The pace of our knowledge on online engagement has not been at par with our need to understand the temporal dynamics of online engagement, the transitions between engagement states, and the factors that influence a student being persistently engaged, transitioning to disengagement, or catching up and transitioning to an engaged state. Our study addresses such a gap and investigates how engagement evolves or changes over time, using a person-centered approach to identify for whom the changes happen and when. We take advantage of a novel and innovative multistate Markov model to identify what variables influence such transitions and with what magnitude, i.e., to answer the why. We use a large data set of 1428 enrollments in six courses (238 students). The findings show that online engagement changes differently —across students— and at different magnitudes —according to different instructional variables and previous engagement states. Cognitively engaging instructions helped cognitively engaged students stay engaged while negatively affecting disengaged students. Lectures —a resource that requires less mental energy— helped improve disengaged students. Such differential effects point to the different ways interventions can be applied to different groups, and how different groups may be supported. A balanced, carefully tailored approach is needed to design, intervene, or support students' engagement that takes into account the diversity of engagement states as well as the varied response magnitudes that intervention may incur across diverse students’ profiles. •Transitions and changes in engagement are largely influenced by instructional variables•Students' subgroups respond differently to instructional variables•Cognitively engaged students persist or improve with cognitively engaging learning resources•Disengaged students fare better with behaviorally engaging resources.•The ability to transition to a favorable state explains performance