Spider preparation
Adult female Araneus diadematus were collected from several locations in Oxfordshire, England. Spiders were stored in our standard Perspex frames (30 cm × 30 cm × 5 cm) separated by greased Perspex sheets. Twice per week, they were watered with a spray gun and fed two Drosophila melanogaster. Spiders had to build three trial webs on consecutive days in standard frames to be able to be selected for experiments.
Experimental method
Spiders were transferred into individual flexible frames which are identical to standard frames but with bendable corners. Four flexible frames were placed in parallel in the frame-shaking tool (henceforth ‘shaker’). An electric motor moved 4 parallel mechanical arms, each of which could be attached to the bottom of a flexible frame with magnets. Under Rigid control conditions ‘R’, the flexible frame was detached from the moving mechanical arm and remained stationary. Under Moving treatment conditions ‘M’, the frame was attached to its arm and was repeatedly moving side to side (Fig. 2). To establish a workable frequency and configure the hardware, we cycled through circa 10 spiders (in addition to those used in the actual experiment) in a set of pilot trials, some of which were used for multiple trial settings. Ultimately, it was not possible to fully and consistently replicate the variability of anchor movements in nature (Online Resource 1), and during trials spiders refused to build webs altogether at frequencies > 0.05 Hz. The results have been interpreted cautiously to reflect these experimental constraints.
The experiments followed a 5-day RRMRM regime. This RRMRM regime is known as (multiple baseline) reversal design and is particularly suitable for small N research designs (Saudargas and Drummer 1996). We selected this design because we predicted that our challenging experiment would result in a small sample size. Put succinctly, the reversal design ensures that observed effects are unlikely to be caused by extraneous factors as it increases internal validity (Saudargas and Drummer 1996)and thus (partly) compensates for a small sample. On days where spiders did not build a web, the same run was repeated up to four consecutive days (e.g. R1R2R2R2R2M3R4M5) after which the spider was removed and the sequence was considered incomplete. Between runs, the webs were sprinkled with water and all radials were cut except for two radials leading North and South to the frame. Spiders were fed 1 D. melanogaster per day of the experimental regime regardless of web building success and experimental conditions.
Building behaviour
Web building was recorded with timelapses; one time-stamped photograph was obtained per ~ 10 s. With a custom tracking programme in Python, the coordinates of spiders and all frame corners were recorded for each photograph. Tracking began when spiders started web construction and terminated when the spider sat still at the hub. Raw tracks were corrected for shaker-induced movements in moving runs—because frame movements were greatest at the bottom, the correction was adjusted for the spider’s latitudinal position in the frame. By scaling relative to the frames, the absolute distances between corrected coordinates were calculated.
From activity plots (Online Resource 3) and established activity signatures from (Zschokke and Vollrath 1995), we identified the photographs in which spiders started and completed construction of the radials, auxiliary spiral and capture spiral. This information was used to calculate four building behaviour variables; time differences between photographs were used to calculate the total web construction time and the percentage of time spent on each web component, and the absolute positional data were used to calculate the total distances covered during total web construction and the percentage of the distance covered per web component.
Web geometry
The resulting webs were photographed (Panasonic LUMIX GH5 digital camera and Nikon AF NIKKOR 50 mm lens) and eight web measurements were obtained in ImageJ (Online Resource 4); measurements scaled relative to frames. These measurements were used to calculate the six wind-affected web features (Table 1). Whilst the total radial lengths and capture spiral length are also affected by wind (Vollrath et al. 1997), we did not include these features in the analysis as they are geometrically related to other included features; radial length is determined by web area and radial count, whilst capture spiral length is determined by capture spiral count, mesh space and capture spiral area.
Table 1 Analysed web design features
Data analysis
The effects of moving anchors on web design features and building behaviour were examined with mixed models (Davies and Gray 2015) in which Spider ID was included as a random effect to adjust for pseudoreplication. The models thus examined whether anchor movement affected building behaviour and each web feature when adjusted for variation that naturally occurs between individuals.
Crucially, spiders may also use the experience of building one web under certain conditions when building the following web e.g. (Venner et al. 2000). Using boxplots, we visually assessed if there was an experience effect within the rigid data (days R1, R2, R3) or within the moving data (days M1, M2) for any of the variables of interest. The boxplots did not indicate a clear and significant experience effect (Online Resource 5–7).
The effects of moving anchors on capture spiral count and radial count data were assessed with generalised linear mixed models (GLMMs); Poisson error distributions. All other dependent variables were continuous and analysed with linear mixed models (LMMs). Several continuous variables analysed with LMMs were log-transformed or inverse transformed to meet the assumption of normality which was verified with histograms, Q-Q plots and Shapiro Wilk tests (Online Resource 8).
Multiple LMMs and the radial count GLMM were overfitted (i.e. not optimally parsimonious (Hawkins 2004)) due to the inclusion of the random effect. However, because overfitting did not change model outputs at all, we opted to retain Spider ID in all models to accurately represent our experimental design throughout. Per the methods in (Thomas 2017), the capture spiral count and radial count GLMMs were also assessed for overdispersion (when variability in the data exceeds that predicted by the Poisson GLMM (Berk and MacDonald 2008)), and no significant overdispersion was detected (ratio = 0.69, p = 0.82 and ratio = 0.34, p = 0.99 respectively). The assumptions of all models were met.
Finally, as we examined four building behaviour variables, a Bonferroni correction was applied and \(p \le 0.013\) would indicate a significant effect. Likewise, because we examined six web design features, \(p \le 0.008\) would indicate a significant effect.