The present study used an elicited imitation paradigm to test the prediction of Schutze & Wexler's (1996) AGREEMENT/TENSE OMISSION MODEL (ATOM) that the rate of non-nominative subjects with agreement-marked verb forms will be sufficiently low that such errors can reasonably be disregarded as noise in the data. A screening procedure identified five children who produced non-nominative subject errors (all her for she) who were then asked to repeat 24 sentences with 3sg feminine pronoun subjects (she) and agreeing main verbs or auxiliaries. All five children produced at least one non-nominative subject (her) with an agreement-marked verb form, and for none of these five children was the non-NOM+AGR rate significantly different to the rate that would be expected by chance, given the independent frequencies of non-nominative subjects and agreement-marked verb forms in their data. The three children for whom this expected (by chance) error rate was significantly greater than 10% (representing an acceptable level of noise in the data) produced non-NOM+AGR errors at a rate significantly greater than 10%, counter to the prediction of the ATOM. These results replicate and extend the naturalistic-data findings of Pine et al. using a different method. They also provide support for the use of elicited imitation as a methodology for assessing children's early grammatical knowledge.