In this study, we investigated the relationship between child mortality and physical access to health care in a low-income setting that has recently seen substantial investments in health infrastructure. We found two-fold variations in observed infant and under-5 mortality across the 900 km2 Epi-DSS area. However, our data did not lend support to the widely held notion that mortality increases with distance to hospitals and vaccine clinics: in children 1 to 4 years old, mortality increased with vehicular travel time to hospitals; in infants under-1 year old, mortality decreased with increasing vehicular travel time to vaccine clinics; in each case the effects were small and no associations were observed in other age groups.
These findings contrast with published analyses from rural Tanzania , the Democratic Republic of Congo  and Burkina Faso  showing a strong relationship between mortality and distance to health facilities but are consistent with another DSS study from the Gambia . In Burkina Faso and the DRC, the density of health facilities was significantly lower than in our setting, with 35% of families residing more than 10 km from the nearest clinic and 65% more than 5 km away, respectively, as compared to 0% more than 10 kms' and 22% more than 5 kms' distance in Kilifi. The study populations in Tanzania and the Gambia had relatively similar physical access to care to the population in Kilifi: in both studies, mortality increased with distance in univariate models; only the Gambia study presented a multivariable analysis in which this effect disappeared. In Kilifi, there was no relationship between travel time to hospital or vaccine clinics and mortality overall, but children living more than 2 hours by vehicle from the hospital had worse survival than those living less than 2 hours away. Together, these findings suggest that the high density of health services available in our study area may explain the lack of an association between travel time to health facilities and mortality. Because the Kilifi DSS is representative of Kenya as a whole in terms of physical access to health care with approximately two-thirds of the population within one hour's walk of a primary care facility , we expect these results to be generalizable to most of the country.
Several sources of bias and confounding may have influenced our results. First, methodological errors may have led us to inaccurately estimate mortality risk in this population. The lack of an association between travel time to hospitals and infant or under-5 mortality and the negative relationship between vehicular travel time to vaccine clinics and infant mortality were partly driven by a high hazard of death in Kilifi Township, which was concentrated in the early neonatal period. We conjectured that women with high-risk pregnancies may migrate from outlying areas into town in order to give birth at the hospital, leading to increased neonatal mortality in town. However, location-specific mortality and survival patterns did not change when we excluded children whose mothers had migrated in the three months prior to giving birth, suggesting that pregnancy-related migrations did not bias this analysis. While we cannot rule out other errors in data collection or cleaning procedures, these are unlikely to vary spatially and should therefore not affect our results.
Second, the assumptions underlying our travel time models may require refinement. Significant effects were seen for vehicular travel time to hospital in older children and for vehicular time to vaccine clinics in infants. Pedestrian and vehicular travel times to both hospitals and vaccine clinics were 70% correlated (Spearman's rank correlation rho = 0.69, p < 0.01). The absence of an effect for pedestrian travel time to hospital or vaccine clinics may reflect high levels of vehicular transport usage in the Epi-DSS. However, data from other Kenyan districts suggest that a majority of patients walk to the hospital, irrespective of distance [13, 27]. The choice of transport mode may depend on a variety of considerations such as distance, availability of disposable income to cover matatu costs (from 20 to 120 Kenyan shillings, or US$0.26 to 1.56 per trip) and perceived severity of a child's illness. Further, even if theoretical travel times to the nearest clinic are accurate, they may not reflect actual travel times, as families are likely to use more distant clinics thought to provide higher quality services based on drug availability, staffing and other factors [27–29]. Detailed studies of matatu, bicycle, and private vehicle usage patterns as well as health facility choice and its relationship to service quality are necessary to improve upon the travel time models proposed here .
Third, we were unable to account for a number of possible confounders of the relationship between travel time and mortality. Data from the antenatal clinic at Kilifi District Hospital have shown an HIV prevalence of 5 to 7% over the past five years, but geographically stratified data are lacking. In most of Africa, HIV prevalence is highest in urban areas [31, 32] and in close proximity to roads [33, 34]. This could drive the high infant mortality rates in Kilifi Township and at shorter vehicular travel times to health facilities, negating or even inversing the effect of distance on mortality. Controlling for HIV in an individual-level analysis would require knowledge of the HIV status of all residents, since sublocation-level variables may mask heterogeneity within small areas; obtaining this information may not be feasible or ethical. Given the higher prevalence of HIV among immigrants from Western Kenya (primarily of Luo origin) than in the local population, adjusting for ethnicity may diminish but is unlikely to fully eliminate this source of confounding. Socio-economic data from the Epi-DSS area were not obtainable for individual residents, and we resorted to sublocation-level maternal education as a proxy variable. Maternal education has been shown to correlate highly with traditional measures of socio-economic status such as income and expenditures. However, we were unable to capture socio-economic inequalities within sublocations, which can be substantial (personal communication: C. Molyneux) and have a strong impact on mortality [9, 10]. Finally, other spatial factors may confound the association between travel time to health facilities and mortality. Ecological features affect the risk of childhood infectious diseases, such as pneumonia, malaria or diarrhea . Socio-behavioral characteristics determine adherence to various risk-reduction interventions, such as the use of insecticide-treated nets [18, 36]. We were unable to correct for these and other sources of clustering in mortality risk in our models. Further analyses adjusting for these factors or stratified by cause-of-death should be conducted when additional individual-level data become available.