The current study analyzes 4,000 tokens of Spanish coda /s/ from both a discrete and continuous perspective. The data, drawn from the Otheguy-Zentella Corpus of Spanish in New York, are described categorically in terms of the presence or absence of frication, i.e., in terms of s-deletion. Data are also described in terms of two instrumentally measured gradient parameters: frication duration (in milliseconds) and spectral center of gravity (COG, in Hz). Results suggest that a unified methodology does more than simply increase the descriptive breadth of the analysis. Instead, it shows that certain patterns of variation are opaque at exclusively categorical or gradient levels. With respect to social factors, the data suggest that Spanish speakers born and raised in NYC produce /s/ in significantly different ways than their recently arrived counterparts from Latin America. The cause of these differences is very likely to be language and dialect contact. That is, generational shifts in /s/ production are due to native New Yorkers’ extensive experience with English and also with multiple varieties of Spanish. However, it is impossible to draw this conclusion purely on the basis of either deletion rates or measurements of the spectrotemporal properties of fricatives. Instead it is necessary to examine both kinds of data to properly assess these trends. Similar findings emerge with respect to linguistic factors. While some linguistic factors robustly constrain deletion rates as well as variation in frication duration and COG, the predictive power of other factors is restricted to either the discrete or continuous domain of variability. For example, the morphemic status of /s/ strongly conditions the presence or absence of frication. However, this same constraint has no bearing on either the duration or COG of fricative moments. Such asymmetries in predictive power suggest that the conditioning effects of linguistic factors can be restricted to specific domains of variability. The theoretical implications of this finding are substantial, indicating that models of socio-phonetic variation that do not distinguish and relate discrete and continuous levels of expression may misconstrue the relative contributions of explanatory factors.