Spatial access inequities and childhood immunisation uptake in Kenya

Background: Poor access to immunisation services remains a major barrier to achieving equity and expanding vaccination coverage in many sub-Saharan African countries. In Kenya, the extent to which spatial access affects immunisation coverage is not well understood. aim of this study was to quantify spatial accessibility to immunising health facilities and determine its influence on immunisation uptake in Kenya while controlling for potential confounders. Methods: Spatial databases of immunising facilities, road network, land use and elevation were used within a cost friction algorithim to estimate the travel time to immunising health facilities. Two travel scenarios were evaluated; (1) Walking only and (2) Optimistic scenario combining walking and motorized transport. Mean travel time to health facilities and proportions of the total population living within 1-h to the nearest immunising health facility were computed. Data from a nationally representative cross-sectional survey (KDHS 2014), was used to estimate the effect of mean travel time at survey cluster units for both fully immunised status and third dose of diphtheria-tetanus-pertussis (DPT3) vaccine using multi-level logistic regression models. Results: Nationally, the mean travel time to immunising health facilities was 63 and 40 min using the walking and the optimistic travel scenarios respectively. Seventy five percent of the total population were within one-hour of walking to an immunising health facility while 93% were within one-hour considering the optimistic scenario. There were substantial variations across the country with 62%(29/47) and 34%(16/47) of the counties with < 90% of the population within one-hour from an immunising health facility using scenarios 1 and 2 respectively. Travel times > 1-h were significantly associated with low immunisation coverage in the univariate analysis for both fully immunised status and DPT3 vaccine. Children living more than 2-h were significantly less likely to be fully immunised [AOR:0.56(0.33-0.94) and receive DPT3 [AOR:0.51(0.21-0.92) after controlling for household wealth, mother’s highest education level, parity and urban/rural residence. Conclusion: Travel time to immunising health facilities is a barrier to uptake of childhood vaccines in regions with suboptimal accessibility (> 2-h). Strategies that address access barriers in the hardest to reach communities are needed to enhance equitable access to immunisation services in Kenya. Health facilities that offer immunisation services were sourced from the Kenya Master Health Facility List (KMHFL) [31] and the Kenya Health Information System (KHIS) [31, 32] based on the District Health Information Systems version 2(DHIS2). Facilities that offer immunisation services based on the reported number of vaccinations during 2012–2014 period were identified from the KHIS and merged to KMHFL list to obtain coordinates. Where coordinates were missing, a previously geocoded health facility list was used [33]. The final list of health facilities obtained covered the whole spectrum of health facility levels and ownership status, comprising both public and private facilities. Relevant ancillary datasets of factors that influence travel speeds including road network, land cover and digital elevation model (DEM) were assembled nationally. Road network data was obtained from the ministry of transport of Kenya that used the gold standard GPS technique to map coverage of roads in 2016 [34]. This was overlaid with roads obtained from OpenStreetMaps (OSM) and Google Map Maker (GMM) [35, 36] and combined using ArcMap version 10.5 (ESRI Inc., Redlands, CA, USA). We eliminated duplicates, corrected for road sections with short connection segments due to digitization and deleted those that extended to water bodies from the resultant vector file. Roads were classified as primary, secondary, county and rural roads [37]. Land cover was based on 2016 Copernicus Sentinel-2 satellites at 20 m × 20 m spatial resolution available from RCMRD GeoPortal [38]. It contained five land cover categories namely; bare areas, built up areas, water bodies, cultivated areas and vegetation cover areas (forests, shrubs and grassland areas). Major rivers and lakes available from global lakes and wetlands database [39] were considered as barriers to movement (except in the presence of bridges informed by the road data set). Forty-nine protected areas [40–42] were considered unpassable and treated as barriers as shown in the additional file 1. The DEM from Shuttle Radar Topographic Mission (SRTM) 30 m × 30 m spatial resolution archived at RCMRD Geoportal [38] was used to account for the influence of topography on walking and bicycling speeds. Several modelled population distribution datasets exist [43]. However, these datasets include covariates such as the density of health facilities and road networks to determine regions likely to be inhabited. To avoid model induced correlation, such as circularity when estimating travel time to health facilities, a population distribution map that excluded density of health facilities and road networks was constructed using dasymetric spatial modelling techniques. Kenya’s 2009 census population data were redistributed at enumeration areas (EA) to 100 m square grids and projected to 2014. During the redistribution, the grids were weighted based on the probability of being inhabited and the relationship between population density [44]. The weights were then used in a random forest technique while adjusting for rural-urban differences to obtain population at the 100 m square grids [45]. Landuse, road network and travel barriers (protected areas and water bodies) were rasterized, resampled to 100 m square grids and combined in ArcMap 10.5 (ESRI Inc., Redlands, CA, USA). The resultant raster was used to generate a cost raster surface with impedance value based on cumulative speeds at each 100 m square grids predetermined spatial grids. Travel time to the nearest health facility was computed using the generated cost surface and locations of the health facilities through each square grids from all areas in Kenya on a regular raster grid using AccessMod 5.0 [46, 47]. Slope derived from DEM was used to adjust walking speeds using Tobler’s formulation [48] and to adjust for bicycling speeds using bicycling power correction [49, 50]. Two possible travel scenarios typically used by the Kenyan population to access health facilities were assessed; one where we assume walking only scenario and a second more optimistic travel scenario that assumed the population walks to the nearest road and takes a different mode of transport immediately available depending on the terrain and the available road infrastructure as shown in additional file 2. The walking scenario was important since most of the population use walking as the main mode of transport, especially in the rural areas [28, 51] where about 73% of people reside in Kenya [52]. In addition, it facilitated the evaluation of the government policy, for a threshold of 90% of the people within an hour of walking to the nearest facility [26]. Input travel speeds for each road type and landcover were adopted from previous work in Kenya [37, 51] and refined through a discussion with the National Vaccination and Immunization Programme (NVIP) staff from five counties and the national offices in Kenya [53]. The output of the accessibility analysis was two continuous surfaces depicting the theoretical time it would take to get to the nearest immunising health facility for walking only and a combination of walking and motorized travel models. The travel time was depicted in minutes at a spatial resolution of 100 m square grids for the entire country. The geographical coordinates for sampled clusters in Kenya Demographic Health Survey 2014 (KDHS 2014) were used to extract travel times for each child needing immunisation. Since the cluster coordinates are randomly perturbed by up to 5 km in rural areas and 2 km in urban areas [54], 5 km and 2 km buffers were drawn around the rural and urban clusters respectively and mean travel times extracted within the buffers. Maps of travel time to the nearest immunising health facility at 100 m square grids and the average time per cluster were then plotted in ArcMap version 10.5 (ESRI Inc., Redlands, CA, USA). Using the continuous travel time surfaces from the walking only and the optimistic cost analysis, we computed the proportion of the total population at county levelwithin 1-h to the nearest immunising health facility. Data on immunisation coverage for children aged 12–23 months and its predictors collected from women aged 15–49 were based on KDHS 2014 conducted between May and October 2014. KDHS 2014 employed a two-stage sampling design where 1612 clusters were selected in the first stage while 40,300 households were selected in a second stage [25]. This is the largest sample household survey to ever be conducted in Kenya. The main outcomes were DPT3 vaccination status and fully immunised child (FIC) status defined by KEPI [55] as having received: one dose each for Bacille Calmette-Guerin (BCG) and measles, DPT3, polio (excluding polio at birth) and pneumococcal vaccines from either the child’s vaccination card or by mother’s recall for children aged 12–23 months. Fully immunised status was used to assess the overall impact of travel time to health facilities on immunization coverage. The antigens contained in FIC status are either delivered through stationary health facilities or via supplemental immuinisation activities. Therefore, the effect of travel time might be diminished. To mitigate against the attenuated effect, DPT3 vaccine was used to validate the impact of travel time to health facilities because it is offered through stationary health facilities which are not influenced by supplemental immunisation activities. Computed travel time to the nearest immunising health facility was the primary explanatory variable of interest. Potential confounders found in the literature related both utilization of healthcare services and travel time to health facilities were identified, reviewed and included based on data availability. Data were abstracted from KDHS 2014. The covariates considered as confounders were mother’s education level, person who decides on mothers/child healthcare seeking, parity, residence type, marital status, mother’s age and household wealth index [56] . Each variable was categorized based on a literature review assessing the association between immunisation coverage and its determinants [11–16]. Among the abstracted confounders, wealth index describing social economic status across households and parity were derived from a combination of several indicators. Wealth quintiles (index) derived from the DHS measures the relative socioeconomic status of households based on household assets and amenities at the time of the survey using principal component analysis. The wealth quintiles were classified into poorest, poor, middle, rich and richest [57]. Since maternal age and parity are highly correlated, parity was categorized into two groups as follows; high parity (if mother’s age < 30 years and has more than two children living in the household or is aged ≥30 years and has more than three children living in the household and low parity otherwise. Although datasets may be available from other sources, we restricted the sample to KDHS 2014 for consistency and a common period. Proportions were computed to describe the characteristics of data in relation to FIC status and DPT3. We estimated crude associations between the dependent or confounding factors and the two outcomes FIC status and DPT3. Confounders that were significant at the cutoff (p < 0·20) in the crude analysis were incorporated into the multivariable regression analysis. The Person who decides on mothers/child healthcare seeking variable was excluded from the analysis as it was asked for a subset of women (married mothers) in the sample. Computed travel time from the optimistic scenario was used in the analysis as it has been shown to provide a more realistic estimate of travel times [58]. We anticipated unmeasured effects at the county, community and individual levels due to the hierarchical data structure of DHS data [57]. A null model with no explanatory variables was fit and the intracommunity correlation coefficient (ICC) was used to assess clustering at county and cluster levels. Therefore, we conducted the analysis with counties as level 1, clusters as level 2 and individuals (children) as level 3. A multi-level logistic regression model was used to assess the relationship between immunisation coverage and travel time to immuinisation facility when controlling for potential confounders. Travel time to health immunising facilities was added first followed by the confounders in an increasing order of their p values from bivariate analysis. We also examined statistical interactions based on a priori hypotheses that the effect of travel time to immunising health facility may be different for rich and poor households as well as between travel time and urban/rural residence. To define the final model, we reported adjusted odds ratios with their 95% confidence intervals, and Wald test (p-value < 0.05) to inform the overall significance of the models. Multicollinearity test to evaluate associations among the independent variables was assessed using Variance Inflation Factor (VIF) at cut off point of 5 [59]. The analyses were done using STATA v.14 (Stata Statistical Software: Release 14. College Station, TX: StataCorp LP) and KDHS 2014 sampling weights were incorporated throughout the analysis.