Adolescence is a key period of biological and social development and household living arrangements in adolescence in sub-Saharan Africa has been shown to be associated with multiple biosocial outcomes. Household is a commonly used term across a wide range of disciplines, however traditional, western-centric definitions are often used which may not capture important, context-specific differences in household membership (who belongs to which group) and composition (how the household members are related). This study used data on adolescents from rural northern Malawi from 2004-2016 to create context-relevant household composition variables using latent class analysis (LCA) with two household membership definitions, ‘immediate’ (as defined within the dataset) and ‘expanded’ (created to include relatives living close by). The extent to which different definitions of household composition alter observed associations with biosocial outcomes was investigated. LCA identified household compositions with greater complexity than those represented in western-centric definitions (LCA classes included ‘brother’s family’, ‘sister’s family’, ‘maternal’ and ‘paternal’), with few individuals living in ‘nuclear’ families. Using the ‘expanded’ household definition created classes which, for example, distinguished between ‘single mother’ households and those with a single mother but living very close to maternal family. LCA was found to be most useful for guiding the creation of manual ‘LCA-guided’ variables to produce household composition definitions which were suitable for use as predictor variables. Compared to western-centric definitions, LCA-guided household composition definitions using both ‘immediate’ and ‘expanded’ definitions provided greater detail about the contribution of household composition to variation in associations with biosocial outcomes: for example female adolescents in ‘maternal’ households had higher odds of a poor educational outcome, while for male adolescents this effect was found in ‘paternal’ households. While potential drawbacks in terms of generalisability and statistical power are recognised, other researchers are recommended, where appropriate, to consider using context-specific household definitions.