Models of L2 pronunciation learning have hypothesized that accurate speech perception promotes accurate speech production. This claim can be evaluated longitudinally by examining the extent to which changes in stop consonant perception predict changes in stop consonant production. Taking a time-sensitive view of the perception-production link, this study used longitudinal data to analyze perception as a time-varying predictor of production accuracy. Mixed-effects models were fit to oddity, delayed word repetition, and picture description tasks to examine how participants' perception and production changed over time. Oddity task perception data were then decomposed into their between- and within-subjects components and integrated into the delayed repetition and picture description production models. Surprisingly, only the between-subjects predictors reached significance, and the strength of the perception--production link varied across production tasks and target phones. The methods used have implications for future research on perception-production links.