You are on page 1of 15

QUALITY AND RELIABILITY ENGINEERING INTERNATIONAL

Qual. Reliab. Engng. Int. 2005; 21:605–619


Published online 10 March 2005 in Wiley InterScience (www.interscience.wiley.com). DOI: 10.1002/qre.678

Research Measurement Uncertainty in the


Calibration of Low-flow Ambient
Air Samplers
Shaun S. Wulff1,∗,† and Mark A. Weitz2
1 Department of Statistics, University of Wyoming, Laramie, WY 82071, U.S.A.
2 Department of Atmospheric Science, University of Wyoming, Laramie, WY 82071, U.S.A.

This paper considers the quality of measurement for a new standard used to assess
ambient airflow measurements. The quality is quantified through the uncertainty
associated in the measurements of this device. While guidelines are available for
these analyses, this case study presents a set of additional challenges since the
flow rate measurements are based on a linear relationship between two functions
where the variables making up these functions are measured with non-negligible
error. In order to estimate the parameters describing this linear relationship, it is
necessary to account for the variability associated with these errors. Measurement
error modeling approaches can be used to estimate the linear regression parameters
and the associated uncertainties. Ordinary least squares procedures are simpler,
but must be used with care since the estimates can be biased and the resulting
uncertainties can be underestimated. Once estimates for the linear relationship are
obtained, measurements from the device are used to calculate a predicted flow rate.
Techniques are presented for estimating the uncertainty associated with this predicted
flow rate. Copyright 
c 2005 John Wiley & Sons, Ltd.

KEY WORDS: flow metrology; error in variables; linear functional model; covariance; correlation; Taylor
series approximation

1. INTRODUCTION

T
he uncertainty in measurement is an unknown parameter that characterizes the dispersion of values for
the quantity being measured. Of particular interest is standard uncertainty or the standard deviation
of the measured quantity (ANSI1 , pp. 10–11). Uncertainty is calculated based on assumed probability
distributions, available knowledge, or is estimated from data. The estimation of uncertainty can be vital in
determining the quality and reliability of a measuring device or process. The American National Standards
Institute (ANSI) has prepared a complete guide for assessing and calculating uncertainty1. Atwood and
Engelhardt2 discussed general techniques for uncertainty analyses and they apply these techniques to a specific
example.

∗ Correspondence to: S. S. Wulff, Department of Statistics, University of Wyoming, Department 3332, 1000 East University Avenue,
Laramie, WY 82071, U.S.A.
† E-mail: wulff@uwyo.edu

Received 9 July 2003


Copyright 
c 2005 John Wiley & Sons, Ltd. Revised 7 January 2004
606 S. S. WULFF AND M. A. WEITZ

In many engineering settings, the relationship between an independent variable and a response variable
is used to calibrate a measurement process. In cases where the independent variable is measured exactly or
where the uncertainty in the independent variable is negligible, standard least squares procedures can be used
to estimate the functional relationship. However, when the uncertainty in the independent variable is non-
negligible, standard least squares procedures are not appropriate.
Measurement error models are useful for this problem as recognized by ANSI1 (p. 78), but the guidelines do
not provide details on how to use these models. Researchers from the National Physical Laboratory (NPL)
or the National Measurement Laboratory of Great Britain3 have begun to use measurement error models
and have developed computational algorithms to estimate model parameters. In uncertainty analyses, not
only are the estimates of interest, but also the uncertainties of these estimates as well as the uncertainties
of the predicted values. There are also practical considerations for applying these models to problems in
metrology.
Through a case study, this paper presents approaches and considerations when there are errors in the
variables used in a calibration procedure. In this study, the calibration relies on a linear relationship between
two variables that have non-negligible uncertainty. Techniques from Fuller4 and Reilman et al.5 are used to
assess the uncertainty of estimates of model parameters as well as provide the uncertainty for the predicted
values. This problem arises in the area of flow metrology with regards to a new generation of ambient air
samplers that which are capable of operating at lower flow rates at or below 16.7 l min−1 . When these air
samplers emerged several years ago, there were limited tools available with which to assess and calibrate their
operating flow rate. In 1996, the StreamlineTM Flow Transfer Standard (SFTS) was developed to assess the
new air samplers. The SFTS, developed by Chinook Engineering, is now a widely used standard with the
establishment in 1999 of a nationwide PM2.5 sampling network which exclusively use low-flow air samplers.
With increased use of the SFTS, questions arose about its accuracy and the uncertainty associated with
its measurements. Since the uncertainty of the SFTS is of primary interest, this paper presents a thorough
discussion of the uncertainty analysis. These details demonstrate the necessary steps and considerations
for assessing measurement uncertainty in the calibration where the variables are subject to non-negligible
error.
Using concepts from flow metrology, Section 2 describes the calibration procedure and the functions that
are of interest in this analysis. Section 3 describes the assessment of uncertainties of the variables making up
two particular functions. Section 4 provides an approximation of the overall uncertainty for these functions.
Section 5 describes the use of measurement error models for assessing the uncertainty in the linear fit involving
the two functions of interest. Section 6 presents the uncertainty for the resulting predicted values.

2. THE CALIBRATION PROCEDURE


The calibration procedure dictates the steps of the uncertainty analysis. The apparatus for the calibration,
illustrated in Figure 1, is designed according to consultation with NIST flow metrologists. Critical flow
venturis (CFVs) or sonic nozzles are used as primary mass flow elements in the calibration apparatus with
associated high-precision temperature and pressure instrumentation. Five measurements of interest include
ambient temperature (Tamb ), ambient pressure (Pamb ), the pressure drop across the SFTS (P), pressure of
the fluid at the inlet of the CFV (Pin ), and temperature of the fluid at the inlet of the CFV (Tin).
The mass flow rate of air through the calibration apparatus can be calculated as in ASME6

Pin AC ∗ Cd
m= √ (1)
RTin

where A is the cross-sectional area of the CFV, C ∗ is a thermodynamic state variable that is a function of the
fluid pressure and temperature in the throat of the CFV, Cd is the discharge coefficient of the CFV, and R is the
gas constant with the appropriate units.

Copyright 
c 2005 John Wiley & Sons, Ltd. Qual. Reliab. Engng. Int. 2005; 21:605–619
MEASUREMENT UNCERTAINTY IN THE CALIBRATION OF LOW-FLOW AMBIENT AIR SAMPLERS 607

Figure 1. Calibration apparatus for the StreamlineTM Flow Transfer System (no scale)

The mass flow rate through the CFV can be converted to a volumetric flow rate by dividing (1) by the density
of ambient air (ρamb ), which gives

m RTamb Pin Tamb


Y= =m =K √ (2)
ρamb Pamb M Pamb Tin

using the ideal gas law where M is the appropriate unit adjustment and K is a variable combining A, C ∗ , Cd ,
R, and M. The calibration is based on Bernoulli’s equation, which describes the theoretical flow rate (Yt ) as
  √  
P AGCd R P Tamb 1/2
Yt = AGCd = √ (3)
ρamb M Pamb

where G is the gas expansion factor. Since the quantity in parentheses is of particular interest, define
 1/2
P Tamb
X= (4)
Pamb

Through the calibrated range of the SFTS, it is assumed that the quantity in brackets in (3) is constant, and so
by Bernoulli’s equation there exists a theoretical linear relationship between Y in (2) and X in (4). This linear
relationship can be summarized as

Y = β0 + β1 X (5)

Values for Y and X are found from the calibration apparatus for seven specific levels of P ranging from
2.91  H2 O (inches of water) to 7.46  H2 O. The ambient pressure and ambient temperature are constant in the
laboratory across the seven levels. Using these values for Y and X, the quantities β0 and β1 in (5) are estimated,
resulting in a unique calibration for each SFTS. The assumed linear relationship in (5) is substantiated by
coefficient of determination (R 2 ) values exceeding 0.99.
Given estimates of β0 and β1 as well as a value for X in (5), the flow rate can be estimated or predicted
based solely on information from the SFTS. The purpose of the uncertainty analysis is to ascertain the overall
standard uncertainty or standard deviation of the predicted flow rate. However, this uncertainty consists of the
uncertainty of the estimates of β0 and β1 along with the uncertainties associated with the measurements of X
and Y . In addition, the estimation and uncertainties in β0 and β1 are affected by the uncertainties in X and Y .
Consideration of the correlations can also be used to reduce the overall uncertainty for the estimated flow rate.

Copyright 
c 2005 John Wiley & Sons, Ltd. Qual. Reliab. Engng. Int. 2005; 21:605–619
608 S. S. WULFF AND M. A. WEITZ

3. ELEMENTAL SOURCES OF UNCERTAINTY


This section details the elemental sources of uncertainty in the variables that form the functions Y and X in (2)
and (4). Careful estimation of these uncertainties indicate the extent of measurement error present in Y and X
as well as potential correlations in the error sources. The variables of interest include K, Pin , Pamb , Tin , Tamb
and P . The true values of the variables are taken to be nominal values that are expected to occur in practice.
Uncertainty in the measurement of these variables arises when the true value cannot be measured directly due
to random errors. For general purposes of discussion, consider some variable with true value V and observable
value v. The error attributable to this variable
√ is εV = v − V . It is assumed that the expected value of εV is zero
and the standard uncertainty is uV = var(εV ).
Another quantity of interest for obtaining the overall uncertainty is the sensitivity coefficient, which describes
how the functions Y or X vary with changes in the values of the variable making up that function. The sensitivity
coefficient is calculated by taking the partial derivative of Y or X with respect to a variable in that quantity and
evaluating the partial derivative at the true values.
The calculation of the uncertainties and sensitivity coefficients is illustrated for each of the six variables of
interest. Table I provides a summary of the units of measurement, nominal values, values of the uncertainties,
and values of the sensitivity coefficients. A description is given to how these values are obtained.

The variable K
The variable K incorporates the variables A, C ∗ , Cd as described by (2). Since these variables relate to the
performance of the CFV, the error in measurement (εK ) arises from the CFV itself. The manufacturer reports
an uncertainty of 0.25%. The nominal value for K is 1.6457 × 10−3 l atm K−1/2 mmHg−1 min−1 , which
is calculated from the nominal values of the variables making up K. The resulting standard uncertainty is
uK = 4.1143 × 10−6 . From (2), the sensitivity coefficient evaluated at the nominal values of the variables is
∂Y Pin Tamb
sK = =√ = 9911.2 mmHg K1/2 atm−1
∂K Tin Pamb

Pressure at the inlet variable


The variable Pin consists of two independent sources of error. The first error source (ε1Pin ) is due
to the calibration of the electronic pressure instrument against a mercury-in-glass absolute barometer.
The manufacturer states an uncertainty of 0.02%, which is interpreted as the maximum bounds for all calibration
values. The probability distribution is conservatively
√ assumed to be uniform, so the uncertainty for a nominal
value of 500 mmHg is u1Pin = (0.0002/ 12) × 500 = 0.028 868 (ANSI1 , p. 13). A second source of error
(ε2Pin ) is due to instrument resolution. The uncertainty is u2Pin = 0.1 mmHg. Assuming these errors are additive
(εPin = ε1Pin + ε2Pin ) and independent, the combined uncertainty is

uPin = (var(ε1Pin ) + var(ε2Pin ))1/2 = (u21Pin + u22Pin )1/2 = 0.104 08 mmHg

Using (2), the sensitivity coefficient evaluated at the nominal values is


∂Y KTamb
sPin = =√ = 0.032 622 l mmHg−1 min−1
∂Pin Tin Pamb

Ambient pressure variable


The uncertainty calculations for ambient pressure are similar to those for inlet pressure. The first source
of error (ε1Pamb ) is due to the calibration of the electronic pressure instrument against a mercury-in-glass
absolute barometer. The manufacturer states an uncertainty of 0.02%, which is interpreted as the maximum
bounds for all calibration values. The probability distribution is conservatively assumed to be uniform, so

Copyright 
c 2005 John Wiley & Sons, Ltd. Qual. Reliab. Engng. Int. 2005; 21:605–619
Copyright 
c 2005 John Wiley & Sons, Ltd.
Table I. Elementary uncertainties and associated sensitivities. When there are two sensitivity coefficients or variance contributions for a single variable, the first is for the
function Y and the second is for the function X

Source
K Pin Pamb Tin Tamb P
Units l atm K−1/2 mmHg−1 min−1 mmHg atm K K  H O
2
Nominal value 1.6457 × 10−3 500.00 0.865 00 294.00 294.00 4.9000
Uncertainty 4.1143 × 10−6 0.104 08 1.4113 × 10−4 0.133 79 0.132 29 7.9051 × 10−3
Sensitivity coefficient 9911.2 0.032 622 −18.857/−23.589 −0.027 740 0.055 479/0.069 404 4.1643
Variance contribution 1.6628 × 10−3 1.1528 × 10−5 7.0821 × 10−6 /1.1083 × 10−5 1.3774 × 10−5 5.3866 × 10−5 /8.4300 × 10−5 1.0837 × 10−3
MEASUREMENT UNCERTAINTY IN THE CALIBRATION OF LOW-FLOW AMBIENT AIR SAMPLERS

Qual. Reliab. Engng. Int. 2005; 21:605–619


609
610 S. S. WULFF AND M. A. WEITZ


the uncertainty for a nominal value of 0.865 atm is u1Pamb = (0.0002/ 12) × 0.865 = 4.9941 × 10−5 atm.
A second source of error (ε2Pamb ) is due to instrument resolution. The uncertainty is u2Pamb = 0.1 mmHg or
1.32 × 10−4 atm as 1 atm = 760 mmHg. Assuming independence, the combined uncertainty arising from the
error εPamb = ε1Pamb + ε2Pamb is

uPamb = (var(ε1Pamb ) + var(ε2Pamb ))1/2 = (u21Pamb + u22Pamb )1/2 = 1.4113 × 10−4 atm

There are two sensitivity coefficients corresponding to Y in (2) and X in (4). Evaluating these at the nominal
values gives
∂Y KPin Tamb
sPamb −Y = =√ = −18.857 l atm−1 min−1
∂Pamb Tin Pamb
2
√ √
∂X P Tamb  √
sPamb −X = =− = −23.589  H2 O K atm−1.5
∂Pamb 1.5
2Pamb

Inlet temperature variable


The variable Tin consists of three independent sources of error. The first source of error (ε1Tin ) is due to the
calibration of the electronic thermometer against a mercury-in-glass thermometer in which the manufacturer
states an uncertainty of 0.3 K. The manufacturer’s claim is interpreted as the maximum bounds for all
calibration values.
√ The probability distribution is conservatively assumed to be uniform, so the uncertainty
is u1Tin = (0.3/ 12) = 0.086 603 K. The second source of error (ε2Tin ) arises from the thermocouple probe,
which is heated by friction from the air passing over it and the heat is dissipated by conduction and radiation
of the surrounding surfaces. A conservative value for this uncertainty is u2Tin = 0.02 K. The last component of
uncertainty (ε3Tin ) corresponds to the instrument resolution which is u3Tin = 0.1 K. Under the assumption that
these errors are additive (εTin = ε1Tin + ε2Tin + ε3Tin ) and independent, the combined uncertainty is

uTin = [var(ε1Tin ) + var(ε2Tin ) + var(ε3Tin )]1/2 = 0.133 79 K

Using (2), the sensitivity coefficient evaluated at nominal values, including the nominal value 294 K for Tin , is
given by
∂Y KPin Tamb
sTin = = − 1.5 = −0.027 740 l K−1 min−1
∂Tin 2Pin Pamb

Ambient temperature variable


The sources of error for the variable Tamb are due to the manufacturer’s uncertainty for the thermometer (ε1Tamb )
and the instrument resolution (ε2Tamb ). These quantities are the same as for the variable Tin . Thus, the overall
uncertainty for the error εTamb = ε1Tamb + ε2Tamb is uTamb = 0.132 29 K. There are two sensitivity coefficients
corresponding to Y in (2) and X in (4). Evaluating these at nominal values, including the nominal value 294 K
for Tamb , are
∂Y KPin
sTamb −Y = =√ = 0.055 479 l K−1 min−1
∂Tamb Tin Pamb

∂X 1 P
= 0.069 404  H2 O atm−1/2 K−1/2
1/2
sTamb −X = = √ √
∂Tamb 2 Pamb Tamb

Change in pressure variable


The error in the variable P consists of two independent sources. The first source of error (ε1P ) arises from the
manufacturer’s stated uncertainty of 0.02% for the liquid manometer. The manufacturer’s claim is interpreted

Copyright 
c 2005 John Wiley & Sons, Ltd. Qual. Reliab. Engng. Int. 2005; 21:605–619
MEASUREMENT UNCERTAINTY IN THE CALIBRATION OF LOW-FLOW AMBIENT AIR SAMPLERS 611

as the maximum bounds for all values. The probability distribution is conservatively
√ assumed to be uniform,
so the uncertainty for a nominal value of 4.9  H2 O is u1P = (0.0002/ 12) × 4.9 = 2.8290 × 10−4  H2 O.
The second error source (ε2P ) is due to the reading of the instrument. Regular observations indicate this
uncertainty is u2P = 0.0079  H2 O. Assuming these errors are independent and additive (εP = ε1P +
ε2P ), the combined uncertainty for the change in pressure is

uP = (var(ε1P ) + var(ε2P ))1/2 = 0.007 9050  H2 O

Using (4), the sensitivity coefficient evaluated at the nominal values is



∂X 1 Tamb
sP = = √ √ = 4.1643 K1/2 atm−1/2  H2 O−1/2
∂P 2 Pamb P

4. UNCERTAINTY IN THE FUNCTIONS FROM THE CALIBRATION


The quantities Y and X in (2) and (4) are nonlinear functions involving the variables K, Pin , Pamb , Tin , Tamb , and
P . The uncertainties in the measurements of these two functions can be approximated using a first-order Taylor
series expansion to linearize these quantities in terms of the variables of interest. This allows the uncertainties to
be combined as well as the correlations to be assessed (ANSI1 , pp. 19, 46; Casella and Berger7, Section 5.5.4).
For ease of notation, denote the true value of the function in (2) by Y , the observed value of this function by
y, and the error by εY = y − Y . Similarly, denote the true value of the function in (4) by X, the observed value
of this function by x, and the error by εX = x − X. The first-order Taylor series expansions of y and x about Y
and X, based on (2) and (4), are given by

y ≈ Y + sK εK + sPin εPin + sPamb −Y εPamb + sTin εTin + sTamb −Y εTamb


(6)
x ≈ X + sP εP + sPamb −X εPamb + sTamb −X εTamb
√ √
The uncertainties in the measurement of Y and X are uY = var(εY ) and uX = var(εX ). The equations
in (6) allow for an approximation of these uncertainties based on the uncertainties and sensitivities of K, Pin ,
Pamb , Tin , Tamb , and P listed in Table I. However, the description of errors for these variables in Section 3
reveals that the variables Pin and Pamb share error sources due to manufacturer specifications and instrument
resolution. This is also true of the variables Tin and Tamb . Thus, the errors for these variables are correlated,
and such correlations should be accounted for when computing uY . These correlations will decrease the overall
uncertainty since sTin × sTamb −Y < 0 and sPin × sPamb −Y < 0. This is true since these variables occur as ratios in
the function Y . Covariance and correlation are discussed for correlated input quantities in ANSI1 (Section 5.2).
Properties of variance and covariance can be found in Casella and Begrer7 (pp. 171, 604) and ANSI1 (p. 52).
Using these properties and the fact that all variables are uncorrelated except Pin and Pamb as well as Tin and
Tamb , the approximate squared standard uncertainty for the measurements of the function Y in (2) is

u2Y = var(εY ) ≈ sK
2
var(εK ) + sP2 in var(εPin ) + sP2 amb −Y var(εPamb ) + sT2in var(εTin ) + sT2amb −Y var(εTamb )
+ 2(sPin )(sPamb −Y ) cov(εPin , εPamb ) + 2(sTin )(sTamb −Y ) cov(εTin , εTamb ) (7)

From Section A.1 (see Appendix A), cov(εTin , εTamb ) = u2Tamb and cov(εPin , εPamb ) is given by u1Pin u1Pamb +
u2Pin u2Pamb . The approximate squared standard uncertainty for the measurements of the function X in (4) is

u2X = var(εX ) = sP


2
var(εP ) + sP2 amb −X var(εPamb ) + sT2amb −X var(εTamb ) (8)

In addition, it is necessary to assess the covariance between εY and εX since these quantities share errors
sources due to Pamb and Tamb and since the error sources of Pamb and Tamb are correlated with those for Pin

Copyright 
c 2005 John Wiley & Sons, Ltd. Qual. Reliab. Engng. Int. 2005; 21:605–619
612 S. S. WULFF AND M. A. WEITZ

and Tin . Using (6) once again gives

uXY = cov(εY , εX ) ≈ (sPamb −Y )(sPamb −X ) var(εPamb ) + (sTamb −X )(sTamb −X ) var(εTamb )


+ (sPin )(sPamb −X ) cov(εPin , εPamb ) + (sTin )(sTamb −X ) cov(εTin , εTamb ) (9)

The covariance terms are as in (7) and are derived in Section A.1.
The sensitivity coefficients and uncertainties for the individual variables given in Table I can be plugged
into (7), (8) and (9) to obtain uY = 0.040 954 l min−1 , uX = 0.034 337  H2 O1/2 K1/2 atm−1/2 , and ρXY =
0.022 248 where ρXY = uXY /uX uY is the correlation between the errors εY and εX . Table I also lists the
variance contribution of each variable defined as the product of the squared uncertainty and the square of the
sensitivity coefficient. These values indicate K contributes the most to the uncertainty in the measurements of
the function Y while P contributes the most to the uncertainty in the measurements of the function X. Without
accounting for the correlations in (7), the standard uncertainty for Y would be 0.041 822 l min−1 or an increase
of roughly 2%.

5. UNCERTAINTY IN THE LINEAR FIT

The previous sections have used standard techniques from ANSI1 to obtain the uncertainty in the functions X
and Y . The challenge presented by this uncertainty analysis is to estimate the linear relationship between these
two functions described by (5), along with the uncertainty associated with the estimates. This challenge will be
addressed using a measurement error modeling approach. This approach is nicely described by Fuller4 , Reilman
et al.5 , and Casella and Berger7 (Section 12.2). Since the use of these models is not well known in the area of
metrology, this section describes the implementation and relevant considerations for this case study.
Recall from the calibration procedure, seven levels of P were specifically chosen to calibrate each SFTS
through a suitable range. Also, the ambient pressure and ambient temperature are constant across the seven runs.
Thus, the quantity X is a function of variables whose true values are considered fixed. Had the levels of P
not been controlled, or if the ambient conditions were not constant during the calibration, it would be necessary
to treat X as a random quantity. This aspect of the calibration procedure affects the estimation and uncertainty
associated with the unknown regression coefficients.
The first step is to define an appropriate measurement error model. Let i index each level associated with a
level of P . The linear relationship in (5) for the true values Y and X can be written as

Yi = β0 + β1 Xi for i = 1, 2, . . . , 7 (10)

Equation (10) can be re-expressed in terms of the observable values and errors so that

yi = β0 + β1 xi + εY,i − β1 εX,i (11)

where yi = Yi + εY,i and xi = Xi + εX,i . It is assumed the errors between levels or runs (i) are independent.
In addition, for each i, the joint distribution of εY,i and εX,i is assumed to be normal with expectation zero and
variance–covariance matrix given by
   2 
ε uY uXY
Cov Y,i = (12)
εX,i uXY u2X

This variance–covariance matrix allows for possible correlation between the two error sources. Since the levels
of the independent variable are fixed across runs, the measurement error model specified by (10), (11) and (12)
is called a linear functional model. On the other hand, when the values of X are assumed random according to
some distribution, the model is called a linear structural model. In addition, these models usually assume u2Y , u2X
and uXY are unknown quantities to be estimated from data. However, in this uncertainty analysis, these values
are known from the calculations in Section 4. It is through the variance–covariance matrix that the magnitude
of the uncertainties in X and Y are manifest.

Copyright 
c 2005 John Wiley & Sons, Ltd. Qual. Reliab. Engng. Int. 2005; 21:605–619
MEASUREMENT UNCERTAINTY IN THE CALIBRATION OF LOW-FLOW AMBIENT AIR SAMPLERS 613

The second step is to estimate the regression coefficients (11) given the above linear functional model. Let the
regression error in (11) be denoted εi∗ = εY,i − β1 εX,i . Using formulas for the variance of a sum (Casella and
Berger7, p. 171), the variance of ε∗ is

var(εi∗ ) = u2Y + β12 u2X − 2β1 uXY (13)

This equation shows that, when there is non-negligible uncertainty in the measurements of the independent
variable (X), the variance of the regression error depends on β1 . Hence, modeling techniques must accommodate
the fact that the variance is a function of the mean of yi . In addition, the error ε∗ is correlated with the observed
values xi as both terms contain the error term εX,i (Reilman et al.5 ). Fuller4 (Section 1.3.3) described a method
that uses generalized least squares to simultaneously minimize the errors εY,i and εX,i . This procedure can be
conveniently carried out in the NLMIXED procedure in SAS8 . This estimate of β1 is chosen to minimize the
quantity

(sY Y + β12 sXX − 2β1 sXY )(u2Y + β12 u2X − 2β1 uXY )−1 (14)

where

7
7
7
sY Y = (yi − ȳ)2 , sXX = (xi − x̄)2 and sXY = (xi − x̄)(yi − ȳ)
i=1 i=1 i=1

When the denominator in (14) does not depend upon β1 , the quantity in (14) is minimized by the ordinary least
squares estimator. Sufficient conditions are for ρXY to be about zero and for |β1 |uX to be sufficiently less than
uY (Mandel9). Even if these conditions are not met, ordinary least squares procedures can be used to estimate
β1 , but this estimate may be severely biased. The amount of bias can be found from equations in Reilman
et al.5 . The ordinary least squares estimate is β̂1 = sXY /sXX , which minimizes (14) when the denominator does
not depend upon β1 . In either case, the estimate for β0 is given by β̂0 = ȳ − β̂1 x̄.
Returning to this case study, the standard uncertainties from Section 4 indicate ρXY is close to zero and
uX /uY = 0.84. Thus, an ordinary least squares approach is reasonable when the estimate of the slope is
sufficiently smaller than one. Recall, the calibration procedure is conducted on an individual air sampler.
The data for the particular SFTS considered in this paper is given in Section A.2. For this data set, the ordinary
least squares estimates are given by β̂0 = −0.787 98 and β̂1 = 0.421 91. Performing Fuller’s approach in SAS
gives nearly identical estimates in Section A.5. Using the equations given by Reilman et al. 5 , which are also
listed in Section A.4, the estimated bias can be found by plugging in the ordinary least squares estimates for β0
and β1 along with the uncertainties from Section 4. This gives

estimated bias for β̂0 = Ê[β̂0 ] − β̂0 = 4.7813 × 10−4


(15)
estimated bias for β̂1 = Ê[β̂1 ] − β̂1 = −1.1671 × 10−5

These values are quite small relative to the magnitude of the estimates. The agreement of the two estimation
methods, the small bias, and the conditions from Mandel9 all suggest the ordinary linear regression approach is
appropriate for this particular example.
The third step is to estimate the uncertainty of the estimates of β0 and β1 in the linear functional model.
There are three ways in which this can be done. First of all, if the estimation approach of Fuller is used4 ,
then the estimated uncertainties consist of the square root of the estimated variance of β̂1 given in Fuller4
(Equation (1.3.31)) and the square root of the estimated variance of β̂0 given in Fuller4 (Equation (1.3.12)).
The estimated covariance between β̂0 and β̂1 is given in Fuller4 (Equation (1.3.12)). These quantities can also
be calculated using SAS as demonstrated in Section A.5.
Secondly, if the ordinary least squares approach is used for estimation, then the estimated uncertainty in
the regression estimates can be found using the bias and variance formulas for the linear functional model
given by Reilman et al.5 . These formulas are also listed in Section A.4 for convenience. Since the least
squares estimators have the potential to be severely biased, the uncertainty in the estimates should include

Copyright 
c 2005 John Wiley & Sons, Ltd. Qual. Reliab. Engng. Int. 2005; 21:605–619
614 S. S. WULFF AND M. A. WEITZ

both the bias and the variability. One approach to do this is through the mean squared error (Casella and
Berger7, Section 7.3.1). The mean square error for the vector β̂ = [β̂0 β̂1 ] is given by the 2 × 2 matrix
MSE(β̂) = Cov(β̂) + (E(β̂) − β)(E(β̂) − β) . The estimated covariance between β̂0 and β̂1 is approximated
by −x̄ v
ar(β̂1 ), which is the same as that used in ordinary least squares estimation and by Fuller4 . Because the
estimates of bias in (15) are so small, the estimate of MSE(β̂) is quite close to the estimate of Cov(β̂). For this
particular SFTS, the estimated covariance matrix is given by
 
0.011 435 −2.7264 × 10−4
Cov(β̂) = (16)
−2.7264 × 10−4 6.6550 × 10−6

The estimated standard uncertainties for β̂0 and β̂1 are ûβ̂0 = [Cov( β̂0 )]1/2 = 0.106 93,

ûβ̂1 = [Cov(β̂1 )] = 0.002 5797, and ûβ̂0 β̂1 = −0.000 272 64. However, if the bias was not small, then
1/2

it would be necessary to take the uncertainties from the estimate of MSE(β̂). Because least squares estimation
is appropriate, these uncertainties do not differ much at all from those obtained by the first method (Section A.5).
When the bias is small, a third method for finding the uncertainty is to use (13) as the error variance from the
regression in the ordinary least square formulas for the variances and covariance of β̂0 and β̂1 . These formulas
are given in Section A.3. Plugging in the appropriate values gives a variance–covariance matrix that is nearly
identical to that in (16). This is due, once again, to the fact that the estimated bias in (15) is so small.
It is important to realize that even though it is appropriate to use ordinary least squares procedures in this
example for purposes of estimating β̂0 and β̂1 , it is not appropriate to use these procedures to estimate the
uncertainty. This is evident since the usual estimate of the residual variance from the regression is 1.3845 × 10−4
compared with 1.8607 × 10−3 obtained from (13) by plugging in the given uncertainties and the estimate β̂1
(Section A.3). This is the case because the uncertainty in the measurement of X is not small compared with the
uncertainty in the measurement of Y . On the other hand, if the residual variance from the regression was larger
than the estimate of u2ε∗ with the given uncertainties, then u2Y should be treated as unknown and estimated from
the data in the usual manner. This is due to the fact that the variability in the linear fit would be sufficiently
larger than the uncertainty in the measurements of X and Y .
The estimated uncertainty of the regression coefficients is for a specific SFTS with the data given in
Section A.2 under laboratory conditions where ambient conditions are constant and the drop in pressure is
controlled. The approaches for assessing the uncertainty described in this section can easily be applied to the
calibration procedure of any SFTS.

6. UNCERTAINTY IN THE PREDICTION

The calibration procedure described in Section 2 is based on (5), which describes the linear relationship between
the measurements of the SFTS and the expected flow rate. Using the estimates for β0 and β1 from Section 5 and
given an observed value for X (x∗ ), predicted flow rates (ŷ∗ ) can be calculated using
ŷ∗ = β̂0 + β̂1 x∗ (17)
It is through (17) that the measurements of the SFTS produce flow rate values. Thus, the final objective of this
uncertainty analysis is to ascertain the uncertainty in the predicted value ŷ∗ .
First, consider the simplified case where the predicted value ŷ∗ is based on a true known value X∗ . Here the
uncertainty depends upon a given value of the independent variable (X) without regards to the measurement
error. This problem is similar to finding the standard deviation of prediction in ordinary linear regression.
Given X∗ , the uncertainty for ŷ∗ can be expressed as
u2 = var(ŷ∗ |X∗ ) = var(β̂0 ) + X∗2 var(β̂1 ) + 2X∗ cov(β̂0 , β̂1 ) (18)
Ŷ∗

Setting X∗ equal to the nominal value of X or 40.810  H2 O1/2 K1/2 atm−1/2 yields ŷ∗ = 16.430 l min−1 . Using
the estimated covariance matrix in (16), the estimated uncertainty of ŷ∗ based on (18) is ū2 = 0.016 309 l min−1
Ŷ∗

Copyright 
c 2005 John Wiley & Sons, Ltd. Qual. Reliab. Engng. Int. 2005; 21:605–619
MEASUREMENT UNCERTAINTY IN THE CALIBRATION OF LOW-FLOW AMBIENT AIR SAMPLERS 615

or a relative uncertainty of about 0.1%. This estimated uncertainty value can be calculated in SAS Section (A.5).
The nominal value of X and the value of the mean of the calibrated data are close (x̄ = 40.968). This minimizes
the uncertainty of the SFTS at the values of its intended use. In addition, accounting for the covariance
term in (18) significantly reduced the uncertainty, as, without the covariance term, the uncertainty would be
ûŶ∗ = 0.150 06. The uncertainty expressed in (18) is a function of the given value X∗ . If the SFTS is to be used
at an extreme change in pressure, then the estimate of uncertainty will be higher. For example, repeating the
calculations for X∗ = 32  H2 O1/2 K1/2 atm−1/2 , which corresponds to a change of pressure of about 3  H2 O
given the same nominal values for Tamb and Pamb , yields ŷ∗ = 12.713 with ū2 = 0.028 302.
Ŷ∗
The more practical case is that (17) is implemented based on observed measurements of the SFTS (x∗ ) which
are subject to error. In (18), the uncertainty in the independent variable is not accounted for. The uncertainty
in x∗ can be incorporated using a first-order Taylor series expansion of (17) about β0 , β1 , and X∗ where X∗
denotes the true value of x∗ . Following similar steps as in Section 4,

u2 = var(ŷ∗ |x∗ ) ≈ var(β̂0 ) + X∗2 var(β̂1 ) + β12 var(εX )


Ŷ∗
+ 2X∗ cov(β̂0 , β̂1 ) + 2β1 cov(β̂0 , εX ) + 2X∗ β1 cov(β̂1 , εX ) (19)

A conservative approach for estimating (19) is to assume a correlation of one between β̂0 and εX as well as
between β̂1 and εX 2 . The uncertainty in (19) can be then be estimated for any observed value using

û2Ŷ = û2β̂ + x∗2 û2β̂ + β̂12 u2X + 2x∗ ûβ̂0 β̂1 + 2β̂1 ûβ̂0 uX + 2x∗ β̂1 ûβ̂1 uX (20)
∗ 0 1

For example, suppose the value x∗ = 40.810  H2 O1/2 K1/2 atm−1/2 is observed based on measurements
by the SFTS. Such a value is close to the expected operating conditions. Plugging in the uncertainty in the
measurements of X from Section 4, the estimated uncertainties from (16), and the least squares estimate of
β1 , the estimated standard uncertainty using (20) is ûŶ∗ = 0.081 39 l min−1 or a relative uncertainty of about
0.5%. Thus, incorporating the uncertainty in the measurements of X has increased the relative uncertainty in the
prediction by the amount of 0.4%.

7. CONCLUDING REMARKS
This paper provides an example where the measurements of an instrument are obtained through an estimated
linear relationship between two functions subject to error. In order to assess the uncertainty in the measurements
of this instrument, it is necessary to incorporate these errors when estimating the linear relationship and for
obtaining the uncertainty of the estimates. As demonstrated in Section 3, the first step is to carefully consider
the individual error sources in each variable making up the functions of interest. In Section 4, these uncertainties
are combined using Taylor series approximations to calculate the uncertainty of the functions of interest. If the
uncertainty of the independent or regressor variable is non-negligible relative to the uncertainty in the response
variable, then it is necessary to account for the given uncertainties in the estimation and uncertainty of the
regression coefficients. A convenient way to address such problems is through measurement error models.
In Section 5, a linear functional model is considered where the true value of the independent variable is fixed
and the uncertainties are given. Ordinary least squares procedures can be used for estimating the regression
parameters, but these estimators can be severely biased. If the bias is large, then it is better to use the generalized
least squares approach described in Section 5 for obtaining the estimates. Even if the bias is small, it may not be
appropriate to use ordinary least squares procedures to estimate the uncertainty of the regression coefficients.
While it is not necessary to assume normality of the measurement errors presented in (11), this assumption
is used for calculating the uncertainties based upon Fuller4 and Reilman et al.5 . The approach in Fuller4 can
be implemented in the NLMIXED procedure in SAS8 (Section A.5) while the equations for the approach in
Reilman et al.5 are available in Section A.4. If the uncertainties in the regressor variables are unknown or
difficult to ascertain, these quantities can be estimated from the data4 . Section 6 details how uncertainty can be

Copyright 
c 2005 John Wiley & Sons, Ltd. Qual. Reliab. Engng. Int. 2005; 21:605–619
616 S. S. WULFF AND M. A. WEITZ

assessed for predicted values with and without incorporating the uncertainty in the independent variable. Fuller
also provides details when the prediction of the true value of the regressor variable is also of interest4 .

Acknowledgements
The authors would like to thank the two anonymous referees and the editor for their helpful suggestions with
regards to this manuscript.

REFERENCES
1. ANSI/National Conference of Standards Laboratories. U.S. Guide to the Expression of Uncertainty in Measurement,
ANSI/NCSL Z540-2-1997 1997. (Note that this is the U.S. version of International Organization for Standardizations
(ISO). Guide to the Expression of Uncertainty in Measurement (GUM). ISO: Geneva, Switzerland, 1995.)
2. Atwood CL, Engelhardt M. Techniques for uncertainty analysis of complex measurement processes. Journal of Quality
Technology 1996; 28:1–11.
3. Forbes AB, Harris PM, Smith IM. Generalised Gauss–Markov regression. Algorithms for Approximation IV:
Proceedings of the 2001 International Symposium, Levesley J, Anderson IJ, Mason JC (eds.). The University of
Huddersfield: Huddersfield, 2002; 270–277. Available at: http://www.npl.co.uk/ssfm/theme1/project1 2/index.html
[6 January 2002].
4. Fuller WA. Measurement Error Models. Wiley: New York, 1987.
5. Reilman MA, Gunst RF, Lakshminarayanan MY. Stochastic regression with errors in both variables. Journal of Quality
Technology 1986; 18:162–169.
6. American Society of Mechanical Engineers (ASME)/ANSI Standard. Measurement of Gas Flow by Means of Critical
Flow Venturi Nozzles, MFC-7M-1987, 1992.
7. Casella G, Berger RL. Statistical Inference (2nd edn). Duxbury: Pacific Grove, CA, 2002.
8. SAS Institute Inc. SAS/STAT User’s Guide, Version 8. SAS Institute: Cary, NC, 2000.
9. Mandel J. Fitting straight lines when both variables are subject to error. Journal of Quality Technology 1996; 16:1–14.
10. Myers RH. Classical and Modern Regression with Applications (2nd edn). Duxbury: Belmont, CA, 1990.

APPENDIX A

A.1.
For the temperature variable described in Section 3, let ε1T and ε2T denote the error sources for manufacturer
specifications and instrument resolution, respectively, where var(ε1T ) = u21T and var(ε2T ) = u22T . Assuming ε1T
and ε2T are uncorrelated, the covariance is
uT in,Tamb = cov(εTin , εTamb ) = cov(ε1T + ε2T , ε1T + ε2T ) = var(ε1T ) + var(ε2T )
= u21T + u22T = u2Tamb .
The variables Pin and Pamb also share error sources due to manufacturer specifications and instrument resolution
as explained in Section 3. The units for the error terms are different where the units of Pin are mmHg and the
units for Pamb are atm where 1 atm = 760 mmHg. Let ε1P denote the manufacturer’s error associated with
the uniform distribution and ε2P denote instrument resolution error in mmHg. Assuming these error sources are
uncorrelated, the covariance can be expressed as
 
Pamb 1
uPin ,Pamb = cov(εPin , εPamb ) = cov (Pin )ε1P + ε2P , ε1P + ε2P
760 760
Pin Pamb 1
= var(ε1P ) + var(ε2p ) = u1Pin u1Pamb + u2Pin u2Pamb
760 760
by converting back to atmospheres for Pamb .

Copyright 
c 2005 John Wiley & Sons, Ltd. Qual. Reliab. Engng. Int. 2005; 21:605–619
MEASUREMENT UNCERTAINTY IN THE CALIBRATION OF LOW-FLOW AMBIENT AIR SAMPLERS 617

A.2.
The following data are for the specific SFTS with serial number 980 822:

X Y
31.5462 12.5130
34.6195 13.8200
37.7814 15.1560
40.9623 16.4910
44.1122 17.8400
47.2375 19.1490
50.5161 20.5090

where x̄ = 40.968, ȳ = 16.497, sY Y = 49.770, sXX = 279.59 and sXY = 117.96.

A.3.
Using results from Myers10 (p. 15) or Casella and Berger7 (p. 553), the variance–covariance matrix of the
ordinary least squares estimator of β̂ = [β̂0 β̂1 ] is
 
1 x̄ 2 x̄
 + − 
 n sXX sXX 
Cov(β̂) = σ 2  
 x̄ 1 

sXX sXX

The usual estimator of σ 2 is given by

1 n
σ̂usual
2
= (yi − (β̂0 + β̂1 xi ))2
n − 2 i=1

which in this example is σ̂usual


2 = 1.3845 × 10−4 . An adjusted estimator of σ 2 is given by (13) where β̂1 and the
2 = 1.8607 × 10−3 . Using σ̂ 2 produces an estimated covariance
given uncertainties are plugged in to obtain σ̂adj adj
matrix of
 
0.011 436 −2.7265 × 10−4
Cov(β̂) =
−2.7265 × 10−4 6.6551 × 10−6

A.4.
For convenience, approximate formulas for the expected values and variances of the ordinary least squares
estimators are listed below for the linear functional model where the correlation between the errors is allowed
to be non-zero. These are taken from (B.2), (B.4), (B.6) and (B.7) of Reilman et al.5 . Using the notation in this
paper gives
  −1
uXY nu2X uXY
E[β̂1 ] = β1 − 2 1+ + 2 (B.2)
uX sXY uX
 
uXY nu2X
E[β̂0 ] = β0 + β1 − 2 x̄ (B.4)
uX sXX + nu2X
 −2  2   
u2X nu2X uXY u2Y u2XY nu2X
var(β̂1 ) = 1+  β1 − 2 + − 4 1+
sXX sXX uX u2X uX sXX

Copyright 
c 2005 John Wiley & Sons, Ltd. Qual. Reliab. Engng. Int. 2005; 21:605–619
618 S. S. WULFF AND M. A. WEITZ

 2 
−2 
nu2 nu2 uXY 
−2 X 1+ X β1 − 2 (B.6)
sXX sXX uX
 
−2  2    
nu2X nu 2 2 u2 u
1 2 u
 β1 − XY + 1 + X Y
2
 + x̄ 2 var(β̂1 )
var(β̂0 ) = uX 1 + − XY (B.7)
n sXX u2X sXX u2X u4X

A.5.
The following program commands are given to illustrate how the NLMIXED procedure in SAS8 can be used
to obtain the estimates and the estimated covariance matrix of the regression coefficients using the approach in
Fuller4 (Section 1.3). Comments are included after an asterisk ‘*’. Relevant output for the data in Section A.2
are also provided.
* Commands for PROC NLMIXED;
proc nlmixed cov ;
parms b0=-1 b1=1 ; * Need good starting values;
vy = 0.001 677 197 328 72; * Known uncertainty values;
vx = 0.001 179 035 374 13;
rho = 0.022 247 883 016 81; c = rho*sqrt(vx*vy);
muy = (b0 + b1*x)/sqrt(ve);
ys = y/sqrt(ve);
model ys ∼ normal(muy,1); * Scaled model to match Fuller4 (Section 1.3);
Xnom = 40.809 723 424 835 42; * Nominal value of X;
predict b0+b1*Xnom out=pred; * Predicted value and associated standard error;
run;
proc print data=pred; * Print out of the prediction output;
run;
* Selected Output from PROC NLMIXED;

Parameter Estimates
Parameter Estimate Error DF tValue Pr > |t| Alpha Lower Upper Gradient
b0 −0.7880 0.1069 7 −7.37 0.0002 0.05 −1.0409 −0.5351 1.276 × 10−9
b1 0.4219 0.002 580 7 163.55 <0.0001 0.05 0.4158 0.4280 6.528 × 10−8

Covariance Matrix of Parameter Estimates


Row Parameter b0 b1
1 b0 0.01144 −0.00027
2 b1 −0.00027 6.655 × 10−6

Obs y x Pred StdErr DF tValue Probt Alpha Lower Upper


1 12.513 31.5462 16.4301 0.016309 7 1007.44 0 0.05 16.3916 16.4687

Under the output titled, ‘Parameter Estimates’, the estimates of β0 and β1 are given. The output titled,
‘Covariance Matrix of Parameter Estimates’, produces the estimated variances and covariances of β0 and β1 .
The last part of the output is the first observation of the data set ‘pred’ generated by the PREDICT statement.
It gives the predicted value in (17) for the nominal value X∗ = 40.810 using the above estimates. The ‘StdErr’
corresponds to the standard uncertainty or the square root of (18).

Copyright 
c 2005 John Wiley & Sons, Ltd. Qual. Reliab. Engng. Int. 2005; 21:605–619
MEASUREMENT UNCERTAINTY IN THE CALIBRATION OF LOW-FLOW AMBIENT AIR SAMPLERS 619

Authors’ biographies
Shaun S. Wulff is an Assistant Professor of Statistics at the University of Wyoming. He received his doctoral
degree from Oregon State University and his current research interests include linear models, mathematical
statistics, and applications to the physical sciences.
Mark A. Weitz is working on a PhD in Atmospheric Science at the University of Wyoming, specializing in
aerosols and air pollution. Mark was the principal design engineer for the Streamline FTS, manufactured by
Chinook Engineering of Sheridan, Wyoming. A registered engineer, he holds a MS in Atmospheric Science
from the University of Wyoming and a BS in Chemical Engineering from Montana State University.

Copyright 
c 2005 John Wiley & Sons, Ltd. Qual. Reliab. Engng. Int. 2005; 21:605–619

You might also like