Filling the gap in functional trait databases: use of ecological hypotheses to replace missing data

Abstract

Functional trait databases are powerful tools in ecology, though most of them contain large amounts of missing values. The goal of this study was to test the effect of imputation methods on the evaluation of trait values at species level and on the subsequent calculation of functional diversity indices at community level using functional trait databases. Two simple imputation methods (average and median), two methods based on ecological hypotheses, and one multiple imputation method were tested using a large plant trait database, together with the influence of the percentage of missing data and differences between functional traits. At community level, the complete-case approach and three functional diversity indices calculated from grassland plant communities were included. At the species level, one of the methods based on ecological hypothesis was for all traits more accurate than imputation with average or median values, but the multiple imputation method was superior for most of the traits. The method based on functional proximity between species was the best method for traits with an unbalanced distribution, while the method based on the existence of relationships between traits was the best for traits with a balanced distribution. The ranking of the grassland communities for their functional diversity indices was not robust with the complete-case approach, even for low percentages of missing data. With the imputation methods based on ecological hypotheses, functional diversity indices could be computed with a maximum of 30% of missing data, without affecting the ranking between grassland communities. The multiple imputation method performed well, but not better than single imputation based on ecological hypothesis and adapted to the distribution of the trait values for the functional identity and range of the communities. Ecological studies using functional trait databases have to deal with missing data using imputation methods corresponding to their specific needs and making the most out of the information available in the databases. Within this framework, this study indicates the possibilities and limits of single imputation methods based on ecological hypothesis and concludes that they could be useful when studying the ranking of communities for their functional diversity indices.

Gap-filling a spatially explicit plant trait database: comparing imputation methods and different levels of environmental information

“…Even if Mean imputations imply the rather naive assumption that species identity may be unknown in a given dataset, it is nonetheless useful to compare Mean imputations against mice and kNN, which use the full trait matrix for prediction. In this case, trait covariation did not improve imputations at high missingness; recent assessments also report that the performance of MICE and kNN notably declines when missingness is ≥ 30 % (Penone et al, 2014;Taugourdeau et al, 2014). Therefore, our results for OrdKrig, compared to those for mice and kNN, show that spatial structure, rather than trait covariation, may provide more accurate trait imputations when gaps are frequent (Fig.…”

“…Stochasticity is introduced in the imputation process because the parameters of the univariate imputation models are drawn from their posterior distributions, obtained using a Gibbs sampler (van Buuren, 2012). Assessments of imputation methods in the ecological literature have not tested the impact of the choice of univariate imputation models within MICE (Penone et al, 2014;Taugourdeau et al, 2014). Here we showed that predictive mean matching (PMM), the default algorithm in the mice package, performed well compared to alternative methods (Sect.…”

“…Complex imputation methods such as kNN, MICE or random forests generally outperform overall mean or species mean imputations (Penone et al, 2014;Taugourdeau et al, 2014). In earlier applications of these methods, it has been common to assume that interspecific trait variability was dominant, compared to intraspecific variability.…”

“…The expected declining performance of imputation methods with increasing missingness levels may be trait and dataset dependent (Penone et al, 2014;Taugourdeau et al, 2014). Moreover, the impact of imputations on altering bivariate trait relationships has only been assessed for single relationships (Penone et al, 2014;Schrodt et al, 2015) and not for the multiple relevant relationships within a plant trait dataset.…”

Integrating data‐deficient species in analyses of evolutionary history loss

“…Trait proximity was previously shown to perform well in imputation of missing data in the assessment of functional diversity indices (Taugourdeau et al. 2014). We chose six species as a compromise between having variability in extinction risks and because the method was less powerful when we chose a higher number of species more different in range size (Appendix S4).…”