation would suggest that other errors are involved and that
these errors have magnitudes comparable to the size of scale divisions on the ruler.

For example, the following results exhibit too much variation to be explained by
resolution error alone

703, 702, 701, 704, 700

In this case a more careful assessment of the influences will be needed. It is possible
that some influences have been overlooked or that some of the influences already
identified have been underestimated.
40 an introduction to measurement uncertainty

10.8 The effect of temperature

The steel ruler was calibrated at 20 °C, but its length will change
and the scale will expand, or contract, if used at a different tempera-
ture.

If the ambient temperature is known, a correction could be applied

to the raw results to account for expansion or contraction. Other-
wise, this error should be considered as an additional influence.

Following step 1 on page 23, a term ET can be incorporated in the

model to represent thermal expansion. This is the amount that the
measured interval increases due to thermal expansion. The model
becomes
L = xB − xA − ( EB − EA ) − ET .

To assess ET , we need to know how the steel in the ruler responds

to temperature and how long the feature of interest is.

Example: thermal expansion

The coefficient of thermal expansion for the steel ruler is

11.55 × 10−6 °C−1 , or 11.55 × 10−3 mm/m/°C .

The ruler was calibrated at 20 °C, but it will be used in an environment where the
temperature could be anywhere from 15 °C to 25 °C. At the limits of this range the
scale would have expanded, or contracted, by a factor of

0.058 mm/m .

So the range of possible error in a scale reading of a 700 mm feature is ±0.040 mm.

We will assume that the actual temperature is equally likely to be anywhere in the
range from 15 °C to 25 °C, so the unpredictable expansion ET can be described by a
uniform distribution.

The distribution is centred on our best estimate of the expansion error, eT = 0.0 mm.
The range of possible error, from −0.040 mm to 0.040 mm, must be converted to a
standard uncertainty using Table 3.1

0.080
u(eT ) = √ = 0.023 mm .
12

Comparing u(eT ) with the standard uncertainty associated with a rounding error,

u(eres ) = 0.41 mm ,

we see that thermal expansion is not a significant concern.

11 Calibrating a thermometer

This chapter looks at the calibration of a mercury-in-glass ther-

mometer, which is an example of a simple instrument that is af-
fected by random influence factors and a fixed offset error.1
1
While the problem is realistic, we
present a simple analysis that is
Calibration is a process intended to check that an instrument is intended to illustrate the general
measuring the quantity of interest correctly. It can often be carried approach to uncertainty analysis.
out by examining the response of an instrument to a well-defined
stimulus, or reference, of the quantity of interest. An uncertainty
analysis of the calibration process must assess how accurately the
reference quantity is known, as well as model the behaviour of the
instrument being calibrated. In some cases, such as the thermome-
ter considered in this chapter, calibration will determine corrections
to be applied to raw instrument indications, and uncertainties in
the values of the corrections.

11.1 Overview

A temperature-controlled bath is used to provide an environment

that has a well-defined temperature. Two thermometers are im-
mersed close to each other in the bath: one is the glass thermometer
being calibrated, the other is an electronic platinum resistance
thermometer. The resistance thermometer is used to measure the
temperature of the bath and the differences between readings of the
glass thermometer and readings of the resistance thermometer are
used to calibrate the glass thermometer.

Mercury-in-glass thermometer indications may deviate from the

2
Hysteresis effects, due to slow
changes in the size of the bulb in
actual temperature for a number of reasons. Broadly speaking, the response to external temperature, can
influences fall into two groups:2 also affect readings after calibration.
However, we do not consider that here.

1. systematic factors, which do not change over long periods of

time;

2. random factors, which can vary over a short time.

42 an introduction to measurement uncertainty

The effect of systematic factors is to introduce a fixed offset error

to thermometer readings. Random factors cause variability in
readings. Thermometer calibration determines a value for the offset
error and collects information about the variability of readings. This
enables reliable measurements of temperature to be made with the
glass thermometer after calibration.

11.2 Measurand during calibration

Glass thermometer indications can be described by a simple obser-

vation equation

wglass [i ] = T + Dglass + Eglass·ran [i ] , (11.1)

where T is the actual temperature in the immediate vicinity of

the thermometer, Dglass is a fixed offset error and Eglass·ran [i ] are
random errors. This equation has the same form as equation (4.2);
however, we use a different symbol here for the fixed offset, because
this term is of central importance to the calibration process.

If the offset were known, the accuracy of measurements made with

the thermometer could be improved by applying a correction to
readings. So, the measurand during calibration is Dglass and the
calibration process is designed to measure this offset.

One way to think about Dglass is to imagine a situation where there

is no variability in thermometer readings (i.e., if Eglass·ran [i ] = 0). If
that were the case, Dglass would be the difference between a ther-
mometer reading and the actual temperature T in the immediate
vicinity of the thermometer.

11.3 Influence quantities

There are many factors that can influence the thermometer cali-
bration. Some are fatal in that, unless remedied, the measurement
should be abandoned.

Fatal flaws: the mercury column may be separated, or there may be

droplets of mercury high in the capillary; the quality of manufac-
ture may not be satisfactory – for example, scale markings may
be badly formed; or the thermometer may be extremely unstable
because it has not been annealed properly.

Random influences: operator error when reading the glass ther-

mometer (e.g., parallax error), the response time between read-
ings, etc; irregularities in the scale markings and bore diameter,
 calibrating a thermometer 43

and stiction (sticking) of the mercury in the bore; randomness in

the PRT display and instrumentation noise; short-term instability
of the bath temperature, etc.

Glass thermometer: factors that would contribute to an offset in the

thermometer reading: incorrect immersion of the thermometer;
incorrect orientation of the thermometer.

Reference thermometer: the accuracy of the platinum resistance ther-

mometer (PRT) used as a reference is characterised by calibration.
However, the thermometer must be used correctly, so, for exam-
ple, attention should also be given to appropriate immersion.

Bath temperature uniformity: a stirred bath produces a volume in

which the temperature is nearly uniform.
The spatial uniformity of temperature can be assessed by regular
measurements at different positions and over a range of tempera-
tures. This should be done as part of the on-going maintenance
of the calibration equipment.

11.4 Control over influence quantities

Many influence factors can be managed effectively by a carefully

designed measurement procedure. For example:

• ensure that the operator waits a sufficient amount of time for the
calibration bath to reach operating temperature before reading
the thermometers;

• ensure the correct immersion of both thermometers and orienta-

tion of the glass thermometer, which should be vertical;

• the glass thermometer body can be tapped lightly before reading

to avoid striction;

• readings should be taken at right angles to the thermometer to

avoid parallax errors;

• ensure that the randomness in readings of the PRT is small

compared to randomness in the glass thermometer readings
(variability in the PRT readings should be < 1/3 that of the glass
thermometer).

11.5 Modelling the calibration procedure

The measurand has been identified as the offset term Dglass in

equation (11.1). So we can begin our model by rearranging the glass
44 an introduction to measurement uncertainty

thermometer observation equation to give

Dglass = wglass [i ] − Eglass·ran [i ] − T . (11.2)

This actually suggests a way of performing a calibration. If the ther-

mometer can be placed in an environment of known temperature T,
the average difference between thermometer indications and T will
provide an estimate of Dglass .

Our measurement procedure uses a heated bath to obtain a nearly

uniform temperature in a volume of liquid. An electronic platinum
resistance thermometer immersed in the bath is used to measure
the temperature.

Readings of the resistance thermometer, just like a glass thermome-

ter, can fluctuate due to random influence factors and they will also
be subject to a small fixed offset, which we attribute to a residual
calibration error Eprt·cal .3 An observation equation for the indica-
tions of a reference thermometer is
3
Raw readings from a calibrated PRT
must be corrected using information
supplied in the calibration report.
wprt [i ] = Tprt + Eprt·ran [i ] + Eprt·cal . (11.3) For simplicity, we assume that those
corrections have already been applied
Note that Tprt is the temperature sensed by the resistance ther- to get wprt [i ].
mometer. The temperature T in the vicinity of the glass ther-
mometer is not necessarily the same as Tprt . To account for this,
we introduce an error term Ebath representing a small temperature
difference between the thermometer locations in the bath, so that

Tprt = T + Ebath . (11.4)

Replacing Tprt with T + Ebath in equation (11.3) we obtain

wprt [i ] = T + Ebath + Eprt·ran [i ] + Eprt·cal

and by rearranging this equation we obtain a model for the heat-

bath temperature in the vicinity of the glass thermometer based on
one reading from the resistance thermometer

T = wprt [i ] − Eprt·ran [i ] − Ebath − Eprt·cal . (11.5)

Equation (11.5) shows how a reading of the resistance thermometer

is related to the temperature T. In this equation, T refers to the
same temperature T as in equation (11.2). So, we can combine these
equations, replacing T in (11.2) with the expression for T in (11.5).
In this way, we obtain a model of how the glass thermometer offset
error is related to indications on each thermometer and on the other
error terms:

Dglass = wglass [i ] − wprt [i ]

+ Eprt·ran [i ] − Eglass·ran [i ]
+ Ebath + Eprt·cal . (11.6)
 calibrating a thermometer 45

It is the difference between readings that is important, so we can

simplify the notaion by introducing a term

d[i ] = wglass [i ] − wprt [i ] ,

as well as a term for the difference between the random errors as

Ed·ran [i ] = Eglass·ran [i ] − Eprt·ran [i ] .

With these changes, equation (11.6) becomes

Dglass = d[i ] − Ed·ran [i ] + Eprt·cal + Ebath . (11.7)

This shows how Dglass is related to the difference between a pair of

4
A range of temperatures will be
used so that the column of mercury
readings on the glass and resistance thermometers. However, by moves through a few scale divisions.
averaging a sample of readings we can reduce the effect of random- This samples the imperfections in
the thermometer that contribute to
ness.4 After evaluating the mean of the sample of differences, randomness in the readings (see §4.5).
N
1
d=
N ∑ d [i ] ,
i =1

our final measurement model of the calibration process is

Dglass = d − Ed·ran + Eprt·cal + Ebath , (11.8)

where Ed·ran represents the error associated with d due to variability

in the sample of differences (see §4.3).

11.6 Estimate influence quantities

11.6.1 Measured data

A sample of thermometer readings and the differences d[i ] =

wglass [i ] − wprt [i ] is

i wglass [i ]/°C wprt [i ]/°C d[i ]/°C

1 99.740 99.740 0.000
2 99.750 99.750 0.000
3 100.050 100.040 0.010
4 100.045 100.042 0.003
5 100.190 100.180 0.010
6 100.190 100.182 0.008

The statistics of this sample of differences are

sample mean = 0.0052

sample standard deviation = 0.0048
sample standard error of the mean = 0.0019
46 an introduction to measurement uncertainty

So
d = 0.0052 °C .

The term Ed·ran in (11.8) represents the error associated with our
value of the sample mean due to variability in the readings. We
expect positive and negative variations to be equally likely, so our
best estimate of this error is

ed·ran = 0.0000 °C .

Variability of the sample mean is characterised by the standard

error of the mean. So the standard uncertainty associated with ed·ran
is
u(ed·ran ) = 0.0019 °C .

11.6.2 Other influence factor assessments

Eprt·cal : this is a residual error representing the difference between

the actual PRT calibration factor and an ideal one: the best
estimate is eprt·cal = 0 °C.
The PRT calibration certificate reports a standard uncertainty
u(eprt·cal ) = 0.0078◦ C for readings after corrections have been
applied (the raw PRT data will have been corrected to give the
wprt [i ] values).
The distribution associated with this error is Gaussian.

Ebath : this is also a residual error with best estimate ebath = 0 °C.
Routine measurements of the bath indicate that the maximum
temperature difference, between any two points located within
100 mm of each other, near the centre of the bath, is 0.01 °C.
If we assume that any temperature difference is equally likely, a
uniform distribution can be associated with this range.
The standard uncertainty associated with the bath non-uniformity
is
0.01
u(ebath ) = √ = 0.0029 °C .
12

11.6.3 Summary

The sample mean

d = 0.0052 °C

and the three other influences, their estimates and uncertainties are
 calibrating a thermometer 47

Influence Estimate Uncertainty

Ed·ran ed·ran = 0.0000 °C u(ed·ran ) = 0.0019 °C
Eprt·cal eprt·cal = 0.0000 °C u(eprt·cal ) = 0.0078 °C
Ebath ebath = 0.0000 °C u(ebath ) = 0.0029 °C

11.7 Evaluate the result and measurement uncertainty

We can evaluate the measured value using the model (11.8) with the
mean of our observations and best estimates of the other influences

dglass = d − ed·ran + eprt·cal + ebath

= 0.0052 °C − 0.0000 °C + 0.0000 °C + 0.0000 °C
= 0.0052 °C .

The measurement model (11.8) is additive, so, from §8.1, the com-
bined standard uncertainty of dglass as an estimate of Dglass is
q
u(dglass ) = u(ed·ran )2 + u(eprt·cal )2 + u(ebath )2 (11.9)
q
= (0.0019)2 + (0.0078)2 + (0.0029)2
= 0.0085 °C .

11.8 Using the glass thermometer

When a calibrated glass thermometer is being used, it is the tem-

perature T in the immediate vicinity of the thermometer that is of
interest. We can model this by rearranging the observation equa-
tion (11.1), in §11.2,

T = wglass [i ] − Eglass·ran [i ] − Dglass . (11.10)

This describes how a single thermometer reading is related to

temperature.

11.8.1 Influence estimates

Dglass : already determined by the calibration procedure

dglass = 0.0052 °C

with
u(dglass ) = 0.0085 °C .

Eglass·ran : the estimate is eglass·ran = 0 °C, because positive or

negative variations are assumed to be equally likely, and there
will be no tendency for one to occur more than the other.
48 an introduction to measurement uncertainty

Because individual thermometer readings are now used to mea-

sure the temperature, the sample standard deviation (evaluated
earlier, see §11.6.1) should be used to characterise the width of
the Gaussian distribution describing Eglass·ran .5 So the standard
uncertainty u(eglass·ran ) = 0.0048 °C.
5
The randomness of PRT readings
also contributed to this variability.
However, the measurement procedure
imposed a condition on the magni-
tude of variability of PRT readings
11.8.2 Evaluating the measurement result compared to the variability of glass
thermometer readings, so the contribu-
tion from glass thermometer readings
The result of a measurement of T from a single thermometer read- will dominate.

ing can be obtained by evaluating equation (11.10) with our best

estimates for each quantity. That is,

t = wglass [i ] − 0.000 − 0.0052 °C .

with a standard uncertainty

q
u(t) = u(dglass )2 + u(eglass·ran )2 (11.11)
q
= (0.0085)2 + (0.0048)2
= 0.0098 °C .

11.9 Deciding what to report

The information of interest to the end-user of the thermometer

is most likely to be a value of dglass , which can be used to correct
for the thermometer offset error, and the uncertainty in a single
corrected reading u(t), from equation (11.11).

Note that one is not the uncertainty of the other!

A calibration laboratory may choose to report dglass and the uncer-

tainty u(t), rather than u(dglass ) from equation (11.9), because the
corrected-reading uncertainty is more useful to the client. When
this is done, the calibration report should make the distinction clear.
12 Final words

This booklet provides an introduction to the subject of measure-

ment uncertainty analysis, with an emphasis on modelling. We
have shown that measurement modelling is important for the data
processing of results and calculation of uncertainty, but models are
useful in other ways too. They can both describe a measurement
process and serve as a tool for understanding, refining and imple-
menting that process; they can offer insight into the strengths and
weaknesses of a particular measurement process and shed light
on ways to improve it; they can serve as a specification, during the
development of data-analysis algorithms; and they can be used as a
reference when validating software.

Measurement models must be documented, and the documen-

tation should be maintained over time; terms should be clearly
defined and assumptions made about influence quantities should
be recorded. All too often, these details are elided, or overlooked
entirely. In accredited quality assurance environments, the valid-
ity of a measurement model will often be closely scrutinised and
adequate evidence to support a model will be required. Model doc-
umentation should be prepared so that it can be readily understood
by other metrology professionals, who are not necessarily experts in
the field.

There are some aspects of uncertainty analysis that have not been
discussed here, such as what to do when only a small sample of
observations is available,1 or when estimates of quantities are corre-
lated. However, the main difference in the way that these situations
1
This is known as finite degrees of
freedom in metrology vernacular.
must be handled is that a slightly more elaborate mathematical
process will be applied to the measurement model to calculate the
uncertainty.2 The calculations involved should present little diffi-
culty as long as the basic ideas described here are well understood.
2
There are software tools available that
take care of the mathematical details.

Uncertainty analysis is not easy, but skill and confidence in analysing

measurement processes will improve with practice and experience.
There is no ‘right’ or ‘wrong’ way of doing it: more or less detail
may be used, different distributions may be selected to describe
influences, and the model may be refined and extended over time.
Index

decision making measurement error, 9 random behaviour, 10

level of confidence, 20 and decision making, 20 rounding error, 12
measurement model, 9, 22, 25, 31
estimate, 9 refining, 23
sample statistics, 14
sample mean, 16 measurement process, 30
sample mean, 15
single observation, 15
sample standard deviation, 15, 17
expanded uncertainty
observation equation, 14, 17 standard error of the mean, 16, 17
and decision making, 20
sampling, 14, 17
coverage factor, 19, 33
standard deviation, see probability
coverage probability, 19 probability, 10
distribution, see also sample
level of confidence, 19 relative frequency, 10
statistics
probability distribution, 11
influence quantity, 9, 30, 32, 33 Gaussian distribution, 11
mean of, 11 uncertainty
mean, see probability distribution, see standard deviation of, 11 evaluating, 25
also sample statistics triangular distribution, 12, 13 expanded uncertainty, 18
measurand, 8, 30, 31 uniform distribution, 12, 13 standard uncertainty, 18