Natural spoken language is full of disfluency. Around 10% of utterances produced in everyday speech contain disfluencies such as repetitions, repairs, filled pauses and other hesitation phenomena. The production of disfluency has generally been attributed to underlying problems in the planning and formulation of upcoming speech. However, it remains an open question to what extent factors known to affect the selection and retrieval of words in isolation influence disfluency production during connected speech, and whether different types of disfluency are associated with difficulties at different stages of production. Previous attempts to answer these questions have largely relied on corpora of unconstrained, spontaneous speech; to date, there has been little direct experimental research that has attempted to manipulate factors that underlie natural disfluency production. This thesis takes a different approach to the study of disfluency production by constraining the likely content and complexity of speakers utterances while maintaining a context of naturalistic, spontaneous speech. This thesis presents evidence from five experiments based on the Network Task (Oomen & Postma, 2001), in addition to two related picture-naming studies. In the Network Task, participants described to a listener the route of a marker as it traverses a visually presented network of pictures connected by one or more paths. The disfluencies of interest in their descriptions were associated with the production of the picture name. The experiments varied the ease with which pictures in the networks could be named by manipulating factors known to affect lexical or pre-lexical processing: lexical access and retrieval were impacted by manipulations of picture-name agreement and the frequency of the dominant picture names, while visual and conceptual processing difficulty was manipulated by blurring pictures and through prior picture familiarisation. The results of these studies indicate that while general production difficulty does reliably increase the likelihood of disfluency, difficulties associated with particular aspects of lexical access and retrieval have dissociable effects on the likelihood of disfluency. Most notably, while the production of function word prolongations demonstrates a close relationship to lexical difficulties relating to the selection and retrieval of picture names, filled pauses tend to occur predominantly at the beginning of utterances, and appear to be primarily associated with message-level planning processes. Picture naming latencies correlated highly with the rates of observed hesitations, establishing that the likelihood of a disfluency could be attributed to the same lexical and pre-lexical processes that result in longer naming times. Moreover, acoustic analyses of a subset of observed disfluencies established that those disfluencies associated with more serious planning difficulties also tended to have longer durations, however they do not reliably relate to longer upcoming delays. Taken together, the results of these studies demonstrate that the elicitation of disfluency is open to explicit manipulation, and that mid-utterance disfluencies are related to difficulties during specific production processes. Moreover, the type of disfluency produced is not arbitrary, but may be related to both the type and location of the problem encountered at the point that speech is suspended. Through the further exploration of these relationships, it may be possible to use disfluency as an effective tool to study online language production processes.