Geographical gradients in network complexity were present many of the analysed network types. Latitude, which I expected to affect network complexity, did so to some extent in all network types, with the exception of food webs. Dimensionality was the only network property which was not significantly affected by any geographical variables. Intervality from both predator and prey perspectives decreased at higher absolute latitudes, especially seed dispersal networks, but also pollination (predator perspective) and antagonistic (prey perspective) networks. This was contrary to my original hypothesis, which expected that higher biodiversity closer to the equator would be associated with more complex species interactions, allowing room for more niches. The unexpected result can be partly explained by interactions with land cover types; predator intervality in pollination networks responded as expected, decreasing with absolute latitude in grasslands, rain forests and varied forests.

Following the hypothesis that specialisation should be more varied closer to the equator, generality distributions widened with proximity to the equator in seed dispersal networks and antagonistic networks. In pollination networks, latitude was only significant when interacting with altitude. Vulnerability distributions in antagonistic networks instead narrowed with higher latitudes. My results of the antagonistic networks were not in line with the findings of a previous study, where no significant latitudinal effect on antagonistic network structure was found. As most studies look at means of different specialisation measures, direct comparisons with my distributions are not obvious. A previous study found that the overall biotic specialisation in mutualistic networks was lower closer to the equator. Wider generality distributions in my seed dispersal networks may support this, as a more varied degree of specialisation also requires generalist species, making the overall specialisation lower. The significance of latitude as a predictor varied among the network types, indicated by the varying degree of relative importance. This, as well as the previously mentioned interactions, suggests that while latitude is often relevant, additional variables are necessary for more accurate models.

Higher altitude was associated with more interval pollination networks from a predator perspective. High altitude can provide an insular effect for ecosystems, especially in stable climates, leading to higher complexity. While the result may first appear opposite of the expected, when taking into account interactions with all forest types, grasslands and croplands, increased altitude instead lowered intervality. However, no models including altitude held high relative importance, meaning that the overall effect, while significant, was not strong. This was also true for latitude and altitude affecting vulnerability in antagonistic networks. Given the multitude of significant land covers and interactions between latitude, altitude and several land covers, the model also showed signs of over-fitting.

Despite their relation, none of the significant geographical patterns detected in intervality could be seen in dimensionality. Dimensionality can be considered a sensitive measurement, where a single link can make the difference in the total number of dimensions. Dimensionality measurements were limited to about half of my networks, restricting the number networks available for this network metric. Possibly as a result of this, geographical models of food webs and seed dispersal networks could not be obtained, and the relative importances of the pollination and antagonistic network models were low.

While especially latitude, but also the other geographical variables, are often significant predictors to the network complexity, there can be room for improvement. It is likely useful to further divide the gradients into separate continents or hemispheres, as has been shown with the LDG. Interactions can also be explored more detailed, e.g. additional combinations of interactions, including between geographical and environmental factors.

Overall, environmental models held higher relative importance than geographical models. Hence, in general, they are more accurate to use as predictors for the complexity measurements that I use in this study. The relative importance, and which network types they were important for, varied largely though.

The most pervasive environmental factor was precipitation average. Higher precipitation average increased dimensionality of food webs, and increased gaps of prey intervality in seed dispersal and antagonistic networks. In seed dispersal and antagonistic networks, higher precipitation average instead lowered predator intervality. Taking interactions into account, the impact direction regarding complexity differed depending on network type. For example, in association with temperature range, predator intervality in pollination networks was more interval. When instead looking at seed dispersal networks with the same factors, they were less interval. This confirmed the purpose of splitting networks depending on interaction type (e.g. pollination networks), rather than sticking with e.g. mutualistic networks. High precipitation average widened generality distributions in seed dispersal and antagonistic networks. In pollination networks, higher precipitation average instead narrowed vulnerability distributions.

Precipitation seasonality and temperature range were used to measure contemporary climate stability. Based on the assumption that biodiversity is higher in more stable climates, though often linked with altitude restricted dispersal, I expected these factors to decrease network complexity and the specialisation distributions. However, lacking a pervasive trend, in four out of six instances where precipitation seasonality was significant, it had an increasing effect on network complexity. A previous study looked at how seasonality affected modularity (how interactions between species are mainly concentrated in smaller subgroups) in food webs and mutualistic networks. They found higher precipitation seasonality to increase modularity in food webs, compared to my study (largely comprised of the same stream food webs) where the same factor increased dimensionality. This opens up an unanswered question; whether the interactions between the subgroups are caused by unique traits for each group, or by larger combinations of shared traits.

In pollination networks, high precipitation seasonality decreased dimensionality while it decreased predator intervality (increased number of gaps). At first glance, this seems counter-intuitive, as you would expect the additional gaps to be explained by more dimensions. However, dimensionality and intervality does not have a fully linear relationship, with intervality at its lowest around 2-3 dimensions. This may explain the opposite effect of the two complexity measurements, where the decrease in intervality may instead be the result of other changes, such as narrower niche breadth. In accordance with my hypothesis, generality distributions tended to widen with higher precipitation seasonality in pollination and antagonistic networks, while vulnerability distributions for antagonistic networks decreased. Temperature range by itself was only significant in two occasions, increasing dimensionality in food webs and increasing prey intervality in pollination networks. Taking interactions into account, temperature range was instead significant in twelve different factors. Similarly, the other environmental factors were also included in a multitude of interactions with each other, indicating that interactions between environmental factors is a complex issue which must be put under careful consideration when modelling.

Fast historical temperature change had a significantly widening effect on generality and vulnerability distributions in pollination and antagonistic networks, meaning the degree of specialisation varied more. The widening effect in faster changing areas was expected based on how biodiversity is generally higher in more historically stable climates, and may provide better conditions for species to specialise. Supporting this, while not always significant, higher numbers of species often indicated narrower generality and vulnerability distributions. A previous study similarly compared plant-pollinator networks with hummingbirds, where they found that higher temperature change velocity decreased network modularity. Lower modularity could mean more generalists in the networks, supporting my results of wider generality and vulnerability distributions. My study only tested interactions between contemporary and historical factors separately, which is likely an important area to expand further.

Given the broad purview of my project, there are plenty of areas which are suitable for further development. For improved accuracy in the models, it may be motivated to apply a similar hemispherical split of networks as has been shown for biodiversity gradients. Variable interactions is also an area which can be expanded, as due to technical and analytical limitations they were somewhat restricted in this study. Looking closer at network types with varying intervality and dimensionality may also reveal key traits behind these differences, as well as the relationship to modularity.