Study design
The dataset used in this study is a cross-sectional data obtained from the 2016 Ghana Malaria Indicator Survey (GMIS). It provides information on malaria prevalence and indicators in the country. The sampling frame used for the 2016 GMIS was stratified and is the frame of the 2010 Population and Housing Census (PHC) conducted in Ghana. A two-stage sampling frame was used to select the sample for the 2016 GMIS. Each region was separated into urban and rural areas. In the first stage, 200 Enumeration areas (EAs), including 93 EAs in urban areas and 107 EAs in rural areas, were selected with probability proportional to the EA size and with independent selection in each sampling stratum. The resulting lists of households then served as a sampling frame for the selection of households in the second stage. EAs that were large were segmented and only one segment was chosen for the survey. In the second stage of selection, a fixed number of 30 households was selected from each cluster to make up a total sample size of 6,000 households and replacements for non-responding households was not allowed. Detailed description of the sampling can be found at https://dhsprogram.com/pubs/pdf/MIS26/MIS26.pdf [21].
Data collection
Information from women aged 15 and 49 years who were either permanent residents or visitors present in the households a night preceding the survey was obtained through the administration of questionnaire. There were three types of questionnaires (Household, Woman’s and Biomaker) and were written in English and three local Ghanaian languages (Akan, Ewe, and Ga). The questions were programmed into tablet computers and so allowed the use of computer-assisted personal interviewing (CAPI) for the survey. Issues that were captured included background characteristics, malaria prevention, malaria in children, malaria knowledge, among others. More details can be found elsewhere [21].
Outcome variable
This study concentrated on women between ages 15 and 49 years who possessed at least one mosquito net. A question which asked about sleeping under a treated mosquito net in the night preceding the survey was categorized as using ITNs while those who did not sleep under ITNs were categorized as not using mosquito nets. This was coded as a binary variable with a value of 1 and 0.
Independent variables
The variables considered included the following: age, education, household wealth, knowledge about causes of malaria, exposure to malaria messages, knowledge about the efficacy of mosquito nets, knowledge about coverage of malaria under National Insurance Scheme (NHIS), number of household members, place of residence, region and socioeconomic disadvantage. Age was categorized into 15–24, 25–34 and 35+ years. Education was defined as no education, primary and secondary or higher. Household wealth was recoded into poor, middle and rich. Knowledge about causes of malaria was defined as ‘Yes’ for women who reported that mosquitoes cause malaria and ‘No’ for those who reported otherwise. Exposure to malaria messages was defined as ‘Yes’ for those who heard or saw messages about malaria 6 months preceding the survey and ‘No’ for those who were not exposed to messages of that nature. Knowledge about the efficacy of mosquito nets was measured as ‘Yes’ for women who agreed that the chances of getting malaria were the same whether mosquito nets were used or not and ‘No’ for those who either disagreed or claimed that they did not know. Knowledge about coverage of malaria under National Health Insurance Scheme (NHIS) was defined as ‘Yes’ for those who agreed malaria was covered under NHIS and ‘No’ for those who disagreed or reported that they did not know. Number of household members was categorized into less than 5 and 5 or more. Socioeconomic status was obtained by applying principal component analysis technique to measure the proportions of individuals who were uneducated and poor within the communities. The resulting values was categorized into tertile 1 (least disadvantaged), tertile 2 and tertile 3 (most disadvantaged). Coding of independent variables were done based on previous works [22] [20].
Statistical analysis
Descriptive statistics
Percentage distribution of all the explanatory variables against the outcome variable were presented in a table. We used Chi-Square test to obtain p-values to indicate the level of significance of each variable.
Modelling approaches/ model fit and specifications
The study adopted the multilevel binary response models to estimate the odds and associated credible intervals of associations with mosquito net usage among women. There were four models in all. The first model had no covariates (Null model) and only showed how the outcome variable reacted. The second model included the independent variables except socioeconomic status (a community level factor). The third model then introduced residence as a random coefficient to allow for urban-rural differentials by communities. Finally, model four then added community socio-economic status to attain the full model. Analysis was done using MLwiN command in Stata version 14.0 and models were fitted using Markov Chain Monte Carlo (MCMC) estimation after using second order penalized quasi-likelihood (PQL2) to obtain sensible starting values for the model parameters [23]. No multicollinearity was identified amongst the covariates. Results were weighted and accounted for complex survey design.
Estimation technique
A two-level women-within- region variance components model was written as:
Net_useij ∼Bernoulli(πij)
logit (πij) = β0 + uj
uj ∼ N (0, σ2u)
Where Net_useij is the binary response for whether a woman i in the region j uses ITNs. β0 is the intercept for the log-odds of using ITN in the average household. Uj is a household level random effect assumed normally distributed with a zero mean and constant variance σ2u. There was no woman-level residual error appearing in the linear predictor.
The next model adjusted for individual-level factors which included age, education, wealth, knowledge about cause of malaria, exposure to malaria messages, chances of getting malaria, NHIS coverage, number of household members and residence. Age, educational level and wealth status were entered into the model as dummy variables
useij ∼Bernoulli(πij)
logit(πij) = β0 +β1age1ij + β2age2ij + β3edu1ij +β4edu2ij +β5wealth1ij +β6wealth2ij +β7ruralij + β8mcmij + β9mediaij + β10cgmij + β11nhisij + β12nhmij +uj
uj ∼ N (0, σ2u)
The third model presents a random coefficient to allow for urban-rural differentials across regions still adjusting for the same set of predictors as before.
The between region variance is now a function of residence in the random effect part of the model.
var (u0j +u7j rural ij) = σ2u0 +2σu07ruralij +σ2u07 rural2ij
The final model then introduces a community level explanatory variable, socio-economic status, to check whether certain regional-level variations in rural-urban areas may be explained.