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

Abstract

Abstract. The ubiquity of missing data in plant trait databases may hinder trait-based analyses of ecological patterns and processes. Spatially-explicit datasets with information on intraspecific trait variability are rare but offer great promise in improving our understanding of functional biogeography. At the same time, they offer specific challenges in terms of data imputation. Here we compare statistical imputation approaches, using varying levels of ecological information, for five plant traits (leaf biomass to sapwood area ratio, leaf nitrogen content, maximum tree height, leaf mass per area and wood density) in a spatially-explicit plant trait dataset of temperate and Mediterranean tree species (Catalonia, north-east Iberian Peninsula, 31900 km 2 ). We simulated gaps at different missingness levels (10% -80%) in a complete trait matrix, and we used overall trait means, species means, k-nearest neighbours (kNN), ordinary and regression kriging and multivariate imputation using chained equations (MICE) to impute missing trait values. We assessed these methods in terms of their accuracy and of their ability to preserve trait distributions, multi-trait correlation structure and bivariate trait relationships. The relatively good performance of mean and species mean imputations in terms of accuracy masked a poor representation of trait distributions and multivariate trait structure. Species identity improved MICE imputations for all traits, whereas forest structure and topography improved imputations for some traits. No method performed best consistently for the five studied traits, but, considering all traits and performance metrics, MICE informed by relevant ecological variables produced globally more plausible datasets. However, at higher missingness (> 30%) species mean imputations and regression kriging tended to outperform MICE for some traits. Spatial patterns of the studied traits in Catalan forests (5495 plots) were largely similar when using species means, regression kriging or the best-performing MICE application. Our results highlight the need to assess imputation quality beyond just imputation accuracy, and show that including ecological information in statistical imputation approaches yields more plausible imputations in spatially-explicit plant trait datasets.