Criteria for identification of children as specifically language impaired (SLI) vary greatly among clinicians and researchers. Standardized psychometric discrepancy criteria are more restrictive and perhaps less sensitive to language impairment than is clinical judgment based on a child’s language performance in naturalistic contexts. This paper examines (a) differences in groups of preschool children clinically diagnosed as SLI who were and were not identified as SLI through standard psychometric discrepancy criteria, and (b) the validity of quantitative measures of mean length of utterance (MLU), syntax, and pragmatics derived from a spontaneous language sample as criteria for discriminating clinically diagnosed preschoolers from normally developing preschoolers. Spontaneous language data indicated that children clinically identified as SLI produced a significantly higher percentage of errors in spontaneous speech than normal children whether they met psychometric discrepancy criteria or not. Logistic regression analysis indicated that a combination of MLU, percent structural errors, and chronological age was the optimal subset of variables useful for predicting a clinical diagnosis of SLI. This combined criterion captured a larger proportion of the clinically identified SLI children than even the best psychometric discrepancy criteria.