Date of Award


Degree Type


Degree Name

Doctor of Professional Studies


Information Management


Jeff Stanton


belief dynamics;Bias;Markov;social media;transition probability


This thesis seeks to determine how the probabilistic dynamics of belief can be estimated. More specifically, it explores how effectively those dynamics can be estimated while accounting for the different nuances that arise from social media data in the framework of belief dynamics. Belief dynamics, a sub-discipline of computational social science dating back to 1964, explores the evolution of individual and group beliefs over time. Belief dynamics research delves into facets such as convergence to shared belief, belief interdependence, and clustering. However, there has been a long-standing wall preventing major progress in fully understanding belief dynamics: available data. Social media has contributed toward tearing down this wall, yet no research seems to have been done applying the existing belief dynamics frameworks to the endless supply of data generated by humans via social media. It can be argued that belief dynamics have the Markov property, which in this case asserts that tomorrow's beliefs depend solely on today's. This approach was first used in 1967, and it provides a foundation for modeling belief dynamics through probabilistic transitions between possible states of belief. Despite numerous models developed in the last two decades, there are challenges in estimating these transitions without complete data. Additionally, the presence of organically developed bias in social media (specifically polarization, popularity, self-selection, status quo, and outlier bias), pose unique challenges as well. Existing literature suggests various approaches for addressing the missing data, yet testing these was done on very small data sets, which social media is not. And there is no mention of bias relative to this venture. Building on this previous literature that assumed the Markov property, many sets of social media data were simulated incorporating missing data and the different forms of inherent bias. Each data set had the transitions estimated via the standard, ``na\"{i}ve'' approach, as well as more specialized techniques. Ultimately, it was determined that the different forms of bias have direct impact on the estimation accuracy, however the nature and magnitude of that impact was not consistent. Additionally, it was revealed that missing data, data volume, and data depth have varying types and magnitudes of influence. Researchers can use this investigation to strategize the best way to estimate the transitions that organically arise in a belief dynamics framework.


Open Access