What works best to model correlates of primary school dropout in the presence of clustering?
MetadataShow full item record
This study establishes what works best to model correlates of primary school dropouts in the presence of clustering. The main objective of this dissertation was to examine the most reliable of the multilevel logistic, probit, complementary log-log models under casewise deletion or imputation to identify variables used to explain dropout. The study used secondary data from the individual and poverty modules collected under the UNHS 2016 by UBOS. Multilevel mixed effects models were used in the analysis of correlates. Findings from the study indicated that the random-effects variance component of the null Multilevel logistic model was large (approximately 4.6), therefore using a random effects analysis, it was found to provide a better appreciation for the uncertainty about the strength of the relationship between the independent and dependent variables following the analysis. In addition, the variable child sex had a random slope of approximately 0.31 which was a stronger variation for modeling of the random slope. Furthermore, the study indicated that; Clustering had an effect on dropout, the multilevel logistic model with casewise deletion was found to be better to model correlates of primary school dropout in the presence of clustering. Furthermore, results indicated that poverty had a significant effect on dropout (OR = 1.95; p = 0.001), household heads above 53 years significantly affect child dropout (OR = 2.39; p = 0.004). Education level of the household head had a significant effect on child dropout. Child relationship with the household head had a significant effect on dropout. Tests also indicated that a multilevel logistic model with casewise deletion was better compared to the model with imputation. The major recommendations from the findings include; hierarchical models should be used to model clustered data, casewise deletion should be used instead of imputation in case the fraction of missing is very small at about 2%.