Background: The majority of maternal deaths, stillbirths, and neonatal deaths are concentrated in a few countries, many of which have weak health systems, poor access to health services, and low coverage of key health interventions. Early and consistent antenatal care (ANC) attendance could significantly reduce maternal and neonatal morbidity and mortality. Despite this, most Kenyan mothers initiate ANC care late in pregnancy and attend fewer than the recommended visits.Methods: We used survey data from 6,200 pregnant women across six districts in western Kenya to understand demand-side factors related to use of ANC. Bayesian multi-level models were developed to explore the relative importance of individual, household and village-level factors in relation to ANC use.Results: There is significant spatial autocorrelation of ANC attendance in three of the six districts and considerable heterogeneity in factors related to ANC use between districts. Working outside the home limited ANC attendance. Maternal age, the number of small children in the household, and ownership of livestock were important in some districts, but not all. Village proportions of pregnancy in women of child-bearing age was significantly correlated to ANC use in three of the six districts. Geographic distance to health facilities and the type of nearest facility was not correlated with ANC use. After incorporating individual, household and village-level covariates, no residual spatial autocorrelation remained in the outcome.Conclusions: ANC attendance was consistently low across all the districts, but factors related to poor attendance varied. This heterogeneity is expected for an outcome that is highly influenced by socio-cultural values and local context. Interventions to improve use of ANC must be tailored to local context and should include explicit approaches to reach women who work outside the home. © 2013 Prudhomme O’Meara et al.; licensee BioMed Central Ltd.
The six districts included in this analysis – Bungoma East, Burnt Forest, Kapsaret, Bunyala, Chulaimbo, and Teso North – are distributed across western Kenya (Figure 1) but are all part of the catchment area of the Academic Model Providing Access to Healthcare (AMPATH). Each district has a district hospital and between 7 to 11 peripheral health facilities within the district. The districts vary in ethnic composition, population density, economic activities, access to health services, and disease burden. Burnt Forest and Kapsaret are more urban than the other districts and are predominantly settled by the Kalenjin people, although the town centers are more diverse. Malaria transmission is low to absent. Chulaimbo is populated mostly by the Luo ethnic group. Bunyala lies on the shores of Lake Victoria and the main economic activity is fishing. Most people belong to the Luhya ethnic group with a smaller number of Luo families. Malaria transmission in Bunyala and Chulaimbo is intense and perennial and the HIV burden is also very high. In Bungoma East, sugar cane is an important cash crop and most families engage in cash crop farming as well as subsistence farming and animal husbandry. The major ethnic group is the Bukusu which is a subtribe of the Luyha. Teso North district borders Uganda and is inhabited by the Teso people. Malaria transmission in Teso North and Bungoma East is also very high and transmission is experienced year round. Unlike Chulaimbo and Bunyala, the HIV burden is much lower. Percentage of women (age 13-45) who report being pregnant, by village. Pregnant women were identified during a large home-based HIV counseling and testing program (HBCT) in six districts in western Kenya conducted between 2009–2011. Women of reproductive age were asked whether they are currently pregnant and if so, whether they had received antenatal care for their current pregnancy. Our sample consists of women over the age of 13 years in each district who self-reported their pregnancy in the survey. At each stage of our analysis, the population of pregnant women is always the denominator. Details of the HBCT program and data collection are presented in detail elsewhere [22]. As part of the survey, basic demographic information, GPS coordinates, and socioeconomic information were collected at each household. Individual information such as age, marital status, and working hours were recorded for those who consented to counseling. HBCT is a home-based public health initiative. All participants gave voluntary informed consent for HIV testing. Consent was obtained verbally prior to data collection or any test being conducted. In the context of a community health initiative, written consent was not considered appropriate. Verbal consent is considered the norm for most clinical care procedures and activities in our region. Documentation of verbal informed consent was collected by recording who had accepted household entry and testing. The Institutional Review and Ethics Committee at Moi University and Moi Teaching and Referral Hospital in Eldoret, Kenya and Duke University Institutional Review Board approved the use of de-identified data from this program for analysis and publication. Spatially derived variables associated with geographic access to ANC care were: calculated distance to the nearest national road (Africover; last updated 2003), distance to the nearest health facility, and the type of facility nearest a woman’s household. Proximity to national roads and health facilities could conceivably have a non-linear relationship with use of antenatal care, therefore we also tested the use of a quadratic distance term as well as a log distance term as alternatives to linear distance, however, the non-linear terms did not improve model fit so a linear term was used. For five districts, distances were calculated as Euclidean distance which was necessitated by a lack of geographic data on local roads and footpaths, the most likely route of travel to a health facility. Although not a precise measure of travel distance, straight-line distance still provides an acceptable proxy for a woman’s proximity to resources for ANC care. In Bunyala, spatially derived variables were more complex. This district is bisected by two main rivers and several areas of the district are located in close proximity to wetland areas likely to limit a woman’s access to care. Based on knowledge of local geography, we assumed that women would not access care by crossing either of the two main rivers going south towards the more rural part of the district, though they may cross the rivers travelling north towards the hospital and main town center. Furthermore, when a river needs to be crossed, an individual must use an established crossing (bridge or ferry point) to safely traverse the river in order to reach a road or a health facility. Thus when the closest road or facility was located across a river to the north, the distance was calculated first from the household to the river crossing, then from the river crossing to the point of interest. If this distance was still the shortest distance to a road or facility, then this was the distance used for the derived variable of interest. The survey included location information such as broad administrative district, location, sublocation, and village, in addition to household GPS coordinates. Because the village unit is considered an important geographic as well as cultural unit, and maps of village boundaries are not available, village boundary maps were constructed by the study team using households’ GPS coordinates and village affiliations listed in the survey. Thiessen polygons were constructed from the household coordinates and then dissolved by village and clipped using the administrative boundaries of each district. Due to error in some GPS coordinates and use of village cultural association in place of geographic village location, village shapefiles were inspected and edited by hand where errors were apparent. All geographic data manipulation was conducted in ArcGIS 10. We conducted a preliminary exploratory spatial analysis by testing the population at risk of ANC non-attendance (in our case the population of pregnant women) in each district for autocorrelation in ANC attendance. Since our outcome is binary, we chose to run a join count test [23,24]. Our version of the join count looks at all pairs of neighboring points (locations at which a pregnant woman resides). The neighboring points will be one of three possible combinations: two pregnant women who both attend ANC, one pregnant woman who attends ANC and one who does not, and two pregnant women who do not attend ANC. Because we are interested in autocorrelation of ANC attendance, the pairs are designated as positive “joins” if they are of the same type (in our case, a pair of women who attended ANC). The test statistic is computed by taking the difference between the expected number of joins (given the total number of pairs of neighboring women and the total ratio of women attending ANC versus not) and the actual number of joins and determining whether the difference deviates significantly from a pattern that would be expected under spatial randomness. A higher than expected number of joins of the same type suggests positive spatial autocorrelation. A weights matrix is required which defines whether a relationship exists between two observations, in other words whether they are defined as ‘neighbors’. Neighborhoods (or definitions of “neighbor”) can be defined by proximity, e.g. some number of nearest neighbors to an observation or a distance band, which defines a set radius around an observation and considers every observation within the area to be a “neighbor.” In computation of autocorrelation statistics such as the join count, it is common to specify multiple levels and types of weight matrices [23]. We defined neighborhood in 4 ways; 8 nearest neighbors, 10 nearest neighbors, the maximum distance to the nearest neighbor, and 1.5 times the maximum distance to nearest neighbor. We used the spdep package in R 3.0 [25] to compute the join count test statistic and p-value which tests only for positive autocorrelation. Our final sample for regression analysis was composed of the population of pregnant women in each district. Bayesian models were specified at each level with and without uncorrelated heterogeneity (an independently distributed random effect) and with or without correlated heterogeneity (spatial autocorrelation term, using Besag’s model). The spatial random effect, via Besag’s model [26,27] defines a spatially structured residual, based on contiguity of village areas, using an intrinsic conditional autoregressive (iCAR) structure. Spatial random effects were specified with binary weights matrices based on queen contiguity (two villages are neighbors if they share at least one point on their borders) of village polygons. The uncorrelated effect adds an unstructured residual with exchangeable correlation between villages. For ease of interpretation of village level heterogeneity, we calculated median odds ratios (MOR) for the random effects of each of the districts. MORs can be interpreted as the odds ratio for the difference in likelihood of adherence between two persons with the same covariates, one from a village of higher adherence and one from a village of lower adherence [28]. The MOR is a more commonly used measure of variation between random effects in regressions with binary outcome variables [29,30] as it provides a more intuitive interpretation than the variance of the random effects. Logistic regression models were estimated using the Integrated Nested Laplace Approximation (INLA) package in R 3.0 [31]. INLA is a relatively new method used to estimate latent Gaussian Markov Random Fields and is a computationally efficient alternative to MCMC methods for estimating Bayesian models including those with spatial heterogeneity [32]. Models were compared using the Deviance Information Criterion (DIC) and Moran’s I test of the deviance residuals. DIC considers the posterior mean deviance and adjusts for model complexity via the numbers of effective parameters. DIC is often used to compare and evaluate the fit of variations of the same regression model in Bayesian analysis and smaller scores are preferred [33]. We used Moran’s I test for autocorrelation of the deviance residuals to determine whether additional spatial autocorrelation remained after adjusting for covariates or adding village-level uncorrelated heterogeneity effects. We also compared these results to Moran’s I test of the deviance residuals of the unadjusted model [34,35].