The nature of Apraxia of Speech (AOS) is poorly understood, particularly in connected speech, and there has been controversy about diagnosis, which has been based on clinical judgements without adequate measures of reliability. There have been few studies of AOS in connected speech rather than single words, fewer about grammar, and no investigations of AOS and grammar in spontaneous speech. The thesis adopts Text units ('TU's, which are smaller than traditional sentences), for segmenting connected speech when looking at AOS, and applies multilevel random-intercept logistic regression modelling to measure many factors concurrently. Three strands of evidence about AOS in connected speech are presented.