Climate change and other human-caused environmental disturbance may lead to declines in biodiversity. Recently, a number of studies have collated large data sets of monitoring time series for selected ecosystem or organism groups and used these data sets to estimate trends in biodiversity, with many studies identifying large declines in biodiversity across a number of organisms or ecosystems. These results are not without controversy however; data selection and quality issues, as well as questions over statistical methodology have lead to vigorous debate.
Typically, trends in biodiversity are estimated using linear effects, via generalized linear mixed (or hierarchical) models to account for site-to-site heterogeneity in temporal trends. Additionally, year-to-year variation may enhance or mask estimated losses or gains in biodiversity over time if the first observation year in a given series is unusually rich or depauperate. Using year random effects has been suggested as a mechanism to account for this potential bias. An alternative way to model trends in biodiversity time series is using penalized splines for the trends, leading to hierarchical generalized additive models (HGAMs). Here, I introduce HGAMs and penalized splines and their use for modelling biodiversity trends, and reanalyze a forest arthropod diversity time series data set.
Results/Conclusions
HGAMs were fitted to the forest arthropod data (150 plots, 9 years per plot) that included a variety of decompositions of biodiversity trends. Of the these, two models had clearly superior predictive ability: (1) a model with region-specific trends plus plot-specific random smooths (AIC 15120), and (2) as model (1) plus year random effects to model the average year-to-year effects (AIC 15122). Both models represent significant improvements over a GLMM with linear trends plus random linear slopes per plot and year-to-year effects (AIC 15467). The main difference between the two HGAMs is the complexity of the region-specific trends: model (2) had simpler regional trends, closer to linear, than model (1), due to the additional year-to-year effect modelling variation ascribed to the regional trends in model (1).
This shows that there may be several ways to decompose trends in biodiversity data that are equally good at describing those underlying trends. Here, the bias in the estimated linear trends due to many series starting in “good” years can be removed either via the year-to-year random effect or region specific smooths, both of which capture the overall tendency for some years to have greater arthropod abundance than others.
