Introducing Empirical Data to the Model
In order to adapt the model to the specific case of the Austrian biotechnology- related industry, we follow an empirical agent-based modelling approach (Smajgl and Barreteau 2014). Apart from the use of expert knowledge during model conceptualisation and scenario formation, this section describes how empirical data is introduced into the model to (1) initialise the agent population (Input), (2) to calibrate the model (Research processes) and (3) to determine the research output (Output).
Initialization of the Agent Population
The model is initialised at the agent level as well as at the system level. For initialisation at the agent level, we include 61 private firms of the Austrian biotechnology-related industry. The sample is composed of firms with active patenting records in patent classes associated with biotechnology (OECD 2008) over the years 2000-2010, extracted from the PATSTAT database. The set of technology classes individually assigned to each agent corresponds to the patent classes in which they actually hold patents in the period of initialisation (i.e. 2000-2010). The patent classes are used on a three-digit subclass level specified by the International Patent Classification (IPC). The number of technology classes in the model results from the most frequently occurring IPC classes on the patents of the firm sample (i.e. 45 technology classes). Additionally, 16 other related patent classes are considered to provide the possibility for further knowledge creation. A description of the technology classes used is given in the Appendix.
For the initialisation of an agent’s knowledge endowment, first, for every kene kj a technology class is randomly drawn from the set of references to IPC classes (from the associated firm’s individual patent stock). Second and third, the respective subfield and expertise level are chosen randomly according to the specifications stated in Eq. (1). Moreover, each agent at is also individually equipped with four empirically based organisational figures: (1) research expenditures R,, (2) number of employees Ц, (3) assets It and (4) age At, taken from a company database and a recent industry study (Schibany et al. 2010; Orbis 2014).
At the system-level, an important aspect of model initialisation is the notion of an empirically based “knowledge space” in which the agents intentionally move around if they achieve new knowledge. To this end, a relational concept is needed that determines the technological distance between the technology classes. This can be obtained using a similarity measure on the set T of technology classes derived from the (empirical) patent stock of the firm sample. Hereby, we define the similarity of two technology classes Ti and Tm (with l, m = 1,2, ..., M) as the Jaccard index Jlm (Rip and Courtial 1984) given by
where clm denotes the number of co-references to technology classes Tl and Tm, while cl and cm denote the numbers of references to technology classes Tl and Tm in the given set of patents, so that 0 < Jlm < 1. By means of this definition, two technology classes are considered more similar the more often they are both mentioned in a patent. In the model, we assume that a new technology class is more easily accessible for an agent if it is more similar to its existing knowledge endowment.