My results indicate that network structure changes with aggregation. There was a clear correlation between the increase in dissimilarity of motif frequencies and the increase in dissimilarity of resolution. I found that aggregation by both trophic similarity and taxonomic groups alter motif frequencies. The change in dissimilarity was significant for all aggregated networks regardless of clustering technique and linkage criterion. 

 

The effects of trophic and taxonomic aggregation on motif frequencies support the notion that network structure is not robust against aggregation. This corresponds to previous studies, which have showed that a wide variety of network measurements are sensitive to aggregation. The conclusion from those studies was that whole-network metrics changes with increasing aggregation. These type of metrics give a single measure which quantifies the overall network structure. This can make two networks appear similar when they in fact have very different detailed network structure. Here I show that the detailed network structure is also sensitive to increasing aggregation by using motif frequencies. Motifs describe the detail of networks by describing a unique configuration of interacting species. Comparing networks based on a detailed network measurement will therefore give a more precise comparison.

 

My results indicate that overly aggregated networks will not be accurate to use when describing the details of the network structure using motifs. They would be inaccurate to use because their motif frequencies differ from the original network, which mean they describe a different configuration of interacting species. This might make conclusions drawn from these networks incorrect. This can be a problem when comparing networks compiled by different research groups which might produce networks with different levels of resolution. The networks might look different because of their resolution rather than a true difference. It is therefore preferable to collect data for networks with as high resolution as possible or, if this is not possible, make it clear how the network is aggregated. When re-using published networks it is important to be aware of the potential differences that can arise with different resolution. To avoid misleading comparisons an idea is to combine several networks covering the same ecosystem to get as high resolution as possible. 

 

In contrary to the findings for trophic and taxonomic aggregation, I found no evidence suggesting that spatial boundaries will affect the network structure. My results showed that motif frequencies did not vary with spatial scale. This result may however differ between network types. When collecting data and compiling networks there is a a trade-off between either sampling a small area very intensively versus sampling many sites or a large area less comprehensively to obtain networks with the best resolution. Only sampling a small area would provide higher-resolution sampling for the same cost and effort. For networks similar to the Baltic Sea (e.g., other marine or freshwater ecosystems), our results suggest that a single area is a good representation for the wider area. In this study I only tested an system with limited variation over spatial scales. This mean that other types of ecosystems might be more sensitive to aggregation and can not be represented by a single area. This can for example be land-based ecosystems where habitats differ more in shorter distances. In a study from the Serengeti ecosystem the group structure of the plant-mammal network represented a mixture of trophic guild structure and spatial patterns which supports the idea of spatial coupling and its effect on the importance of persistence. Sampling many sites might take longer and cost more money but give a higher chance of capturing spatial coupling that can exist in some ecosystems. Spatial coupling can occur due to migration and seasonal turn-over, which forces both intermediate species and top predators to seek different food-resources during the year. Overall, my result suggest that it is most important to compile networks with as high resolution as possible. Using a single area to describe the bigger network is preferable due to cost and time perspectives of sampling larger areas.