«To refine the modeling of the adoption process, we will include an awareness stage that separates the spread of information via observation or direct communication from the actual adoption of the innovation»

Jan Schalowski y Christian Barrot
«The Long-term Diffusion of Digital Platforms — An Agent-based Model»
Fortieth International Conference on Information Systems, Munich 2019
AIS eLibrary (Association for Information Systems, AIS)

«A Whole Market in an Agent-based Model

»In the second step of our analysis, we establish an agent-based model capable of capturing diffusion processes at the individual level with variable characteristics. Our aim is to identify the specific effect of varying levels of market fluctuations on the platform diffusion and to describe the deviations from a standard Bass model specification. For this purpose, we use our real-world data on the market for used cars in Germany. Hence, we incorporate approximately 60,000 agents based on the original locations and characteristics of the real-world dealerships and also account for the annual fluctuations in our model.

»The agent-based model intends to cover the whole car dealer market in Germany on the adopter side. On the innovator side, we include the first mover in Germany and its main competitor. Based on the number of adopters, both platforms have today a nearly complete coverage (>97 %) of all professional car dealers registered in Germany. The code for the diffusion process in the agent-based model is adapted from Netlogo’s simple viral marketing model (Rand and Wilensky 2012; Stonedahl et al. 2010).

»The transition rule for agents to change their status from non-adopter to adopter is in accordance with the Bass model: There is a small probability for the agents to adopt due to external influence and a larger probability to adopt the innovation contingent on the number of adopters in the neighborhood of the focal agent that have already adopted. As the model for our study consists of a vast number of agents that interact with each other and as the environment is based on geographical data, we opt for GAMA (Taillandier et al. 2018) as our modeling platform. GAMA, being also open source, can handle large models comprising up to millions of agents and is specialized on operating with GIS data. The idea is to enrich the Bass model with social networks that are based on the geographic distance between any two agents.

»Spatial proximity increases the probability for social interaction, observation and learning between local car dealers (Rangaswamy et al. 2018). Due to zoning laws, car dealerships are often geographically clustered, for example along major roads (so called “Automeile” or auto malls). These are often located in industrial areas with a limited number of catering outlets. Indeed, many restaurants in the neighborhood of auto malls are hotspots for social interactions and technical discussions. In many cities, local dealers form an advertising association, a joint committee to coordinate events (for example special sales weeks, street festivals, or car fairs) and to discuss business matters (Rangaswamy et al. 2018).

»Consequently, aftersales-service surveys, in which new platform adopters were called one month after signing up for the service, confirm such social contagion effects: asked how they became aware of the platform and why they signed up for the service, 54.3% of the adopting car dealers named referrals from other car dealers, customers, or suppliers as key sources (Rangaswamy et al. 2018). Thus, the adoption process is significantly driven by word-of-mouth, facilitated by the (professional) social network of a car dealer. A network that is based merely on geographic distance is highly clustered, and the characteristic path length is high and information needs to be passed through local subsets of nodes.

»The speed of information dissemination is inversely correlated with the characteristic path length, which would lead to a slower diffusion process, and some distant nodes, especially in sparsely populated areas in Eastern Germany, would never be affected by social contagion at all. Following the more realistic idea of small world networks (Watts and Strogatz 1998), we include short cuts between distant agents, either randomly, or based on observable criteria, such as the affiliation with a certain car brand. These distant links decrease the characteristic path length while preserving a large amount of clustering. In the model, we instantiate all car dealers at their actual geographical location and, depending on the actual entry respectively exit date, agents are activated or deactivated at the according time step.

»In a first stage, we model the empirical diffusion process to assess the occurrence of peer effects and adopter clusters. Figure 3 visually confirms one of the key factors to choose agent-based models for the simulation of the diffusion processes in our study. When plotting the actual adoption data (originally in much finer resolutions), we clearly observe a distinct pattern of adopter clusters, very similar to the phenomenon described by Garber et al. (2004). These clusters can be confirmed also analytically with divergence measures (here, we use the cross-entropy), while controlling for aspects such as population density or distribution of car dealerships. The existence of clusters in the adoption process signals the influence of contagion effects (word-of-mouth) among nearby car dealers. Thus, it is critical for the individual adoption process where a car dealer is located and whether a drop-out or market-entry occurs in proximity. This can only be captured in individual-level simulations and not in aggregate models.

»Next Steps

»In the next stage of the agent-based model development, we will implement scenarios of different degrees of fluctuation in the adopter population, simulate the resulting diffusion processes, and evaluate the effects of ongoing changes in the market population. We argue that market entry and exit does not only alter the population quantitatively, but also qualitatively with regard to the mixture of young and old firms. Depending on the firm age, different degrees of exposure to marketing measures, e.g., advertising stock, will hold in the population. To refine the modeling of the adoption process, we will include an awareness stage that separates the spread of information via observation or direct communication from the actual adoption of the innovation.

»We aim to quantify the effects of fluctuation in terms of change in adoption probabilities (as used as a dependent variable by Bollinger and Gillingham 2012) or time between starting up a car dealership until the adoption event. A core focus will be if and to what extend fluctuations within the potential market alter the diffusion curve of a digital platform, thereby testing the hypotheses that higher levels of fluctuation lead to a longer time to take-off, a flatter and more prolonged shape of the S-curve, and a longer time period to complete the diffusion process.

»Furthermore, we aim to have a closer look at local sub-clusters of adopters, auto malls, to study the mechanisms of social influence on a meso-level as an intermediate between individual adoptions (microlevel) and economy-wide diffusion (macro-level). Finally, we strive to quantify the value of a “lost” adopter due to market exit.

»These findings will be relevant for existing and future online platforms with a focus on markets with significant levels of fluctuation (for example, the current generation of innovative payment services), and may also provide insights for businesses marketing to customers that enter a cohort only for a limited time, associated likewise with constant inflows and outflows (e.g., students)».

.../... Read all on AIS eLibrary (Association for Information Systems, AIS)