Stage models of language production (e.g., Garrett, 1975, Psychology of Learning and Motivation, Vol. 9. New York: Academic Press; Levelt, 1983, Cognition, 14, 41–104) maintain that lexical retrieval occurs in two independent stages: retrieval of semantic and syntactic information and retrieval of phonological form. Interactive activation production models (e.g., Dell, 1986, Psychological Review, 93, 283–321), on the other hand, assume that the processes involved in the retrieval of semantic and phonological information are nonindependent, and that these two factors interact with syntactic and frequency variables to influence lexical retrieval. Predictions of each model regarding the relative influence of semantic and phonological variables on lexical retrieval were tested in an analysis of laboratory-induced naming errors. In Experiment 1, color- and shape-naming errors were elicited in a paradigm introduced by Levelt (op. cit,) in an attempt to replicate his findings, which supported an independent levels model of lexical retrieval. The results were consistent with Levelt's findings that substitution errors are influenced by semantic similarity, but not by phonological similarity. Also consistent with Levelt's results, no frequency effect was obtained for color errors. Frequency effects were obtained for shape-naming errors, but this effect may have been confounded by semantic and phonological similarity of two low frequency shape targets. In Experiment 2, the same paradigm was used to elicit errors, but stimuli were pictures of objects rather than colored shapes. The composition of the picture patterns was controlled to provide subjects with opportunities to make errors reflecting both individual and interactive influences of semantic and phonological information. The combined effects of semantic and phonological similarity were demonstrated to be interactive and not additive, indicating that semantic and phonological processes do not operate independently on lexical retrieval. These results, which were most consistent with an interactive model of lexical retrieval, were replicated with a different set of materials in Experiment 3.