Spatial models for the rational allocation of routinely distributed bed nets to public health facilities in Western Kenya

Long-lasting insecticidal nets (LLINs) are distributed as the main tool of vector control for malaria prevention in Africa, through mass campaigns and routine distribution systems [1]. Mass campaigns, target population at community level in a single time-limited operation every 3 years. Routine systems deliver LLINs through public health sector all year round to sustain coverage in the interval between mass campaigns to pregnant women and infants who are the most vulnerable. The World Health Organization (WHO) recommends both channels for countries in order to achieve and maintain universal coverage [1, 2].

At the launch of Kenya’s national malaria strategy in 2001, the aim was to achieve 60% insecticide-treated nets (ITNs) coverage of at risk children and pregnant women by 2006 [3]. Prior to this, access to ITNs was limited to the private sector, a few research projects and non-governmental organizations. Different channels of ITNs delivery were explored including commercial retail sector ITNs and subsidized ITNs through public health sector. These approaches did not succeed in reaching the rural poor and achieving maximum coverage of the population at risk [4]. In 2004, free routine LLIN distribution in antenatal care clinics (ANC) began and in 2006 the first free mass-campaign was conducted. Subsequent free mass campaigns were conducted in 2011, 2012, 2014 and 2015 in malaria endemic areas of Kenya [5].

Routine distribution of LLINs in Kenya are targeted to high and moderate malaria transmission areas, where pregnant women are provided with a free LLIN during their first ANC visit so that the mother and the unborn baby are protected at the earliest chance. Infants are provided with a free LLIN during expanded programme on immunization (EPI) [6, 7]. There is no specific WHO recommendation as to when LLINs should be distributed during EPI. However, a range of time-points is used in practice, from birth (tuberculosis vaccine) to 9 months (measles vaccine). Diphtheria-tetanus-pertussis-1 vaccination at 6 weeks is the most common distribution point [7, 8]. ANC clinics that distribute LLINs have increased rapidly from 1000 nationally in 2004 to over 4000 in 2015, with a total 23.3 million LLINs distributed [5].

Pregnant women receive a package of interventions during ANC visits which play an important role in ensuring a healthy mother and baby during pregnancy and after delivery [6, 9]. Healthcare utilization for ANC services is associated with intrapersonal, institutional, health systems and social demographic factors [10, 11]. Similarly, the ownership and use of LLINs during pregnancy is associated with household wealth, distance to health facility, residence, parity, marital status, and education among others [12]. It is key, therefore, to include these spatial variables to account for the uneven healthcare utilization rates within facility catchment areas within a spatial modelling framework [13, 14].

The quantity of LLINs distributed to clinics is quantified at the central level using average monthly LLINs consumption data and the overall size of the population in targeted counties. However, this approach has resulted in frequent stock-outs [7, 15] with subsequent low population coverage of LLIN among pregnant women and infants [8, 15, 16]. In this study, a spatial modelling framework was proposed and used to quantify LLINs need by ANC clinic to allow more efficient targeting. A combination of modelled spatial access to clinics, household survey data on utilization and high resolution population grids were used to quantify the population of pregnant women and infants in 2015 in need of LLINs. The estimated need was compared to the actual allocation in 2015 and clinics with fewer (under-allocation), more (over-allocation) and enough (matched-allocation) LLIN relative to the estimated need identified. A fourth category is those facilities with zero distribution (non-allocation). Approaches to minimize targeting inefficiencies were then proposed.

The final list contained 1271 public health facilities out of which 888 had distributed LLINs in 2015. The analysis of the KDHS 2014 data showed that all 277 cluster points were within approximately 1-h of the nearest public clinic distributing LLINs (Fig. 3). ANC utilization decreased with increasing travel time, with attendance declining rapidly for distances greater than 40 min away from the nearest clinic on the decay curve. Catchment areas defined to include all those locations within 40 min of a health facility for the 888 public clinics distributing LLINs (exclusive of 13 clinics in islands of Lake Victoria) are shown in Fig. 4 and contained over 97% of pregnant women and infants in 2015 in the study area.

In the study counties, among women who had at least one live birth in the 5 years preceding the survey, over 97% had at least one ANC visit. As expected and exhaustively documented in literature within the Kenyan context [53‐55], the final model of variables associated with ANC utilization showed that women of higher education and wealth quintile, those who were married, or had low parity or lived closer to a clinic were associated with higher odds of attending ANC (Table 2). All these variables varied spatially and this structure was used in interpolating them across the study area with 81–97% of pregnant women attending ANC clinics.

In 2015, 730,947 pregnant women and infants resided in the eight study counties with 97% (706,450) residing within modelled clinics catchment areas. The estimated LLINs need per clinic for 2015 was between 15 and 16,900, with 48% having a need of less than 500 while 23% with a need of more than 1000 LLINs annually. Among pregnant women and infants in catchment areas, 97% (682,377) were likely to have attended a clinic distributing LLINs representing 93% of the total population of pregnant women and infants in the study area.

In 888 public health clinics, 591,880 LLINs were allocated in 2015. The number of LLINs allocated per clinic was between 10 and 6720. 53% of the clinics had less than 500 LLINs allocation per clinic while 19% had allocations greater than 1000 LLINs. Relative to an estimated overall need for 682,377 LLINs, the un-met need for 2015 was 90,497. The model suggested that 380 (43%) clinics were under-allocated by 255,628 LLINs (Fig. 4). The under-allocation per clinic was between 100 and >2000 LLINs (Fig. 5b). In 321 (36%) clinics (Fig. 4), 164,241 LLINs were over-allocated by up-to over 2000 LLINs (Fig. 5a) while 21% (187) of the clinics had a matched-allocation accounting for 86,990 LLINs (Fig. 4).

Using the spatial modelling framework to simulate an efficient LLIN allocation system, the 164, 241 over-allocated LLINs were redistributed reducing the shortfall from 255,628 to 91,387. Of the 95 areas that were outside 40 min and thus not covered by the LLIN distribution in 2015 (Fig. 4), 34 had a population of ≤100 and could be feasibly absorbed within the existing neighboring clinics resulting in increasing allocation of about 1606 LLINs to these clinics. 43 clinics not allocated LLINs in 2015, were within the determined 40 min cut-off time and would cater for 43 of remaining polygons equivalent to 17,703 extra LLINs. The remaining un-covered areas (18 polygons) all located between 40 and 60 min of the closest clinic distributing LLINs would require 5187 LLINs and accounts for less than 1% of those likely to attend a clinic.

A spatial modelling framework that can be adapted for the rational allocation of routinely distributed LLINs in Kenya was developed. LLIN need at each clinic was quantified by using a combination of modelled accessibility, population and ANC utilization and compared this to the actual allocations in 2015. Overall, the unmet need of LLINs for all pregnant women and infants and those likely to attend public clinics was 139,067 and 90,497 respectively. The current quantification approach of LLINs need covers approximately 88% of those likely to attend ANC clinics and about 81% of all pregnant women and infants in the study area. This shows that the routine distribution system in Western Kenya has a wide reach and is consistent with the Government of Kenya target of at least 80% of at risk population using appropriate malaria preventive intervention [56]. However, the facility level analysis also shows that there are important inequities in the distribution of LLINs to pregnant women and infants. The model provides a means for reducing such inequities and can be upgraded as consumption data improves.

Currently, the Ministry of Health relies on consumption data reported from health facilities and adjusted for estimated census projections of pregnant women and infants to quantify routine LLIN distribution. At delivery points, mechanisms for rational distribution have been rolled out including verification that a client qualifies for an LLIN, stamping the recipients’ booklet to avoid repeat collection from the same or another clinic and improved overall training of health workers on the management of commodities, regular auditing at the clinics and monthly reports. Recently, with the introduction of DHIS 2 [21, 57] in Kenya, LLINs consumption data are now reported through this system. This could potentially improve rational allocation, especially if predefined catchment areas and population maps such as those developed in this analysis are used as additional information to check whether distributions match the need for LLINs.

The definition of catchment areas could either be theoretical and based on simple proximity of the population to a health facility [58, 59] or could be improved with the available empirical data on utilization of services derived from household survey data rates [13, 41]. Previous studies have used only travel time and its relationship with reported utilization to define the probability of service use [13, 32]. However, in this study, other determinants of utilization such as wealth and maternal education, were used to capture both the spatial and social, demographic and economic factors that influence the use of ANC clinics. Arguably, this approach offers a more complete framework for quantifying the catchment population (likely ANC clients) with a broader catchment area [14].

Some important limitations are worth noting in the modelling framework used. First, the use of the nearest clinic when modelling travel times was assumed. This is because household survey data does not allow for the identification of the actual facility used by women and infants. It is likely that a proportion of these populations often use facilities that are not the closest for a variety of reasons, including the quality of services on offer [60]. In addition, all pregnant women were assumed to have an equal probability of getting an LLIN irrespective of the number of ANC visits (1+ ANC), however, pregnant women with 2 or more ANC visits may have an increased chance of receiving an LLIN in subsequent visits. Second, a threshold of 40 min was imposed in the definition of facility catchment areas but it is likely that this threshold is variable by facility and future models should account for them. Finally, there are uncertainties in the pregnancies and births density maps [61] used into compute the LLIN need. Live births were assumed to be are equivalent of infants, however, a neonatal mortality of 22 per 1000 live births in 2015 in Kenya [30] equivalent to 6429 deaths (2.2%) of all live births and 1.1% of all LLINs distributed in the study counties may empirically bias the estimation of need. However, the national policy planning makes allocation of live saving services on the basis that all avoidable deaths are averted and therefore services, such as LLINs, will not and should not be discounted on the basis on anticipated neonatal deaths. To ensure that all uncertainties are propagated within the modelling approach, Bayesian inference methods may provide a useful alternative, but these advantages must be balanced with the simplicity required in spatial modelling techniques in routine programme planning.

The spatial modelling approach presented here offers a rational basis for spatial allocation of routine LLINs that is applicable to malaria endemic settings and, potentially, other maternal and child related commodities. The analysis shows that with a few minor adjustments in distribution and better rationalization, equity in LLIN distribution through the routine system could be achieved in Western Kenya where the threat of malaria is the greatest.