We compare the predictions of two different accounts of first language acquisition by investigating the relative contributions of abstract syntax and input frequency to the elicited production of main and embedded questions by 36 monolingual English-speaking toddlers aged 3;00 to 5;11. In particular, we investigate whether children’s accuracy rates across different interrogative structures (main vs. embedded, yes/no vs. wh-, argument vs. adjunct) can be explained by difference in terms of input frequency in parental speech or whether abstract structural factors are needed to account for such asymmetries. In main-clause questions, children correctly invert the order of the subject and auxiliary more often with yes/no than wh-questions, despite a higher input frequency of uninverted yes/no questions. Furthermore, in main-clause wh-questions, inversion rates are higher for argument than adjunct wh-questions, independent of input frequencies. Finally, in embedded-clause questions, children correctly avoid inversion more often in yes/no than wh-questions and show no effect of input frequency or type of wh-word. A significant positive correlation between (correct) inversion rates in main and (incorrect) inversion rates in embedded questions suggests that inversion in embedded contexts stems from rule overgeneralization. Taken together, the results highlight the importance of abstract structural factors in children’s production, above and beyond the role of frequency distributions in the input.