ARTICLE IN PRESS Journal of Theoretical Biology 228 (2004) 149–153 Testing the allometric scaling laws Pekka Kaitaniemi* Hyytial 124, FIN-35500 Korkeakoski, Finland . a. Forestry Field Station, University of Helsinki, Hyytial . antie . Received 2 October 2003; received in revised form 8 December 2003; accepted 12 December 2003 Abstract Allometric scaling laws have received increasing attention due to the recent theoretical advancements. However, existing evidence suggests that the scaling relationships may vary a lot without much consistency, which poses a challenge to the applicability of general theories. In this report, I demonstrate that much of the discrepancy may be an artefact caused by the limited use of methods for estimating the parameters in the allometric scaling equations. I suggest alternative procedures that can be utilized to avoid biased interpretations. The comments are largely applicable to any research that involves parameterization of equations. r 2003 Elsevier Ltd. All rights reserved. Keywords: Allometry; Model; Parameter estimation; Regression 1. Introduction Various size measures of organisms have been long known to exhibit specific allometric scaling relationships. Recently, the theoretical studies by West et al. (1999b) suggested a new mechanistic explanation to these scaling laws, often empirically detected in organisms. The theory predicts numerous scaling relationships at multiple levels of biological organization, and the authors claim it can provide a general framework for linking the structure and function of organisms from cells to ecosystems (West et al., 1999b; Enquist, 2002). Therefore, it is likely that there will be an increasing number of studies designed to test the validity of the theoretical predictions for different organisms and traits. However, previous evidence suggests that the scaling relationships may, in fact, vary a lot without much consistency. For plants, Niklas (1994, p. 314) states that ‘each scaling relation is functional only within a narrow domain of size and form’, and for animals, Dodds et al. (2001) claim ‘after a systematic review of the available empirical data and theoretical arguments, we find evidence that there may not be a simple scaling law for metabolic rate’. It thus appears that the applicability of a single general theory might become easily challenged. However, as I will demonstrate in this paper, *Tel.: +358-3-3355217; fax: +358-3-3355555. E-mail address: pekka.j.kaitaniemi@helsinki.fi (P. Kaitaniemi). 0022-5193/$ - see front matter r 2003 Elsevier Ltd. All rights reserved. doi:10.1016/j.jtbi.2003.12.007 much of the discrepancy in the previous results may be an artefact caused by the methods used to determine the parameters in the allometric scaling equations. 2. Allometric scaling laws Allometric scaling laws have the form Y ¼ bX a ; and they represent the dependence of the variable Y on the variable X as a function that involves the normalization constant b and the scaling exponent a: The constant b may be specific for a particular system whereas the scaling exponent a is expected to follow the theoretical predictions (West et al., 1999b). A ‘traditional’ method for determining the scaling exponent is log-transforming the original variables and then fitting a linear reduced major axis (RMA) regression to the data. Although better alternatives exist (Isobe et al., 1990), this method is acceptable, if it is not clear which variable should be treated as the independent and which as the dependent, and if both variables are subject to measurement error. An ordinary leastsquares (OLS) regression, in turn, is recommended for predictive purposes where the aim is to illustrate the relationship (Niklas, 1994, p. 333). If a model predicts that one variable clearly depends on the other (e.g. West et al., 1999a), then OLS regression should be preferred even if there exists measurement error in the values of the variables (Isobe et al., 1990). ARTICLE IN PRESS 150 P. Kaitaniemi / Journal of Theoretical Biology 228 (2004) 149–153 However, besides linear regression, other methods for parameter estimation do also exist, and it is not obvious which of them produces the best fit (Haefner, 1996). It is also a known fact that the number of possible solutions increases as the number of parameters in an equation increases. As I will show, even the two parameters in the allometric equation leave scope for alternative values that can all be accepted using standard selection criteria. 3. Methods As an example, I use the analysis of data described in Kaitaniemi and Ruohom.aki (2003). It consists of allometric scaling relationships between branch and stem diameter, shoot number (comparable to leaf number), tree height, and branch and tree biomass in the mountain birch, Betula pubescens ssp. czerepanovii (Orlova) H.amet-Ahti. Figs. 9 and 10 in Kaitaniemi and Ruohom.aki (2003) show the data and the allometric relationships I investigated. In the present paper, I used four different methods for determining the values of a and b: First, I calculated both the OLS regression and the RMA regression for data that were first log-transformed. These calculations were made with the procedures REG and CORR in the SAS statistical analysis software (SAS Institute Inc., Cary, NC, USA). Second, I determined both parameters with untransformed data using the SAS procedure NLIN and selecting the Marquardt method for iteration. Finally, I set the value of the scaling exponent according to the theoretical value given by West et al. (1999a), and then used the procedure NLIN for determining the value of the coefficient b alone. To compare the resulting four parameterizations, I used the multiple working hypotheses approach (e.g. Haefner, 1996, pp. 22–24). In the present study, it meant that alternative values for model parameters were compared in parallel with each other to find the values that statistically best fit the data at hand. The conceptual basis of the model itself remained the same. However, the same approach would have been appropriate, if I had compared alternative structural formulations of a model instead of different parameter values. In that case, I would have judged the fit of data with different mechanistic models that are based on different hypotheses about the causal relationships, and that way predict different parameter values for the allometric scaling equations, or perhaps predict completely different equations. Two different indices of model fit were used in the comparisons, the coefficient of determination r2 ; and the Akaike’s Information Criterion (AIC, Akaike, 1973). The former simply states the proportion of variation in the response variable that can be explained by the regression equation. The latter is a more abstract measure which describes the likelihood that the data fits a specific model. It is defined as AICi ¼ 2 log Li þ 2Vi ; where Li is the maximum likelihood for the candidate model i; and Vi is the number of parameters estimated from the data for the model i: AIC belongs to a larger group of information criteria that can all be used as criteria for selecting the best from among the set of alternative models (Zucchini, 2000). The model that exhibits the lowest absolute value of AIC is judged the best, but if the difference in the AIC values is low for the alternative models then the alternatives also deserve serious consideration. A rough rule of thumb is that models with a difference less than two have substantial support, and models with a difference larger than ten might be omitted from further consideration (Burnham and Anderson, 2002, p. 446). The requirement is that the alternative models are compared with an identical set of data. The number of parameters in the different models may vary, but increasing their number adds a penalty to the value of AIC. In the present study, the number of parameters was constant for all models. To obtain r2 ; the values predicted by different parameterizations were correlated with the observed values. To obtain AIC for the fit, the predicted values were regressed against the observed values with no intercept term in the procedure REG. This procedure checked the deviation of the match between the variables from the perfect 1:1 relationship. In addition to the above criteria, I calculated the 95% confidence limits for the scaling exponent a to get an indication on what the conclusions would have been, if a statistical criterion had been used to compare the exponents in different models. 4. Results and discussion The four parameterization methods produced quite variable parameter values (Table 1). When the results of the purely empirical methods were compared with each other, the values of the scaling exponent, which are of interest here, showed up to 1.5-fold differences. There were also noticeable differences in the values of r2 ; although they generally indicated a reasonably good fit in all cases. What is most important, is that none of the methods was able to consistently produce the best fit. The largest values of r2 and the smallest values of AIC were not associated with each other, or with a particular parameterization method. This shows that the final conclusions may depend on the choice of method, if the researcher decides to use just one of the alternatives. The confidence limits of the scaling exponents, in turn, indicated that in more than half of the cases the theoretical value was not included within the limits, even when it provided an acceptable fit. For example, as ARTICLE IN PRESS P. Kaitaniemi / Journal of Theoretical Biology 228 (2004) 149–153 151 Table 1 Parameters and fit statistics of allometric scaling equations fitted to data by Kaitaniemi and Ruohom.aki (2003; Figs. 9 and 10). Four different linear or nonlinear regression methods were used to fit the parameters of equations for different branching orders. D=diameter, M=dry mass, N=shoot number, H=tree height. OLS refers to ordinary least-squares regression and RMA to reduced major axis regression. Value of a is represented with 95% confidence limits. r2 is coefficient of determination and AIC is Akaike’s Information Criterion a 95% CL for a r2 a 3.13 2.89 2.74 2.45 0.38 0.39 0.39 0.41 — 0.34–0.44 0.36–0.42 0.38–0.45 0.87 0.87 0.87 0.87 235.0 233.1 225.8 229.7 Theoretical, nonlinear Empirical, nonlinear Empirical, linear, OLS Empirical, linear, RMA 1.84 1.90 2.24 1.85 0.38 0.37 0.30 0.38 — 0.35–0.39 0.28–0.33 0.35–0.40 0.72 0.72 0.71 0.72 800.0 785.9 685.3 804.9 0 0 0 0 Theoretical, nonlinear Empirical, nonlinear Empirical, linear, OLS Empirical, linear, RMA 0.33 4.73 1.47 0.99 2.00 1.31 1.61 1.73 — 1.05–1.57 1.45–1.77 1.56–1.90 0.67 0.73 0.71 0.70 721.6 701.9 703.0 716.4 (d) N ¼ bDa >0 >0 >0 >0 Theoretical, nonlinear Empirical, nonlinear Empirical, linear, OLS Empirical, linear, RMA 1.58 8.09 3.51 1.42 2.00 1.43 1.68 2.23 — 1.31–1.55 1.54–1.81 2.05–2.41 0.60 0.61 0.61 0.59 4128.0 4100.5 3986.0 4663.5 (e) N ¼ bM a 0 0 0 0 Theoretical, nonlinear Empirical, nonlinear Empirical, linear, OLS Empirical, linear, RMA 3.79 2.54 4.96 4.66 0.75 0.81 0.70 0.72 — 0.75–0.86 0.67–0.74 0.68–0.75 0.95 0.96 0.95 0.95 584.2 583.2 582.3 588.7 (f) N ¼ bM a >0 >0 >0 >0 Theoretical, nonlinear Empirical, nonlinear Empirical, linear, OLS Empirical, linear, RMA 7.69 9.00 5.80 5.58 0.75 0.72 0.82 0.84 — 0.71–0.74 0.81–0.84 0.82–0.85 0.96 0.96 0.95 0.95 3227.9 3207.0 3429.8 3536.7 (g) H ¼ bM a 0 0 0 0 Theoretical, nonlinear Empirical, nonlinear Empirical, linear, OLS Empirical, linear, RMA 69.00 40.58 33.27 29.15 0.25 0.22 0.25 0.27 — 0.19–0.25 0.22–0.28 0.24–0.30 0.81 0.81 0.84 0.80 522.2 418.5 421.5 427.2 (h) H ¼ bDa 0 0 0 0 Theoretical, nonlinear Empirical, nonlinear Empirical, linear, OLS Empirical, linear, RMA 33.48 33.24 22.71 17.07 0.25 0.46 0.56 0.65 — 0.43–0.49 0.54–0.59 0.61–0.68 0.72 0.70 0.69 0.67 2632.4 3206.0 3228.6 3363.1 Equation Branching order Method 0 0 0 0 Theoretical, nonlinear Empirical, nonlinear Empirical, linear, OLS Empirical, linear, RMA (b) D ¼ bM a >0 >0 >0 >0 (c) N ¼ bDa (a) D ¼ bM a a b AIC This table gives r2 based on correlation coefficient whereas Kaitaniemi and Ruohom.aki (2003) give the values of adjusted R2 : regards the relationship between the main stem diameter and tree height (Table 1(h)), the theoretical value 0.25 was outside the 95% confidence limits for all the empirically fitted scaling exponents, yet it provided the best fit on the basis of AIC and r2 : As regards the relationship between branch diameter and dry mass for higher order branches (Table 1(b)), the theoretical value would have been accepted if r2 was the criterion, but rejected if AIC was the criterion. In addition to differences in parameter values, the values predicted by the corresponding scaling equations showed considerable deviation from each other depending on the value range of variables (Fig. 1). This indicates the potential size of error as regards the predictive use of equations. On the basis of this exercise alone, it is not possible to make any generalizations regarding the choice of best parameterization method, or the best criteria for evaluating models with alternative parameterizations. In any case, since the scaling laws are essentially models that produce predictions (West et al., 1999a), I would prefer the use of AIC or other model validation techniques for selecting the best parameter values (Haefner, 1996; Zucchini, 2000), instead of making simple statistical comparisons of a against the theoretical value. For the same reason, I would also reject the ARTICLE IN PRESS P. Kaitaniemi / Journal of Theoretical Biology 228 (2004) 149–153 152 (a) 100 100 10 10 1 1 0.1 0.1 1 10 100 1000 (b) 0.1 0.1 1000000 10000000 100000 1000000 10000 100000 1000 10000 1 10 100 1000 0.1 1 10 100 1000 0.1 1 10 100 1000 1000 100 100 10 10 1 1 0.1 0.1 0.01 0.01 0.001 0.1 (c) 0.001 1 10 100 1000 (d) 1000 10000 1000 100 100 10 10 1 (e) 1 0.1 0.1 0.1 1 10 100 1000 (f) 10000 1000 1000 100 100 10 10 1 1 (g) 0.1 1 10 100 1000 (h) 0.1 0.1 Theoretical, nonlinear Empirical, nonlinear Empirical, linear, OLS Empirical, linear, RMA 1 Fig. 1. Allometric scaling equations of Table 1 plotted on a logarithmic scale. 10 100 1000 ARTICLE IN PRESS P. Kaitaniemi / Journal of Theoretical Biology 228 (2004) 149–153 use of RMA regression, which is better suited to the examination of the scaling exponent a alone (Sokal and Rohlf, 1981; Niklas, 1994), and only when there is no information about the predictive relationship between the variables (Isobe et al., 1990). To search the values of a and b that produce a model with the best fit, a safe procedure is probably a thorough experimentation with a sufficiently wide range of parameter values during model fitting. The use of efficient parameter estimation methods, such as evolutionary algorithms, would also be preferable. As long as these are not readily available, I would recommend, as a precaution, the simultaneous use of the linear and nonlinear methods presented in this paper, and perhaps using additional, biological criteria for selecting the range of data where the quality of predictions is particularly important. A technique called model averaging can also be applied to base the inferences on the entire set of models instead of a single best model (Burnham and Anderson, 2002). It was clear that, if we stick without good justification to one biased estimate only, it is likely that the conclusions will be misleading, and do not help in developing the theories further. Acknowledgements The study was funded by the Academy of Finland grants 44141 and 201997. I thank K.J. Niklas for helpful comments on the previous version of the manuscript. 153 References Akaike, H., 1973. Information theory and an extension of the maximum likelihood principle. In: Petrov, B.N., Caski, F. (Eds.), Proceeding of the Second International Symposium on Information Theory. Akademia Kiado, Budapest. Burnham, K.P., Anderson, D.R., 2002. Model Selection and Multimodel Inference: A Practical Information-theoretic Approach. Springer, New York. Dodds, P.S., Rothman, D.H., Weitz, J.S., 2001. Re-examination of the ‘‘3/4-law’’ of metabolism. J. Theor. Biol. 209, 9–27. Enquist, B.J., 2002. Universal scaling in tree and vascular plant allometry: toward a general quantitative theory linking plant form and function from cells to ecosystems. Tree Physiol. 22, 1045–1064. Haefner, J.W., 1996. Modeling Biological Systems. Chapman & Hall, New York. Isobe, T., Feigelson, E.D., Akritas, M.G., Babu, G.J., 1990. Linear regression in astronomy. I. Astrophys. J. 364, 104–113. Kaitaniemi, P., Ruohom.aki, K., 2003. Factors controlling resource allocation in mountain birch. Perspect. Plant Ecol. Evol. Syst. 5, 231–249. Niklas, K.J., 1994. Plant Allometry: The Scaling of Form and Process. The University of Chicago Press, Chicago. Sokal, R.R., Rohlf, F.J., 1981. Biometry 2nd Edition. W.H. Freeman and Company, New York. West, G.B., Brown, J.H., Enquist, B.J., 1999a. A general model for the structure and allometry of plant vascular systems. Nature 400, 664–667. West, G.B., Brown, J.H., Enquist, B.J., 1999b. The fourth dimension of life: fractal geometry and allometric scaling of organisms. Science 284, 1677–1679. Zucchini, W., 2000. An introduction to model selection. J. Math. Psychol. 44, 41–61.
© Copyright 2026 Paperzz