Probability Distributions For Measurement Uncertainty - Isobudgets
Probability Distributions For Measurement Uncertainty - Isobudgets
Menu
Home
Probability Distributions for Measurement
Uncertainty
Calculators
Uncertainty
Blog
Downloads
About
October 28, 2015 by Richard Hogan
Contact
You are here:
Search...
1. Home
Search
2. ISOBudgets blog by Rick Hogan
3. Probability Distributions for Measurement Uncertainty
Probability distributions are a part of measurement uncertainty analysis that people continually struggle with.
Today, my goal is to help you learn more about probability distributions without having to grab a statistics
textbook. Although there are hundreds of probability distributions that you could use, I am going to focus on
the 6 that you need to know.
If you constantly struggle with probability distributions, keep reading. I am going to explain what are probability
distributions, why they are important, and how they can help you when estimating measurement uncertainty.
What is a probability distribution
Simply explained, probability distributions are a function, table, or equation that shows the relationship between
the outcome of an event and its frequency of occurrence.
http://www.isobudgets.com/probabilitydistributionsformeasurementuncertainty/ 1/18
6/10/2017 Probability Distributions for Measurement Uncertainty | isobudgets
Probability distributions are helpful because they can be used as a graphical representation of your measurement
functions and how they behave. When you know how your measurement function have performed in the past,
you can more appropriately analyze it and predict future outcomes.
Before jumping headfirst into the different types of probability distributions, let’s first learn a little more about
probability distributions. In the next few paragraphs, I am going to explain some characteristics that you should
know.
Histogram
A histogram is a graphical representation used to understand how numerical data is distributed. Take a look
below at the histogram of a Gaussian distribution.
Look at the histogram and view how the majority of the data collected is grouped at the center. This is called
central tendency.
Now look at height of each bar in the histogram. The height of the bars indicate how frequent the outcome it
represents occurs. The taller the bar, the more frequent the occurrence.
Skewness
Skewness is a measure of the probability distributions symmetry. Look at the chart below to visually understand
how probability distributions can skew to the left or the right.
http://www.isobudgets.com/probabilitydistributionsformeasurementuncertainty/ 2/18
6/10/2017 Probability Distributions for Measurement Uncertainty | isobudgets
Kurtosis
Kurtosis is a measure of the tailedness and peakedness relative to a normal distribution. As you can see from the
image below, distributions with wider tails have smaller peaks while distributions with greater peaks have
narrower tails. Do you see the relationship?
Why is it important
I know it seems like I am making you read more information that you want to know, but it is important to know
these details so you can select the appropriate probability distribution that characterizes your data.
If you are uncertain how your data is distributed, create histogram and compare it to the following probability
distributions.
Most commonly used
The most commonly used probability distributions for estimating measurement uncertainty are;
http://www.isobudgets.com/probabilitydistributionsformeasurementuncertainty/ 3/18
6/10/2017 Probability Distributions for Measurement Uncertainty | isobudgets
Normal
Rectangular
UShaped
Triangle
LogNormal
Rayleigh
Below you will find a list of the most common probability distributions used in uncertainty analysis. After
reading this article, you should be able identify which probability distributions you should use and how to
reduce your uncertainty contributors to standard deviation equivalents.
Gaussian (a.k.a. Normal) Distribution
The Normal distribution is a function that represents the distribution of many random variables as a symmetrical
bellshaped graph where the peak is centered about the mean and is symmetrically distributed in accordance
with the standard deviation.
The normal distribution is the most commonly used probability distribution for evaluating Type A data. If you
do not know what Type A data is, it is the data that you collect from experimental testing, such as repeatability,
reproducibility, and stability testing.
To get a better understanding, imagine you are going to collect 100 measurement samples and create a histogram
graph with your results. The histogram for your data should resemble a shape close to a normal distribution.
The more data that you collect, the closer your histogram will begin to resemble a normal distribution.
Now, I do not expect you to collect 100 samples every time you perform repeatability and reproducibility test.
Instead, I recommend that you begin by collecting 20 to 30 samples for each test. This should give you a good
baseline to begin with, and allow you to characterize your data with a normal distribution.
To reduce normally distributed data to a standard deviation equivalent, use the following equation. The
variable U will be the value of your uncertainty contributor and k is the value of your expansion factor.
http://www.isobudgets.com/probabilitydistributionsformeasurementuncertainty/ 4/18
6/10/2017 Probability Distributions for Measurement Uncertainty | isobudgets
For example, if you collect 20 samples for a repeatability experiment and calculate the standard deviation, the
value of k is 1. If you are wondering, it is equal to 1 because your standard deviation is already at the 1sigma
level (i.e. 68.27% confidence).
So, if your calculated standard deviation is 1 ppm, then;
When using Microsoft Excel to calculate measurement uncertainty, use the following equation:
=[Cell1]/1
For the next example, imagine you are evaluating the measurement uncertainty from your calibration report.
Most likely, it is reported to 95% confidence where k equals 2 (I am sure that you have read this somewhere
before). If your reported uncertainty is 1 ppm, then;
When using Microsoft Excel to calculate measurement uncertainty, use the following equation:
=[Cell1]/2
Rectangular (a.k.a. Uniform) Distribution
The Rectangular Distribution is a function that represents a continuous uniform distribution and constant
probability. In a rectangular distribution, all outcomes are equally likely to occur.
The rectangular distribution is the most commonly used probability distribution in uncertainty analysis. If you
http://www.isobudgets.com/probabilitydistributionsformeasurementuncertainty/ 5/18
6/10/2017 Probability Distributions for Measurement Uncertainty | isobudgets
are wondering why, it is because it covers the majority of uncertainty factors where the evaluator says, “I am not
sure how the data is distributed.”
When you are not confident how your data is distributed, it is best evaluate it conservatively. In this situation,
the rectangular distribution is a great default option which is why most ISO/IEC 17025 assessors recommend it.
So, make sure to pay attention, you will be using this probability distribution a lot.
To reduce your uncertainty contributors to standard deviation equivalents, you will want to divide your values
by the squareroot or 3.
For example, if you performing measurement uncertainty analysis and evaluating the contribution of a factor
that has an influence of 1 partpermillion and you propose that the data is uniformly distributed, then;
When using Microsoft Excel to calculate measurement uncertainty, use the following equation:
=[Cell1]/SQRT(3)
Ushaped Distribution
The Ushaped Distribution is a function that represents outcomes that are most likely to occur at the extremes of
the range. The distribution forms the shape of the letter ‘U,’ but does not necessarily have to be symmetrical.
The ushaped distribution is helpful where events frequently occur at the extremes of the range. Consider the
thermostat that controls the temperature of your laboratory. If you are not using a PID controller, your thermostat
controller only attempts to control temperature by activating at the extremes.
http://www.isobudgets.com/probabilitydistributionsformeasurementuncertainty/ 6/18
6/10/2017 Probability Distributions for Measurement Uncertainty | isobudgets
For example, imagine that your laboratory thermostat is set at 20°C and controls temperature to 1°C. Most
likely, your thermostat does not activate your HVAC system until the laboratory temperature reaches either 19°C
or 21°C. This means that your laboratory is not normally at 20°C. Instead, your laboratory temperatures are
floating around the limits of the thermostat’s thresholds before activating or deactivating.
For this reason, it is best to characterize your laboratory temperature data using a ushaped distribution.
To reduce your uncertainty contributors to standard deviation equivalents, you will want to divide your values
by the squareroot or 2.
So, if you performing measurement uncertainty analysis and evaluating the contribution of a factor that has an
influence of 1 partpermillion and you propose that the data for this factor is ushaped distributed, then;
When using Microsoft Excel to calculate measurement uncertainty, use the following equation:
=[Cell1]/SQRT(2)
Triangle Distribution
The Triangle Distribution is a function that represents a known minimum, maximum, and estimated central
value. It is commonly referred to as the “lack of knowledge” distribution because it is typically used where a
relationship between variables is known, but data is scare.
Additionally, the triangle distribution is commonly used where the data collection is difficult or expensive.
http://www.isobudgets.com/probabilitydistributionsformeasurementuncertainty/ 7/18
6/10/2017 Probability Distributions for Measurement Uncertainty | isobudgets
For a real world example, image your laboratory is temperature controlled using a PID thermostat controller.
The PID thermostat controller is constantly trying to achieve the target temperature setpoint. For this reason,
the temperatures in your laboratory are constantly floating around 20°C and rarely reach the temperature
thresholds (i.e. limits) of a typical thermostat controller.
What this means is that most of your laboratory’s temperature data is centered around your set temperature.
Therefore, it is best characterized by a triangular distribution because we know the limits and the estimated
mean but we are unsure how the data is distributed between these points.
To reduce your uncertainty contributors to standard deviation equivalents, you will want to divide your values
by the squareroot or 6.
So, if you performing measurement uncertainty analysis and evaluating the contribution of a factor that has an
influence of 1 partpermillion and you propose that the data for this factor is triangle distributed, you would
reduce it’s value by using the equation above.
If you using Microsoft Excel to calculate measurement uncertainty, use the following equation:
=[Cell1]/SQRT(6)
LogNormal Distribution
The lognormal distribution is a function of a natural logarithm that is normally distributed.
The lognormal distribution is a distribution that is commonly encountered but rarely used. Most of the time it is
the result of lack of knowledge or failure to develop a histogram for your data.
http://www.isobudgets.com/probabilitydistributionsformeasurementuncertainty/ 8/18
6/10/2017 Probability Distributions for Measurement Uncertainty | isobudgets
For example, if you are performing measurements that are finite, such as length, height, weight, etc., you are
most likely going to end up with data that resembles a lognormal distribution. It is most common in
dimensional and mechanical metrology.
To give a better understanding, think of calibrating a gage block. Before you begin calibration, you know the
target length. If you perform repeated measurements at the single point on the gage block, the majority of your
measurement results will be centered around the actual length of the gage block. Some measurement results will
be larger than the actual value of the gage block, and much fewer measurement results will be less than the
actual value of the gage block.
The reason this happens is your measurement results are limited by the length of the gage block. Realistically,
you cannot measure less than the length of the block; so, your measurement results are finite or limited.
Make sure to consider the lognormal distribution next time you are performing measurements that are finite. It
may prevent you from encountering measurement errors and miscalculated uncertainties.
To reduce your uncertainty contributors to standard deviation equivalents, you will want use the following
equation.
Where,
m = median
q = limit
Rayleigh Distribution
Rayleigh distributions are used when the magnitude of a vector is associated with it’s directional components
(e.g. x and y), which can also be real and imaginary components (e.g. i and j).
When directional components are orthogonal and normally distributed, the resulting vector will be Rayleigh
distributed.
Rayleigh distributions are commonly used in electrical metrology for RF and Microwave functions.
Additionally, they are commonly used in mechanical metrology where vectors are involved.
http://www.isobudgets.com/probabilitydistributionsformeasurementuncertainty/ 9/18
6/10/2017 Probability Distributions for Measurement Uncertainty | isobudgets
For example, when wind velocity is analyzed by its 2 dimensional vector components, x and y, the resulting
vector is Rayleigh distributed. For this to happen, x and y must be orthogonal and normally distributed.
Reducing uncertainty components to standard deviation equivalents is tricky with the Rayleigh distributions.
You will need to know the standard deviation of each directional component to calculate the measurement
uncertainty of the vector component. Afterward, you can use the equation below to reduce your uncertainty
component to a standard deviation equivalent.
For a better explanation, click the link below to read this paper by Michael Dobbert from Agilent (now
Keysight).
Revisiting Mismatch Uncertainty with the Rayleigh Distribution
Conclusion
Probability distributions are an important part of understanding the behavior of functions, analyzing data, and
predicting future outcomes. This is why they are a critical component of uncertainty analysis. If you are
estimating measurement uncertainty without considering probability distributions, you are going to make
mistakes. So make sure to use this guide as a reference when calculating uncertainty.
Additionally, it never hurts to use this chart.
http://www.isobudgets.com/probabilitydistributionsformeasurementuncertainty/ 10/18
6/10/2017 Probability Distributions for Measurement Uncertainty | isobudgets
I hope that you have found this article helpful for your uncertainty analysis. Leave me a comment and tell me
probability distributions you use in your uncertainty analysis.
References
Castrup, H. (2007). Distributions for Uncertainty Analysis. Bakersfield, CA: Integrated Sciences Group.
Castrup, H. (2009). Error Distribution Variances and Other Statistics. Bakersfield: Integrated Sciences Group.
DeCarlo, L. T. (1997). On the Meaning and Use of Kurtosis. American Psychological Association Inc., 292302.
Dobbert, M., & Gorin, J. (2011). Revisiting Mismatch Uncertainty with the Rayleigh Distribution. Santa Rosa:
Agilent.
Petty, N. W., & Dye, S. (2013). Triangular Distributions. Christchurch: Statistics Learning Centre.
If I have left anything out of this article or if you can think of any additional tips that would improve this list,
please leave a comment or contact me to share your advice.
Posted in:
Uncertainty
Tagged with:
lognormal distribution
measurement uncertainty
normal distribution
probability distributions
rayleigh distribution
rectangular distribution
See more
1. Prev:3 Ways to Combine Measurement Uncertainty with Different Units of Measure
http://www.isobudgets.com/probabilitydistributionsformeasurementuncertainty/ 11/18
6/10/2017 Probability Distributions for Measurement Uncertainty | isobudgets
2. Back: All Posts
3. Next:8 Sources of Uncertainty in Measurement for Every Uncertainty Budget
About the Author
Richard Hogan
Richard Hogan is the CEO of ISO Budgets, L.L.C., a U.S.based consulting and data analysis firm. Services
include measurement consulting, data analysis, uncertainty budgets, and control charts. Richard is a systems
engineer who has laboratory management and quality control experience in the Metrology industry. He
specializes in uncertainty analysis, industrial statistics, and process optimization. Richard holds a Masters degree
in Engineering from Old Dominion University in Norfolk, VA. Connect with Richard on Google+ and LinkedIn.
Related entries
5 Reproducibility Tests You Can Use For Estimating Uncertainty
How to Calculate Sensitivity Coefficients for Measurement Uncertainty
15 Places to Find Sources of Uncertainty in Measurement
http://www.isobudgets.com/probabilitydistributionsformeasurementuncertainty/ 12/18
6/10/2017 Probability Distributions for Measurement Uncertainty | isobudgets
How to Add An Uncertainty Calculator to Your Calibration Checklist
7 Comments
Kiwi Tan
1. November 1, 2015 at 3:42 am
Hello Mr. Rick Hogan,
It is a very nice explanation.
I am try to figure out how the measurement result can be linked to the distribution. I have the following
would appreciate for your kind advise.
I have a Calibrator 5720A, Calibrated by Fluke with 95% of confidence level, k=2.
With refer to the Calibration Data, example DC Voltage Accuracy Test, 220mV Range, test point
100.00000mV, result is 99.99880 and expanded uncertainty is 4.8e007V.
How would this uncertainty 4.8e007V is link to the belt curve (normal distribution)? the formula is U/k
where 4.8e007/2 = 2.4e07.
Appreciate for your kind explanation.
Thank you.
Reply
Richard Hogan
November 1, 2015 at 12:59 pm
Hi Kiwi,
In probability theory and statistics, there is a theorem known as the Central Limit Theorem. When
you combine multiple uncertainty sources, each with their own probability distribution, the
combined result approaches an approximated normal distribution. This is why the uncertainty
reported in your calibration report is normally distributed. Since it is expanded to 95% confidence,
this is why you divide it by 2 to reduce your calibration uncertainty to a standard deviation
equivalent.
http://www.isobudgets.com/probabilitydistributionsformeasurementuncertainty/ 13/18
6/10/2017 Probability Distributions for Measurement Uncertainty | isobudgets
I hope this helps.
Best regards,
Rick Hogan
Reply
2. Peter Benko
November 4, 2015 at 1:36 pm
Dear Kiwi,
Let me bring some light into this topic. You got calibration certificate with result of calibration which
states at conventionally true value (I believe) 100.00000 V the expected value on your instrument is
99.99880 V ± 0.048x10E05 V with 95% probability of normal distribution, coverage coefficient k=2?.
Now when you take reading with your calibrated voltmeter, you should create your own uncertainty
budget, where one of the uncertainty contribution is from calibration of voltmeter which is given by
calibration laboratory written on your certificate. Now! Each of your contribution has own probability
distribution, probability distribution of your calibrated voltmeter has normal distribution expressed as an
expanded uncertainty, so therefore you have to bring it down to the level of standard uncertainty which
probability is approximately 66%. That means you have divide expanded uncertainty by coverage
coefficient u=U/k Then you can combine with other contributions and express your own measurement
uncertainty.
Kind regards,
Peter
Reply
3. Kiwi
November 12, 2015 at 7:26 pm
Thanks Rick and Peter.
Reply
Richard Hogan
November 12, 2015 at 10:43 pm
You are welcome Kiwi!
Reply
4. sakir
http://www.isobudgets.com/probabilitydistributionsformeasurementuncertainty/ 14/18
6/10/2017 Probability Distributions for Measurement Uncertainty | isobudgets
December 7, 2015 at 2:36 am
Nice Document
Reply
Richard Hogan
December 19, 2015 at 11:13 am
Hi Sakir,
Thank you. I am glad that you like my article. Please share it with people that you like would like it
too.
Reply
Leave a Reply
Your email address will not be published. Required fields are marked *
Comment
Name *
Email *
Website
Post Comment
About Rick Hogan
http://www.isobudgets.com/probabilitydistributionsformeasurementuncertainty/ 15/18
6/10/2017 Probability Distributions for Measurement Uncertainty | isobudgets
Richard Hogan, M.Eng.
An Engineer, Metrologist, and Manager who answers questions and delivers solutions to ISO 17025 accredited
testing and calibration laboratories. Learn more about me here.
Email: [email protected]
Call: 1.757.525.2004
http://www.isobudgets.com/probabilitydistributionsformeasurementuncertainty/ 16/18
6/10/2017 Probability Distributions for Measurement Uncertainty | isobudgets
More About Uncertainty
5 Reproducibility Tests You Can Use For Estimating Uncertainty
Proficiency Testing and Interlaboratory Comparisons: The Ultimate Guide to ISO/IEC 17025
How to Calculate Sensitivity Coefficients for Measurement Uncertainty
15 Places to Find Sources of Uncertainty in Measurement
8 Free Uncertainty Calculator Software You Can Download Now
Learn to Calculate Measurement Uncertainty
I'm sharing everything I know about Measurement Uncertainty! Get updates when I publish new articles.
Your Email
Subscribe
About Us
ISOBUDGETS is a consulting firm specializing in the analysis of uncertainty in measurement. Our consulting
services are targeted to assist calibration and testing laboratories to attain and retain ISO/IEC 17025:2005
accreditation.
Services
Uncertainty Budgets
Uncertainty Training
ISO 17025 Consulting
Guides & Calculators
Contact Info
ISOBUDGETS LLC
P.O. Box 738
Williamsburg, VA 23188
757.525.2004
[email protected]
Subscribe to my Blog
http://www.isobudgets.com/probabilitydistributionsformeasurementuncertainty/ 17/18
6/10/2017 Probability Distributions for Measurement Uncertainty | isobudgets
Get updates when new blogs are posted.
Your Email
Subscribe
Home
Uncertainty
Guides
Blog
About
Contact
© 2017 ISOBUDGETS
Back to Top
http://www.isobudgets.com/probabilitydistributionsformeasurementuncertainty/ 18/18