Examining the effect of repeating stimuli on brain activity is important for theories of perception, learning and memory. Functional magnetic resonance imaging (fMRI) is a non-invasive way to examine repetition-related effects in the human brain. However the Blood-Oxygenation Level-Dependent (BOLD) signal measured by fMRI is far removed from the electrical activity recorded from single cells in animal studies of repetition effects. Despite that, there have been many claims about the neural mechanisms associated with fMRI repetition effects. However, none of these claims has adequately considered the temporal and spatial resolution limitations of fMRI. In this thesis, I tackle these limitations by combining simulations and modelling in order to infer repetition-related changes at the neural level. I start by considering temporal limitations in terms of the various types of general linear model (GLM) that have used to deconvolve single-trial BOLD estimates. Through simulations, I demonstrate that different GLMs are best depending on the relative size of trial-variance versus scan-variance, and the coherence of those variabilities across voxels. To address the spatial limitations, I identify six univariate and multivariate properties of repetition effects measured by event-related fMRI in regions of interest (ROI), including how repetition affects the ability to classify two classes of stimuli. To link these properties to underlying neural mechanisms, I create twelve models, inspired by single-cell studies. Using a grid search across model parameters, I find that only one model (“local scaling”) can account for all six fMRI properties simultaneously. I then validate this result on an independent dataset that involves a different stimulus set, protocol and ROI. Finally, I investigate classification of initial versus repeated presentations, regardless of the stimulus class. This work provides a better understanding of the neural correlates of stimulus repetition effects, as well as illustrating the importance of formal modelling.