The independent effect of living in malaria hotspots on future malaria infection: an observational study from Misungwi, Tanzania

Misungwi district (lat 2.85000 S, long 33.08333 E) is located 60 km from Mwanza town in the northwest of Tanzania at an altitude of 1,178 m above sea level. Details of the study site have been previously described [10]. In brief, the district is rural with moderately intense malaria transmission; the overall prevalence of infection in the region is estimated to be 31.4% by microscopy in children six to 59 months of age [11].

This is a cross sectional study, which was conducted twice in year 1 in 2010 and year 2 in 2011. The exposure of interest was whether a person resided in a malaria hotspot or not in the first year of the study. Hotspots were defined by SaTScan and Kernal density method using infection status derived using nested polymerase chain reaction (nPCR) and AMA-1 sero status in year 1, as previously reported [10]. The methods are explained briefly below. The outcome of interest for this analysis was infection status at the individual level by nPCR (infected/not infected) in the survey taken in the second year. In brief, the SaTScan software (SaTScan, version 8.2.1) used a spatial scan statistic using the Bernoulli model to identify clusters of significant high (hotspot) and low (coldspot) risk of infection [12]. Using the SaTScan method, SaTScan cold spots were coded as 0, hotspots as 1 and everything else as 0.5. The kernel method of household clustering of both nPCR and AMA-1-positive individuals was estimated using Kernel density smoothing. Kernel density estimates, for any given point, the density of events within a predefined window, with the influence of events weighted according to the distance from the centre of the window. The weight assigned to each event is derived from the kernel function applied. Details of these methods have been described previously [10]. Using the Kernel method, each household was assigned a value between 0 (least exposed households) and 1 (most exposed households). Households for which data were only available in the second year were assigned a hotspot score based on infection in neighboring households only.

A census of four villages in a single ward of the Misungwi district of Tanzania was carried out in the dry season, in two consecutive years 2010 and 2011 between the months of August and November. Trained interviewers administered a structured questionnaire to consenting household heads. Information gathered included morbidity, demographics and data on potential risk factors. Latitude and longitude of each household was taken using a GPS device. Data were recorded electronically using personalized digital assistants and were downloaded each evening of the survey onto a desktop computer to a master Microsoft access database. Also, every consenting individual was asked whether they slept under a long-lasting insecticidal-treated net (LLIN) the previous night. A finger-prick blood sample was collected and was spotted onto Whatman® standard 3 mm filter paper for parasite detection by nPCR and for serology analysis (AMA-1).

Statistical analysis was performed using STATA (version 12, College Station, TX, USA). Mixed effect logistic regression models were used to evaluate whether residing in a hotspot in the first year was predictive of subsequent malaria infection in the second year, controlling for other potential risk factors for malaria infection. The model adjusted for possible clustering of malaria cases within households.

Summary contingency tables, graphs and scatter plots with lowess curves were used to explore the relationship with potential risk factors and the outcome (malaria infection defined by nPCR in the second year). How well the linear and quadratic terms of these variables fitted was also explored. Variables with a non-linear relationship to the outcome (age, SES, number of cattle sleeping outside the household, and number of people sleeping in the household) were categorized. Age was categorized into five groups: under four years, five to 15 years, 16–25 years, 26–35 years and over 36 years (Table 1); number of cattle into three groups: none, one to ten and ten + cattle. SES was based on wealth index, which was a weighted sum of data on household possessions and utilities, according to principal component analysis. SES was categorized by dividing wealth index into four wealth quartiles, from the poorest to the least poor. Principle component analysis was conducted using a set of household construction materials (wall material, roof material, floor material, presence of eaves, and whether windows were screened or not) to define household quality. The household quality index was categorized by dividing the index into tertiles. Presence of ponds, rice plantations, water in clay pots, old tires, garbage, and any kind of stagnant water around the house were considered a mosquito breeding site. The presence of breeding site within 100 m (which was manually checked by the study team) around the household was chosen as this is considered to be the distance with abundant vector densities, and also vector densities decline rapidly away from the breeding sites [9]. Euclidean (straightline) distance to health facility from each surveyed household was calculated using coordinates of the households and that of health facility the household attended. Guided by lowess curves distance to the health facility was categorized into four groups, <1 km, 1-2.5 km, 2.6–3.5 km, 3.6+ km.

All variables were analysed individually for an association with the outcome (malaria by nPCR infection in year 2) using logistic regression. A household-level random effect was included to account for correlation between individuals within the same household. All variables showing evidence for a possible association with malaria risk (p < 0.1) were included in the preliminary main effect multivariate logistic regression model. A forward stepwise approach was then followed to exclude any variable that showed a lack of effect on malaria risk (p > 0.05). Hotspots defined by SaTScan and kernal are fitted in different multivariate models.

Table 1 presents the results of the univariate associations with individual infection status in year 2. These univariate estimates were adjusted for possible household level clustering. Residing in hotspots defined by malaria infection and AMA-1 sero status were associated with higher odds of malaria infection. Children between the age of five and 15 years and males had significantly higher odds of malaria infection. Higher wealth status, use of LLIN and mother’s education were associated with lower odds of malaria infection. Households with poor quality of construction materials, presence of a breeding site near the household and greater distance to the health facility were associated with higher odds of malaria infection.

Table 2 presents results from the multivariable analysis to determine the independent risk of malaria infection associated with residing in a malaria hotspot, adjusting for other risk factors for malaria infection and for household clustering. Only residing in malaria infection hotspots, using SaTScan method to define hotspots, was predictive of increased odds of malaria infection in year 2. For example, individuals residing in infection hotspot defined by SaTScan were three times more likely to have malaria infection after controlling for other factors (OR 3.11; 95% CI 1.57, 6.18).

However, hotspots defined by both nPCR and AMA-1 using kernel method did not appear to be independent risk factors for future malaria infection after controlling for other factors in the multivariable analysis, OR 1.52; 95% CI 0.87-2.66 and OR 0.99; 95% CI 0.49-2.00, respectively.

Apart from residing in a malaria infection hotspot being an independent risk factor for malaria infection in the second year, other factors were also independently associated with increased risk of malaria infection in the second year in all multivariable hotspot models. These were age, gender, mother’s education, using LLIN, presence of breeding sites, longer distance to a health facility, and lower quality of houses (Table 3).

Individuals in the age group of five to 15 years had more than five times the odds of infection than those who were in younger age group (OR 5.04; 95% CI 3.82-6.64). There was borderline evidence that the risk of malaria infection was higher for males than females (OR 1.24; 95% CI 1.01-1.53).

Individuals living within 100 meters of a mosquito breeding site had increased odds of malaria infection compared to those not living near a mosquito breeding site (OR 1.59; 95% CI 1.11-2.29). Individuals living in poor-quality households had increased odds of malaria infection compared to those who were living in good-quality households (OR 1.53; 95% CI 1.01-2.32). Distance to the health facility was strongly associated with risk of having malaria infection. Individuals residing in households based further from health facilities had increased odds of malaria infection (OR 1.34; 95% CI 1.03-1.76).

There was strong evidence that sleeping under LLIN was associated with decreased odds of infection (OR 0.40; 95% CI 0.25-0.63). Individuals whose mother had primary education or more had decreased odds of malaria infection compared to those individuals whose mother had not gone to school (OR 0.66; 95% CI 0.45-0.95).

Wealth, IRS, number of cattle, and number people sleeping in the house were not significantly associated with the risk of malaria infection in multivariate analyses.