Statistical summary representations (SSRs) are thought to be computed by the visual system to provide a rapid summary of the properties of sets of similar objects. Recently, it has been suggested that a change in the statistical properties of a set can be identified even when changes to the individual items comprising the set cannot. Haberman and Whitney (2011) showed that subjects were correctly able to report which of two consecutively presented sets of faces was, on average, happier, even when participants were unable to localize any of the items contributing to this change. In this paper, we revisit this conclusion, and suggest that the results supporting it may be an artifact of the paradigm used. In four experiments we find little evidence to suggest that subjects can reliably detect a change in the average size or emotion of an array of faces when they are unable to localize changes to individual items. The results are well accounted for by assuming that observers are selectively attending to individual items and then inferring the direction of the overall change based on the behaviour of the attended items. We suggest that this occurs because change localization requires focused attention to individual items, impeding calculation of SSRs which requires global attention to the entire set. We conclude that there is currently little evidence that SSRs can facilitate change detection when individual change localization fails.