Abstract:
|
Multiple imputation (MI) 'fills in' partially observed data. MI software often uses the untestable missing at random (MAR) assumption without exploring inference sensitivity. Here we demonstrate multilevel MI with ORiEL data with tipping point sensitivity analysis using the new JOMO R package. The substantive model compared change in depression over one year for adolescents in schools within and outside the Olympic intervention area. Change in depression was categorised: not depressed, not depressed to depressed, depressed to not depressed, depressed. Using MAR imputed data, the intervention was significant at 5% in each group compared with not depressed after adjusting for risk factors including being bullied (RR (95%CI): 1.52 (1.07-2.15), 1.61 (1.08-2.39) and 2.21 (1.09-4.48) respectively). Crudely changing all imputed values for the being bullied indicator to bullied caused the intervention to become not significant in depressed RR (95%CI): 2.02 (0.97-4.20), indicating a tipping point in the departure from MAR which would affect the significance of the intervention on depression. Tipping point sensitivity was shown through JOMO; inferences were robust within realistic settings.
|