Study area and data
The data for this study was collected as part of the Zackenberg Basic monitoring programme. The sampling was carried out at Zackenberg, North-east Greenland (74°28'N; 20°34'W) which is in the high-arctic climatic zone. The mean summer (June through August) temperature during the years 1996–2005 was 4.3°C. At Zackenberg, June, July and August include on average 85% (range: 73%–93%) of the annual number of days with positive average temperatures. A climate station within 600 metres from all sampling plots provided data on ambient air temperature two metres above the ground, solar radiation (W/m2, SR), precipitation (mm, PREC) and wind speed (m/s) [16]. Wind speed was recorded every ten minutes and other weather variables were recorded hourly throughout the study period [17]. We calculated thawing degree days (TDD) by letting the daily contribution to TDD equal the mean of all hourly air temperature measurements where recordings of subzero temperatures were set to zero. To accommodate for the anticipated non-linear response of arthropods to wind speed, wind data were converted to the number of ten minute intervals per day with wind speed higher than 3 m/s (WIND, Fig. 1).
Arthropods were monitored during 10 consecutive years (1996–2005) with samples from one window trap plot (plot 1) and six pitfall trap plots (plot 2–7) collected weekly during June, July and August. All plots were operated during the period 1996–2005 except for plots 6 and 7 which were operated during the periods 1996–1998 and 1999–2005, respectively. Each pitfall trap plot (10 × 20 m2) consisted of eight pitfall traps and each window trap plot consisted of two window traps [18]. The window trap plot was located next to a pond. The pitfall traps were yellow plastic cups 10 cm in diameter and each window trap consisted of two chambers and the two traps were placed perpendicular to each other [see [19] for details]. Trapping started in June once the snow at each trap had melted. All specimens caught in all years were sorted to the family level except for collembolans and mites who were only counted. Here, we focus on: Chironomidae (Diptera), Muscidae (Diptera), Sciaridae (Diptera), Ichneumonidae (Hymenoptera), Nymphalidae (Lepidoptera), Lycosidae (Aranea), Linyphiidae (Aranea), collembolans and mites. The family Nymphalidae is represented by two species (Boloria chariclea, (Schneider) and B. polaris, (Boisduval)), the Lycosidae by one species (Pardosa glacialis, (Thorell)) and the Linyphiidae by five species (Collinsia thulensis, (Jackson), Hilaira vexatrix, (O. P.-Cambridge), Erigone arctica (White), Erigone psychrophila, Thorell and Mecynargus borealis, (Jackson)) [20]. The data set included a total of 531,036 individuals, each belonging to one of the nine taxa. This corresponds to 93.6% of all arthropod specimen caught during the study period. Traps were occasionally flooded, emptied by arctic foxes (Alopex lagopus, (Linneaus)), or trampled by muskoxen (Ovibos moschatus, (Zimmermann)), so the capture numbers in each plot were transformed to individuals caught per trap per day for each trapping period [7].
Statistical analyses
Our aim was to model the concurrent influences of ambient weather and arthropod phenology on capture numbers. Since we anticipated a non-linear phenological development through each season but had no a priori assumptions about its exact shape, we used partial smoothing splines in generalized additive models (GAM) [21] to model the seasonal development in capture rates. This family of models identifies the most likely relationship between parameters based on a non-parametric back-fitting algorithm [15] and so is particularly useful in situations where a non-linear relationship is anticipated but its form is unknown. We developed full models of the following general structure:
Where Y
i
are the log10-transformed number of individuals per trap per day, β0 is the intercept and trapping date (DAY) is modelled by a spline function s(·). The term WEATHER refers to one of the four different weather variables (TDD, SR, PREC and WIND) and was modelled as a linear continuous predictor and ε
i
is the error term. The weather variables were calculated as daily averages for each trapping period. We repeated the model estimation for each species group (n = 9) in each plot (n = 7 for flying arthropods, n = 6 for surface-dwelling arthropods, because this group was excluded from the window trap plot) in each year (n = 10) with each of the weather parameters (n = 4) as linear predictors and date as a spline function (with df = 4). This resulted in a total of 500 sets of four competing models. Some taxa were only caught in small numbers in some plots in some years, and we therefore restricted our analyses to years and plots where at least 100 individuals were caught except for the Nymphalidae and the Ichneumonidae where the limit was set to 50 due to the larger size of individuals in these families. This reduced the number of sets of models to 385. We had an average of 11.5 trapping periods within each season. Although it is likely that the capture numbers were affected by several weather variables, we included only one in each model to reduce the risk of over-parameterization. Since all models with the same response variable had the same number of degrees of freedom and the same null deviance, the lowest residual deviance indicated the best fit to data. In this way, we identified for each species group for each plot for each year the weather variable that was best able to explain variation in capture numbers. A previous study [14] separated seasonal patterns of abundance from activity by first fitting a Gaussian curve to the data and using weather variables to explain variation in the residuals. In addition to fitting GAM models we adopted this approach. Hence, we fitted Gaussian curves to the entire set of 500 time series and identified the weather variable that best explained variation in the residuals from linear regression models.