Background Health at midlife is associated with early-life socioeconomic status (SES), intelligence and education; the latter often used as a marker of SES in adulthood. SES is typically modelled as a latent construct with reflective indicators (where the construct causes the variables) but may be better operationalised as a formative trait (where the variables cause the construct). In this report, the authors explain the difference between these two approaches and evaluate reflective and formative modelling for the prediction of health outcomes. Methods The Aberdeen Children of the 1950s Study comprises 12150 children from the Aberdeen area of Scotland, of whom 7183 completed a follow-up questionnaire in middle age. Data were available on indicators of parental SES at birth of the participant, childhood intelligence at age 11 and education and self-rated health at midlife. Results Childhood SES predicted childhood intelligence and educational attainment in adulthood, both of which partially mediated the effects of SES on self-rated health. Both approaches produced well-fitting models (CFI>0.99, TLI>0.99 and RMSEA<0.03), detected indirect effects from parental SES to health (all ps<0.001) and accounted for similar proportions of variance in health. Conclusions In these data, there was little difference between reflective and formative models of SES in their ability to elucidate pathways from childhood SES to adult health. Formative approaches to modelling SES seem appropriate theoretically, yet reflective models were equally informative. The outcome chosen here was selfrated health, but the models are extensible to other health outcomes. Results may differ in other data sets, suggesting that it is first appropriate to compare both strategies.