See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/6393791

Precision of Measurement as a Component of Human Variation

Article in Journal of PHYSIOLOGICAL ANTHROPOLOGY · April 2007

DOI: 10.2114/jpa2.26.253 · Source: PubMed

CITATIONS READS

146 639

2 authors:

Goto Rie Nick Mascie-Taylor

University of Cambridge University of Cambridge
23 PUBLICATIONS 1,188 CITATIONS 227 PUBLICATIONS 9,298 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Nick Mascie-Taylor on 01 March 2016.

The user has requested enhancement of the downloaded file.

 Precision of Measurement as a Component of Human Variation

Rie Goto and C.G. Nicholas Mascie-Taylor

Department of Biological Anthropology, University of Cambridge

Abstract This paper describes the main quality control Variance Components
methods for determining human observer measurement error
and instrument error focusing on intra-observer and inter- Usually for a continuous (quantitative) trait such as weight
observer technical error of measurements (TEMs) and relative or height the variance (s2) or standard deviation (s) would be
TEM, and the coefficient of reliability (R). R values above 0.95 the statistic of choice to quantify the extent of the variation
are indicative of small errors. To compare variances between (Zar, 1999). Suppose, for example, the variance in weight of a
different variables, the coefficient of variation, a measure of random sample was determined. It would be erroneous to
relative variability, is used. The total variation of a character assume that this calculated variance necessarily reflected the
can be partitioned into genetic, environmental, and error ‘real’ variance of the character under investigation. This
components. Determination of the genetic component of depends on how precisely the weight is measured. In other
variation (heritability) is usually obtained from twin studies or words, errors may occur because of human measurement error
other family studies. A good environment improves the or through instrument error.
population mean, whereas a poor environment can lower the
population mean without any change in the underlying genetic Thus, ‘Total’ variance
‘Real’ varianceHuman measurement
structure. Worked examples of how to calculate TEM, R, and errorInstrument error.
heritability are provided. J Physiol Anthropol 26(2): 253–256, Measuring the magnitude of these errors is a major element
2007 http://www.jstage.jst.go.jp/browse/jpa2 of quality control.
[DOI: 10.2114/jpa2.26.253]
Quality Control Determination
Keywords: quality control, technical error of measurement,
coefficient of reliability, heritability Human measurement error
Human measurement error or observer measurement error is
indicative of whether research assistants measure a trait
Introduction consistently. Imagine that ten children have had their weight
and height independently measured by two research assistants
The study of similarities and differences between human (A and B) as shown in Table 1.
populations has been investigated by scientists for centuries,
and the notion of different human races derived from the Table 1 Weights and heights (recumbent length) of 10 children
incorrect supposition that there was greater variation between measured by two assistants, A and B (Goto, 2006)
populations than within populations (Cavalli-Sforza and
Weight (kg) Height (cm)
Bodmer, 1971). Nowadays it is widely acknowledged that Child
upwards of 95% of the variation of a trait is within a Number
Assistant A Assistant B Assistant A Assistant B
population and only 5–10% of the variation is between
populations (Cavalli-Sforza and Bodmer, 1971). These within 1 5.50 5.45 64.50 64.00
2 7.70 7.60 71.00 71.50
and between population differences reflect underlying genetic
3 5.35 5.35 58.00 59.00
and environmental variation. However, if errors have occurred 4 5.00 5.05 58.00 58.00
in the measurement of the trait under consideration (either due 5 8.95 8.95 70.50 71.50
to instrument or human measurement errors), these can 6 6.15 6.15 69.00 67.50
compromise many statistical methods and may falsify the 7 6.70 6.75 63.00 64.00
8 5.85 5.85 65.00 64.50
partitioning into genetic and environmental components.
9 7.35 7.35 62.00 62.00
10 6.95 6.95 68.00 68.00
254 Precision of Measurement

There are a number of ways in which the consistency or TEM

inconsistency in weight measurements between assistants can %TEM
* 100
X
be examined. For example, the correlation coefficient (r)
provides a measure of the association (consistency) of two The %TEMs were 0.0687 and 0.8433 for weight and height
continuous normally distributed variables, which in this sample respectively.
is very high for both weight and height (r
0.999 and 0.985, From TEM, the coefficient of reliability (R) can be
respectively) and both are highly significant (p0.001). determined, which ranges from 0 (not reliable) to 1 (complete
However the correlation coefficient does not provide a measure reliability), where SD is the standard deviation of all
of how dissimilar two measurements are; for example if each measurements.
child’s weight recorded by Assistant B was 1 kg more than
 (TEM )2 
indicated in Table 1, the correlation coefficient would be 1 2 
unchanged at 0.999! Inconsistency between two measurements  SD 
can be assessed using a paired t-test, which determines
whether the mean difference is significant. In these examples So in these examples, (reducing the decimal places to 4 for
the mean differences were not significant. If three or more simplicity) R
1(0.00452/1.18662)
0.999986 for weight and
assistants had measured the same children, then a repeated R
1(0.54772/4.57942)
0.985686 for height. Although there
measures analysis of variance would have been used to are no recommended values for R, Ulijaszek and Kerr (1999)
examine the extent of the within-subject variance. suggested that a cut-off of 0.95 be used (i.e. a human
Alternatively the technical error of measurement (TEM) can measurement error of up to 5%). So a reduction in error
be determined which is an accuracy index and measures indicates improvement in measurement technique between
the standard deviation between repeated measures. The assistants, and greater quality control.
formulation of TEM depends on how many assistants have
taken the measurement. If the same assistant has measured the Instrument Error
children on two occasions (a measure of intra-TEM) or two The extent to which a machine or instrument gives the same
assistants have measured the same children (inter-TEM as in reading provides an indication of instrument error. For
Table 1), then the formula for TEM is example, one of us (Goto, 2006) conducted a longitudinal
study in which infants had their weight measured at monthly
∑D 2 intervals. The two weighing scales were checked every time
before use (134 times) with 5 different weights. The inter-TEM
2N
was determined (0.0443) and the coefficient of reliability was
where D is the difference between the two measurements, and 0.999924 (99.9924%), indicating that the two scales were very
N is the sample size. For weight the inter-TEM
0.0045 and consistent.
for height TEM
0.5477. The much higher height (recumbent Consequently, in percentage terms, combining the results of
length) TEM value is because the unit of measurement was in human measurement error (i.e. 100%R
100%99.9986%

centimetres. 0.0014%) and instrument error (100%R
100%
If three or more assistants are involved, then the formulation 99.9924%
0.0076%) for weight gives
is more complicated: ‘Real’ variance

‘Total’ variance
 
2  (Human measurement errorInstrument error)
∑ M ( n)
K
 
N   
∑ ∑  M ( n)2   
100%(0.0014%+0.0076%)
K 1
1  1  K 
99.9910%
 
 
N ( K1) Relative Variability

where N is the sample size, K the number of assistants, and M As a measure of variability, the variance (and, of course,
the measurement, and M(n) is the nth replicate of measurement standard deviation) have magnitudes which are dependent on
(see Ulijaszek and Lourie, 1994 for a review of anthropometric the magnitude of the data. For example, elephants have ears
measurement error). that are 100 times larger than those of mice. If elephant ears
It is also possible to compute the relative TEM (%TEM), were no more variable, relative to their size, than mouse ears,
which provides an estimate of the error magnitude relative to relative to their size, the standard deviation of elephant ear
the size of the measurement (expressed as a percentage) and is lengths would be 100 times as great as the standard deviation
analogous to the coefficient of variation (see below). of mouse ear lengths—and the variance of the former would be
10,000 times (1002) the variance of the latter! The coefficient
of variation (CV) expresses sample variability relative to the
 Goto, R and Mascie-Taylor, CGN J Physiol Anthropol, 26: 253–256, 2007 255

Table 2 Absolute (SD) and relative (CV) variability of height, weight, differences of the character within each type of twin pair. So,
plasma albumin, and haemoglobin concentrations from the for example, a USA study (Newman et al., 1937) found that
longitudinal study (Goto, 2006)
the variance of the difference in stature between dizygotic
Variable Mean SD CV (%) twins was 14.79, and for monozygotic twins 2.61. So,
VDZ VMZ (V V )VE VG
Height (cm) 67.28 4.89 7.27 h2
≡ G E

Weight (kg) 7.06 1.20 17.00 VDZ VG VE VT
Plasma albumin (g/l) 42.37 7.00 16.52
Haemoglobin (g/l) 91.83 12.57 13.68 14.792.61 12.18
h2


0.82
14.79 14.79
mean of the sample. Thus 82% of the stature variation in dizygotic twins is due to
genetic differences and 18% to the environment.
SD *100
CV
Just because the heritability is high does not imply that the
X mean value for a character cannot change. If individuals are
The coefficient of variation may be calculated only for ratio brought up in a poor environment they will be smaller than
scales (i.e., those with a true zero value). It is not valid to their genetic potential. For example, when Japanese migrants
calculate coefficients of variation for interval scales (e.g., to Hawaii and their first generation descendants were
temperature data using the Celsius Scale, or z-scores of height- compared, the descendants were over 4 cm taller (Shapiro,
for-age or weight-for-height). 1939), presumably because of the better environmental
Table 2 presents information on the mean and standard conditions in Hawaii at that time, compared with Japan. In
deviation (SD) for 4 nutritional variables. The standard developing countries the average stature and weight of adults is
deviation of haemoglobin is more than 10 times larger than the generally lower than in developed countries. These reductions
SD for weight, but after computation of the coefficient of are not caused by genetic factors. Rather, the poor environment
variation, weight has a larger CV than height (because the (food insecurity and poor nutritional intake, and high levels of
mean height is so much larger that the mean weight). Simpson infectious disease and stress) leads to child growth retardation
et al. (1960) present a good discussion of the coefficient of that starts as early as 3 months of age and continues
variation and its biological applications. throughout life.

Causes of Human Variation Discussion

Assuming that quality control has been good, what are the High levels of error can render statistical comparisons
causes of variation of trait within a population? Geneticists invalid, and errors can artificially inflate the variance
partition the variation of a trait into various components. associated with a particular measurement (Bailey and Byrnes,
Simplistically (ignoring errors):- 1990). Furthermore, empirical data have shown the
measurement error can compromise a wide variety of
Total Variation (VT)
Genetic Variation (VG) univariate and multivariate statistical methods, particularly
Environmental Variation (VE) those that rely on correlation, regression, and covariance (Liu,
1988; Bailey and Byrnes, 1990). Good quality control,
and heritability (h2)
VG/VT. Heritability defines the involving both reduction in instrumentation and human
proportion of the total variation accounted for by genetic measurement error, are therefore prerequisites for increased
differences between people within a population. Heritability precision and improved interpretation of results.
ranges from 0 (all variation environmental) to 1 (all variation A number of methods of measuring inconsistency are
due to genetic factors). More complicated formulations include available but the preferred method involves calculation of
taking into account, for example, genotype-environment either intra- or inter-technical error of measurement (TEM)
interactions and correlations (see Plomin et al., 2001 for a full initially, relative TEM, and then the coefficient of reliability
discussion). (Ulijaszek and Kerr, 1999). TEM gives information on the
Twin studies, comparison of non-identical (dizygotic, DZ) error margin of a trait and therefore is an accuracy index.
and identical (monozygotic, MZ) twin pairs, and other family Determination of acceptable levels of measurement errors are
studies, yield heritabilities of weight between 0.5 and 0.75 and not straightforward and relate to the variable being studied as
height between 0.7 and 0.9 (Cavalli-Sforza and Bodmer, 1971), well as the age of the subjects (Ulijaszek and Kerr, 1999).
suggesting that between 50% and 75% of the variance in Even so, coefficients of reliabilities above 0.95 are indicative
weight is caused by inheritance (nature), and only 25% to 50% of good quality control.
by the environment (nurture). Trait variation within and between populations results from
For twins, heritability can be calculated using the formula genetic differences between people as well as environmental
shown below, where VDZ and VMZ refer to the variance of the effects. Although determination of the genetic component
256 Precision of Measurement

within a population commonly relies on the twin method, York

which has limitations (Plomin et al., 2001), for many Simpson GG, Roe A, Lewontin RC (1960) Quantitative
anthropometric traits the heritability is high (0.5). Even so, Zoology. Harcourt, Brace, Jovanovich, New York
high heritabilities do not imply fixation of a trait. In a poor Shapiro HL (1939) Migration and environment: a study of the
environment, good genes, for example for stature, will not be physical characteristics of the Japanese immigrants to
able to show their full potential and so the population mean Hawaii and the effects of environment on their descendants.
will be reduced. In a good environment, the full genetic Oxford University Press, Oxford
potential will be realised and the overall population mean will Ulijaszek SA, Lourie JA (1994) Intra- and inter-observer error
be higher. in anthropometric measurement. In Ulijaszek SJ and
Mascie-Taylor CGN eds. Anthropometry: the individual and
References the population. Cambridge University Press, Cambridge,
30–55
Bailey RC, Byrnes J (1990) A new, old method for assessing Ulijaszek SA, Kerr DA (1999) Anthropometric measurement
measurement error in both univariate and multivariate error and the assessment of nutritional status. Br J Nutr 82:
morphometric studies. Syst Zool: 39, 124–130 165–77
Cavalli-Sforza LL, Bodmer WF (1971) The Genetics of Zar JH (1999) Biostatistical Analysis (4th edition). Prentice
Human Population. W.H. Freeman and Company, San Hall International Inc, New Jersey
Francisco
Goto R (2006) Effects of anti-Giardia and antihelmintic This article was presented at the 8th International Congress of
treatment on infant nutritional and biochemical status and Physiological Anthropology, 2006 (ICPA 2006), in Kamakura,
intestinal permeability in rural Bangladesh. PhD Thesis, Japan.
University of Cambridge Received: September 22, 2006
Liu K (1988) Measurement error and its impact on partial Accepted: November 17, 2006
correlation and multiple linear regression analyses. Am J Correspondence to: Dr. R. Goto, Department of Biological
Epidemiol: 127, 864–874 Anthropology, University of Cambridge, Downing Street,
Newman HH, Freeman FN, Holzinger KJ (1937) Twins: A Cambridge, CB2 3DZ
Study of Heredity and Environment. University of Chicago Phone: 44–1223–335462
Press, Chicago Fax: 44–1223–335460
Plomin R, DeFries JC, McClearn GE, McGuffin P (2001) e-mail: [email protected]
Behavioural Genetics (4th edition). Worth Publishers, New

View publication stats