This study aimed to identify determinants of a late autism spectrum disorder diagnosis, including diagnoses made 'very late' (i.e., in adolescence), using the Millennium Cohort Study, a nationally representative population-based cohort in the United Kingdom. Children diagnosed with autism spectrum disorder by age 14 (N = 581) were included and grouped by the parent-reported timing of diagnosis: before school (up to age 5), during primary school (age 5-11) and during secondary school (age 11-14). Predictors of diagnostic timing, at the child, family and school levels, were investigated using multinomial logistic regression. Most (79%) children with autism spectrum disorder were diagnosed after school entry, and 28% were not diagnosed until secondary school. Among those not diagnosed until secondary school, 75% had been identified at age 5 years by a parent and/or teacher as having socio-behavioural difficulties. Being diagnosed after starting school was predicted by living in poverty (adjusted relative risk ratio: primary = 1.90, 95% confidence interval: 1.03-3.53; secondary = 2.15, 1.05-4.42) and/or having no initial parental concerns (primary = 0.32, 0.15-0.70; secondary = 0.19, 0.09-0.43). Having typical-range intelligence also predicted diagnosis during secondary school. The result indicates that those without cognitive delays and poorer children were at risk of 'very late' (i.e. adolescent) diagnosis. Strategies to promote earlier identification, targeting age at primary school entry, could help those more likely to be diagnosed late.