Accounting for correlated error structure within phenological data: A case study of trend analysis of snowdrop flowering

Abstract

Given that phenological studies can provide insight into some of the climate change driven alterations in global ecosystems, easily understood but valid statistical analyses are paramount. Results from studies of trends in phenophases at a regional level provide more powerful evidence of climate change; and such studies require observations from multiple locations. However, data containing phenophase time series from multiple locations has an inherent correlated error structure which may render standard statistical methods invalid. This chapter explores the problems in statistical inference that can arise from applying naïve techniques to data containing correlated error and provides two alternative modelling approaches for more valid analyses. These alternative modelling approaches - data resolution and random effects modelling - are extensions of simple linear regression. These modelling approaches are described in detail, compared and discussed in the context of potential questions, data and analysis issues in phenological research. A case study of trends in flowering of the spring bulb snowdrop (Galanthus nivalis L.) across England is used to demonstrate these alternative modelling approaches. For the period 1952-2000, snowdrop flowering advanced approximately 6.5 days (SE = 0.10) per decade across England. Snowdrop flowering was also estimated to advance approximate 3.9 days (SE = 0.52) for every 1°C increase in mean January air temperatures during the same period.