Methods/Statistics
Issues in measurement and causal modelling of complex social exposures: examining typologies of violence in young people in Luwero Uganda Daniel J Carter* Daniel Carter Amiya Bhatia Jenny Parkes Dipak Naker Karen Devries
Background. Like many exposures in social epidemiology, operationalising measures of interpersonal violence is complex. Experiences of different acts of violence from different perpetrators co-occur across an individual’s life course, particularly in high-prevalence settings. Using measures such as binary or sum scores may result in a wide array of experiences being classified as the same exposure to violence, biasing estimates of the consequences of violence as the exposure is not well-defined.
Methods. We draw on three waves of data from the Context of Violence in Adolescence Cohort (CoVAC), a prospective cohort study of 3431 young people in Luwero, Uganda, to examine how approaches like latent class analysis (LCA) could offer more consistent measurement for violence exposure. We fit LCAs at each wave to estimate underlying typologies of violence in early adolescence, late adolescence, and early adulthood. We investigated solutions ranging from two to nine classes and compared fit statistics and qualitative class cohesion to select the optimal number of classes.
Results. Three-class, six-class and five-class solutions were returned in early adolescence, late adolescence, and early adulthood respectively. Lowest violence classes were characterised by experiencing no or common acts of violence, often in school. Highest violence classes were characterised by violence across many places and perpetrators, including at work, and from caregivers and partners, and including acts of severe violence and sexual violence. Middle classes reflected both common acts of violence and violence experienced in a specific domain. Movement between classes was fluid across the lifecourse.
Conclusion. Measurement accounting for co-occurence demonstrates distinct patternings of violence. Such operationalisation of complex exposures may reduce misclassification and help mitigate consistency assumption violations, not just in violence research, but across the field of social epidemiology.