Original Article

An improved procedure for testing for assay linearity

HA Ross1,2 and CGJ Sweep 1

Abstract
Addresses Background Statistical testing of assay linearity using replicate measurements
1
Department of Chemical Endocrinology made on binary mixtures of samples provides no information about the significance
2
Department of Endocrinology
of any detected deviations from linearity, which may often be desirable.
University Medical Centre Nijmegen
PO Box 9101 Method Performance of additional statistical analysis of the assay data.
6500 HB Nijmegen
Result Objective measures for non-linearity and the minimum detectable deviation
The Netherlands
from linearity of the assay are obtained.
Correspondence Conclusion A few additional calculations from the available data make it possible
Dr HA Ross to assess whether or not an assay method needs improvement to achieve adequate
E-mail: [email protected]
performance within its intended application.
Ann Clin Biochem 2003; 40: 75–78

Introduction Another problem with this approach lies in the

interpretation of the test result. High assay precision
Many laboratories use standard protocols for evalu- permits detection of very subtle deviations from lin-
ating new assay methods, kits or analysers. The earity, so that the analyst may be penalized for good
National Committee for Clinical Laboratory Standards assay design and performance. The original procedure
(NCCLS) issued a number of evaluation protocols that does not include calculation of the magnitude of this
have been widely applied. One of these protocols, EP6,1 deviation, so that it is not possible to judge whether or
is designed to assess assay linearity, and the statistical not it is relevant within the context of the assay and its
theory behind it can be found in standard texts.2 intended application. On the other hand, when non-
The protocol is based on replicate measurements linearity is undetectable, possibly due to assay impre-
performed in binary mixtures of samples di¡ering in cision, it might wrongly be assumed that linearity has
the concentration of the analyte of interest. A series of been demonstrated. In order to avoid this, it would be
mixtures is prepared so that, under the null hypoth- useful to know what degree of non-linearity would
esis of perfect linearity, equal concentration intervals have been detectable, given the precision with which
are obtained. A classical least-squares linear regres- the test was performed.
sion analysis is performed on the data in which the In this article a few simple methods are presented
expected values are assigned to the independent vari- that aim to address these problems.
able and the actual measurements are assigned to the
dependent variable. Under the null hypothesis, the
scatter of the data points about the regression line is Methods
completely due to replicate error. To detect non-
linearity, the variance of the distance of data points to Linearity tests were performed according to the
the line (residual variance) is tested against the repli- principle of the NCCLS Evaluation Protocol 6 (EP6).1
cate variance using F statistics. Binary mixtures in proportions 0 +1, 1+3, 1+1, 3+1
It is common practice to calculate the values of the and 1+0 were analysed in quadruplicate, so that a
independent variable from actual estimates of the two total of 20 equally spaced data points, at ¢ve di¡erent
unmixed samples and the proportions in which these levels, were obtained. The data in the sample calcula-
are mixed. Thus, the independent variable depends on tion originated from a serum total thyroxine (T4)
the uncertainty in the two measurements of the pure assay performed on an automated analyser.
samples, although classical linear regression analysis The dependent variable y consisted of the actual
presupposes an error-free independent variable. measurements performed in quadruplicate on the

© 2003 The Association of Clinical Biochemists 75

76 Ross and Sweep

mixtures. The data points of the independent variable Table 1. Result of replicate measurements (y1–y4) of serum
x were obtained in one of two ways: total thyroxine on Žve mixtures of samples a and b
(a) The usual way in which the in-between expected b a+3b a+b 3a+b a
values were calculated from the average of
measurements of the two pure serum samples; y1 31¢7 99¢5 147 217 271
y2 31¢5 97¢6 146 208 265
(b) By using dimensionless, equally spaced numbers, y3 31¢3 95¢6 144 203 265
in this case 0, 0¢25, 0¢5, 0 ¢75 and 1. y4 30¢6 93¢7 144 198 264
The rationale of the latter approach is the following : Mean y 31¢28 96¢60 145¢2 206¢5 266¢2
Var y 0¢2292 6¢273 2¢25 65¢67 10¢25
x1 31¢28 90¢02 148¢8 207¢5 266¢2
The expected value y’ for the concentration of a x2 0 0¢25 0¢5 0¢75 1
binary mixture of samples with concentrations a and
b, when mixed in the proportions x and 17x (05x51) Var y ˆ replicate variances calculated for each mixture; x1 ˆ values for
is: the independent variable calculated from estimates of a and b;
x2 ˆ dimensionless values of the independent variable.
y0 ˆ x:a ‡ (1 ¡ x):b ˆ x:(a ¡ b) ‡ b

This means that, independent of the actual values of calculated, given the actual replicate variance and a
a and b, minimization of squares of (y7y’) leads to a critical value for F:
regression line with a slope that is an estimate of the Non-linearitycrit ˆ
concentration di¡erence (a7b) and an intercept that
estimates the concentration b. H(replicate variance £ (F…3,15)crit ¡ 1)=6) (3)
It should be noted that the values for a and b are also
A brief explanation of the statistics is given in
obtained by substituting the measured averages of
Appendix 1.
the pure samples in the regression equation that uses
the conventional independent variable.
Calculations of slope, intercept and residual error Results
were performed by standard algorithms contained in The results of the quadruplicate measurements of
Microsoft Excel.3 The replicate variance was calcu- total T4 in samples a and b and the ¢xed mixtures (in
lated by averaging the ¢ve variances obtained for each nmol/ L) are given in Table 1.
mixture. For the EP6 experimental design, with 18 Performing linear regression analysis on these data
degrees of freedom for the residual variance and 15 for using values calculated from the estimates in pure a
replicate variance, the formula in Selvin (1995) 2 for and b as the independent variable x1 resulted in a slope
the F-value for linearity testing must be rewritten as: of 0 ¢987 and an intercept of 23¢3 nmol/ L.
F ˆ …6 £ residual variance With the dimensionle ss variable x2, a slope of 231 ¢9
and an intercept of 33¢2 nmol/ L was obtained (see
¡5 £ replicate variance†=replicate variance (1)
Fig. 1).
The procedure was extended by estimating the The residual error of 5 ¢ 023 nmol/ L (Excel algo-
contribution of non-linearity to residual variance. rithm) was squared to give the residual variance of
Therefore, the notion of `non-linearity’ was intro- 25¢23 and replicate variance was 16 ¢93 (var y
duced, which is de¢ned as the square root of the averaged) implying a replicate error of 4¢115 nmol/ L.
di¡erence between residual and replicate variances. Using (1), an F value of 3¢94 was found, which is
Thus, it is the root of the residual variance that would signi¢cant with a P value of 0¢029, implying that non-
have been obtained had there been no replicate error. linearity is proven. In order to ascertain whether this
Non-linearity is conveniently expressed in concentra- is relevant in the context of this assay, additional
tion units and calculated as the square root of the calculations according to (2) and (3) were made.
di¡erence between residual and replicate variances: Non-linearity was estimated from (2), yielding
2 ¢88 nmol/ L. This is 1 ¢2% of the total concentration
Non-linearity ˆ range (231¢9 nmol/ L). Since the T4 range for
H…residual variance ¡ replicate variance† (2) euthyroid subjects is about 100 nmol/ L wide, this
deviation from linearity should be considered as irre-
Alternatively, it may be expressed in a relative levant for the usual applications.
manner by dividing by the concentration range to The critical value of F(3, 15) at P50¢05 is 3¢29.With
which it applies (ja ^ bj). Whether or not the null the current replicate variance, the critical value for
hypothesis is rejected, an estimate of the magnitude non-linearity is 2 ¢54 or 1¢1%. This means that the
of the minimally detectable non-linearity can be precision with which the linearity experiment was

Ann Clin Biochem 2003; 40: 75–78

 Testing for assay linearity 77

However, choosing a dimensionless independent

variable has no e¡ect on linearity testing. This is
because concentration intervals in both approaches
are equally spaced, resulting in equal values for resi-
dual variance.
We consider calculation of the presently de¢ned
`non-linearity’ as a valuable addition to the linearity
test. Together with the `minimum detectable non-
linearity’, it o¡ers the possibility of:

(a) Assessing whether any observed statistically sig-

ni¢cant deviation from linearity is relevant con-
sidering the purpose of the assay in the clinical
laboratory. The criteria used to determine this are
beyond the scope of this article;
(b) Assessing whether the power of the linearity test
is su¤cient to detect deviations from linearity that
are considered as relevant to the application of the
assay. A high value for the (relative) minimum de-
Figure 1. Graphical representation of evaluation of assay tectable non-linearity may mean that assay preci-
linearity using a dimensionless independent variable. T4 ˆ total sion is insu¤cient or the concentration range is
serum thyroxine. too small. It also means that failure to detect non-
linearity must not be taken as a reassurance that
performed allows detection of deviations from lin- the assay is acceptable.
earity as small as 1 ¢1%. Had the linearity test not led to
signi¢cance, this low value shows that the experiment Recently, this approach was applied by us in the
was sensitive enough to detect deviations that do evaluation of an assay method, based on high perfor-
matter. mance liquid chromatography and electrochemical
detection, for urinary metanephrines (metanephrine,
normetanephrine and 3-methoxytyramine).4 The
Discussion concentrations ranged from low normal to high
normal. In all three instances signi¢cant non-
Classical least-squares linear regression analysis is linearity was detected. Without calculation of the
based on the assumption of an error-free independent actual magnitude of the non-linearity, doubt could
variable and a dependent variable that is subject to have arisen with respect to all three parameters.
random (measurement) error. In practice, this Actual values for non-linearity of 1¢8, 7¢1 and 0¢6%,
assumption is considered to be ful¢lled when the error respectively, were found, with minimum detectable
in the independent variable is negligible or small deviations from non-linearity of 1 ¢3, 2¢9 and 0 ¢3%. In
compared with the error in the dependent variable. In the case of metanephrine and 3-methoxytyramine,
the original EP6, the error in the independent variable the deviation was considered as not relevant. However,
not only depends on accurate and precise pipetting to the results for normetanephrine illustrated the
yield the correct proportions, but also on the extreme sensitivity of the test as a consequence of very
measurement error involved in determining the high assay precision, and the fact that a detected
concentration of the pure samples. Therefore, the deviation may be statistically signi¢cant but otherwise
actual estimates for the slope (which should approx- totally irrelevant. The deviation of 7¢1% in the norme-
imate 1) and the intercept (which should approximate tanephrine led to further investigation of possible
0) will be di¤cult to interpret. causes, which could be explained. Had the test not
In contrast, a dimensionless independent variable is resulted in detection of non-linearity, it would have
subject only to pipetting error, which can be kept been considered as adequate for the purpose since a
su¤ciently low, so that the interpretation of parameter deviation of 2 ¢9% would have been detected.
estimates is unequivocal and straightforward ; esti-
mates and con¢dence intervals of the analyte
concentration in the pure samples are obtained using References
all available data, although the same estimates for the 1 Passey RB, Bee DE, Caffe A, Erickson JM. Evaluation of the
pure samples can be obtained from the usual regres- Linearity of Quantitative Analytical Methods. Villanova, PA: National
sion equation as well, since both approaches use all Committee for Clinical Laboratory Standards, 1986; 6:18 (Document
data. EP6-P)

Ann Clin Biochem 2003; 40: 75–78

78 Ross and Sweep

2 Selvin S. Practical Biostatistical Methods. Belmont CA: Duxbury LOF. Thus, LOF/3 (3 df) can be tested against the repli-
Press, 1995: 45–51 cate variance (15 df) using F-statistics.
3 Microsoft1 Excel 97 SR-2, Microsoft Corporation 1985–97
4 Willemsen JJ, Ross HA, Wolthers BG, Sweep CGJ, Kema IP. (LOF)=3
Fˆ
Evaluation of speciŽ c high-performance liquid-chromatographic replicate variance
determinations of urinary metanephrine and normetanephrine by
comparison with isotope dilution mass spectrometry. Ann Clin since LOF ˆ SSresidual ¡ SSreplicate
Biochem 2001; 38: 722–30

Accepted for publication 19 September 2002 F ˆ (SSresidual ¡ SSreplicate )=3

(1)
replicate variance

SSresidual
and residual variance ˆ
18

(and) so SSresidual ˆ 18 £ residual variance (2)

Appendix 1. Explanation of statistics
The scatter of data points about the ¢tted line is SSreplicate
and replicate variance ˆ
expressed in terms of the sum of squared deviations 15
from the line (SSresidual) which, after division by the (and) so SSreplicate ˆ 15 £ replicate variance (3)
number of degrees of freedom (df; in this case18, i.e. 20
data points minus 2 calculated parameters) yields the Substituting (2) and (3) in (1):
residual variance. This variance comprises both the
(18£ residual variance¡15 £ replicate variance†=3
variance due to replication and, if the relationship is Fˆ
not linear, the average of squared deviations of the replicate variance
expected values of the responses from the line.
At each concentration level an estimate of the repli- 6 £ residual variance ¡ 5 £ replicate variance
Fˆ
cate variance can be obtained from the sum of squared replicate variance
deviations of the data from their mean value. ˆ expression (1) in `Methods’ section, above.
Summing the variances over the ¢ve levels multiplied
by 3 (3 df) yields the replicate sum of squares A detailed Excel spreadsheet illustrating the calcu-
(SSreplicate). The replicate variance is obtained by divi- lations used may be obtained on request from the
sion by the appropriate number of df (15, i.e. 3 df at author.
each of ¢ve levels of four replicates). The de¢nition of non-linearity is derived from the
In the EP6 document, the `lack of ¢t’ (LOF) is de¢ned part of the SSresidual that cannot be ascribed to replicate
as the di¡erence between SSresidual and SSreplicate. Under error. The magnitude of the contribution of replicate
the null hypothesis of perfect linearity, residual error to SSresidual is 18/15£ SSreplicate, or 18 £ replicate
variance is completely explained by replicate variance. variance. The remainder is the SSresidual that would
Then, according to the de¢nitions given above, each of have been obtained without any replicate error. This
SSresidual, SSreplicate and LOF would be multiples of the `reduced’ residual variance is obtained by division by
replicate variance, with factors 18, 15 and 3 df respec- 18. The term `non-linearity’ is introduced in this
tively. Non-linearity adds to SSresidual and therefore to article for the square root of this reduced variance.

Ann Clin Biochem 2003; 40: 75–78