Towards achievement of Sustainable Development Goal 3: multilevel analyses of demographic and health survey data on health insurance coverage and maternal healthcare utilisation in sub-Saharan Africa

BACKGROUND: Improving maternal health and achieving universal health coverage (UHC) are important expectations in the global Sustainable Development Goals (SDGs) agenda. While health insurance has been shown as effective in the utilisation of maternal healthcare, there is a paucity of literature on this relationship in sub-Saharan Africa (SSA). We examined the relationship between health insurance coverage and maternal healthcare utilisation using demographic and health survey data. METHODS: This was a cross-sectional study of 195 651 women aged 15-49 y from 28 countries in SSA. We adopted bivariable and multivariable analyses comprising χ2 test and multilevel binary logistic regression in analysing the data. RESULTS: The prevalence of maternal healthcare utilisation was 58, 70.6 and 40.7% for antenatal care (ANC), skilled birth attendance (SBA) and postnatal care (PNC), respectively. The prevalence of health insurance coverage was 6.4%. Women covered by health insurance were more likely to utilise ANC (adjusted OR [aOR]=1.48, 95% CI 1.41 to 1.54), SBA (aOR=1.37, 95% CI 1.30 to 1.45) and PNC (aOR=1.42, 95% CI 1.37 to 1.48). CONCLUSION: Health insurance coverage was an important predictor of maternal healthcare utilisation in our study. To accelerate progress towards the achievement of SDG 3 targets related to the reduction of maternal mortality and achievement of UHC, countries should adopt interventions to increase maternal insurance coverage, which may lead to higher maternal healthcare access and utilisation during pregnancy. The study utilised pooled data from the most recent DHS of 28 countries in SSA. The data were extracted from the women’s files (IR Recode) of the selected countries. The DHS is a nationally representative survey usually conducted every 5 y in >85 LMICs.27 The survey employed a structured questionnaire to collect data from the respondents on health indicators such as maternal healthcare utilisation.27 The survey utilised a cross-sectional design in obtaining the data from the respondents. A two-stage sampling technique was employed to collect data from the respondents. The survey methodology and sampling process have been detailed elsewhere.28 In this study, a total of 195 651 women aged 15–49 y were included in the final analysis. Appendix 1 contains the sample distribution per country. We relied on the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) statement in writing the manuscript.29 The dataset is freely available for download at https://dhsprogram.com/ (accessed 28 May 2021). ANC attendance, SBA and PNC attendance were the outcome variables in this study. From the DHS, the women were asked about the number of ANC visits they made during their recent pregnancies. The responses were recoded as 0–3 ‘No’ and 4 and above ‘Yes’. For SBA, the women were asked ‘Who assisted [NAME] during delivery?’ The response options were regrouped into ‘Traditional Birth Attendant/Others’ and ‘SBA/Health professionals’. Regarding PNC attendance, the women were asked ‘Did [NAME] go for postnatal checks within 2 months?’ The response options to this question were ‘Yes’, ‘No’ and ‘Don’t know’. Those who responded ‘Don’t know’ were dropped. The recoding and categorisation used in the study were informed by previous literature.30–34 The explanatory variables included in the study were based on the review of pertinent literature and centred on the conceptual framework adopted for the study.30,32 Also, selection of the variables was based on their availability in the DHS dataset. Based on the constructs of the model, the predisposing variables consisted of the age of the woman, educational level, religion, marital status and parity. The enabling factors consisted of health insurance coverage, mass media, wealth status, place of residence, current working status, permission to go, money for treatment, distance to facility and geographical subregions. The morbidity and mortality to the child or the mother is the need for care factor. The explanatory variables have been further divided into the key explanatory variable and the covariates with their accompanying categorisation, as shown below. The key explanatory variable was health insurance coverage. This variable was assessed in the DHS using the question ‘Are you covered by any health insurance?’ The types of health insurance found in the DHS, which vary per country, include mutual health insurance, national or district health insurance, employer-based health insurance, social security and privately purchased commercial insurance, among others. The response options were ‘No’ and ‘Yes’. Studies have utilised this variable either as a key explanatory variable or as a covariate in determining maternal healthcare utilisation.1 A total of 13 variables were studied as covariates in the study. These variables were selected from the predisposing and enabling factors and were categorised into individual- and contextual-level factors, as shown below. Maternal age, level of education, marital status, religion, current working status, parity, getting medical help: permission to go, getting medical help: money for treatment and getting medical help: distance to the facility were considered as the individual-level factors. Except for parity, which was recoded as ‘1’, ‘2’, ‘3’ and ‘4 or more’, we utilised the existing coding for the remaining variables as found in the DHS datasets. The contextual variables were wealth index, mass media exposure, place of residence and geographical subregions. From the DHS, the wealth index was coded as ‘poorest’, ‘poorer’, ‘middle’, ‘richer’ and ‘richest’. The place of residence was coded as ‘urban’ and ‘rural’. Exposure to mass media was assessed using three variables (frequency of watching television, frequency of reading newspaper/magazine and frequency of listening to the radio). The response options in each of these variables were ‘not at all’, ‘less than once a week’, ‘at least once a week’ and ‘almost every day’. The response was recoded into ‘No’ (those who responded not at all and less than once a week) and ‘Yes’ (those that responded at least once a week and almost every day). An index variable called mass media exposure was created using the recoded responses from the three variables. Any woman whose response option was ‘Yes’ in each of the variables after the recoding was said to have been exposed to mass media. Those that responded ‘No’ in all three had no exposure to mass media.33 Also, the studied countries were further regrouped into the geographical subregions (Eastern, Central, Western and Southern) and used as a contextual-level factor. Data analysis was carried out using Stata software version 16.0 (Stata Corporation, College Station, TX, USA). The analysis was carried out in three levels. First, percentages were used to present the results of the health insurance coverage and maternal healthcare utilisation (ANC, SBA and PNC) (Table 1). Later, we performed a cross-tabulation to examine the distribution of ANC, SBA and PNC across health insurance coverage, individual- and contextual-level factors, as well as an estimated Pearson’s χ2 test of independence at p<0.05 to show significant variables (Table 2). The statistically significant variables from the χ2 test were placed in the regression model. A multilevel binary logistic regression using five models (Models O–IV) was used to examine the association between health insurance coverage and each of ANC, SBA and PNC, controlling for individual- and contextual-level factors (Tables 3–5). Prevalence of maternal healthcare utilisation and health insurance coverage Bivariable analysis of health insurance coverage and maternal healthcare service utilisation among women in SSA *p-values obtained from χ2 test. Fixed and random effects analysis on the association between health insurance coverage and ANC among women in SSA Abbreviations: AIC, Akaike's information criterion; ICC, intra-class correlation; LR test, likelihood ratio test; PSU, primary sampling unit. * p<0.05; ** p<0.01; *** p<0.001; 1=reference. Fixed and random effects results on the association between health insurance and PNC among women in SSA Abbreviations: AIC, Akaike's information criterion; ICC, intra-class correlation; LR test, likelihood ratio test; PSU, primary sampling unit. * p<0.05; ** p<0.01; *** p<0.001; 1=reference. Model O showed the variance in ANC, SBA and PNC attributed to the clustering of the primary sampling units (PSUs) without health insurance coverage and studied covariates. Model I was fitted to contain health insurance coverage alone. Model II was fitted to contain the individual-level factors. Model III contained the contextual-level factors. Model IV was finally fitted to comprise health insurance coverage, individual- and contextual-level factors. The Stata command ‘melogit’ was used in fitting the five models. Fixed and random effects were included in all five models. The fixed effects represented the relationship between the explanatory variable and/or covariates and the outcome variable, whereas the random effects represented the measure of variation in the outcome variable based on PSU, as measured by intra-cluster correlation (ICC). As a result, the ICC was used to quantify the differences between clusters in the sample used for the analysis. Finally, Akaike's information criterion (AIC) was used to examine the model fitness, or how the various models were fitted with the data. From the models, the one with the least AIC value was selected as the best-fitted model. Thus, the final models (IV) in Tables 3–5 were chosen for forecasting the association between ANC, SBA and PNC and health insurance coverage. All the regression results were presented using crude ORs (cORs) and adjusted ORs (aORs) at a 95% CI. The women's sample weight (v005/1000 000) was applied to the data to cater for the complex nature of the DHS dataset. The Stata command ‘svy’ was used to adjust for the disproportionate sampling and non-response and to improve the generalisability of the findings.