To understand how organisms adapt, researchers must link performance and microhabitat. However, measuring performance, especially maximum performance, can sometimes be difficult. Here, we describe an improvement over previous techniques that only consider the largest observed values as maxima. Instead, we model expected performance observations via the Weibull distribution, a statistical approach that reduces the impact of rare observations. After calculating group-level weighted averages and variances by treating individuals separately to reduce pseudoreplication, our approach resulted in high statistical power despite small sample sizes. We fitted lizard adhesive performance and bite force data to the Weibull distribution and found that it closely estimated maximum performance in both cases, illustrating the generality of our approach. Using the Weibull distribution to estimate observed performance greatly improves upon previous techniques by facilitating power analyses and error estimations around robustly estimated maximum values.
Studies of ecological morphology and evolution often link an organism's morphology, performance and ecology to suggest adaptation (Wainwright and Reilly, 1994). These studies typically assume accurate measurements of performance, which can be difficult to obtain, sometimes with a few observations contributing to a final point estimate of performance, especially when maximum performance values are of interest (Anderson et al., 2008; Garland and Losos, 1994). For the initial approval of projects by animal use and collection review boards, it is also necessary to plan for the number of individuals to be tested, the testing procedure for each individual and the resulting statistical power of the results. Power analyses can describe the number of trials and individuals needed, but require an explicit statistical distribution of results, which is difficult to estimate when quantifying maximum performance. Here, we describe a modeling approach using the Weibull distribution. In this approach, maximum performance is estimated as a parameter of the Weibull distribution. This approach is more robust than current methods of quantifying performance, which consider only a subset of maximum observations. The Weibull approach minimizes the effect of rare events and allows the use of additional statistical tools.
The Weibull distribution is valid for data that are likely to be non-normal, such as estimates of a maximum value, which would be expected to produce a skewed distribution of observations. While the Weibull distribution has been used occasionally to model behavior (Britten et al., 1992; Davis, 1996; Pugno and Lepore, 2008; Simpson and Ludlow, 1986), it is most often used in material science to predict mechanical failure. In this scenario, the Weibull distribution is modeling the likelihood of an observed event (i.e. mechanical failure) as it relates to some other factor (i.e. applied force or time), but this does not need to be the case. The Weibull distribution is highly flexible and can model many different patterns (Cornwell and Weedon, 2014; McCool, 2012; Weibull, 1951; Yang and Xie, 2003). This differs from the exponential distribution, which assumes a constant event rate and the normal distribution, which can only model non-skewed data. The Weibull distribution includes a scale parameter (λ) and a shape parameter (k, also known as the Weibull modulus). The scale parameter (λ) has the same units as the modeled data and dictates the distribution's location on the x-axis. λ is closely related to the distribution's mode when k>1 (Fig. 1). The dimensionless shape parameter (k) loosely controls the width and shape of the distribution for a given λ (Fig. 1). With these parameters, the Weibull distribution can model left and right skewed distributions (when k>1) as well as exponential distributions (when k≤1). Here, we use the λ parameter as our performance value estimate.
To investigate the use of the Weibull distribution, we quantified the adhesive performance of gecko lizards by measuring toe detachment angle (Fig. 2; Autumn et al., 2006; Hagey et al., 2014). We first compared the fit of our observations to multiple distributions. We then considered how to model multiple observations from multiple individuals to avoid pseudoreplication and produce group-level estimates. We also conducted a power analysis to identify the minimum number of trials per individual and number of individuals needed to detect differences between groups. Lastly, we explored the generality of our approach by using the Weibull distribution to analyze previously published lizard bite force data.
MATERIALS AND METHODS
To evaluate the use of the Weibull distribution, we used two empirical toe detachment datasets from the lizard Gekko gecko Linnaeus 1758. We used a multi-individual dataset that included 206 observations from 13 individuals with an overall average of 15.4 trials per individual (max: 40 trials per individual; min: four trials per individual). Our second dataset, a subset of the first, comprised 40 observations from a single individual. To quantify toe detachment angle, we suspended live lizards from a single rear center toe (Schulte et al., 2004; Wang et al., 2010; Zani, 2000) from a slowly inverting glass microscope slide, recording the surface angle at which the lizard spontaneously detaches (Fig. 2; Autumn et al., 2006; Emerson, 1991; Hagey et al., 2014). As the glass surface inverts, the probability of detachment increases, making angle of toe detachment an excellent assay to be modeled by the Weibull distribution.
We preliminarily compared the fit of the normal, Weibull, exponential, gamma and log normal distributions qualitatively using Q–Q plots (Fig. 3). Previous studies have shown that the Weibull distribution can sometimes be difficult to distinguish from other distributions (Bain and Engelhardt, 1980; Fearn and Nebenzahl, 1991). We also compared the fit of these distributions to our data employing a bootstrapped Kolmogorov–Smirnov (K–S) test with 5000 bootstraps to alleviate issues with repeated values (Table 1; Sekhon, 2011). The K–S test evaluates the probability that our observed values could have been drawn from each considered distribution. We also compared Akaike information criterion weights with a correction for small sample size (AICc; Burnham and Anderson, 2002).
To evaluate multiple observations from multiple individual lizards, we needed to determine how variable our observed individuals were and whether they could be modeled together under one distribution. We fitted the Weibull distribution to our multi-individual dataset (our null model, all observations lumped together) or allowed each individual lizard to have its own set of parameters (our alternative model). We then conducted a likelihood ratio test and used the chi-squared distribution to determine if we should reject our null model. Finally, we fitted a parametric survival regression model using the Weibull distribution with individual as a factor, predicting detachment angle using the R library ‘survival’ (Therneau and Grambsch, 2000).
It is worth noting that when the individual errors (σi2) are all equal, Eqn 3 simplifies to the standard error around an unweighted mean.
We then produced multiple datasets of simulated toe detachment data using the Weibull distribution. Each dataset was composed of simulated data from two groups with different assigned λ parameter values to see if we could significantly differentiate small differences between groups. The differences between groups were one, two, three or five degrees (i.e. our effect sizes). We chose to compare groups with λ values of 15 deg vs 15 deg plus an effect size and 25 deg vs 25 deg plus an effect size based on empirical observations. We also evaluated comparisons in which groups had 20 trials per group distributed across one individual, two individuals and four individuals or 50 trials per group distributed across two individuals, five individuals and ten individuals (Fig. 4). To statistically compare our estimated weighted averages between simulated groups, we tested whether either of the groups' means were within 1.96 standard deviations (square root of the estimated mean variance multiplied by 1.96) of the other. We then calculated the percentage of our 1000 replicate trials that produced significantly different comparisons (Fig. 4). All analyses were conducted using RStudio statistical software (v.0.98.501; https://cran.rstudio.com).
RESULTS AND DISCUSSION
We qualitatively compared our data with the normal, Weibull, exponential, gamma and log normal distributions (Fig. 3). Our Q–Q plots strongly suggest that our data differ from what would be expected under the exponential distribution because high angle observations were rare. The fitted normal, gamma and log normal distributions all yielded similar results, suggesting that low-angle observations were more common than if our data were drawn from these distributions, i.e. our data are left-skewed, which is to be expected when attempting to observe a maximum value. The Q–Q plots suggest that the Weibull distribution closely approximates our observed data. An important distinction between the Weibull and normal distributions is that the Weibull predicts rare early failure events, producing a negatively skewed distribution of observations, whereas the normal distribution does not. We also evaluated the goodness of fit for normal, exponential, Weibull, gamma and log normal distributions using a bootstrap version of the Kolmogorov–Smirnov test (Sekhon, 2011). Our multi-individual dataset was significantly different from all the distributions considered, whereas our individual dataset was not significantly different to any of the distributions considered, except the exponential distribution (Table 1). These results might be due to the fact that our multi-individual dataset represents a collection of observations from multiple individuals, each with their own performance distribution (see below). Conversely, our individual dataset might be too small to distinguish between distributions (see the Materials and methods). AICc weights, corrected for small sample sizes (Burnham and Anderson, 2002) provided more clear results, suggesting the Weibull distribution is the best fitting distribution considered with an AICc weight of 0.94 and 0.81 for our multi-individual and individual datasets, respectively (Table 1).
To estimate a group-level detachment angle, we needed to consider multiple observations from multiple individuals, while also limiting pseudoreplication. We found significant support to reject our null hypothesis (pooling all our observations, D=100.3, d.f.=24, P<0.0001). We also fitted a parametric survival regression model using the Weibull distribution in which individual, as a factor, significantly predicted detachment angle (P<0.0001), again suggesting it is better to model performance observations of each individual lizard separately (Therneau and Grambsch, 2000). By fitting each individual to a distribution and using each individual's parameter values and error to estimate weighted group averages, we prevent pseudoreplication and reduce the impact of unbalanced sampling, penalizing individuals with large error estimates due to small sample sizes or erratic observations.
We also investigated the trade-off between the number of trials per individual and the number of individuals tested using power analyses. We found abundant power to detect small differences between groups (effect sizes of one to two degrees) even with relatively small datasets, i.e. few total trials, few individuals, or few trials per individual (Fig. 4). In our simulations using 50 observations, we observed an overall increase in power, as to be expected with more data, regardless of partitioning. Fitting the Weibull distribution to datasets with five or fewer observations per individual, especially if there is little variation among the observations, can hinder the estimation of parameter estimates. Although our power analyses suggested we could detect group differences with few trials or individuals, we recommend six to ten trials per individual to assure a successful fit of the Weibull distribution. Using data from two individual lizards with 10 trials each, we consistently had enough power to detect a difference of 1 deg 80% of the time, although we recommend sampling more individuals to better capture inter-individual variation.
In addition, we did not observe differences in power between datasets considering low detachment angles (15 deg) and high detachment angles (25 deg). We believe this is due to the fact that we simulated our data using values of k that complemented λ and the number of trials per individual (see Materials and methods). We conducted similar analyses with a constant k value regardless of λ and number of trials per individual and found that power decreased when considering higher values of λ. This is because at higher λ values, with a constant k, the Weibull distribution has a larger variance (Fig. 1), reducing the power to detect differences between groups.
To investigate the generality of the ability of the Weibull distribution to model animal performance, we also compiled published bite force data from Anolis carolinensis (provided by A. Herrel; Irschick et al., 2005a,,b). This dataset comprised 381 individuals, with 5 trials for each. We found that an individual's estimated λ values were similar to the mean of the best three trials with the added benefit of being robust to extreme values (Fig. 5).
In conclusion, using the Weibull distribution as an expected distribution of observed performance is a more robust approach compared with use of only a subset of trials. We strongly encourage researchers to investigate the fit of the Weibull distribution to their performance datasets. With the addition of estimating weighted group means, comparisons between treatment and control groups, or comparisons between species are easily facilitated. If making comparisons across species, once species means have been estimated, phylogenetic non-independence can then easily be taken into account.
We thank our anonymous reviewers for their helpful comments, Anthony Herrel for providing published bite force data, Matt Pennell for helpful discussions, Jon Boone for access to animals, Bobby Espinoza, J. R. Wood, Jesse Grismer and Mat Vickers for field logistics, and Andrew Schnell, Jorn Cheney, Scott Harte, Jonathan Losos, Anthony Herrel, Shane Campbell-Staton, Kristi Fenstermacher, Hannah Frank, Martha Munoz and Paul VanMiddlesworth for help collecting data in the lab and field.
All authors contributed significantly in reviewing and editing the manuscript. T.J.H. analyzed the data and drafted the manuscript. J.B.P. initially conceived the use of the Weibull distribution and provided relevant input in its implication. K.E.C. collected the G. gecko performance data. K.A. supervised the live animal experiments. L.J.H. provided guidance in the analyses.
This work was funded by the National Science Foundation [0844523 to Aaron Bauer and Todd Jackman, NSF-IOS-0847953 and NSF Special Creativity Award to K.A. and NSF DEB 1208912 to L.J.H.]; the National Geographic Society Waitts Institute [W216-12 to T.J.H.]; the BEACON Center for the Study of Evolution in Action [Request 302, 429 to T.J.H.]; and Sigma XI [G200803150489 to K.E.C.].
The authors declare no competing or financial interests.