MULTIPLE IMPUTATION MODELS FOR MISSING DATA: A CASE STUDY OF MODELING PSYCHOSOCIAL FUNCTIONING AMONG ADOLESCENTS IN IKERE-EKITI LOCAL GOVENMENT AREA, EKITI STATE

Abstract

Missing data present a challenge to health researchers in particular as incomplete data violate the complete-case assumption. A study about modeling Adolescents Psychosocial Functioning (APF) in Ekiti State presents such occurrence. Improper approaches to these missing data such as listwise deletion and mean imputation can lead to biased statistical inference using complete case analysis. This study presents the multiple imputation (Ml) method, a technique based on Bayesian inference, and Fully Conditional Specification approach to imputing the missing values in the APF dataset. A secondary dataset consisting of a random sample of 490 students from secondary schools in Ikere-Ekiti Local Government Area of Ekiti State participated in a study that seeks to know the effect of psychosocial well-being on depression using a combination of Rosenberg Self-Esteem Scale (RSES), Strength and Difficulty Questionnaire, and Center for Epidemiology Studies Depression Scale. Missing items on RSES ranges from 16 (3.3%) to 25 (5.1%). Hence, RSES was imputed using STATA mi command. Pattern of missingness found in the dataset was arbitrary. Also, the data provided sufficient evidence against the MCAR assumption. Indeed, on the basis of their religion, students who were satisfied with themselves (item RI of RSES) significantly differ from those without responses (x2 =5.836, p < 0.05). Furthermore, a multiple logistic regression model estimation showed that the effects of religion (P = 1.549, p < 0.05) and father's education (P= 1.672, p < 0.05) on probability of nonresponse to RI are significant. A linear regression model of self-esteem scores on the socio-demographic variables revealed more precise estimates when nonresponse is accounted for. For example, SSS1` students had significantly higher self-esteem score before imputation (P = 6.930, s.e. = 1.217, p < 0.0 I) and after imputation (P= 6.671, s.e. = 1.138, p < 0.001) than the SSS 2 students with a relative reduction in standard error (s.e.) of about 6%. Also, effects that were not significant prior to imputation became significant after imputation. Consequently, Ml is a missing data technique that allows for valid statistical inference with complete case statistical analysis. Therefore, health researchers should consider conducting proper missing value analysis so as to achieve substantial inference.