This study was a population based cross sectional study conducted in high malaria burden areas of Myanmar in 2019.
Study location
The study townships were 20 townships from 6 States and Regions with highest number of total malaria positive cases in 2018. They were 10 from Sagaing, 1 from Mandalay, 4 from Tanintharyi, 2 from Kayin, 1 from Chin and 2 from Kachin (Fig. 2). The study townships were the overlapping townships of the two criteria such as (1) hot spot townships, as defined by National Malaria Control Programme where the total number of positive cases was above 1000 in the year 2018, and (2) townships with at least 50 positive cases in the year 2018 as reported in PSI/Myanmar Management Information System data. Townships were the third level administrative divisions in Myanmar. A typical township consisted of urban wards and rural villages. Rural villages, primary place of residence for forest goers, were selected as primary sampling units. A total of 40 villages were selected from 20 townships where 1 village was selected from 6 townships, 2 villages were selected from 9 townships, 3 villages were selected from 4 townships and 4 villages were selected from 1 township. When selecting the villages, villages that were within 2 kms away from forest were screened first and probability-proportional-to-size (PPS) sampling was used to select the villages.
Study procedure
The study consisted of two phases. In the first phase, a census mapping of all the health service providers (public, private, informal or traditional) in the selected villages was conducted. This was followed by a quantitative household survey in the second phase (Fig. 1).
Health service provider mapping
Census exercise was completed in all selected villages so as to link the available services in these villages with the health-seeking for fever episode. Upon arrival, community leaders and gatekeepers were contacted by the survey team to identify the formal and informal health services available within the selected villages. Then, a road by road census of the village was conducted to identify all the service providers and confirmed the service provider type.
Household survey
After mapping exercise in each selected village, random households were selected using systematic random sampling and screened for eligibility of having at least one forest goer. For the purpose of the study, forest goers were defined as: adult men or women aged 18 and above, who spent at least 1 night in the forest during the past 4 weeks. An interview was conducted in the household with at least one forest goer. Informed consent was obtained from selected participants before the interview. In the household with more than one forest goers, the one who had stayed in the forest for the longest period was selected for the interview. Among 1,680 households screened, 713 (42.4%) eligible households were visited by the study team. Total of 479, 67.2% of eligible households completed interviews where at least one household member was eligible for the interview. Eligible participants were not available at their residences for interview in 209 households, 3 households refused to participate in the study and interviews could not be completed in 22 households because they did not come back from forest for interviews.
Eligibility for each section of fever
If any member of household had fever within past one month, he or she was eligible for recent fever section. If a forest goer had fever, he or she was eligible for both recent and past fever sections. Therefore, more than one forest goer completed the past fever section if they were eligible. Past fever episode was any fever episode in the past occurred in residential village and working in forest. If there was any RDT testing experience during the past, it was explored for residential village and working in forest for forest goers.
Data collection
Data were collected by PSI/Myanmar research team in October and November, 2019 with instruments developed by GEMS Program Team and PSI/Myanmar Research Department. For health service provider mapping, mapping tools were used to collect data from all the fixed (venue or facility based) and mobile health services within the village. For the household survey, a structured household survey questionnaire was used to collect demographic information of members, forest goer information and their knowledge of malaria transmission and prevention. Household socio-economic status information such as household assets, housing materials, drinking water source etc. was collected through a short version of equity tool [24]. For exploring health-seeking behaviors of any household members, experience of having a febrile illness during past one month and how a care was sought was asked in the survey. To provide an insight for differences in decision making based on fever locations, forest goers in every household were asked about how they received care during past fever episodes as well as RDT testing providers if there was any. All data collection tools were developed in English and translated into Myanmar. Face-to-face interviews were conducted in Myanmar language. Data was collected by tablets using C. S Pro version 7.1. Field data were uploaded to the server daily and data quality checks were done for completeness and consistency of the quantitative data as stated in the study protocol.
Data preparation and variable selection
To find the association between choice of providers and potential predictors based on fever episodes and RDT testing, analyses were done for 5 scenarios; recent fever episode for any member of household, past fever episode in residence, that in forest for forest goers as well as RDT testing in residence and that in forest for forest goers. Dependent variable was a polychotomous variable reflecting the four provider alternatives: i. Private, semi-private and informal providers, ii. Public facility and health staff and iii. Community health worker or volunteer and iv. Outside providers-all type of health care providers located outside of the village boundary. The provider categorization was based on the common types of providers that patients seek care for malaria febrile illnesses in Myanmar. The outcome variable for fever had four categories (Public, Semi-private, CHV/CHW and Outside providers) and the outcome variable for RDT testing had three categories (Non-CHV/CHW, CHV/CHW and utside providers) depending on the choices that the respondents made for each occasion. The questionnaire and codebook were reviewed in order to identify questions and variables that would be useful in identifying choice of health care provider as well as potential predictors of their health-seeking practice.
Independent variables included in the analyses were characteristics such as age (continuous variable), gender (male and female), relationship with household head (household head himself or family member) and having correct knowledge of malaria transmission and prevention (correct and incorrect knowledge). In addition, access to health care providers, which was reflective of respondent’s accessibility to health care providers in their residential village, was also included (access to 1 to 3 providers and access to 4 and above providers). The minimum number of providers that respondents accessed to was 1 and the average number of providers in the villages was 3 in the study. Household socio-economic status (5 quintiles from poorest to richest) and education level of main income earner (illiterate,primary and secondary and above) were also treated as independent variables. Household socio-economic status was derived using a set of household assets questions [24]. Tool measures relative wealth of a household, in which household assets were transformed into a composite score and a cut-off used to derive five wealth quintiles [25]. The cut-offs applied were those from national quintiles, therefore, the wealth of studied households represented their relative wealth with reference to national wealth quintiles [26].
Data analysis
Data cleaning, management and analyses were performed using STATA version 14 [27]. Percentages and 95% Confidence Intervals (CI) were computed for all categorical variables of interest. Means and standard deviations were calculated for all continuous variables. Bivariate regressions were done to investigate the bivariate relationship between each potential predictor variable and the outcome of interest which was the provider choice variable. Statistical significance was set as p = 0.05. To control the effect of confounding variables, multinomial logistic regressions were done. P value and 95% CI were used to interpret the findings.
Multinomial logistic model
The multinomial logistic model used the following equation. J + 1 is the number of distinct categories in the dependent variable and assume that the category 0 is selected as the base category. Then the probabilities given by the multinomial logistic function are:
$${\text{p (Y = j) = }}\frac{{{\text{Exp (}}\beta^{\prime}_{j} {\text{x}}_{{\text{i}}} )}}{{1 + \sum\limits_{{{\text{k}} = 1}}^{{\text{J}}} {{\text{Exp (}}\beta^{\prime}_{k} {\text{x}}_{{\text{i}}} )} }}\quad {\text{for j = 1,}} \ldots {\text{, J and}}$$
$${\text{P}}\,(Y = 0) = \frac{1}{{1 + \sum\limits_{{{\text{k}} = 1}}^{{\text{J}}} {{\text{Exp}}\,{(}\beta \prime_{k} {\text{x}}_{{\text{i}}} )} }}\quad {\text{for}}\,{\text{the}}\,{\text{base}}\,{\text{category}}.$$
where \(\beta^{\prime}_{j}\) is the vector of estimated coefficients for the jth category and \({\text{x}}_{{\text{i}}}\) is the ith case (row) of the data matrix.
The relative risk ratio for case i relative to the base category is:
$$\frac{{{\text{p}}_{{{\text{ij}}}} }}{{{\text{P}}_{{{\text{i}}0}} }} = {\text{Exp}}\,{(}\beta^{\prime}_{j} {\text{x}}_{{\text{i}}} )\quad for\,j = 1, \ldots ,J\,and\,i = 1, \ldots ,n$$
The first step in building the multivariable multinomial logistic model involved conducting simple multinomial model between each of the potential predictors, and the polychotomous choice of health service provider variable. Age and gender were added in the model with a priori belief that they might have influence on provider choice as well as they found to be associated with malaria treatment-seeking behaviour [16, 33, 51]. For other variables, those significantly associated with the outcome (p < 0.05) in bivariate analyses were considered for inclusion in the multivariate model. However, to avoid highly correlated predictor variables, two-way correlations between the predictor variables were assessed using Pearson’s correlation coefficient. Relative risk ratios (RRR) and their 95% confidence intervals (CI) were then computed for all variables in the final model. Model goodness-of-fit was assessed using the Stata command mlogitgof and post-test diagnostics were done using mlogtest [28, 29].
Ethical statement
Ethical approval was obtained from Population Services International Research Ethics Board with the approval number 40.2019, and Institutional Review Board-1 of Ministry of Health and Sports Myanmar with the approval number IRB 1/2019–2.