We have presented BEEHAVE, the first honeybee model that integrates processes within the hive and in the landscape and thereby allows representation of various stressors and their interactions in a more realistic way than previous models. There is also the option of including the effects of other stressors in the future, either by changing parameter values accordingly, or coding further modules. This model, if it is well tested and captures the most important processes in real honeybee colonies, will be an invaluable tool for exploring the relative importance of stressors to devise management and policy advice designed to reduce losses of honeybee colonies.
Indicators of structural realism
The colony dynamics of the BEEHAVE model are within the range of experimental data and indicate how BEEHAVE is driven by the availability of nectar and pollen within the hive and in the landscape. The level of pollen and nectar stores in the colony affects the age at which workers are sent out to forage, and this can have knock on effects on colony size as mortality of foragers is higher than for in-hive bees. BEEHAVE captures the internal relationships between season, colony size, and honey and nectar stores well.
We used AFF and the average life span of worker bees as indicators to demonstrate that the most important feedback mechanisms within a colony have been captured sufficiently well, because they are affected by a large number of factors which interact in a complex way (see 'Materials and methods'). Survival of the brood is affected by the number of in-hive bees available for nursing and the protein content of jelly fed by nurse bees, and survival of foragers is mainly determined by their activity level (total foraging time). Because of these complex relationships, it would have been impossible to directly impose AFF and average life span by choosing the correct phenologies of environmental drivers. They emerge from the model and, while the magnitudes and turning points of average AFF and life span over time do not match the empirical data well (possibly because environmental conditions differ between simulations and the Neukirch 1982 experiment), the overall trends are similar and the correlation between these indicators is well captured. Thus, a central feedback mechanism of real honeybee colonies, the change of AFF in response to the colony's state and demands, is realistically represented in the model. Although it would be possible, via calibration, to obtain better fits between these data and the model output, this would defeat the purpose of the model to mechanistically represent the internal mechanisms of honeybee colonies and how they respond to changes in the environment (rather than mimic a single environmental setting).
The foraging model was tested against the results of Seeley, Camazine and Sneyd (1991). This feeder experiment has become a standard test for individual-based foraging models of honeybees (e.g. De Vries & Biesmeijer 1998, 2002; Johnson & Nieh 2010; Schmickl, Thenius & Crailsheim 2012) because the colony's behaviour observed in the experiment is the summary outcome of a complex network of processes, including searching, recruitment of foragers and accounting for the energetic efficiency of foraging. Hence, reproducing the results of the feeder experiment is generally taken as a strong indicator that a foraging model is structurally realistic. BEEHAVE captured the switch of foragers between feeders very well. BEEHAVE is the first integrated colony model that can reproduce these patterns, even under the default parameterization which was not optimized to closely fit the empirical data.
When varroa mites were included, the model output compared well with predictions of the models developed by Martin (1998, 2001). If DWV-carrying mites are present in BEEHAVE, the mites will start affecting the colony in the third year and severely damage it in the fourth year, resulting in the death of the colony in the fourth or fifth winter. In contrast, colonies with DWV-carrying mites survive in Martin's (2001) model for only two summers and die during their second winter or the following spring. This seems to be surprising, as we developed our varroa model on the basis of Martin's (2001) model. However, Martin uses the critical colony size of 4000 bees (threshold for colony failure) throughout the winter, whereas in our simulations, this same critical colony size is implemented on the last day of the year. If we apply the same criterion for colony survival as Martin (2001), then most varroa-infested colonies actually die in the spring of the third year (data not shown). Without varroa, nine of 10 colonies survive for at least 5 years under this tightened survival criterion. Empirical survival experiments by Fries, Imdorf and Rosenkranz (2006) with non-treated, varroa-infested colonies show that most colonies died during their third and fourth winter after varroa infestation. From our results, we conclude that the varroa model is sufficiently realistic. The criterion used for assuming when a colony essentially fails can have a strong influence on predicted extinction dynamics. Predictions regarding colony losses over the course of time should be considered as relative, not absolute predictions.
BEEHAVE is thus shown to give realistic and robust predictions of colony dynamics, under different conditions. However, there are processes which are not yet included in the model, such as dynamic task allocation and temperature regulation within the hive, and the effects of bacterial and microsporidian pathogens. The model is also not able to predict synergistic effects between stressors on an individual. Elucidation of such interactions should be carried out experimentally or using toxicokinetic/toxicodynamic models or similar; once quantified they can later be included in BEEHAVE if desired. In its current form, BEEHAVE also allows us to address some problems, which are not explicitly implemented. Nosema infections for example could be simulated by increasing the background mortality (Higes, Martín-Hernández & Meana 2010) and the energy consumption rates (Naug & Gibbs 2009) of bees, by modifying AFF towards an earlier onset of foraging (Mayack & Naug 2009; Dussaubat et al. 2013) and by increasing the foraging mortality (Kralj & Fuchs 2010). With respect to future development of the model, some factors can be taken into account via parameter variation (as above), some might be considered in future modules, but some would require a fully individual-based design of the model, with internal states and decision-making of bees explicitly represented. The development of all of these is likely to require further data.
In our sensitivity analysis, we varied one parameter at a time, but over large ranges. This goes far beyond local sensitivity analysis, the most common type of sensitivity analysis in ecological modelling, where parameters are changed by only 5–10%. We thus performed 61 ‘sensitivity experiments’ (Railsback & Grimm 2012), which give a quite comprehensive overview of how single parameters affect model behaviour. Our analysis did not, however, cover interaction between parameters, which would have required a global sensitivity analysis based on some systematic sampling of parameter space (Saltelli et al. 2008; Saltelli & Annoni 2010). Although such an analysis might be desirable, it requires running the model for a very large number of parameter combinations, with significant run time and computing power implications.
Initial applications of BEEHAVE
Here, we provide evidence that BEEHAVE is realistic enough for exploring the impact of various stressors on honeybee colonies and provides a sample of applications to demonstrate how BEEHAVE can be used to address practical questions.
BEEHAVE output demonstrated that a single annual treatment with an effective hypothetical acaricide can protect varroa-infested colonies from failure over 5 years. This indicates how beekeeping interventions can be implemented within BEEHAVE to explore the relative effects of different mite management options on colony, mite and virus dynamics. If required, resistance of varroa to an acaricide could be addressed by reducing its efficacy.
We also found increasing distance to forage decreased survival time of varroa-infested colonies. The increased distance led to higher foraging costs in terms of energy expenditure and forager mortality, which hugely reduced honey stores. Certainly, the default landscape setting used is simplified, and further scenarios representative of realistic nectar and pollen landscapes will be needed to better understand the impact of landscape structure and dynamics on the colony. However, it does indicate that with multiple stressors, the increase in one (distance) can lead to greater sensitivity to another (varroa).
The final scenario simulated doubled forager mortality with an exposure period of 30 days as was simulated by Henry et al. (2012) who used the simple Khoury, Myerscough & Barron (2011) model. Again, adding another stressor (food availability) reduced the resilience to the existing stressor (pesticide). Thus, with poor forage conditions, average colony size was markedly reduced and colonies were lost if exposure leading to doubled forager mortality was repeated year-on-year during the most sensitive months. The results of these simulations do not lead to colony losses as rapidly as the simulations by Henry et al. (2012), most likely because BEEHAVE captures more of the processes and feedbacks within the colony, so that resilience of the colony emerges in a more biologically realistic fashion. Cresswell and Thompson (2012) find less severe colony effects than Henry et al. (2012) while Guez (2013) questions the calculation of the homing failure in Henry et al. (2012).
Overall, our results indicate that the effect of stressors such as forager losses, varroa and poor forage can build up over several years, particularly as in these simulations, beekeeping interventions were lacking. While it is immensely challenging to test such interacting stressors in controlled multiyear experiments, the study of interacting stressors is feasible with BEEHAVE. Moreover, our results indicate that the timing of a stressor may be as important as the magnitude of the stressor and that release from one stressor may mitigate the effects of other stressors. Lastly, our results indicate the importance of looking at possible effects affecting the colony repeatedly over several years, which are historically not captured in pesticide risk assessments.
Synthesis and recommendations
We conclude that BEEHAVE is ready to be used to tackle basic and applied questions regarding honeybees, their functioning and their decline. BEEHAVE was designed to be used by all those who are willing to invest time in understanding the model and the netlogo program and could be a valuable tool for scientists, pesticide regulators, land-based industries and beekeepers. We chose netlogo because it is freely available, easy to learn and comes with powerful and flexible tools for visualizing model output (including a link to GIS data and R statistical package). The model, its underlying assumptions and its biological basis are fully described using the ODD protocol (Grimm et al. 2006, 2010), which is a standard format that can be read by anybody, not just modellers. Moreover, the Supporting Information includes a User Manual and a Guided Tour which should enable non-modellers to understand how the program works and how it can be used. We are also maintaining a website (http://beehave-model.net/) supporting the use of BEEHAVE. To make sure that all users are working with the same version of BEEHAVE, we also provide means for version control and this will require that publications based on BEEHAVE include evidence that the correct version has been used. This is critical to ensure that results obtained with BEEHAVE can consistently be related to each other. Beehave (2013)©, the implementation of the model BEEHAVE is copyrighted to Matthias Becher and licensed under the GNU General Public License.
There is an urgent practical need for a model to provide biologically realistic predictions of honeybee colony dynamics, growth and survival in complex and changing environmental conditions, so that we can understand and manage the effects of emergent diseases, parasite pressure, changing landscapes and multiple pesticide exposures (Osborne 2012; Becher et al. 2013; 2013a-2013c; Vanbergen and the Insect Pollinators Initiative 2013). The tests and applications illustrated here demonstrate that BEEHAVE provides a robust platform, with sufficient complexity to simulate realism, to be developed and used to explore a range of practical management questions, of relevance to beekeepers (e.g. Application 1: acaricide), land managers (e.g. Application 2: forage and varroa) and risk assessors (e.g. Application 3: pesticide exposure and forage). Two further examples of such applications are (i) how colonies respond to different proportions and locations of planted forage mixtures that are used within agri-environment schemes and (ii) contributing to higher tier risk assessments of agrochemicals (2013a-2013c) using realistic projections of time and space. Such simulation experiments could save substantial time and resources, allowing scientists to focus field experiments on those factors and interactions which seem to be having the strongest effects in the simulations. We therefore recommend that BEEHAVE is used to explore the complex and urgent problems underlying honeybee colony failure and also to find and test alternative management techniques for the landscape and for the colonies themselves to improve their health and survival.