Spatial and multilevel analysis of unskilled birth attendance in Chad

Background: Unskilled birth attendance is a major public health concern in Sub-Saharan Africa (SSA). Existing studies are hardly focused on the socio-demographic correlates and geospatial distribution of unskilled birth attendance in Chad (a country in SSA), although the country has consistently been identified as having one of the highest prevalence of maternal and neonatal deaths in the world. This study aimed to analyse the socio-demographic correlates and geospatial distribution of unskilled birth attendance in Chad. Methods: The study is based on the latest Demographic and Health Survey (DHS) data for Chad. A total of 10,745 women aged between 15 and 49 years were included in this study. A multilevel analysis based on logistic regression was conducted to estimate associations of respondents’ socio-demographic characteristics with unskilled birth attendance. Geographic Information System (GIS) mapping tools, including Getis-Ord Gi hotspot analysis tool and geographically weighted regression (GWR) tool, were used to explore areas in Chad with a high prevalence of unskilled birth attendance. Results: The findings show that unskilled birth attendance was spatially clustered in four Chad departments: Mourtcha, Dar-Tama, Assoungha, and Kimiti, with educational level, occupation, birth desire, birth order, antenatal care, and community literacy identified as the spatial predictors of unskilled birth attendance. Higher educational attainment, higher wealth status, cohabitation, lowest birth order, access to media, not desiring more births, and higher antenatal care visits were associated with lower odds of unskilled birth attendance at the individual level. On the other hand, low community literacy level was associated with higher odds of unskilled birth attendance in Chad whereas the opposite was true for urban residency. Conclusions: Unskilled birth attendance is spatially clustered in some parts of Chad, and it is associated with various disadvantaged individual and community level factors. When developing interventions for unskilled birth attendance in Chad, concerned international bodies, the Chad government, maternal health advocates, and private stakeholders should consider targeting the high-risk local areas identified in this study. The study was based on secondary data obtained from the Chad 2014–2015 demographic and health survey (DHS) conducted from October 2014 to April 2015. The survey included a nationally representative sample of 17,719 women aged 15–49 years selected from 17, 233 households. Respondent selection was based on a two-stage stratified cluster sampling procedure. For the first stage, 626 enumeration areas were selected from a list of clusters nationwide [11]. Households were then selected from the complete list of households in each selected cluster during the second stage. For this study, the analysis excluded women who had never given birth within the five years preceding the survey in 2014–2015. The analytic sample, thus, was made up of 10,745 women of reproductive age whose last birth occurred during the five years preceding the survey. The outcome variable – unskilled birth attendance – was constructed from the categories used in the DHS concerning the person who assisted with the delivery of the respondent’s last child. In the DHS, the respondents were asked to identify the personnel that assisted them during delivery such as a doctor, nurse or midwife, community health officer or nurse, traditional birth attendant, traditional health volunteer, community or village health volunteer, relative, other, or no one. A binary outcome was constructed by categorising baby deliveries performed by traditional birth attendants, traditional health volunteers, community/village health volunteers, relatives, and others as unskilled birth attendance and the remaining health personnel as skilled birth attendance. The independent variables of the current study were mainly informed by findings of existing studies on unskilled birth attendance [12–14]. The independent variables considered in the current study are both individual and community level variables measured at the cluster level. The individual-level variables comprised socio-demographic characteristics such as age (15–19, 20–24, 25–29, 30–34, 35–39, 40–44, 45–49), educational level (No formal education, Primary, Secondary, Higher), wealth index (Poorest, Poorer, Middle, Richer, Richest), marital status (Never in a marital union, Married, Cohabitation, Widowed, Divorced, Separated), occupation (Not working, Working), birth order (1, 2–3, 4 +), media exposure (No, Yes), desire for birth (Then, later, no more), and antenatal care visits (Less than 4, 4 or more). Community variables were computed at the primary sampling unit level. Community socioeconomic status was computed from occupation, wealth, and education of study participants who resided in each sampled cluster as all community variables are for only women. Principal component analyses were applied based on women who were unemployed, uneducated, or from poor households. A standardised rating was derived with an average rating (zero) and standard deviation [15]. The rankings were then segregated into tertiles where the lower scores (tertile 1) denote high socioeconomic status, and average scores (tertile 2) as moderate socioeconomic status while the higher scores (tertile 3) denote lower socioeconomic status. Respondents who attended higher than secondary school were considered literate while all other respondents were given a sentence to read and were considered literate if they could read all or part of the sentence. Therefore, if respondents had higher than secondary education or had no school/primary/secondary education but could read a whole sentence, it was considered as high literacy. Medium literacy represented respondents who had no school/primary/secondary education and could read part of the sentence. Low literacy comprised respondents who had no school/primary/secondary education and could not read at all. Thus, socioeconomic status and literacy level were measured as low, moderate, and high at the primary sampling unit level (cluster) while place of residence was measured as urban or rural. The explanatory variables of the current study were mainly informed by findings of existing studies on unskilled birth attendance [12–14]. Two levels of analysis were conducted. A descriptive analysis was performed by calculating the proportion of respondents’ socio-demographic characteristics. Multilevel logistic regression analysis was conducted to estimate associations of respondents’ socio-demographic characteristics with unskilled birth attendance. The Model 1 (null model) was estimated to establish whether there is a significant variance in unskilled birth attendance at the cluster level to justify the multilevel analysis, while Model 2 comprises the fixed effects analysis of individual-level socio-demographic factors and assumes that these socio-demographic factors have fixed or constant association with unskilled birth attendance across all the primary sampling units. Model 3 estimated the effects of the community-level variables and assumed that the effects may vary across the primary sampling units. Model 4 contains the full model (Models 1, 2, and 3) and the mainstay of the analysis. Adjusted odds ratios and 95% confidence intervals were calculated for the variables. All analyses were conducted with Stata software (Version 14) and the results were weighted to cater for potential over-sampling and under-sampling. Regarding the spatial analysis, the coordinates of the surveyed respondents were obtained from the DHS website. These data were projected to UTM Zone 33 N to aid the spatial analysis. Relying on just the surveyed point was insufficient, so Chad’s departments’ shapefile was obtained from the Humanitarian Data Exchange for the spatial analysis [16]. The departments’ shapefile was also projected to UTM Zone 33 N. This dataset had 70 departments, but 55 departments were found to have had respondents sampled for the survey. Therefore, the 15 additional departments were excluded from the analysis. As part of the dependent variable’s data processing, birth attendance was coded 0 for skilled birth attendance and 1 for unskilled birth attendance. The independent variables were in their original categories, but proportions were generated from the lower categories at each sub-regional level. Therefore, the extracted surveyed data was linked with its corresponding coordinates. These data were then merged with the sub-regional shapefile using the join tool. In joining, the average computation system was adopted to estimate averages of unskilled birth attendance at the sub-regional level (departments). After obtaining the sub-regional level estimate of unskilled birth attendance, the independent variables’ proportions were added to asssist with the computation of the geographically weighted regression (GWR). The actual data analysis began with using the spatial autocorrelation tool (Moran’s I) to assess the distribution of unskilled birth attendance in Chad. The spatial autocorrelation assesses the distribution of a phenomenon being studied; thus, either random, dispersed, or clustered. The output does not show the exact sub-regional distribution of unskilled birth attendance. Therefore, the Getis-Ord Gi hotspot analysis tool was used to examine areas that had a higher tendency of experiencing unskilled birth attendance. A limitation of the hotspot tool is that it considers areas with high values of the dependent variables to create hot and cold spots. However, it is advisable to use the Anselin Local Moran’s I cluster and outlier analysis tool for policy implications to explore other areas that may not be identified as a hotspot but have a high incidence of the phenomenon being studied. The final analysis was to determine the spatial explanation of the independent variables by running the GWR. GWR is a spatial regression technique that focuses on the spatial differentiation in the explanatory power of independent variables. This is done by estimating separate regression equations between the dependent and independent variables to every feature in the dataset. Before running the GWR, significant independent variables were identified using the exploratory regression analysis. This regression method was adopted because it assesses all possible combinations of the independent variables that best explain the dependent variable. Thus, it looks for models that meet all of the requirements and assumptions of the OLS method. The first model of the fifth category was accepted since it had the criteria [AICc = -52.19, JB = 0.08, K(BP)0.02, VIF = 1.71, SA = 0.46] that contributed to the highest adjusted R2 (0.32). This revealed the best combination of independent variables that are significant predictors of unskilled birth attendance in Chad. This allowed the use of the identified independent variables that were used for the GWR. In conducting the GWR, the identified independent variables found to be significant predictors of unskilled birth attendance in Chad were used as the explanatory variables with kernel type being fixed as well as bandwidth method using the AICc. The obtained adjusted R2 was about 0.309 percent which implies that the results explain about 31% of the entire data. All the data processing and analyses were conducted in ArcGIS version 10.7. The Likelihood Ratio (LR) test was used to evaluate the fitness of all the models. Before fitting the models, the presence of multicollinearity between the independent variables was evaluated. The variance inflation factor (VIF) test found the absence of high multicollinearity between the variables (Mean VIF = 1.94).