The Case of Self-transitions in Affective Dynamics

Abstract

Affect dynamics, the study of how affect develops and manifests over the course of learning, has become a popular area of research in learning analytics. Despite some shared metrics and research questions, researchers in this area have some differences in how they pre-process the data for analysis [17]. Specifically, researchers differ in how they treat cases where a student remains in the same affective state in two successive observations, referred to as self-transitions. While most researchers include these cases in their data, D’Mello and others have argued over the last few years that these cases should be removed prior to analysis. While this choice reflects the intended focus in their research paradigm on the transitions out of an affective state, this difference in data preprocessing changes the meaning of the metric used. For around a decade, the community has used the metric L to evaluate the probability of transitions in affect. L is largely believed to have a value of 0 when a transition is at chance, and this is true for the original use of the metric. However, this paper provides mathematical evidence that this metric does not have a value of 0 at chance if self-transitions are removed. This shift is problematic because previously published statistical analyses comparing L values to the value at chance have used the wrong value, incorrectly producing lowered p values and in many cases reporting transitions as significantly more likely than chance when they are actually less frequent.