Early identification is important for improving reading outcomes for children at risk for reading difficulties, but current methods tend to perform sub-optimally in identifying true risk. One possible reason is that whereas most prediction models assume linear relationships among risk and reading outcome measures, qualitatively different patterns of association may exist among the measures at different skill levels. We implemented dynamic non-linear modeling to test this possibility in two distinct samples of children: 1) 225 pre-kindergarten and kindergarten with concurrent data and 2) 104 children with longitudinal (pre-)kindergarten and second-grade reading scores. Cusp catastrophe modeling was used to evaluate the moderating effects of rapid automatized naming (RAN) and phonological processing, two foundational pre-reading skills, on the concurrent and longitudinal relationship between letter identification and word reading. We further tested whether RAN and phonological processing have independent non-linear effects on the letter-word reading relationship above and beyond that of a single skill. Deficits in RAN and phonological processing beyond a critical level were associated with non-linear changes in the prediction of word reading from letter knowledge, both concurrently and longitudinally, fully supporting the cusp model over the competing models. These findings demonstrate the importance of implementing non-linear models for predicting risk for reading difficulties. There was no evidence for the interactive effects of RAN and phonological processing on reading. Instead, in accordance with the basic tenants of the double-deficit hypothesis, current results suggest that the constructs represent two salient but separable causes of reading impairment, even at the earliest stages of reading ability. These findings suggest that models predicting which at-risk children will develop poor reading must diverge from assumptions of relationships observed in typical reading.