The present study applies single-subject experimental design to examine (a) the acquisition and generalization of complex sentence production in agrammatism using Linguistic Specific Treatment (LST) and (b) the utility of syntactic theory in guiding hypotheses of treatment effects. LST trains construction and production of complex sentence structures. Four sentence types were selected for study: object clefts and object-extracted matrix and embedded questions (which are noncanonical with wh-movement), and embedded actives (which are canonical with no overt movement). All sentences contain overt material in the complementizer phrase (CP) of the syntactic tree. Three of five participants (1, 2, and 3) demonstrated generalization from object cleft treatment to production of matrix questions. Thus, LST was effective in improving their ability to generate less complex sentences with wh-movement. Once production of object clefts and matrix questions was acquired, all 5 participants demonstrated generalization from treatment to improved production of embedded questions and/or embedded actives. This generalization involved improved ability to generate embedded clausal structure to form complex sentences but continuing inability to express overt material in CP. Finally, direct treatment for embedded questions did not result in accurate production of embedded actives or vice versa. There were no trends across participants toward improved production of morphosyntactic behaviors in narrative. Persons 1, 2, and 3 showed generalization to increased informativeness and efficiency of expression and were judged by independent listeners to improve in content, coherence, and fluency of spontaneous production. The remaining two participants showed no change or a decline in performance in narrative language production (4 and 5, respectively). These participants demonstrated more severe Broca's aphasia at pretesting compared to Persons 1, 2, and 3, with greater impairments in auditory comprehension, naming, and reading. Etiology and size of lesion did not appear to account for the different behavioral patterns. This study supports the use of LST, which applies syntactic theory to predict patterns of generalization, as an effective treatment approach.