Determining malaria risk factors in Abuja, Nigeria using various statistical approaches
Malaria is a widespread infectious disease in the tropics with significant health threat to the inhabitants, especially, the low income earners residing in the remote areas. Studies have identified many driving factors of malaria prevalence; however, there are still missing links as malaria remains...
Saved in:
| Main Author: | |
|---|---|
| Format: | Thesis |
| Language: | English |
| Published: |
2018
|
| Subjects: | |
| Online Access: | http://psasir.upm.edu.my/id/eprint/79271/1/IPM%202019%2013%20ir.pdf |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Malaria is a widespread infectious disease in the tropics with significant health threat to the inhabitants, especially, the low income earners residing in the remote areas. Studies have identified many driving factors of malaria prevalence; however, there are still missing links as malaria remains endemic in developing nations despite various interventions instituted against it especially in Sub-Saharan Africa (SSA). Previously, several decision support tools have been used to support expert-based opinion in the field of sciences for malaria modelling. However, these conventional techniques have problems of incorporating prior belief, uncertainty representation, missing values, expert judgement and hierarchical relationships among variables. These problems, however, can be overcome using belief networks, which has not been fully explored. Therefore, this current study is aimed at deriving a Bayesian belief network (BBN) for malaria epidemic in Abuja, Nigeria. A total of 384 respondents took part in the present study which is comprised of two phases: preand post-intervention. Participants were randomly selected based on stratified sampling according to residential areas. Data on socio-demographic characteristics, knowledge, attitude and practice (KAP) related to malaria and intervention measures were collected using validated questionnaires. The climate data were obtained from the data bank of Agricultural Development Project Gwagwalada, Abuja, Nigeria (1997-2014). Data collected were used for the multilevel analysis, Markov Chain Monte Carlo (MCMC) simulation via WinBUGS algorithm and influence diagrams for BBNs. A non-informative prior was assumed for Bayesian logistic regression and posterior samples generated at different sizes. The spatial heterogeneity in malaria incidence patterns at various sites were estimated with Moran I and semivariogram models. Using BBNs, different learning strategies were explored and compared with k-fold using negative entropy loss. The optimal model based on the parsimony principles was obtained from the hill climbing algorithm with score metrics. The results revealed a high reported cases at the baseline data collection possibly occasioned by the observed low malaria KAP levels at the pre-intervention study. A Wilcoxon signed rank test showed a significant change in incidence scores of households in the district considered for pre- and post-test interventions. The study revealed that there was an association between level of usage of the intervention measures and malaria cases. The non-usage increases the odds of the disease by 1.79 ([95% CI: 1.06, 3.03]; p=0.028) and 1.67([95% CI: 1.06, 2.64]; p=0.029) for insecticide-treated nets (ITNs) and window and door nets (WDNs) respectively. The multilevel analysis based on logistic regression identified gender, socio-economic status (SES), household size and intervention measures as predictors while Monte Carlo study of local malaria predictors’ results was comparable to logistics especially when n is large (150,000) and with a lower precision level (0.000001). The Moran I index using distance decay method to generate the weight gave a spatial autocorrelation of -0.33. The spatial outliers (a high-low and low-high outliers) of Moran I results reflected an alternation in incidence patterns. To account for spatial random effects in the models, semivariogram autocorrelation models were incorporated and Moran hypothesis tested. The results revealed that the model without autocorrelation structure has the lowest Bayesian information criterion (BIC) value of -6.31, while the highest value of 0.66 was observed with autoregressive error structure of order 1, hence, there was no spatial autocorrelation in the malaria incidence in the study area (p=0.328). Therefore, this was not incorporated in BBN models. Based on cross-validation analysis, the score-based algorithm outperformed the constraint-based algorithms in the structural learning. Using hill climbing from search and score algorithms, the Bayesian network analysis revealed that there were associations among the network covariates, while cofounding effects of SES were observed. The BBNs developed revealed that SES, household size and education level have the highest influence on reported cases as variations in response due to global sensitivity of network nodes. Based on the data, an empty graph (a network representing models with the usual independent assumption) was also learned and the results compared with BBNs. However, the loglikelihood and other metrics scores of an empty graph were lower than that of the BBNs. Thus, the BBNs represent the dependencies in the variables better than assuming independence of all the variables. This present study provides a detailed record on the epidemiology of malaria in the study area, Abuja, Nigeria. This in turn could be used to formulate effective control and preventive measures for malaria. The study shows that there is a significant influence of household characteristics on the incidence of malaria. The BBN is expected to contribute to the existing literature on malaria epidemic and in identifying the significant predictors of malaria at household and community levels within the study area. |
|---|