Statistical aspects of quantitative real-time PCR experiment design
Robert R. Kitchen a
, Mikael Kubista b,c
, Ales Tichopad c,d,*
a
School of Physics, University of Edinburgh, 10 Crichton Street, Edinburgh EH8 9AB, UK
b
TATAA Biocenter, Odinsgatan 28, 411 03 Göteborg, Sweden
c
Institute of Biotechnology AS CR, Lb Building, 2nd Floor, Videnska 1083, 142 20 Prague 4, Czech Republic
d
Technical University Munich, Physiology Weihenstephan, Weihenstephaner Berg 3, 85354 Freising-Weihenstephan, Germany
a r t i c l e i n f o
Article history:
Accepted 18 January 2010
Available online 28 January 2010
Keywords:
Real-time PCR
qPCR
Gene expression
Experiment design
Sampling plan
Error propagation
Statistical power
Prospective power analysis
Nested analysis of variance
powerNest
a b s t r a c t
Experiments using quantitative real-time PCR to test hypotheses are limited by technical and biological
variability; we seek to minimise sources of confounding variability through optimum use of biological
and technical replicates. The quality of an experiment design is commonly assessed by calculating its prospective
power. Such calculations rely on knowledge of the expected variances of the measurements of
each group of samples and the magnitude of the treatment effect; the estimation of which is often uninformed
and unreliable. Here we introduce a method that exploits a small pilot study to estimate the biological
and technical variances in order to improve the design of a subsequent large experiment. We
measure the variance contributions at several ‘levels’ of the experiment design and provide a means of
using this information to predict both the total variance and the prospective power of the assay. A validation
of the method is provided through a variance analysis of representative genes in several bovine tissuetypes.
We also discuss the effect of normalisation to a reference gene in terms of the measured variance
components of the gene of interest. Finally, we describe a software implementation of these methods,
powerNest, that gives the user the opportunity to input data from a pilot study and interactively modify
the design of the assay. The software automatically calculates expected variances, statistical power, and
optimal design of the larger experiment. powerNest enables the researcher to minimise the total confounding
variance and maximise prospective power for a speciﬁed maximum cost for the large study.
Ó 2010 Elsevier Inc. All rights reserved.
1. Introduction
1.1. The importance of experiment design
The typical quantitative real-time PCR (qPCR) experiment is designed
to test the hypothesis that there is no difference in the
expression of a gene between two or more subpopulations; this
is based on experiments performed on representative groups of
biological subjects that, for example, exhibit different phenotypic
traits or have been exposed to different treatments [1–5]. If this
hypothesis is unlikely to be true, the alternative hypothesis is supported
stating that there is a difference between the subpopulations.
The ability of the researcher to obtain a statistically
signiﬁcant result from the testing of these hypotheses is governed
by three factors:
(1) the ‘treatment effect’, that is the magnitude of the mean differential
expression between the chosen subpopulations;
(2) the inherent and expected ‘biological variability’ in the
expression of the gene between subjects randomly selected
from within each subpopulation;
(3) the ‘technical noise’ introduced through measurement error.
The larger the treatment effect, the easier it becomes to resolve
from the confounding noise. Biological variability is generally
unavoidable, but one can seek to minimise its impact by randomly
selecting large numbers of subjects (biological replicates) from
each subpopulation. Technical noise can be minimised through
careful lab practice, the use of technical replicates, and the addition
of appropriate controls [6].
The concept of treatment effect and measurement variability is
the basis of statistical power. The power of a statistical test is the
probability of rejecting the null hypothesis, given that the null is
false and the alternative hypothesis is true [7]. In other words,
the power is the biological resolution of the experiment; it quantiﬁes
the likelihood of being able to resolve any differential expression
between treatment groups based on the variance of available
measurements. Power increases with increasing magnitude of the
differential expression, increasing number of biological replicates,
increasing measurement precision, and decreasing biological variability.
The objective is to maximise the statistical resolution of the
1046-2023/$ - see front matter Ó 2010 Elsevier Inc. All rights reserved.
doi:10.1016/j.ymeth.2010.01.025
* Corresponding author. Address: Technical University Munich, Physiology
Weihenstephan, Weihenstephaner Berg 3, 85354 Freising-Weihenstephan,
Germany. Fax: +49 8161 71 4204.
E-mail address: ales@tichopad.de (A. Tichopad).
Methods 50 (2010) 231–236
Contents lists available at ScienceDirect
Methods
journal homepage: www.elsevier.com/locate/ymeth
assay, by minimising the confounding variance in the measured
experiment data, such that a determination of the treatment effect
can be more conﬁdently reported.
It may sometimes be the case that results collected with an assay
appear more reproducible where small numbers of biological
and technical replicates are employed. This apparent increase in
precision and power is illusory, however, and signiﬁcant results
may simply reﬂect the chance ﬂuctuations in the particular subjects
or measurement processes chosen for the experiment [8]. It
is generally considered good experimental practice to vary the conditions
of the assay, by sampling multiple subjects and analysing
multiple technical replicates, to increase the chance that the statistical
signiﬁcance of the results obtained is real and reproducible in
different settings [9].
1.2. qPCR experiment design and error propagation
Between the selection of subjects from the subpopulations and
the gathering of expression data by qPCR there are several steps of
sample-preprocessing that are necessary to prepare the genetic
material for analysis. These procedures, illustrated in Fig. 1, are
typically:
(1) the sampling of material from each subject and the extraction
of the nucleic acid;
(2) in the case of RNA analysis, the reverse-transcription (RT) of
RNA to convert it into cDNA;
(3) the ampliﬁcation of the cDNA by qPCR.
Some protocols may include additional steps such as ﬁxation of
the sample, transportation, and storage. All of these procedures are
susceptible to the introduction of error [10] and, combined, they
represent the technical noise in the obtained RT qPCR measurement.
In addition to the biological variability between subjects,
these sources of technical noise all contribute to the total variance
of the measured expressions reported by the qPCR. The minimisation
of this variance can be achieved through effective, informed
experiment designs and sampling plans that employ replicates
where they are expected to have the greatest beneﬁt. The challenge
is therefore to design experiments with the optimal number of biological
and technical replicates such that the statistical power is
maximised and sufﬁcient to test a biological hypothesis, while
maintaining an affordable and realistic sampling plan.
It is assumed that technical noise introduced into the experiment
from each subsequent stage in the sample-preprocessing procedure
is independent and, as such, the effect on the overall noise of the assay
is additive. Namely, the magnitude of error introduced due to
pipetting, uncertainties in instrument readings, and chemical noise
in the different processing steps are not considered to be interdependent.
There are, however, a few exceptions where this assumption is
invalid; for example interference due to the presence of an inhibitor
may not be independent if the same inhibitor impairs the performance
of several steps of the sample processing.
1.3. Focus of the paper
Due to the large scope for the introduction of error into qPCR
measurements and results, it is not only essential that experiments
are well conducted and validated, but also that they are carefully
designed and documented; this enables the researcher to maximise
the likelihood of accurately and reproducibly reporting interesting
biological phenomena [11]. Power analyses afford the
researcher a valuable tool with which to estimate the resolution
of the assay, in terms of its ability to test a speciﬁed hypothesis
while the experiment is still in the design phase. The importance
and utility of these prospective power analyses is universally accepted,
however, the calculation of power is often reduced to mere
guesswork due to the fact that, by deﬁnition, the magnitude of the
effect to be studied and the measurement variation in the prospective
assay cannot be precisely known [12]. After the assay has been
performed and data are available, however, these variables are
known (at least in terms of the set of samples analysed) allowing
a more accurate calculation of power. Such retrospective power
calculations have been shown to be useful, although the measured
effect size is often less informative than the variance estimated
from the data [13].
In this paper we describe a method of estimating the components
of biological variation and technical noise directly from qPCR
measurements. This is achieved through the exploitation of a small
number of biological and technical replicates at each stage of the
sample processing procedure; these stages being the inter-subject,
inter-sample, inter-RT, and inter-qPCR. These biological replicates
form a pilot study to a larger, prospective investigation and, as
such, should be drawn randomly from a larger cohort of subjects
that are to be used as the basis of the future investigation. The variance
components estimated from this small pilot study are used to
determine the optimal experiment design and sampling plan for
the subsequent prospective study. We further exploit these measured
variances by including them in the prospective power calculation
for the future study, providing a more accurate, evidencebased,
estimate of the expected experiment error.
As a validation of the method, variance components are estimated
for several genes from a number of bovine tissue-types
and contrasted with the components from the same data following
normalisation to a reference gene. We use these data to qualitatively
assess the utility of technical replication at only the qPCR level;
a common practice in qPCR assay design and one for which the
rationale is unclear, except perhaps as a low-cost insurance against
a failed PCR.
Finally, we present powerNest; a software implementation of
these methods that provides an intuitive and efﬁcient means of
optimising the sampling plan given the data from such a pilot
study.
2. Description of method
2.1. Model
We deﬁne the model for any given Cq measurement by qPCR
based on four processing steps that account for both the biological
Fig. 1. Example 2 Â 3 Â 3 Â 3 experiment design. Example experiment design for 2
subjects belonging to a single subpopulation. Three qPCRs are performed for each of
3 RTs of 3 samples from each subject. The result is 27 Cq measurements for each
subject, 54 in total for the subpopulation. From this design, variance components at
each stage of sampling can be estimated. Mouse image appears courtesy of the
Wellcome Trust Sanger Institute.
232 R.R. Kitchen et al. / Methods 50 (2010) 231–236
variability and the technical noise that inﬂuence the measured value.
These are the choice of subjects from the subpopulation, the
replicate samples extracted from each subject, the replicate RTs
of mRNA from the same sample, and the replicate qPCRs of cDNAs
from the same RT tube. These effects are all assumed to be independent
and randomly drawn from normal distributions on the
logarithmic scale [14,15].
Although the introduction of biological and technical noise at
each of the sampling levels is independent, the observed variances
are not. The variation introduced at a given level propagates additively
throughout subsequent levels, allowing these variance contributions
to be modelled. All factors therefore meet in unique
combinations and so a nested, or hierarchical, model of additive
noise is applied to the measured Cq values [16] such that any given
measured Cq can be expressed as
Cqgijkl ¼ lg þ agi þ bgij þ cgijk þ dgijkl; ð1Þ
where lg is the geometric-mean expression of the gene in the gth
subpopulation (which is equivalent to the arithmetic average of
the fold-change or expression of the gene on the log-scale), agi is
the random effect of the ith
subject in the gth
subpopulation, bgij is
the random effect of the jth
sample extracted from the ith
subject
in the gth
subpopulation, cgijk is the random effect of the kth
RT reaction
from the jth
sample extracted from the ith
subject in the gth
subpopulation,
and dgijkl is the random effect of the lth
qPCR from the kth
RT of the jth
sample extracted from the ith
subject in the gth
subpopulation.
This model was previously justiﬁed in terms of its application to
qPCR experiment design in [17]. The total variance of any given Cq
measurement follows directly from the model in Eq. (1) and is deﬁned
as
r2
Cq ¼ r2
g þ r2
i þ r2
j þ r2
k þ r2
l : ð2Þ
In simpler terms, the expected variance of each measurement
can be divided into two categories; the ﬁrst is the treatment variation
between subpopulations that is expressed by the r2
g term;
the second is confounding biological variance and processing noise
that is encompassed by the sum of the remaining variance components
corresponding to inter-subject, inter-sample, inter-RT, and
inter-qPCR variation. To maximise the statistical power of the assay
one should minimise the confounding variance to be able to
accurately resolve the treatment effect.
The variance model in Eq. (2) is used to deﬁne a nested analysis
of variance (nested-ANOVA) that produces estimates of each of the
four modelled components of variance. The calculation of these
variance components is performed as described in [18], by a procedure
essentially based on the subtraction of the sum-squared variations
of each level from that of the respective immediate higher
level.
The relative contribution of each component, vcx, to the total
variance is expressed as a percentage:
vcx ¼ 100 Â r2
x =ðr2
i þ r2
j þ r2
k þ r2
l Þ; ð3Þ
where x = i, j, k, or l.
2.2. Experiment optimisation
In terms of the optimisation of the experimental design, it is the
objective to minimise the total expected technical and biological
variation within each treatment group, g, which is deﬁned as
^r2
Cqg ¼ s2
i =ni þ s2
j =ninj þ s2
k=ninjnk þ s2
l =ninjnknl ð4Þ
where s2
i ; s2
j ; s2
k ; and s2
l are the standard deviations of the subject,
sample, RT, and qPCR levels, respectively, estimated from the pilot
data. Additionally ni is the number of subjects, nj is the number of
replicate samples from each subject, nk is the number of replicate
RTs from each sample, and nl is the number of replicate qPCRs from
each RT. By varying the n replicates at each level the ^r2
Cqg can be
changed. The optimal design is the one in which ^r2
Cqg is minimised.
The inclusion of a ﬁnancial cost into the calculation of the optimal
design is trivial; the total cost of the experiment is
CT ¼ cini þ cjninj þ ckninjnk þ clninjnknl ð5Þ
where ci, cj, ck, and cl are the costs of producing a subject, sample,
RT, and qPCR.
2.3. Statistical power
The power of a statistical test is the probability of rejecting the
null hypothesis, given that the null is false and the alternative
hypothesis is true. Power is simply a restatement of the Type II error
rate, b, of falsely accepting a null hypothesis; power = (1 À b).
The power depends on two factors; the signiﬁcance criterion
and the effect size. The signiﬁcance criterion, a, is the Type I error
rate of falsely rejecting a null hypothesis and must be speciﬁed before
the power can be calculated. The a is often referred to as the
rate of false-positives and the b as the rate of false-negatives. The
a and b symbols used in terms of the signiﬁcance criterion and
power bear no relation to the agi and bgij introduced in Eq. (1).
For the purposes of this method only two classes of power calculation
are considered; that used for the testing of the average expression
of a single subpopulation in terms of a difference from a prespeciﬁed
value, and that used for the testing of the means of two
subpopulations in terms of the difference from each other.
The effect size, d1, in the case of a comparison of the mean
expression of a single subpopulation from some pre-determined
value, c, is simply
d1 ¼ ðmA À cÞ=rA ð6Þ
where mA and rA are the mean and standard deviation of the subpopulation,
respectively, and correspond directly to lg in Eq. (1)
and ^r2
Cqg in Eq. (4).
The effect size, d2, in the case of a comparison of two subpopulations
with unequal variances is deﬁned as the difference between
the means of the subpopulations divided by the precision of the
measurement of each, thus
d2 ¼ jmA À mBj=½ðr2
A þ r2
BÞ=2ð1=2Þ
ð7Þ
where mA and r2
A are the mean and variance of one subpopulation
and mB and r2
B are the mean and variance of the second
subpopulation.
Given the number of samples in the subpopulation(s) and the
desired signiﬁcance criterion the power can either be determined
from a table, as found in [7] for example, or calculated from the
cumulative distribution function of the t-distribution. N.B. a compensation
is required when using a table to ﬁnd the power of a test
using a single subpopulation such that the effect size, d1, should be
multiplied by
p
2 to compensate for the fact that the c is a hypothetical
population parameter without any associated sampling
error.
2.4. Software implementation
Here we present a software implementation of this model. The
software has been customised for use with small pilot datasets for
which the variance structure of the experiment design is
estimated.
Cq data can be entered into the software as MS Excel spreadsheets
or plain text ﬁles. The Cq values must be allocated to the
correct position in the experiment design hierarchy so that the
R.R. Kitchen et al. / Methods 50 (2010) 231–236 233
software can determine to which subject, sample, and RT replicate
each given data point belongs. This allocation can be performed
manually in the software or be pre-speciﬁed in the input data ﬁle.
In the case of the latter, a template ﬁle is available; the use of
which enables the software to automatically parse the experiment
design. The user-interface for data input, design speciﬁcation, and
results analysis is shown in Fig. 2.
Once the data have been allocated to their position in the experimental
hierarchy, the nested-ANOVA automatically performs the
analysis to determine estimates for the variance components of each
of the four levels. These components are reported in terms of the relative,
fractional contributions to the total variance in the data as detailed
in Eq. (3). In the situation where experiment data are
unavailable, a facility to manually input error estimates is also provided
for each level. The results generated by using this facility must
be interpreted with caution, depending on the researcher’s conﬁdence
about the quality of the input variance estimations.
Once the data are allocated to their correct position within the
design of the pilot experiment the user is provided the opportunity
to modify the design and sampling plan by adding and removing
biological and technical replicates at each level of the design. Using
the measured variance structure, error estimates and statistical
power are automatically calculated and displayed for each of the
modiﬁed designs. A facility is also provided for inputting the
approximate ﬁnancial cost of performing a single replicate for each
level. If this information is available, the software will display the
total cost of each design as well as the expected total error.
Given the variance and costing information, the software is
capable of determining an experiment design that minimises the
total variance for a speciﬁed ﬁnancial cost. This is achieved through
an implementation of Eqs. (4) and (5). The user can choose from
various designs such as those optimised for cost-performance, for
the minimisation of biological and technical error, or for the maximisation
of statistical power.
2.5. Power calculation
For a single dataset from a single treatment group, the power of
different experiment designs can be estimated in terms of the
difference of the mean of the given data compared to a pre-speciﬁed
value. The population variance is estimated either from the
variance of the input data, or by manual estimate.
Data can be entered for multiple treatment groups such that the
entire experiment design can be optimised based on the observed
variances. Given this information, the software provides an automatic
power calculation such that the statistical resolution of the
assay for the desired contrast can be maximised before the experiment
is performed. The automatic optimisation of the entire
experiment design is capable of producing designs where the replicate
structure of each treatment group is unique, enabling the
overall error of the entire experiment to be minimised (i.e. different
designs for each subject group depending on the result of the
nested-ANOVA).
The power is calculated based on the measured variance structure
of the input data for the treatment group(s) using the effect
size formulae deﬁned in Eq. (6) or Eq. (7). For each design, the
power is calculated using the number of biologically distinct observations
(usually subjects, sometimes samples), the difference between
the means of the treatment groups, and the precision of
the measurements. The difference between the means can be
either speciﬁed manually (preferred) or estimated from the data.
The software can also plot a graph of the number of biologically
distinct observations vs. estimated power, examples of which are
illustrated in Fig. 3.
2.6. Experimental application
This method and the software were ﬁrst used in the analysis of
several different types of biological material by Tichopad et al. [17],
in which the relative contributions of each of the processing levels
to the total variance were estimated for bovine liver, blood, and
culture samples for a number of different genes. The sampling
plans for each of the sample types in this study were designed to
include sufﬁcient biological and technical replication to allow the
estimation of the variance components by the nested-ANOVA,
Fig. 4(A). Here we extend this analysis to estimate variance components
of the same data normalised to the reference gene, ActB, in
each of the three tissues, Fig. 4(B). We use these data to compare
Fig. 2. The powerNest software user interface. The main interface of the powerNest software. Cq data for a number of subject groups from a pilot study can be entered and
grouped to provide estimates of the variance components of the experiment. Subjects within each group are assumed to be biological replicates; qPCRs from the same RT, RTs
from the same sample, and samples from the same subject are assumed to be technical replicates.
234 R.R. Kitchen et al. / Methods 50 (2010) 231–236
the measured variance structures before (Cq) and after (DCq) normalisation
to the reference gene.
Prior to normalisation, the analysis of the liver tissue revealed
substantial variation with an average total standard deviation that,
in terms of the Cq, corresponds to a 2.6-fold variation between
measurements. In blood, the noise arising from sampling and
extraction was consistently small across all of the studied genes,
both before and after normalisation to the reference gene, indicating
that this step is very reproducible for such samples. The cell
culture samples were found to exhibit the lowest overall confounding
variation, attributable to the clonal nature of these
cultures.
In all studied genes, with the exception of the low-expressed
FGF7 in liver and IFNc in blood, the magnitude of variance attributed
to the RT step was reduced after normalisation. Excluding
FGF7 and IFNc, the estimated standard deviations at the RT step
ranged between 0.18 and 0.46 cycles with a mean of 0.31 cycles
in raw data, and were reduced to 0.03–0.25 cycles with a mean
of 0.17 cycles following normalisation. The total standard deviations
observed in blood and culture samples were only marginally
affected by normalisation, while the total standard deviations of
genes in liver (excluding FGF7) were dramatically reduced. The total
standard deviation in FGF7 more than doubled following normalisation
due to a large increase in the variance attributed to
both the sampling and RT steps; the reason for this is unknown
and with only a single observation we cannot speculate as to the
signiﬁcance of this result.
Many published reports have described the use of experimental
protocols that perform only qPCR replicates. On the basis of the
variance contributions we have estimated for the 3 studied sample
types, we are able to evaluate the importance of qPCR replicates.
Again excluding the low-expressed genes, FGF7 and IFNc, we
found the standard deviations in raw data at the qPCR level to be
0.07–0.21 cycles, with a mean of 0.13 cycles; similar to previous
ﬁndings [14]. We conclude that a qPCR standard deviation of
0.13 cycles is a good estimate for genes that are expressed at reasonable
levels and assayed with a protocol that yields at least some
25 template copies per qPCR.
3. Concluding remarks
The powerNest software application was speciﬁcally developed
to implement the method presented in this article; it calculates
the biological and technical variance components for a given dataset
and can deliver cost-optimal, variance-minimising experiment
designs. Multiple datasets can be analysed simultaneously such
that an estimate of statistical power can be calculated for a speciﬁed
contrast between them. There are currently several published
algorithms and software tools that address the analysis of gene
expression with data generated by qPCR experiments; these include,
among other things, different approaches to normalisation,
the use of reference genes, and clustering of multiple targets and
samples [19]. In addition, generalised software implementations
of the nested-ANOVA and power calculations are also available
[20,21]. powerNest, however, represents the ﬁrst dedicated tool to
assist the researcher throughout the planning phase of an experiment
and is available online at www.powernest.net.
General results for each of the pre-processing levels in the
experiments described here highlight the importance of choosing
the correct design for the speciﬁc environment of the experiment,
such as the tissues and genes to be analysed. Across all of the tissues
and genes analysed, the variance contribution from the qPCR
step was only around 10% of the total and the contribution from
the RT was found to exhibit about 2-times this variability, a result
that is in agreement with earlier ﬁndings [22]. Along with the RT
step, the variance of the qPCR replicates was found to be independent
of the gene being assayed. We conclude that the use of technical
replicates at the qPCR level have minimal impact on the
precision of the estimated Cq value, in agreement with previous
ﬁndings [17,23]. In almost all observations, normalisation to the
reference gene reduced the variance attributable to the RT step
and the total variance was reduced in cases where the variance
structure of the reference was similar to that of the gene of interest.
The variability between sample replicates was found to be
highly tissue-dependent and inconsistent estimates of the intersubject
variation in blood and culture tissues suggest that this variation
may be gene-dependent.
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
4 20 36 52 68 84 100 116 132 148 164 180 196 212 228 244 260 276 292
Power
Total number of samples
x=0.05
x=0.1
x=0.2
x=0.3
x=0.5
x=1.0
x=1.5
x=2.0
x=3.0
x=5.0
Fig. 3. Example power-curves. Example power-curves for a number of theoretical experiments at various measurement precisions at an alpha-level of 5%. The power is
calculated for unpaired, two-tailed t-tests between two groups of samples of equal size and equal variance. For each curve, the variable ‘x’ is deﬁned as the fraction of the
standard deviation of the two groups compared to the difference in the means of the two groups. For example x = 0.5 corresponds to a standard deviation that is equal to half
the difference in the means, x = 1.0 corresponds to a standard deviation that is equal to the difference in the means, and x = 3.0 corresponds to a standard deviation that is
equal to three times the difference in the means.
R.R. Kitchen et al. / Methods 50 (2010) 231–236 235
It should be highlighted that in order for the technique described
here to be valid it is essential that the subjects, samples,
and pre-processing procedures used for the pilot are representative
of those taken forward to the larger assay. It is obvious that the
likelihood of the pilot being representative is increased through
the use of larger numbers of biological and technical replicates;
however, a sensible compromise must be made to limit the size
and cost of the pilot study. We would generally recommend that,
for the pilot to offer meaningful variance estimates, no fewer than
three replicates are used at each level. In addition, although the use
of technical replicates increases the statistical power of the assay
by increasing the precision of the measurements, technical replicates
are not independent and do not increase the number of biological
observations of the given subpopulation.
When measurement is expensive and/or the individual measurements
are very precise it is preferable to add biological replicates
rather than technical replicates. In conditions, exempliﬁed
by the bovine liver described above, where the dominant source
of variability is between measurements rather than between the
biological replicates, the use of technical replicates will be very
effective in increasing precision. In general, however, the most
effective means of increasing the power and validity of qPCR
experiments is to increase the number of independent biological
replicates randomly selected from within each subpopulation.
Acknowledgements
We thank Becky Carlyle and Jano van Hemert for helpful discussions.
This work was in part funded by EU Framework VI FP6-2004IST-NMP-2:
SmartHEALTH and GAAV IAA500520809. RRK is
funded by the UK EPSRC.
References
[1] S.A. Bustin, J. Mol. Endocrinol. 25 (2) (2000) 169–193.
[2] M. Kubista, J.M. Andrade, M. Bengtsson, A. Forootan, J. Jonák, K. Lind, R.
Sindelka, R. Sjöback, B. Sjögreen, L. Strömbom, A. Ståhlberg, N. Zoric, Mol.
Aspects Med. 27 (2006) 95–125.
[3] T. Nolan, R.E. Hands, S.A. Bustin, Nat. Protocol 1 (3) (2006) 1559–1582.
[4] M.L. Wong, J.F. Medrano, BioTechniques 39 (1) (2005) 75–85.
[5] K.J. Livak, T.D. Schmittgen, Methods 25 (4) (2001) 402–408.
[6] R.A. Fisher, The Design of Experiments, ninth ed., Hafner Press, 1971. p. 248.
[7] J. Cohen, Statistical Power Analysis for the Behavioral Sciences, second ed.,
Hillsdale, N.J., 1988. p. 567.
[8] G.A. Churchill, Nat. Genet. 32 (Suppl.) (2002) 490–495.
[9] P.R. Rosenbaum, Replicating effects and biases, Am. Stat. 55 (3) (2001) 223–
227.
[10] S.A. Bustin, T. Nolan, J. Biomol. Tech. 15 (3) (2004) 155–166.
[11] S.A. Bustin, V. Benes, J.A. Garson, J. Hellemans, J. Huggett, M. Kubista, R.
Mueller, T. Nolan, M.W. Pfafﬂ, G.L. Shipley, J. Vandesompele, C.T. Wittwer, Clin.
Chem. 55 (4) (2009) 611–622.
[12] R.V. Lenth, Am. Stat. 55 (3) (2001) 187–193.
[13] L. Thomas, Conservation Biol. 11 (1) (1997) 276–280.
[14] A. Ståhlberg, J. Håkansson, X. Xian, H. Semb, M. Kubista, Clin. Chem. 50 (3)
(2004) 509–515.
[15] E. Limpert, W.A. Stahel, M. Abbt, Bioscience 51 (5) (2001) 341–352.
[16] A.L. Oberg, D.W. Mahoney, Methods Mol. Biol. 404 (2007) 213–234.
[17] A. Tichopad, R. Kitchen, I. Riedmaier, C. Becker, A. Ståhlberg, M. Kubista, Clin.
Chem. 55 (10) (2009) 1816–1823.
[18] G.W. Snedecor, W.G. Cochran, Statistical Methods, eighth ed., Iowa State Univ
Press, 1989, p. 503.
[19] M.W. Pfafﬂ, J. Vandesompele, M. Kubista, Data analysis software, in: J. Logan,
K. Edwards, N. Saunders (Eds.), Real-Time PCR: Current Technology and
Applications, Caister Academic Press, 2009. ISBN: 978-1-904455-39-4.
[20] R Development Core Team, R: A language and environment for statistical
computing, R Foundation for Statistical Computing, ISBN: 3-900051-07-0,
URL: http://www.R-project.org, 2008.
[21] MultiD Analyses, GenEx software, URL: http://www.multid.se, 2009.
[22] A. Ståhlberg, M. Kubista, M. Pfafﬂ, Clin. Chem. 50 (9) (2004) 1678–1680.
[23] M. Bengtsson, M. Hemberg, P. Rorsman, A. Ståhlberg, BMC Mol. Biol. 9 (2008)
63.
ActB
H
istone
IL8
Bcl2
ActB
C
asp
3
IL-1b
IFN
g
0
0.5
1
1.5
2
2.5
3
ActB
C
asp
3
IL-1b
FG
F7
Variance
Subjects
Samples
RTs
qPCRs
0
0.5
1
1.5
2
2.5
3
C
asp
3
IL-1b
FG
F7
Variance
C
asp
3
IL-1b
IFN
g
H
istone
IL8
Bcl2
Liver Blood Cell
culture
A
B
Fig. 4. Estimated confounding variation contributed by the sample processing
steps. The estimated variance contribution of each of the four sampling levels to the
overall variance in the measurements for several bovine tissues and genes. The top
row of plots (A) illustrates the variances of raw Cq data while the second row (B)
illustrates variances of DCq data after normalising to a reference gene-ActB.
236 R.R. Kitchen et al. / Methods 50 (2010) 231–236