Comparison of Bayesian and numerical optimization-based diet estimation on herbivorous zooplankton

Abstract

Consumer diet estimation with biotracer-based mixing models provides valuable information about trophic interactions and the dynamics of complex ecosystems. Here, we assessed the performance of four Bayesian and three numerical optimization-based diet estimation methods for estimating the diet composition of herbivorous zooplankton using consumer fatty acid (FA) profiles and resource library consisting of the results of homogeneous diet feeding experiments. The method performance was evaluated in terms of absolute errors, central probability interval checks, the success in identifying the primary resource in the diet, and the ability to detect the absence of resources in the diet. Despite occasional large inconsistencies, all the methods were able to identify the primary resource most of the time. The numerical optimization method QFASA using χ 2 (QFASA-CS) or Kullback­–Leibler (QFASA-KL) distance measures had the smallest absolute errors, most frequently found the primary resource, and adequately detected the absence of resources. While the Bayesian methods usually performed well, some of the methods produced ambiguous results and some had much longer computing times than QFASA. Therefore, we recommend using QFASA-CS or QFASA-KL. Our systematic tests showed that FA models can be used to accurately estimate complex dietary mixtures in herbivorous zooplankton. This article is part of the theme issue ‘The next horizons for lipids as ‘trophic biomarkers': evidence and significance of consumer modification of dietary fatty acids'.

Evaluating the performance of the Bayesian mixing tool MixSIAR with fatty acid data for quantitative estimation of diet

“…In our example, there was a small proportion of herring estimated (i.e., 0.009 or 0.9%) which was the smallest proportion of all the sources. Litmanen et al 30 found that both Bayesian and QFASA methods had more problems identifying absent sources for herbivorous zooplankton when the number of absent sources was higher; however, there were exceptions. Although here MixSIAR did produce a small estimate for herring, more systematic testing is needed in order to understand when an absent source can be correctly identified.…”

“…Given the complexity of biological systems, particularly that they are often influenced by multiple factors, it would be advantageous to use modelling approaches that allow for the inclusion of ecological information. MixSIAR models diet proportions as a function of ecologically meaningful factors 11 ; previous knowledge can be incorporated into models as informative priors, as can the uncertainty in source values, unlike QFASA which can use only mean source values 30 . Despite the promise of MixSIAR with FA data, it has not been widely implemented, hence, its performance is still an open question.…”

“…To date, the most popular and well evaluated model to estimate diet using FAs is the numerical-optimization based method QFASA 30 . The QFASA approach has limitations associated with not being able to account for multiple sources of uncertainty, as well as the inability to incorporate ecological mechanisms into the models 25 .…”

“…The case in the figure applies for a consumer fed 100% a solely prey. Images by Alicia Guerrero. only mean source values 30 . Despite the promise of MixSIAR with FA data, it has not been widely implemented, hence, its performance is still an open question.…”

The critical importance of experimentation in biomarker-based trophic ecology

“…Litmanen et al [54, this issue] provides some guidance on this, with an initial critical comparison of the two classes of methods. These authors found that while all methods mostly performed well, QFASA was favoured in the scenarios they evaluated, in part because it had significantly shorter computation times and less ambiguity in the results [54]. The value in combining data from bulk stable isotope values and fatty acids in mixing model analyses has long been recognized, and modelling approaches that can incorporate both types of data are now available [49,55].…”