The use of residuals for longitudinal data analysis: the example of child growth

Health impact evaluations often measure changes in health status over part of a total life experience. The effects on health up to and including the start of the evaluation, which are embodied in the measure of initial health status, need to be removed while examining the effects that other variables exert during the evaluation period on final health status. Statistical models, which include initial health status as a covariate while examining the effects of other variables during the evaluation, confound the effects of a determinant during evaluation with preevaluation effects, because they do not differentiate between effects produced at different times by the same determinant. A residual model removes the preevaluation effects by regressing final health status on initial health status. The residuals from this regression are then regressed on the other predictor variables. In this paper, standard covariate adjustment, which includes all effects simultaneously, is compared with a two-part residual model using child growth as an example. The simultaneous model over- and underpredicts growth relative to the residual model depending on the age and initial body size of a child. In general, whenever initial (preintervention) and final (postintervention) measures of health outcome exist, the residual model should be considered on the basis of biologic and epidemiologic consideration, not solely on statistical optimality.