Mass Validation of Opinion Dynamics Models

Many social issues, such as vaccination and climate change, depend on people’s opinions. For this reason, it would be extremely valuable to have models that allow to simulate how a certain population would respond to a specific communication campaign, as this would allow to assess risk and design effective strategies. While the field of opinion dynamics produces models that may seem able to achieve this, most of these models have never been tested against real world data. Even models that claim to be more “realistic” often have been only fitted (i.e. calibrated) on data or tested (i.e. validated) against a single time point.

To test the potential opinion dynamics models, in this study we tested multiple models against multiple time waves and datasets. Specifically, in phase (1) we selected 4 “toy” models and other 2 “realistic” models. Then, we use each one to produce some simulated dynamical data, and then we use these data to train and test each model. This validation step allows us to say that, if we were in posses of data that exactly reproduces the model dynamics, the process of training and test would be successful. In phase (2) we repeat the previous step, but adding some noise, so that the simulated dynamical data does not reproduce perfectly the model dynamics. In this case, the process of train and test still shows the ability of partially reproducing the model dynamics, even if with much lower explained variance. Finally, in phase (3) we move from simulated to real world data. Indeed, we use the European Social Survey (ESS) containing opinion data for European countries spanning over 2 decades. This allows us to test how much of the variance in the real-world data can be actually explained by each of these models.