Policy and Complex Systems
messages is what drives the model . When agents receive a message or a referral , they update their BI ( t ) and schedule themselves to update again at some random time in the future . If they stop receiving messages and referrals , they will ultimately stop scheduling to update themselves and the simulation will stop . This propagation feature of the model allows us to test the salience of different messaging campaigns , since those with greater impact will run longer . Recall that cumulative reinforcement decays over time , so all configurations of the model will ultimately pass below the threshold for sending referrals and come to a stop . In the next section , we describe our test scenarios for the model and present the results .
Results and Analysis
To illustrate the practical application of our model , we designed three separate scenarios that test three hypotheses about the effectiveness of a particular persuasive messaging strategy . The purpose of these scenarios and the results that follow are intended to provide insights into how such a model , once properly calibrated , could assist in the design of persuasive messaging campaigns .
Scenarios and Experimental Design
Our first hypothesis is that the presence of messaging will have no effect on the average behavioral intent to perform a particular behavior within the population . Given the heterogeneous nature of our agents , this result is not trivial . If an agent evaluates the outcome negatively , increased reinforcement from messaging will have a negative impact on BI ( t ). The heterogeneity of agents and the fact that belief decay occurs in the model also lead to a natural decline in BI ( t ) in the absence of messaging . As a result , we varied the starting time of the messaging campaign in order to analyze the impact given different initial levels of the average population behavioral intent . We chose four different starting times while holding the interval of messaging constant ( see Table 1 ).
We ran each experiment for 30 replications and collected statistics on the time average behavioral intent for each run . All 30 time series for each of the 4 experiments are shown in Figure 6 . In the first experiment , messaging starts at t = 0 and ends at t = 100 as illustrated by the shaded area . It can be seen from the plot that the initial conditions of the model are impacting the level of BI ( t ) and making it difficult to determine what impact is due to messaging and what
Table 1 . Message starting and stopping times
211