Association Between Prevalence of Anaemia and Malaria in Under-Five of Uganda: A Comparison of Ordered and Binary Logistic Multilevel Models

Abstract Background: several analyses of anemia prevalence in multilevel settings have dichotomized anemia as the response variable, ignoring its distinct ordered levels, leading to loss of information that may produce biased estimates. This study aimed at examining estimates from two multilevel models and the association between anemia and malaria in under-five of Uganda. Methods: a sample of 7,632 children, aged 0-59 months from the 2018-19 malaria indicator survey was analyzed using multilevel mixed effects logistic regression models. In one model, anemia was modeled as an ordinal categorical variable (no anemia, mild, moderate and severe) while in another model as a binary variable (no anemia and anemia). The main explanatory variable was the result of malaria test. Five models were fitted using each method (Model-1: the null model, Model-2: with malaria test result as the only covariate, Model-3: with individual level factors, Model-4: with household level factors, Model-5: with cluster level factors). The final model from each method was selected using Akaike’s information criterion (AIC). The model with more reliable estimates was selected basing on standard error (SE) values. Random effects variables were region, cluster and household. Results: anemia prevalence was (overall=51.62%; severe=1.36%; moderate=25.00%; mild=25.26%). The likelihood of a child who tested positive for malaria being anemic was about 5 folds in both models (OR=5.005; 95%CI, 4.369 – 5.733; p<0.001) and (OR=4.697; 95%CI, 3.983 – 5.538; p<0.001) for the ordered and binary logistic model respectively. Generally; sex, age, and wealth index were associated with a lower likelihood of being anemic. Although AIC values were lower in the binary model compared to the ordered model, the former was considered as a better model as it yielded lower standard errors for all estimates hence more reliable values.Conclusion: analysts should use modeling techniques that account for the order aspect in categorical outcome variables and avoid dichotomization. Choice of the best model should be based on standard errors besides AIC. Large standard errors are associated with large confidence intervals that may lead to type II error. Malaria increases risk of anemia.