Study sites
This analysis used data from health facility-based malaria surveillance systems in two Ugandan sub-counties: Kihihi sub-county, Kanungu district and Nagongera sub-county, Tororo district. Both sub-counties are rural; at the time of the study, Kihihi exhibited moderate transmission intensity (annual entomological inoculation rate [aEIR] 2011-2013 = 32.0) and Nagongera high transmission intensity (aEIR = 310) [15]. Both regions experienced two annual peaks in malaria burden following the rainy seasons.
From 2012-2014, the government of Uganda carried out a universal distribution of free long-lasting insecticide-treated nets (LLINs) with the goal of achieving one net per two people in each household. Nagongera sub-county received nets in November 2013 and Kihihi sub-county received nets in June 2014.
Health facility-based data
Enhanced malaria surveillance was established via the Uganda Malaria Surveillance Program (UMSP) MRCs in 2006, as previously described [16]. UMSP conducts sentinel surveillance in 70 level III and IV public outpatient facilities in Uganda, including Kihihi Health Centre IV and Nagongera Health Centre IV. At each MRC, individual-level outpatient department records are entered into an electronic MS Access (Microsoft Corporation, Redmond, WA) database for all individuals presenting to the outpatient departments of the health facilities using a standardized format. Data collected includes patient demographics (age, gender, and village of residence), results of laboratory tests (rapid diagnostic test or microscopy), diagnoses given, and treatments prescribed. UMSP provides laboratory support and quality control training to ensure high quality diagnostic testing. Data are sent to the UMSP data centre and cleaned before transfer to Stata (Stata Corp, College Station, TX) for analysis. This analysis uses 3 years of health-facility based surveillance data from the two MRCs (September 2011-August 2014). These months were selected given the low level of missingness (<30%) for village of residence. This analysis was restricted to patients aged 6 months through 10 years to make them comparable to cohort data described below.
Cohort data
Dynamic cohort studies were conducted in children aged 6 months through 10 years from 100 households randomly selected from the two study sub-counties, as previously reported [15]. In summary, eligible children from selected households were followed from August 2011 through June 2017. At enrollment, parents/guardians provided written informed consent and received an LLIN. Cohort participants received free medical care at designated study clinics located at the same MRCs where UMSP data were being collected; parents/guardians were encouraged to bring their children to the clinic any time they were ill. Children who presented with a fever (tympanic temperature ≥ 38.0°C) or history of fever in the previous 24 hours had a thick blood smear performed. If the blood smear was positive by microscopy, the child was diagnosed with malaria and provided treatment. Episodes of uncomplicated malaria were treated with artemether-lumefantrine; complicated or recurrent malaria occurring within 14 days of prior therapy was treated with quinine.
Measures
Malaria suspected. Health-facility based surveillance recorded all outpatients as “malaria suspected” or “malaria not suspected.” Malaria suspected was defined as patients who a) underwent a laboratory test for malaria (microscopy or rapid diagnostic test); or b) were given a clinical diagnosis of malaria in the absence of laboratory testing. Any record that did not meet these criteria was considered “malaria not suspected.”
Malaria cases. At MRCs, malaria cases were defined as patients with laboratory-confirmed malaria diagnoses (by microscopy or rapid diagnostic test).
Gold standard incidence. Malaria incidence measured through dynamic cohorts was considered the gold standard. Incidence was defined as the number of new episodes of malaria divided by the total person time observed. New episodes of malaria were defined as any episode of malaria not preceeded by another episode in the prior 14 days. A secondary definition using a parasite threshold of 2000 parasites/μL was also applied as a sensitivity analysis.
Statistical analysis
Travel time estimation. Villages located within Kihihi and Nagongera sub-counties were mapped during cross-sectional enumeration surveys conducted in 2009-2010 [15]. These village shapefiles were linked to unique identifiers of villages found in the UMSP database. Villages of residence for all outpatients living within the MRC subcounty were identified and mapped.
Travel times were calculated using Malaria Atlas Project’s friction surface 2015 raster file obtained through Google Earth Engine, available at 1-kilometer resolution [17]. The authors of this friction surface combined datasets on roads, railways, water bodies, slope and elevation, landcover, and borders to calculate a nominal overall speed of travel across each pixel, in units of minutes of travel time per metre. Travel times represent Uganda-specific mean travel times associated with the road types in the pixel, or, in pixels where no roads are present, walking times. The malariaAtlas R package was used to calculate the mean travel time from each outpatient’s village to the MRC of interest, in addition to the travel times to all nearest level III and IV health facilities [18]. Travel times were defined as the minimum travel time between two points.
Care-seeking model. Observations were restricted to those residing in villages whose nearest level III or IV health facility was the MRC of interest, assuming that individuals attend their nearest health facility. These villages were defined as the MRC’s “catchment area” [13]. Since not all individuals seek care when ill, and this care-seeking behaviour is driven in part by distance to the health facility, this analysis sought to account for this distance-specific care-seeking rather than using the raw population of the catchment area as a denominator for incidence. The probability of seeking care at the MRC was expected to decay as function of travel time to the facility. Relative village-level care-seeking probabilities were modelled and estimated as a function of travel time to the facility from each village within the catchment area. These probabilities were then used to down-weight village populations when estimating incidence. For example, if care-seeking from a particular village was estimated to be 80%, the population seeking care from that village was estimated to be 80% of the total population.
The care-seeking model was restricted to outpatients for whom malaria was not suspected. This group was used because their probability of attendance should be minimally biased by heterogeneity in malaria incidence across villages. Because this population represents a range of diagnoses, spatial bias is expected to be minimal. By using this population to model care-seeking, this analysis assumed that differences in care-seeking for outpatients not suspected of having malaria over space was driven solely by travel time to the health facility.
For each MRC, non-linear Poisson generalized additive models (GAMs) were specified to estimate the relationship between mean travel time from village i to the MRC and the count of outpatients not suspected of having malaria who visited the MRC from village i from September 2011-2014. GAMs are a class of generalized linear models that allow for the relationship between the outcome and predictor to be estimated using smooth functions of the predictor variables [19]. A non-parametric smooth function was applied to the travel time predictor, as the relationship between travel time to the facility and attendance was hypothesized to be non-linear. An offset for the logged population from village i derived from the High Resolution Settlement Layer [20] was included. To calculate relative village-level probabilities of attendance, predicted counts were estimated using the model described above holding the village population size constant. These counts were rescaled to relative probabilities by dividing the predictions by the predicted count in the village where the MRC is located. Calculating the relative probabilities in this way assumes that individuals living in the same village as the MRC have a probability of seeking care of 1.
In order to evaluate the sensitivity of these findings to the aforementioned assumptions, models were re-specified restricting outpatients to the top 5 diagnoses (including malaria) to determine whether the relationship between travel time and attendance differed across indications. In addition, stratified analyses were performed based on age category (6 months to < 5 years, 5 years to < 11 years) and gender. Models were also specified using straight-line distance from the centroid of the village of residence to the MRC as predictor and compared to results using travel time as predictor.
Incidence estimation. HMIS data were used to estimate malaria incidence in two ways. First, incidence was estimated by dividing malaria cases over the catchment area denominator (including all villages for which the MRC is the closest health facility) without down-weighting for travel time, hereafter called unweighted catchment incidence. Second, malaria incidence was estimated by dividing malaria cases by a weighted denominator using the weights described above to adjust village-level populations, hereafter called weighted catchment incidence. All populations were set to grow at a fixed rate each month based on the World Bank’s estimate of population growth during the study window (0.29% monthly) [21]. Both of these HMIS-derived measures were compared to metrics of gold standard (cohort) incidence by generating plots over time, calculating measures of pair-wise correlation by month, and comparing aggregated estimates of malaria incidence over the three year study window. This method assumes that relative treatment-seeking behaviour for non-malarial illness is the same as for malaria.