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

net/publication/11253164

Avoiding Bias in Calculations of Relative Growth Rate

Article in Annals of Botany · August 2002

DOI: 10.1093/aob/mcf140 · Source: PubMed

CITATIONS READS
321 754

2 authors:

William Arthur Hoffmann Hendrik Poorter

North Carolina State University Forschungszentrum Jülich
161 PUBLICATIONS 6,000 CITATIONS 169 PUBLICATIONS 25,656 CITATIONS

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Gestion Durable des Ecosystèmes de Savane: cas d'une savane guinéenne (Réserve de Lamto) et d'une savane soudanaise (Parc National de la Comoé) en Côte d'Ivoire,
Afrique de l'Ouest View project

Testing potential of perennials as alternative biofuel species - using intercroping with legumes View project

All content following this page was uploaded by William Arthur Hoffmann on 31 March 2014.

The user has requested enhancement of the downloaded file.

Annals of Botany 80: 37±42, 2002
doi:10.1093/aob/mcf140, available online at www.aob.oupjournals.org

Avoiding Bias in Calculations of Relative Growth Rate

W I L L I A M A . H O F F M A N N 1 , * and H E N D R I K P O O R T E R 2
1Departamento de Engenharia Florestal, Universidade de BrasõÂlia, Caixa Postal 04357, BrasõÂlia, DF 70919-970,

Brazil and 2Plant Ecophysiology, Utrecht University, PO Box 800´84, 3508 TB Utrecht, The Netherlands

Received: 3 December 2001 Returned for revision: 22 February 2002 Accepted: 14 March 2002

In classical growth analysis, relative growth rate (RGR) is calculated as RGR = (ln W2 ± ln W1)/(t2 ± t1), where
W1 and W2 are plant dry weights at times t1 and t2. Since RGR is usually calculated using destructive harvests
of several individuals, an obvious approach is to substitute W1 and W2 with sample means W 1 and W 2. Here we
demonstrate that this approach yields a biased estimate of RGR whenever the variance of the natural logarithm-
transformed plant weight changes through time. This bias increases with an increase in the variance in RGR, in
the length
ÿ of the interval

between harvests, or in sample size. The bias can be avoided by using the formula
RGR ˆ ln W2 ÿ ln W1 = t2 ÿ t1 †, where ln W1 and ln W2 are the means of the natural logarithm-transformed
plant weights. ã 2002 Annals of Botany Company

Key words: Relative growth rate, growth analysis, methodology, bias.

INTRODUCTION Hereafter, we will refer to this as estimator 1.

Alternatively, RGR can be calculated from the mean
Growth analysis is a widely used analytical tool for
characterizing plant growth. Of the parameters typically natural logarithm-transformed plant weights:
calculated, the most important is relative growth rate ln W2 † ÿ ln W1 †
(RGR), de®ned as the parameter r in the equation: rÃ2 ˆ 4†
t2 ÿ t1
W2 ˆ W1 er t2 ÿt1 † 1†
where ln W†t is the mean of the ln-transformed plant
where W1 and W2 are plant dry weights at times t1 and t2. weights at time t. We will refer to this as estimator 2. In the
Rearrangement of terms yields the equation used to ®rst estimator, plant weights are averaged before ln-
calculate RGR in what is called the classical approach transforming, whereas in the second estimator, the weights
(Hunt, 1982): are ln-transformed before averaging. These estimators
almost always yield different values due to the fact that
ln W2 † ÿ ln W1 † ln W† is not equal to ln W† if there is variation in plant
rˆ 2†
t2 ÿ t1 weight among individuals (Aitchison and Brown, 1976).
Additionally, RGR can be calculated using the pairing
Since eqn (1) describes the growth of a single individual, method or the functional approach. In the pairing method
eqn (2) provides a formula for calculating the RGR of a (Evans, 1972), plants are grouped into pairs of similarly
single individual. Destructive harvesting is required to sized individuals before the ®rst harvest. One plant of each
determine plant dry weight, so in practice RGR is calculated pair is harvested on the ®rst harvest date, and the other plant
from samples of individuals from the same cohort at two is harvested on the second date. RGR is then calculated for
points in time (Evans, 1972). While eqn (2) is widely each pair, and the values averaged over all pairs. We show
presented in studies utilizing classical growth analysis, it in the Appendix that the estimate provided by this approach
has generally been overlooked that this formula does not is identical to that of estimator 2. Alternatively, in the
provide an unequivocal interpretation of how RGR should functional approach, a curve is ®t to the ln-transformed plant
be calculated. Several possible interpretations of eqn (2) weights through time and RGR at a particular time is
arise when calculating RGR from samples of individuals. calculated as the slope of the curve. When applied to
The ®rst, and most obvious, approach is to simply substitute harvests made at only two points in time, the results are
W1 and W2 with sample means W 1 and W 2, estimating RGR identical to estimator 2. While these two methods offer the
as: advantage of providing estimates of the variance in RGR,
ÿ
ÿ
we do not consider either of these further because the
ln W 2 ÿ ln W 1 estimates of RGR do not differ from estimator 2.
rÃ1 ˆ 3† As far as we know, little attention has been given to the
t2 ÿ t1
fact that two interpretations of eqn (2) do exist, and that they
might yield different estimates of RGR. As a consequence,
* For correspondence. Fax 00 55 61 347 5458, e-mail [email protected] recommendations vary among authors as to which estimator
ã 2002 Annals of Botany Company
38 Hoffmann and Poorter Ð Estimating Relative Growth Rate
should be used. For example, McGraw and Garbutt (1990) Plant weights for the second harvest were generated as
recommend the ®rst approach, Venus and Causton (1979) W2 ˆ W10 ert , where r was generated from a normal distri-
and Causton and Venus (1981) recommend the second, and bution with a mean of 0´05 g g±1 d1. For these latter weights,
Radford (1967) and Evans (1972) suggest the pairing W10 was generated from the same probability distribution as
method. Other texts offer no insight as to which form to use the W1 of the ®rst harvest. Here we distinguish between W1
(Hunt, 1982, 1990; Chiariello et al., 1991). and W10 to simulate a realistic case in which different
In most research reports there has been a similar lack of individuals are sampled on the two harvest dates. The term
attention to this issue. Examining issues of Annals of Botany W1 denotes the weight at time 1 of individuals that were
from 1993 to 2001, we identi®ed 28 papers applying harvested at time 1, whereas W10 is the weight at time 1 of
classical growth analysis. Of these, one utilized estimator 1, individuals that were harvested at time 2. Cases were run for
®ve used estimator 2 and in the remaining 22 studies it was a range of values of sRGR from 0 to 0´03, a range of harvest
not possible to infer which equation was used. Clearly, this intervals from 5 to 200 d, and a range of sample sizes from
is an aspect that has not received enough attention. one to 40 individuals per harvest. For each set of conditions
Therefore, in this paper, we analyse the performance of tested, 10 000 simulations were run.
these two estimators to permit well-founded recommenda- In the above simulations, we assumed that plant weight is
tions for the correct choice of estimator. lognormally distributed and that RGR is normally distrib-
uted, so we ran a third set of simulations to test whether the
performance of the two estimators depends on the prob-
MA TE R IA L S A N D ME T H O D S
ability distributions used to generate plant weight and RGR.
To determine if these two estimators yield unbiased It is not feasible to examine all possible probability
estimates of RGR, i.e. that their expected values are distributions, so we limited our study to the normal,
identical to the true mean RGR, we attempted to derive lognormal and exponential distributions. Normal and expo-
analytically the expected values of the two estimators. nential distributions were chosen to provide cases that are
While this analytical approach was possible for the second either unskewed or more strongly right-skewed than the
estimator, for the ®rst it was necessary to resort to a second- lognormal distribution. Probability distributions were gen-
order approximation of the expected value. erated using the algorithms of Press et al. (1989).
Since we were unable to ®nd an exact form for the
expected value of both equations, we relied on simulations
to substantiate the differences between these estimators and R E SU L T S A N D D IS C U SS IO N
to con®rm the generality of the results. In the ®rst set of In agreement with Causton and Venus (1981), we demon-
simulations we generated data corresponding to a typical strate that estimator 2 is unbiased; i.e. the expected value of
experiment in which ten individuals were harvested on each this estimator (Ãr2 ) is equal to the true mean RGR. In
of two dates. Since plant weight is commonly lognormally contrast, estimator 1 is biased. As shown in the Appendix,
distributed (Poorter and Garnier, 1996), we used this when plant weight is lognormally distributed, the bias is
distribution in our simulations. To generate an individual approximately equal to:
plant weight, the algorithm of Press et al. (1989) was used to !
generate a ln W value from a normal distribution. This value 1 e
s2ln W s2
2 ÿ e ln W1
was then transformed with the exponential function to arrive rÃ1 ÿ r  s2ln W2 ÿ s2ln W1 ÿ 5†
2 t2 ÿ t1 † n
at W. The variance of the underlying normal distribution is
s2ln W , which is the variance of the ln-transformed plant
weights. where n is the number of individuals sampled per harvest
Since it is variation in plant weight that causes the two date and r is the true mean RGR. A lognormal distribution
estimators to yield different estimates of RGR, we examined in plant weight develops naturally when RGR or germin-
the effect of the variances s2ln W1 , s2ln W2 as well as RGR on ation time is normally distributed (Poorter and Garnier,
possible bias in the RGR estimates. Using an initial true 1996), so the assumption of lognormality is probably
mean weight (W1) of 1 g and a harvest interval of 50 d, we appropriate in most applications.
generated cases in which we varied s2ln W1 , s2ln W2 and the In the ®rst set of simulations, where s2ln W1 , s2ln W2 and
true mean RGR independently of each other. We used a mean RGR were varied independently of each other, we
fully factorial array of cases using six levels of RGR (0´05, con®rm that the ®rst estimator is biased but the second is not
0´10, 0´15, 0´20, 0´25, 0´30 g g±1 d±1) and eight levels each of (Fig. 1). The size of the bias depends on the values of plant
sln W1 and sln W2 (0´1, 0´2, 0´3, 0´4, 0´5, 0´6, 0´7, 0´8). For variability. If s2ln W remains constant through time, the bias
each combination of parameters, 10 000 simulations were is nil, but the larger the difference between s2ln W1 and s2ln W2 ,
run. For each simulation, RGR was calculated from the the larger the error. Consequently, the relationship between
generated data using both estimators and results were s2ln W2 ± s2ln W1 † and the bias is well described by a single
compared with the true mean RGR. curve, regardless of the value of RGR. The approximate
In a second set of simulations, we examined the effects of bias, as estimated by eqn (5), closely ®ts the simulated
increasing variance in RGR, time between harvests and values (Fig. 1).
sample size on the performance of these two estimators. In A change in s2ln W through time can occur only if there is
these simulations, plant weights (W1) for the initial harvest some variation in RGR among individuals, so it is useful to
were randomly generated from a lognormal distribution. examine how variation in RGR affects the bias of the ®rst
 Hoffmann and Poorter Ð Estimating Relative Growth Rate 39

F I G . 1. Relationship between the difference in variance at two harvest

times and simulated bias of the two RGR estimators. Results are from a
fully factorial array of simulations using six levels of RGR (0´05, 0´10,
0´15, 0´20, 0´25, 0´30 g g±1 d±1) and eight levels each of and (0´1, 0´2,
0´3, 0´4, 0´5, 0´6, 0´7, 0´8). Each point represents the mean of 10 000
simulations with a sample size of ten individuals in each harvest. The
continuous line shows the bias as estimated by eqn (5).

estimator (Ãr2 ). By applying rules for variances of linear

combinations of variables to eqn (1), it can be shown that
when RGR and W1 are independent of each other then

s2ln W2 ÿ s2ln W1 ˆ t2 ÿ t1 †2 s2RGR 6†

Therefore, the bias of the ®rst estimator is expected to

increase in response to increased time between harvests or
increased variance in RGR, as con®rmed by the second set
of simulations (Fig. 2A and B). These simulations also
demonstrate that the bias increases with increasing sample
size, as to be expected from eqn (5) (Fig. 2C).
In general, if there is any variation in RGR among
individual plants, s2ln W2 will be larger than s2ln W1 , and
estimator 1 will have a positive bias, provided that RGR and
W1 are independent variables. In reality, RGR and W1 need
not be independent of each other. We might expect a
negative covariance between RGR and W1 due to the
frequently observed decline in RGR through plant devel-
opment. If larger plants at t1 are those individuals that have
reached a more advanced stage of development, they may
subsequently exhibit lower growth rates than smaller
individuals. Such a decline could be due to developmental
changes in allocation or photosynthesis, or could arise from
an increasingly limiting nutrient supply. Similarly, there is
often a negative correlation between seed size and RGR, at F I G . 2. Simulation results demonstrating the effect of variation in RGR
least among species (Swanborough and Westoby, 1996). If a (A), time between harvests (B) and sample size (C) on the bias of the
two RGR estimators. In A, RGR was ®xed at 0´05, with a harvest
similar relationship occurs within a species, a negative interval of 50 d and a harvest size of ten individuals. In B, RGR was
covariance between W1 and RGR could emerge. In any case, ®xed at 0´05 with a standard deviation of 0´01 and a sample size of ten
a negative correlation between W1 and RGR would reduce individuals. In C, RGR was ®xed at 0´05 with a standard deviation of
the tendency of s2ln W to increase through time and thereby 0´02 and harvest interval of 100 d. For each set of parameters, 10 000
reduce the bias of estimator 1. If this negative covariance is simulations were run.
40 Hoffmann and Poorter Ð Estimating Relative Growth Rate
TA B L E 1 . Simulation results using different probability distributions for initial weight and variance of RGR

Bias of Bias of
Initial weight Initial plant weight RGR RGR (g g±1 d±1) estimator 1 estimator 2
distribution (g) (mean + s.d.) distribution (mean + s.d.) (g g±1 d±1) (g g±1 d±1)

Normal 1 (0´2) Normal 0´05 (0´01) 0´0043 0´0000

Normal 1 (0´2) Lognormal 0´05 (0´01) 0´0024 0´0000
Normal 1 (0´2) Exponential 0´05 (0´05) 0´0753 0´0000
Lognormal 1 (0´2) Normal 0´05 (0´01) 0´0022 0´0000
Lognormal 1 (0´2) Lognormal 0´05 (0´01) 0´0024 0´0000
Lognormal 1 (0´2) Exponential 0´05 (0´05) 0´0583 0´0001
Exponential 1 (1) Normal 0´05 (0´01) 0´0038 0´0000
Exponential 1 (1) Lognormal 0´05 (0´01) 0´0023 0´0002
Exponential 1 (1) Exponential 0´05 (0´05) 0´0708 0´0007

Simulations were run for 100 d, with mean initial weight of 1 g and mean RGR of 0´050 g g±1 d±1.
Note that for an exponential distribution, the standard deviation is constrained to be equal to the mean, so when this distribution is used, it was not
possible to utilize the same standard deviation as was used for the others.

large enough, it could result in a decline in s2ln W through suggest the use of estimator 2, as given in eqn (4),
time, causing a negative bias, as appears in Fig. 1. exclusively, as this equation yields an unbiased estimate
In contrast, there could be a positive correlation between of RGR under all conditions.
W1 and RGR due to genetic variation in RGR, whereby
plants attaining higher W1 due to higher RGR continue
growing at a greater RGR. Alternatively, in experimental A C K N O W L E D G E ME N T S
situations permitting competition among individuals, larger We thank Feike Schieving for checking the mathematical
individuals may gain a competitive advantage and therefore derivations and Danny Tholen, Rens Voesenek, Adrie van
maintain higher RGR than competitively suppressed indi- der Werf and Kaoru Kitajima for comments on the
viduals. Either of these situations could result in a positive manuscript.
covariance between W1 and RGR, thereby accentuating the
quantity s2ln W2 ± s2ln W1 and consequently the bias of
estimator 1. L IT E RA TU R E C I TE D
How important is the bias in RGR when using the wrong Aitchison J, Brown JAC. 1976. The lognormal distribution. Cambridge:
estimator? We used experimental data to demonstrate that Cambridge University Press.
the bias can be large enough to be of concern. In a study Causton DR, Venus JC. 1981. The biometry of plant growth. London:
involving 18 tree and shrub species from the cerrado Edward Arnold.
Chiariello NR, Mooney HA, Williams K. 1991. Growth, carbon
savannas of Brazil (W.A. Hoffmann, unpubl. res.), mean allocation and cost of plant tissues. In: Pearcy RW, Ehleringer J,
sample variance, s2ln W , increased from 0´17 at 50 d to 0´18 at Mooney HA, Rundel PW, eds. Plant physiological ecology. London:
100 d and to 0´27 at 150 d. We estimated the bias of the Chapman and Hall, 327±366.
calculated RGR for each species using eqn (5). On average, Evans GC. 1972. The quantitative analysis of plant growth. Oxford:
Blackwell Scienti®c.
the bias was estimated to be +1 % and +13 % for the ®rst and
Hunt R. 1982. Plant growth curves. London: Edward Arnold.
second intervals, respectively, demonstrating that the bias is Hunt R. 1990. Basic growth analysis: plant growth analysis for beginners.
substantial. London: Unwin Hyman.
With experimental data, we cannot know the true bias McGraw JB, Garbutt K. 1990. Demographic growth Analysis. Ecology
since we must rely on s2ln W , which is an estimate of the true 71: 1199±2004.
Poorter H, Garnier E. 1996. Plant growth analysis: an evaluation of
s2ln W based on ®nite samples. Similarly, we must depend on experimental design and computational methods. Journal of
a second-order approximation of the bias. The simulations Experimental Botany 47: 1343±1351.
indicate that this approximation provides a reliable estimate Press WH, Flannery BP, Teukolsky SA, Vetterling WT. 1989.
of bias, at least for reasonable values of s2ln W1 and s2ln W2 . Numerical recipes in Pascal. Cambridge: Cambridge University
Another uncertainty with experimental data is that plant Press.
Radford PJ. 1967. Growth analysis formulae ± their use and abuse. Crop
weight may not be lognormally distributed. However, the Science 7: 171±175
third set of simulations indicates that this bias is not limited Rice JA. 1988. Mathematical statisitcs and data analysis. Patricia Grove:
to lognormally distributed data. Regardless of the probabil- Wadsworth and Brooks/Cole.
ity distribution used to generate plant weight and RGR, Swanborough P, Westoby M. 1996. Seedling relative growth rate and its
estimator 1 was biased whereas estimator 2 was not components in relation to seed size: phylogenetically independent
contrasts. Functional Ecology 10: 176±184.
(Table 1). Venus JC, Causton DR. 1979. Plant growth analysis: a re-examination of
In conclusion, estimator 1 presented in eqn (3) is biased the methods of calculation of relative growth and net assimilation
and should therefore be avoided in RGR calculations. We rates without using ®tted functions. Annals of Botany 43: 633±638.
 Hoffmann and Poorter Ð Estimating Relative Growth Rate 41
A P PE N D IX
E ln W10 erDt †† ÿ E 1n W1 ††
Proof that the pairing method is equivalent to estimator 2 E rÃ2 † ˆ
Dt
Here we demonstrate that the pairing method (Evans, 1972)
is equivalent to the second estimator [eqn (4)]. In the pairing E ln W10 † ‡ ln erDt †† ÿ E 1n W1 ††
method, plants are grouped into pairs of similarly sized ˆ
Dt
plants. One plant of each pair is harvested at time t1, and the
other at time t2. RGR is then calculated for each pair j, and E ln W10 †† ‡ E ln erDt †† ÿ E 1n W1 ††
then these values are subsequently averaged over all pairs. ˆ
Dt
RGR is therefore estimated as:
 
1 X ln W2j ÿ ln W1j Since W1 and W10 are identically distributed random
Ãr3 ˆ
n j t2 ÿ t1 variables, their expected values are identical. Therefore
E rÃ2 † ˆ Dt1 E ln erDt †† ˆ Dt1 E rDt† ˆ E r† ˆ r where r, is
the true mean RGR. Since the expected value of the
where n is the number of pairs and Wij is the dry weight of estimator rÃ2 is equal to the expected value of RGR, we can
individual j of harvest i. conclude that rÃ2 is an unbiased estimator.
This reduces as follows:
1 Xÿ

rÃ3 ˆ ln W2j ÿ ln W1j Proof that rÃ1 is a biased estimator of RGR
n t2 ÿ t1 † j
Here we demonstrate that rÃ1 ˆ ln W 2t2†ÿln
ÿt1
W1†
is a biased
P P  estimator of RGR. The expected value of this estimator is
1 j ln W2j j ln W1j
ˆ ÿ
t2 ÿ t1 n n
ln W 2 † ÿ ln W 1 †
E rÃ1 † ˆ E †:
ln W2 † ÿ ln W1 † Dt
ˆ
t2 ÿ t1
We were unable to derive an exact solution for this
This is therefore equivalent to eqn (4). estimator so we used the second-order approximation:

Proof that rÃ2 is an unbiased estimator of RGR s2W

E 1n W††  ln mw † ÿ
Here we demonstrate that 2m2W

ln W2 † ÿ ln W1 †
rÃ2 ˆ (Rice, 1988; p. 143), where2 mw is the true population mean
t2 ÿ t1 plant dry weight and s W2 ˆ snW is the true variance of the mean
of plant dry weight, based on some sample size n. If W is
is an unbiased estimator of RGR. The weight at time 2 of an lognormally distributed, we know that
individual i can be expressed as W2i ˆ W1i eri Dt , where t2 ± t1
is replaced with Dt. The estimator can now be rewritten as 2
mW ˆ eln W‡0:5 sln W
ln W10 erDt† ÿ ln W1 †
rÃ2 ˆ :
Dt and
We distinguish between W1 and W10 because different
s2W ˆ eln W‡0:5 sln W †2 esln W ÿ 1†
2 2
individuals are sampled on the two harvest dates. The term
W1 denotes the weight at time 1 of individuals that were
harvested at time 1, whereas W10 is the weight at time 1 of the
individuals that were harvested at time 2. We assume that (Aitchison and Brown, 1976), so
the individuals harvested at the two times are chosen 2
esln W ÿ 1
randomly from the same cohort, so W1 and W10 are E ln W††  ln W ‡ 0:5s2ln W ÿ
identically distributed random variables. 2n
Using theorems for linear combinations of random
variables (Rice, 1988; pp. 109±112), the expected value of and
is:
 42 Hoffmann and Poorter Ð Estimating Relative Growth Rate

   The ®rst term of the right-hand side of this equation is the

ln W 2 † ÿ ln W 1 † 1 expected value of RGR, so the bias is approximately
E  E ln W2 †† ‡
Dt Dt  
s2 s2
 1 e ln W2 ÿ e ln W1
2
s2ln W2 esln W2 ÿ 1 s2 s2
e ln W1 ÿ 1 s2ln W2 ÿ s2ln W1 ÿ
ÿ ÿ E ln W1 †† ÿ ln W1 ‡ 2Dt n
2 2n 2 2n

E ln W2 †† ÿ E ln W1 †† 1
ˆ ‡ s2ln W2 ÿ s2ln W1 ÿ
Dt 2Dt
s2ln W s2ln W 
e 2 ÿe 1

View publication stats