Traditional allometric analysis fails to provide a valid predictive model for mammalian metabolic rates

The field of biological allometry was energized by the publication in 1997 of a theoretical model purporting to explain 3/4-power scaling of metabolic rate with body mass in mammals (West et al., 1997). The model generated a wave of new interest in the discipline (Agutter and Wheatley,2004; Glazier,2005; da Silva et al.,2006), and it also re-opened the long-simmering debate about the`true' value for the scaling factor in the allometric equation(Hoppeler and Weibel, 2005). However, this debate about the value for the scaling exponent may be premature, because investigators on all sides of the issue have for years unknowingly used an unreliable procedure to estimate parameters in the equation (see below). This problem has gone largely unrecognized because allometric equations typically have not been validated in the scale of measurement.

The aforementioned (`traditional') approach to allometric analysis has not changed appreciably since the time of Kleiber, Benedict and Brody(Kleiber, 1932; Benedict, 1938; Brody, 1945). The approach nonetheless is beset by a variety of problems, most of which result from the use of logarithmic transformations. First, transformation profoundly alters the relationship between predictor and response variables(Emerson and Stoto, 1983; Jansson, 1985; Osborne, 2002), so influential outliers may go undetected, remain in the data set, and bias parameter estimates in the fitted statistical model(Packard and Boardman, 2008a). Second, the two-parameter power function (Eqn 1) underlying the traditional allometric analysis may not provide a good fit to the data (Zar,1968; Albrecht and Gelvin,1987; Albrecht,1988; Packard and Boardman,2008a; Packard and Boardman,2008c), in which case parameter estimates again may be inaccurate and misleading. Third, the statistical model obtained by back-transformation from logarithms is one that predicts geometric means for Y instead of arithmetic means (Miller,1984; Smith, 1993; Hayes and Shonkwiler, 2006). And fourth, a straight line fitted to logged values may undergo distortional rotation owing to the fact that squared residuals are not equivalent for large and small values of the original response variable(Zar, 1968; Jansson, 1985; McCuen et al., 1990; Pandy and Nguyen, 1999; Packard and Boardman,2008b).

We re-examined data for the basal metabolic rate (BMR) of 626 species of mammals (Savage et al., 2004)to illustrate how applying the traditional method for allometric analysis can result in biased and misleading estimates for parameters in a two-parameter allometric equation. Species represented in this comprehensive sample varied in size from a 2.4 g shrew to a 3672 kg elephant. We focused on a modified data set created by binning logged values for body mass [Appendix 2 in the study by Savage and colleagues (Savage et al., 2004)], because the resulting estimate of 0.737 for the allometric exponent is regarded by many workers as providing strong support for the concept of 3/4-power scaling (Brown et al., 2004; Farrell-Gray and Gotelli, 2005; West and Brown,2005). The dimension for each of the bins was 0.1 in logged units for body mass, and each bin yielded one representative (`average') value irrespective of the number of species assigned to it. Binning was undertaken by Savage and colleagues (Savage et al.,2004) to prevent the preponderance of small species in the full sample from exerting undue influence on estimates for parameters in the allometric equation; but the use of binned values also facilitates graphical analysis by avoiding the visual clutter that would accompany the display of more than 600 values in a single plot. Preliminary examination of values for the full 626 species, coupled with the results of an independent study of the same data set (Hui and Jackson,2007), indicates that none of our conclusions was affected by using binned values.

We do not address in our study subsidiary issues like phylogenetic independence of measurements for different species(Garland et al., 2005) or assumptions of least-squares regression(Warton et al., 2006). Instead, our treatment follows the same general approach that was used by Savage and colleagues (Savage et al.,2004), thereby enabling us to make detailed comparisons of their findings with our own. We carried all calculations to six decimal places before rounding to three.

Data for 52 samples expressed in both logarithmic and arithmetic scales first were displayed in bivariate scatterplots and examined for patterns and trends (Anscombe, 1973). This step was followed by preliminary statistical analyses(Packard and Boardman, 2008a)wherein a straight line was fitted to logarithmic transformations by ordinary least squares and a two-parameter power function was fitted to values in the arithmetic scale by non-linear regression (Motulsky and Christopoulos,2004). Computations were performed in SigmaPlot (version 10.0 from Systat Software, Inc., San Jose, CA, USA), which uses the Marquardt–Levenberg algorithm to fit non-linear functions by an iterative process that minimizes the sum of squares for residuals(Marquardt, 1963). Allometric exponents estimated by the two procedures were quite different: 0.737 from the slope of the straight line fitted to logarithms (R²=0.986)and 0.909 from the two-parameter power function fitted to values in the arithmetic scale (R²=0.997). However, logarithmic transformation failed to linearize the data (e.g. a plot of residuals for logBMR against predictions yielded a pattern in the form of an inverted parabola) (see also Kozlowski and Konarzewski, 2005), and the equation resulting from back-transformation seriously underestimated metabolic rate for the elephant,the largest species in the sample. In addition, neither of the analyses passed the test for constancy of variances (Spearman rank correlation between absolute values for residuals and observed values for the predictor variable)(Kutner et al., 2004).

The unbalanced distribution of arithmetic values for body mass was also of concern. The elephant (which was the only species represented in that bin) was nearly an order of magnitude heavier than the next species in the sample. This observation, coupled with the decidedly different estimates for allometric exponents, raised the possibility that the elephant was an unduly influential outlier in the non-linear regression(Anscombe, 1973; Stevens, 1984; Osborne and Overbay, 2004). We subsequently discovered that Cook's Distance, which is a sensitive measure of the influence of a data point on parameters in the fitted model(Kutner et al., 2004), was an extraordinary 4600 for the elephant. Any data point for which Cook's Distance exceeds 4 is likely to exert undue influence, so we treated the elephant as a statistical outlier and removed it from the data set.

A straight line fitted to the remaining 51 logged values yielded a statistically significant equation (Fig. 1A), even though a plot of residuals against predicted values indicated once again that a straight line was not an appropriate model(Fig. 1B). Nonetheless, R² was extraordinarily high(Fig. 1A), and the estimate of 0.728 for the allometric exponent is similar to the estimate from examination of the full data set (i.e. exclusion of the elephant had little effect on the outcome). The analysis passed the test for normality (P=0.288 by Kolmogorov–Smirnov test) (Kutner et al., 2004) but it failed the one for constancy of variances(P=0.002). Consequently, confidence limits for the slope and intercept are unlikely to be reliable(Myers, 1986; Finney, 1989).

A plot of data in the original scale shows an unbalanced distribution for body mass, but the imbalance is not as extreme as it was when the elephant was included in the sample (Fig. 1C). A two-parameter power function fitted to the data by non-linear regression yielded an estimate of 0.686 for the allometric exponent(Fig. 1C). Although the model fitted by non-linear regression was statistically significant and R² was high (Fig. 1C), the analysis failed tests for both normality of residuals and homogeneity of variances (P<0.001). A plot of residuals shows that variances were related to body mass in a manner that commonly is associated with multiplicative error (Fig. 1D). Thus, reliable confidence limits cannot be computed for the parameter estimates a and b(Myers, 1986; Finney, 1989). However, the parameter estimates themselves may be better than those obtained by back-transformation (Myers,1986), and the predictive equation is efficient in terms of minimizing residual variance (Asselman,2000).

Next, we back-transformed the equation for the line fitted to logarithms and displayed the resulting function on bivariate graphs together with the function obtained by non-linear regression(Fig. 2A,B). The line from back-transformation is a good descriptor for values in the logarithmic scale(Fig. 2A), but it fails to predict values for large animals in the arithmetic scale(Fig. 2B). In contrast, the non-linear regression predicts consistently higher values for the response variable in the logarithmic domain (Fig. 2A) while performing much better than the alternative model for large animals in the arithmetic scale (Fig. 2B).

Because of the unbalanced distribution for the predictor variable in the arithmetic scale (Fig. 2B), we displayed arithmetic values for BMR against body mass on a log scale in order to better visualize goodness of fits to data for smaller species(Fig. 2C). This display facilitated examination of BMR by expanding the apparent distribution for body mass at the low end of the scale without introducing transformation bias(Finney, 1989). The equation from back-transformation is the better predictor of BMR for animals between 10² and 10⁴ g in body mass(Fig. 2C, inset), but clearly is the inferior function for predicting BMR of larger mammals(Fig. 2C). Thus, the equation from back-transformation is useful for predicting metabolic rates of animals weighing less than 10⁴ g but the function is not valid over the full range in body size. The function fitted by non-linear regression is a better descriptor over the entire range in body size, even though it is more biased than the function from back-transformation in the mid-range for body size. (Part of the bias in the function fitted by non-linear regression is the result of forcing the line to pass through the origin and could be eliminated by fitting a three-parameter function instead. However, the scaling exponent in the three-parameter function differs only slightly from that for the two-parameter function, i.e. 0.667 vs 0.686, respectively.)

The scaling exponent for small species is predicted to be smaller than that for large species (Savage et al.,2004) – a prediction that seems to be confirmed by the observed curvilinearity (concave upward) in the relationship between logBMR and log body mass (Fig. 1A,B). Such a curvilinear pattern of variation in log-transformed data should be cause for concern, because it calls into question the underlying allometric model (Eqn 1). Nevertheless, we examined data for large and small mammals separately to see whether the aforementioned prediction was realized. Binned data for mammals weighing less than 260 g were taken to represent small species whereas those for mammals weighing more than 260 g were taken to represent large ones(Savage et al., 2004). The elephant was omitted from the analyses because of the likelihood that it is an outlier.

Straight lines fitted to transformed values for both small and large species yielded statistically significant equations with high values for R² (Fig. 3A,B). The analysis of values for small species passed tests for normality (P=0.065) and homogeneity of variances (P=0.059). The analysis for large species, however, passed the test for normality(P=0.298) but not that for homoscedasticity (P=0.016). Scaling exponents for small and large species were 0.678 and 0.797,respectively, which seemingly confirmed expectation(Savage et al., 2004).

Non-linear regression on the two sets of values in the original scale also yielded statistically significant fits and high values for R² (Fig. 3C,D). The analysis on the subset of values for small species met assumptions for both normality (P=0.922) and homogeneity of variances(P=0.381), whereas the treatment of large animals failed the test for normality (P=0.008) as well as that for homoscedasticity(P<0.001). Scaling exponents were estimated to be 0.656 and 0.686,respectively, for small and large species.

The alternative methods for fitting the allometric equation to data for small mammals yielded functions that are reasonably good visual fits to values in the original scale (Fig. 3E). The non-linear function fitted to values for large mammals is also a good fit graphically (Fig. 3F) but the equation obtained by back-transformation seriously underestimates BMR for species with masses between 75 and 150kg(Fig. 3F).

Scaling exponents estimated for large and small species by the traditional method are quite different (Fig. 3A,B) whereas exponents estimated by non-linear regression are quite similar (Fig. 3C,D). Indeed, the 95% confidence interval for the exponent estimated by non-linear regression for small mammals (i.e. the group for which such limits can be reliably computed) is 0.615–0.697, which includes the exponent estimated for large mammals (Fig. 3D) as well as the one for all species exclusive of the elephant(Fig. 1C). Thus, analyses of different subsets of the data by non-linear regression lead to a common estimate for a scaling exponent (in the range 0.656–0.686) – not to the different exponents predicted by the aforementioned theoretical model(Savage et al., 2004).

Why do the two-parameter functions estimated by back-transformation differ so much from those estimated by non-linear regression? We suggest that the disparity in this instance is largely the result of fitting straight lines to logarithmic transformations, which are inherently non-linear and which consequently change the relationship between predictor and response variables(Emerson and Stoto, 1983; Jansson, 1985; Osborne, 2002). Logarithmic transformation results in an overall compression of distributions for the variables, but the compression is greater at the high ends of the scales than at the low ends (Emerson and Stoto,1983; Jansson,1985; Osborne,2002). This disproportionate compression causes small values for the variables to have a large effect on parameters in the fitted equation and large values to have a small effect(Glass, 1969; Jansson, 1985; Packard and Boardman,2008b).

By way of example (Jansson,1985), consider a straight line fitted to logarithms of 0.9 and 1.1 at one level for X and to 1.9 and 2.1 at a higher level for X. Predictions for logs of the response variable Y are 1.0 and 2.0, respectively, with all residuals having absolute values of 0.1. Such a balanced distribution of residuals indicates that logged values for Y were weighted equally in fitting the line by ordinary least squares.

Back-transformation of the predicted values yields a geometric (not arithmetic) mean of 10 at the first level for X and 100 at the second level for X. Observed values corresponding to the prediction of 10 are 7.9 and 12.6 whereas those corresponding to the prediction of 100 are 79.4 and 125.9. Thus, absolute values for residuals expressed on the scale of measurement are 2.1, 2.6, 20.6 and 25.9, despite the fact that all residuals were ±0.1 in the log domain. This difference between the scales is important because the fitted line minimizes the sum of the squared residuals regardless of the scale in which the data are expressed(Zar, 1968). Whereas the squares for the residuals in the log domain are identical (i.e. 0.01), the square for the largest of the values in the arithmetic domain (i.e. 670.8) is more than two orders of magnitude larger than that for the smallest value(i.e. 4.4).

At each level for X, the smaller of the two measurements lies below the fitted line and the larger one lies above it. Consequently, the smaller values in the arithmetic scale have a disproportionate influence on the elevation of the line fitted to logarithms because residuals for large and small values are identical in the logarithmic domain. Additionally, the smaller of the two measurements lying above (or below) the fitted line is associated with the lower level for X and the larger with the higher level. Thus, the smaller value in the arithmetic scale has a disproportionate influence on the slope of the line fitted to logarithms (again, because residuals are identical in the log domain).

Depending on the distributions of the variables, both the slope and intercept of the straight line may be affected in unexpected ways(Glass, 1969; Jansson, 1985; McCuen et al., 1990; Pandy and Nguyen, 1999; Packard and Boardman,2008b), and these effects later are transmitted by back-transformation to the two-parameter allometric equation. This disparate influence of small and large values is apparent in the current study in graphs of the alternative equations in both logarithmic and arithmetic scales(Fig. 2). The linear regression on transformed values was rotated in a counter-clockwise direction(Fig. 2A), and the result was a poor fit of the back-transformation to data for large animals(Fig. 2B). The general problem outlined here probably occurs commonly in data sets that include animals spanning large ranges in size (Glass,1969; Packard and Boardman,2008b): the traditional procedure provides good predictions for small animals but poor predictions for large ones.

Savage and colleagues used the binning procedure in an attempt to reduce the disproportionate influence of the many small species in the sample and thereby obtain a more reliable estimate for the scaling exponent(Savage et al., 2004). On the other hand, Glazier argued that binning actually caused values for large species to have too great an influence on parameters in the allometric equation owing to an increase in proportional representation for large species(Glazier, 2008). Both these suggestions, however, are based on a misunderstanding of the logarithmic transformation. First, binning by logs for body mass had the effect of expanding the scale at the lower end of the distribution and compressing it at the upper end, thereby maintaining a skew in the distribution of masses expressed in grams and causing small species to be over-represented in the 51 bins exclusive of the elephant. For example, the arithmetic mean for 51 back-transformed values for body mass is 31,364 g. A total of 42 bins (of which one was empty) were available to accommodate species with masses below the average, and 12 (of which two were empty) were available to accommodate species with masses above the average. Small species consequently continued to be `over-represented' in the data set. Second, the line fitted to logarithms was `transformation biased' by the undue influence of the small species(Jansson, 1985; Packard and Boardman, 2008b),leading to rotation of the line, to underestimation of the allometric coefficient from back-transformation, and to overestimation of the allometric exponent (Fig. 2B). The minor influence of large species is why deletion of the elephant from the data set had little effect on the allometric exponent estimated by the traditional method.

Logarithmic transformations have a long history of use in allometric analyses, so it is useful here to consider briefly the reasons for performing such transformations and to ask whether the transformations continue to have application.

1. Logarithmic transformations were used for many years to linearize data and thereby promote graphical display and statistical analysis(Smith, 1984; Smith, 1993). This application of log transformations was based on the implicit assumption that the data conformed with a two-parameter power function(Smith, 1984). The assumption was seldom addressed, and inappropriate equations were sometimes fitted(Albrecht and Gelvin, 1987; Albrecht, 1988; Packard and Boardman, 2008a; Packard and Boardman, 2008c). However, the form of the underlying equation now is moot because the advent of sophisticated PCs and software for both graphical and statistical analysis has rendered linearization unnecessary(Motulsky and Christopoulos,2004).

2. Logarithmic transformations commonly produce symmetrical distributions for data that have unbalanced distributions in the original scale(Emerson and Stoto, 1983). The main problem with unbalanced distributions is that the largest animals in the sample may exert undue influence on parameters that are estimated by methods based on least squares. Indeed, the problem of imbalance was apparent in data examined here (Fig. 1C). Whereas transformation may produce a better distribution of values(Fig. 1A), it also introduces a bias that may be more serious than the problem that the transformation was intended to correct (Jansson,1985). The bias results from the change in error structure of the statistical model on back-transformation to the arithmetic scale(Glass, 1969; Manaster and Manaster, 1975; Miller, 1984; Jansson, 1985; Smith, 1993; Hayes and Shonkwiler, 2006)coupled with the distortion that is introduced by the non-linear relationship between logs and values in the original scale(Jansson, 1985; McCuen et al., 1990; Pandey and Nguyen, 1999; Hui and Jackson, 2007; Packard and Boardman, 2008b). A better way than transformation for addressing the problem of unbalanced distributions might be to make greater use of regression diagnostics and model validation (Anscombe, 1973; Snee, 1977; Kutner et al., 2004).

3. Allometric data often are heteroscedastic, and transformation may create a new distribution that meets the assumption of homogeneity of variances(Emerson and Stoto, 1983; Smith, 1984). Unfortunately,few authors examine the transformed variables for heteroscedasticity, so the resulting analysis may fail to correct the problem that led to transformation in the first place; the logged data examined here provide but one example of the failure of transformation to achieve its intended goal(Fig. 1B). Even when the original data are heteroscedastic and transformation results in a homoscedastic distribution, the potential exists for introducing a transformation bias that is evident only on back-transformation to the arithmetic scale. However, the problem of heteroscedasticity probably will become moot as more investigators discover statistical procedures that accommodate different assumptions concerning distributions of data in the original scale (Lane, 2002; Cox et al., 2008).

Regardless of the means by which an allometric equation is fitted to data,it is essential that the model be validated(Snee, 1977; Emerson and Stoto, 1983; Myers, 1986; Finney, 1989; Kutner et al., 2004). A graphical display is the most effective way to verify that the fitted model actually describes the data on which it is based(Anscombe, 1973). Unfortunately, validation in the traditional allometric analysis typically is limited to a display of values in the logarithmic scale, which often has little bearing on the relationship between metabolic rate and body mass(Fig. 2A,B). A good fit of a linear function to logarithms does not imply a good fit of the re-expressed equation to data in the arithmetic scale(McCuen et al., 1990; Pattyn and Van Huele, 1998). Thus, proper validation requires that the allometric equation be shown against the backdrop of data in the scale of measurement(Emerson and Stoto, 1983; Myers, 1986; Finney, 1989).

The data set compiled by Savage and colleagues is widely regarded to be one of the very best (Savage et al.,2004). It comes as no surprise, therefore, that the statistical analysis performed on those data by Savage and colleagues(Savage et al., 2004) is viewed by many in the scientific community as offering strong evidence in support of the concept of 3/4-power scaling for metabolic rate on body mass in mammals (Brown et al., 2004; Farrell-Gray and Gotelli,2005; West and Brown,2005). However, Savage and colleagues(Savage et al., 2004) seem to have omitted three critical steps from their investigation: they apparently did not (1) examine their data for potential outliers, (2) test assumptions underlying their statistical analysis, or (3) validate the allometric model in the original scale. Our re-analysis of data exclusive of the elephant, which was an apparent outlier, revealed that a linear equation fitted to log–log transformations failed tests for both linearity and constancy of variances, and that the two-parameter power function estimated by back-transformation did not predict metabolic rates of large animals in the sample. Consequently, the earlier estimate of 3/4-power scaling is not well supported, thereby calling into question the validity of theoretical models that purport to explain such a scaling factor in mammals (e.g. West et al., 1997; Banavar et al., 1999; Darveau et al., 2002).

The traditional approach to allometric research is to fit a straight line to logarithmic transformations and then back-transform the resulting equation to the arithmetic scale (Smith,1984; Agutter and Wheatley,2004; Glazier,2005; da Silva et al.,2006). Consequently, biases of the kinds shown in the current study and elsewhere (Glass,1969; Hui and Jackson,2007; Packard and Boardman,2008a; Packard and Boardman,2008b; Packard and Boardman,2008c) are likely to occur commonly in published research on allometry. For this reason, scaling coefficients and exponents reported in the literature should be interpreted with a healthy dose of skepticism.

We thank four anonymous referees for constructive criticism that helped us to improve this report.