On the hypothesis that the position constraint on segmental interactions in speech errors reflects some aspect of lexical access during speech production planning, we have analyzed corpora of both spontaneous and elicited errors to determine the unit over which the position constraint operates. In spontaneous errors, two interacting segments are likely to occupy similar positions in both their respective words and syllables. Moreover, since most of the interaction errors in the MIT error corpus occur in initial position in words with first-syllable stress, it is difficult to determine which unit (word or stressed syllable) is the most appropriate one over which to state the constraint. Results of error elicitation experiments suggest that the word is the appropriate unit, since, for example, /j/ and /d/ are more likely to interact in the tongue twister “July dog dock Gillette” (where /ǰ/ and /d/ are both word-initial) than they are in the twister “legit dog dock largesse” (where /ǰ/ and /d/ are both prestressed). Implications for models of lexical access during production planning will be discussed.