Heterogeneous exposure and hotspots for malaria vectors at three study sites in Uganda

Data sources

Entomological surveillance was conducted in three areas of Uganda: between October 2011 and March 2015 for Walukuba subcounty, Jinja District (0°26΄33.2˝ N, 33°13΄32.3˝ E) and Kihihi subcounty, Kanungu District (0°45΄3.1˝ S, 29°42΄3.6˝ E); and between October 2011 and September 2016 for Nagongera subcounty, Tororo District (0°46΄10.6˝ N, 34°1΄34.1˝ E). Details of this study, including the overall study design, study sites, a detailed description of the entomological methods, and ethical approval have been described elsewhere^18,23. See Figure S1 to Figure S3 for the maps of the three study sites (Supplementary File 1).

Jinja District is located in the east central region bordering Lake Victoria and has a population of roughly 470,000 with 63% of the population living in rural area²⁴. Jinja town was characterized historically by moderate malaria transmission intensity, but it was the site with lowest transmission in our study. Kanungu District is a rural area in south-western Uganda bordering the Democratic Republic of Congo and has a population of approximately 250,000 with 80% of the population living in villages²⁴. Malaria in Kanungu has been characterized by relatively low transmission intensity and low endemicity, but it had moderate transmission in our study. Tororo District is located in south-eastern Uganda on the Kenyan border and has a population of about 520,000 with 86% of the population living in villages²⁴. Tororo is characterized by very high malaria transmission.

CDC light traps (Model 512; John W. Hock Company, Gainesville, FL, USA) were placed in 330 randomly-selected households; 116 in Jinja, 107 in Kanungu, and 107 in Tororo. The traps were positioned one meter above the floor at the foot of the bed where a study participant slept. Traps were set on one day each month at 19:00 h and collected the following morning at 07:00h. All anophelines were identified taxonomically based on morphological criteria according to established taxonomic keys²⁵. Up to 50 anophelines per household were tested for presence of Plasmodium falciparum sporozoites using ELISA²⁶. The total number of anophelines trapped in Jinja per household each day ranged from 0 to 121 with a median of 0; ranged from 0 to 306 with a median of 0 in Kanungu; and ranged from 0 to 1011 with a median of 9 in Tororo.

There are typically two rainy seasons in Uganda (March to May and August to October) with annual rainfall of 1,000–1,500 mm. Residents of all study households were provided with long-lasting insecticidal bednets (LLINs) at enrollment. Over the course of the study, compliance (defined as sleeping under an LLIN the previous night at the time of each clinic visit) was over 98% for all three sites. Thus LLIN coverage within the study households (where mosquitoes were collected) was very high and consistent over time and across the three sites. In terms of LLIN coverage in the surrounding communities, coverage levels varied across the sites and over time based on repeated cross-sectional surveys. Changes in coverage over time were primarily due to a national LLIN distribution campaign conducted in November 2013 at two sites (Jinja and Tororo) and June 2014 at the third site (Kanungu). However, in a separate time series analysis²⁷, there was no significant difference in human biting rates (mosquito abundance) before and after community LLIN distribution for all three sites. Indoor residual spraying of insecticide (IRS) was only used in Tororo, where 3 rounds of the carbamate insecticide Bendiocarb were initiated in December 2014, June 2015, and December 2015 in Nagongera.

Statistical analyses

Here we partitioned the observed variance of mosquito counts at the three Ugandan study sites among the sources of heterogeneity attributed to individual households, seasonality, and environmental noise or measurement error. Here, ‘partitioning’ refers to attributing proportions of the total variability to individual factors. Partitioning the total variance of a response variable into its component sources is common in ecological studies^28,29 as it can help inform a wide variety of research and management questions^30,31. For instance, the variance partitioning approach developed by 31 for a negative binomial mixed model was a useful method for assessing the response of a variance structure to large-scale ecological changes. On the other hand, as observed in our mosquito data, overdispersion is commonly exhibited in ecological count data, and can be modeled effectively using a variety of methods, including a negative binomial distribution.

The mosquito counts consisted of a large proportion of zero counts; 70% in Jinja, 54% in Kanungu, and 21% in Tororo; and showed some degree of overdispersion. As a general rule, the presence of over 30% of zeros in the data would require a zero-inflated model. Due to the presence of excess zeros in the Ugandan mosquito count data (a common feature of malaria vector data), a range of models capable of handling excess zeros were considered, including zero-inflated Poisson, zero-inflated negative binomial (ZINB), Poisson hurdle, and negative binomial hurdle regression models. A simulation study was designed to determine the most robust and best-fitted model for quantifying household-level heterogeneity in malaria vector density. The ZINB model³² outperformed the other three models in the simulation study. The details and the discussion of the results of the simulation study are described in Supplementary File 2.

A Bayesian framework was adopted to estimate model parameters, using the integrated nested Laplace approximation (INLA) approach³³ implemented via the R-INLA package. By fitting the ZINB regression models to the mosquito counts, we estimated household biting propensities (ω), seasonal signal (S), and noise (e) at each of the study sites. For mosquito counts Y = {y₁, y₂, . . . , y_n}, the ZINB distribution can be written as

where τ > 0 is a shape parameter which quantifies the amount of overdispersion. The mean and variance of the ZINB distribution are E(Y_i ) = (1 − π_i )λ_i and var(Y_i ) = λ_i (1 − π_i )(1 + π_i λ_i + λ_i /τ), respectively.

For each of the study sites, the expected count for household j on day i, λ_ij, was modeled on a set of explanatory variables with the aid of a log link function, while the probability π_ij was modeled on the same set of explanatory variables using a logit link function. The covariates of interest included household identifiers (ID) and sampling days (t). Household identifiers were treated as fixed effects in the regression model, whereas daily seasonal signal and random noise were treated as random effects. A range of Bayesian prior distributions were imposed on sampling days (t), including first- and second-order random walks and autoregressive processes of order 1 and order 2³⁴. The random noise was assumed to be independent and identically distributed (i.i.d.), i.e., Gaussian i.i.d.. For consistency and convenience, the default prior specifications in R-INLA have been chosen for each of the prior distributions. The Watanabe-Akaike information criterion³⁵ was used as the model comparison criterion in order to select the model that produced the best estimates of ω, S, and e for each of the three study sites in Uganda. Note that the estimated ω for each of the following scenarios had been scaled to have a mean of one so that they reflected relative attractiveness of household mosquito biting with respect to other households. See Supplementary File 3 for the R code for estimating household biting propensities and seasonal signal using a ZINB model.

The model for estimating the overall household biting propensities (ω_overall) is as follows,

                                                                                             logit(π_ij) = a_j · ID_j + f (t_i) + 𝜖_ij,

                                                                                                 log(λ_ij) = a_j · ID_j + f (t_i) + 𝜖_ij.

Here a_j quantifies the effects of household biting. The random noise, 𝜖_ij, are assumed to be i.i.d., and t_i, the temporally structured random effects, are assigned one of the four Bayesian temporal smoothing prior distributions. Using the estimated parameters, we obtained the household biting propensities, ω_overall = exp(a_j) and the seasonal signal, S = exp(f (t_i)). The noise, e = exp(𝜖_ij), accounted for additional variation among mosquito counts that was not accounted for by the covariates or by Poisson (random) variation around the mean ω S.

We quantified differences in household biting propensities due to different seasons and vector control interventions using Kendall’s coefficient of concordance (or Kendall’s W)³⁶. To study the impact of different seasons on household biting, we explored biting propensities during dry (ω_dry) and rainy seasons (ω_rainy) for all three sites. For Jinja and Kanungu, we evaluated changes in transmission in the first and last half of the study by estimating biting propensities in the first half period (ω_{1st half}) and the second half period (ω_{2nd half}) of surveillance. This provides a rough idea of how the use of LLIN over time has impacted on mosquito abundance in Jinja and Kanungu. For Tororo, the impact of IRS on household mosquito biting was examined by computing biting propensities before (ω_{before IRS}) and after (ω_{after IRS}) the deployment of IRS.

Household biting propensities during dry (ω_dry) and rainy (ω_rainy) seasons, during the first half period (ω_{1st half}) and the second half period (ω_{2nd half}) of surveillance, and before the deployment of IRS (ω_{before IRS}) and after the deployment of IRS (ω_{after IRS}) were estimated using one of the following sets of equations,

                                                                                             logit(π_ij) = f (ID_dry) + f (ID_rainy) + f (t_i),

                                                                                                 log(λ_ij) = f (ID_dry) + f (ID_rainy) + f (t_i),               or

                                                                                             logit(π_ij) = f (ID_{1st half}) + f (ID_{2nd half}) + f (t_i),

                                                                                                 log(λ_ij) = f (ID_{1st half}) + f (ID_{2nd half}) + f (t_i),      or

                                                                                                 logit(π_ij) = f (ID_{before IRS}) + f (ID_{after IRS}) + f (t_i).

                                                                                                     log(λ_ij) = f (ID_{before IRS}) + f (ID_{after IRS}) + f (t_i).

Here, ID_dry denote household IDs during the dry season, ID_rainy denote household IDs during the rainy season, and so forth. The random effects f (ID...) was assumed to be i.i.d.. By modeling the biting propensities of different scenarios as random effects, it allowed for estimation and comparison of two scenarios within one model, instead of having to split the datasets into two subsets. The noise term, 𝜖_ij, was not required in the model because the focus here was to estimate household-level effects. Hence, the biting propensities for different scenarios are such that: ω_dry = exp( f (ID_dry)); ω_rainy = exp( f (ID_rainy)); ω_{1st half} = exp(f (ID_{1st half})); ω_{2nd half} = exp(f (ID_{2nd half})); ω_{before IRS} = exp(f (ID_{before IRS})); and ω_{after IRS} = exp( f (ID_{after IRS})).

We also conducted hotspot analysis on ω using the Getis-Ord ($G_i^{*}$) statistic³⁷ for all three sites to identify hotspots of vector abundance. The $G_i^{*}$ statistic is a local statistic that identifies variation across the study area, by focusing on individual features and their relationships to nearby features. It also identifies statistically significant hot spots and cold spots, given a set of weighted features. To be statistically significant, the hot spot or cold spot will have a high or low value and be surrounded by other features with high or low values. The value of the target feature is included in analysis. The $G_i^{*}$ statistic is a z-score and given as:

where ω_j is the attribute value for feature j (biting propensities in this case), k_i,j is the spatial weight between feature i and j, n is the total number of features while:

The Getis-Ord test was performed using the localG function in the spdep package in R³⁸. At a significance level of 0.001, a z-score would have to be less than −3.29 or greater than 3.29 to be statistically significant.

High resolution climatic and environmental covariates (100m × 100m) known to interact with mosquito density, along with housing covariates for all the sampled households at the three study sites were assembled (Table 1). The housing covariates (roof type, wall type, floor type, eaves, air bricks, and house type) were all available as categorical variables. To explore the association of household biting with housing covariates, a linear regression model was fitted between ω and the housing covariates of various study sites and scenarios, which resulted in multiple correlation coefficient for each pair. For each of the environmental covariates, we calculated Pearson correlation coefficients for the covariate and ω of various scenarios to understand the association of a covariate with household biting propensities. The associations of household biting propensities with environmental and housing covariates were presented on a correlation heatmap.

For each study site and for all three study sites combined, a negative binomial generalized linear regression was used to model the relationship between the six categorical housing covariates and the total number of Anopheles mosquito caught per household by light trap catches, with the number of sampling nights included as an offset term in the model. The analyses were performed using the glm.nb function in the MASS package in R (R code available in Supplementary File 4). This allowed us to evaluate the overall effect of the various housing covariates at all study sites as well as their effects separately at each study site. In a separate analysis to assess the impact of environmental factors on mosquito abundance, a Bayesian ZINB geostatistical spatio-temporal framework, along with environmental covariates, were used to predict mosquito density across the whole study site (details in 39).

Quantifying heterogeneous household biting in mosquito abundance

Using pseudo-datasets that mimicked the Ugandan data (i.e., presence of excess zeros and Poisson distributed) and several model selection criteria in the simulation study, we found the ZINB model to be the most robust and best-fitted method for estimating household biting propensities (ω), seasonality (S), and environmental and measurement noise (e) for each study site (Figure 1).

Figure 1.

Simulation study: (a) The estimates of seasonal signal (Ŝ_P) for pseudo-dataset D3 reconstructed using the best-fitted model, i.e. the zero-inflated negative binomial model. Smoothing technique S1 (a Gaussian kernel was used to smooth the counts prior to model fitting) produced the worst fit while smoothing technique S3 (a second-order random walk prior distribution imposed on temporal random effects) produced the best fit. (b) The scatter plot shows log(Ŝ_P/S_P): ZINB S3 resulted in a much better fit of Ŝ_P than ZINB S1. (c) The ZINB S1 model produced the worst fit when Ŝ_P (estimated) is fitted against S_P (the truth) on a simple linear regression; coefficient of determination, R², is around 80%. (d) The ZINB S3 model produced the best fit when Ŝ_P is fitted against S_P on a simple linear regression; R² is close to 1.

The overall household biting propensities for the entire duration of surveillance were estimated for each of the study sites, denoted ω_overall. We found patterns among households as well as spatial patterns at a larger scale. Although household biting propensities are closely correlated (by design) with the average number of mosquitoes caught in a light trap per household per night at each study site, they are not identical (Figure 2). Households receiving a larger amount of mosquito bites tend to have a larger biting propensity but this is not always true due to other factors such as seasonality and the environment. We also note that there were differences in the distributions of biting propensities across sites. A large variation in biting propensities was observed in households across the three study sites, as shown in panel (a) of Figure 3, Figure 4, and Figure 5 for Jinja, Kanungu, and Tororo, respectively. Households with the largest and smallest ω_overall in Jinja appeared to be located away from the centre of the study region (Figure 3(a)), which is highly urbanized⁸. Households at low elevations in Kanungu (northern part of the study region) generally had a larger ω_overall while households at high elevations and less rural part of Kanungu (southeast part of the study region) had a much smaller ω_overall (Figure 4(a)). In Tororo, some of the households with the largest and smallest ω_overall were found to be located around the border of the study region (Figure 5(a)).

Figure 2. Household biting propensities for the entire duration of surveillance (ω_overall) are plotted against the normalized average mosquito density per household per night for each of the study sites.

The red line denotes the 1:1 line. Evidently, ω_overall had a linear relationship with mosquito density at each study site.

Figure 3.

Jinja (a): Household biting propensities for the entire duration of surveillance. Pink circles denote households with ω_overall > 1 with the size of the circles showing the magnitude of ω_overall; blue circles denote households with ω_overall < 1 with the size of the circles showing the magnitude of 1/(3×ω_overall) to make their sizes comparable to the pink circles; 10% of households with the largest ω_overall are labelled. The largest pink circle indicates a household with the largest ω_overall while the largest blue circle indicates a household with the smallest ω_overall. Households with the largest ω_overall include HH31, HH29, and HH40. Grey dots in the background denote all enumerated households at Walukuba subcounty, Jinja District. (b) Ratios of household biting propensities during dry (ω_dry) and rainy seasons (ω_rainy). Each plotted upward pointing blue triangle represents a household with a larger ω during rainy season compared to dry season; a blue filled triangle denotes a hotspot for the ratio of ω_rainy/ω_dry. Each plotted downward pointing red triangle represents a household with a larger ω during dry season compared to rainy season; a red filled triangle denotes a hotspot for the ratio of ω_dry/ω_rainy. (c) Ratios of household biting propensities in the first half period of surveillance (ω_{1st half}) and the second half period (ω_{2nd half}). Each plotted upward pointing blue triangle represents a household with a larger ω in the second half period of surveillance compared to the first half period; there were no hotspots for the ratio of ω_{2nd half}/ω_{1st half}. Each plotted downward pointing red triangle represents a household with a larger ω during the first half period of surveillance compared to the second half period; a red filled triangle denotes a hotspot for the ratio of ω_{1st half}/ω_{2nd half}. (d) Time series of mosquito count data for the entire duration of surveillance, where each grey circle denotes an observation of a household. The blue solid line denotes the estimated seasonal signal and cyan dashed lines denote the 95% Bayesian credible interval for the seasonal signal. The red vertical dashed line denotes the cut-off for the first half and the second half periods. Rainy season is highlighted in pink and dry season is highlighted in light green. A weak seasonal signal was observed in Jinja. (e) Scatter plot of log(ω_dry) and log(ω_rainy) along with measures of R², correlation, and Kendall’s W. For ease of interpretation, we plotted the ω on the logarithmic scale. The log transformation also preserved the order of the observations while making outliers less extreme. (f) Scatter plot of log(ω_{1st half}) and log(ω_{2nd half}) along with measures of R², correlation, and Kendall’s W.

Figure 4. Kanungu.

The layout of the figure is the same as in Figure 3 with some exceptions as follows. (a): Blue circles denote households with ω_overall < 1 with the size of the circles showing the magnitude of 1/(5 × ω_overall) to make their sizes comparable to the pink circles. Households with the largest ω_overall include HH23, HH18, and HH58. (d) A moderate seasonal signal was observed in Kanungu.

Figure 5. Tororo.

The layout of the figure is the same as in Figure 3 with some exceptions as follows. (a): Blue circles denote households with ω_overall < 1 with the size of the circles showing the magnitude of 1/(2 × ω_overall) to make their sizes comparable to the pink circles. Households with the largest ω_overall include HH75, HH50, HH56, and HH52. (c) Ratios of household biting propensities before the deployment of IRS (ω_{before IRS}) and after the deployment of IRS (ω_{after IRS}). Each plotted upward pointing blue triangle represents a household with a larger ω after the deployment of IRS compared to before the deployment of IRS; a blue filled triangle denotes a hotspot for the ratio of ω_{after IRS}/ω_{before IRS}. Each plotted downward pointing red triangle represents a household with a larger ω before the deployment of IRS compared to after the deployment of IRS; a red filled triangle denotes a hotspot for the ratio of ω_{before IRS}/ω_{after IRS}. (d) The red vertical dashed lines denote the deployment of IRS which were six months apart. A strong seasonal signal was observed in Tororo. (f) Scatter plot of log(ω_{before IRS}) and log(ω_{after IRS}) along with measures of R², correlation, and Kendall’s W.

Seasonally varying hotspots of mosquito abundance

In Figure S4 (Supplementary File 1), we plotted ω for different scenarios against ω_overall, changes of transmission over time was confirmed in Jinja while the application of IRS in Tororo had affected the ω. Using the $G_i^{*}$ statistic, we found seasonal varying hotspots of mosquito abundance at the household level for different scenarios at the three sites. Here, a seasonal varying hotspot refers to a household (HH) with drastic differences in ω under different scenarios and was surrounded by other households with less drastic changes in ω. In Jinja, three households showed considerable differences during dry and rainy seasons while three households were identified as seasonal varying hotspots when compared in the first and last half of the study (Figure 3(b–c)). Being the study site with the lowest malaria transmission intensity, Jinja exhibited a weak seasonal signal in mosquito counts over time (Figure 3(d)). It is noteworthy that HH29, located next to a swampy area near Lake Victoria, had a much larger ω during the rainy season compared to the dry season. Most households in Jinja behaved similarly during dry and rainy seasons, whereas greater differences in ω were shown between the two time periods (Figure 3(e–f)).

In Kanungu, there was a seasonal varying hotspot for the dry vs. rainy seasons comparison while three households were identified as seasonal varying hotspots when compared in the first and second half of the study (Figure 4(b–c)). The intermediate transmission intensity found in Kanungu in previous studies^23,46 could be explained by the moderate seasonal signal in mosquito counts over time (Figure 4(d)). Most households in Kanungu behaved very similarly during different seasons and time periods despite some slight differences in the first half and the second half periods of surveillance (Figure 4(e–f)).

In Tororo, a seasonal varying hotspot was identified for the dry vs. rainy seasons comparison while six households were identified as seasonal varying hotspots before the deployment of IRS but not after IRS (Figure 5(b–c)). Tororo, the study site with highest transmission intensity (add REF, please), exhibited a strong seasonal signal in mosquito counts over time (Figure 5(d)). It was clear that mosquito counts peaked at the end of the rainy season and at the very beginning of the dry season and reduced remarkably after the deployment of IRS. Most households in Tororo behaved very similarly during dry and rainy seasons but showed large differences in ω before the deployment of IRS and after (Figure 5(e–f)).

Association of heterogeneous household biting with environmental and housing covariates

The environmental covariates did not have strong associations with household biting heterogeneity at all three sites (Figure 6). Nevertheless, households in Kanungu had greater associations of biting propensities with environmental covariates including land surface temperature (day time, night time, and diurnal difference), precipitation, and elevation, compared to two other sites. Kanungu was a rural area of rolling hills in western Uganda located at an elevation of 1,310 m above sea level and it was the study site with the highest elevation and greatest within-site elevation change among the three sites²³. The greater variation in environmental characteristics across households in Kanungu due to its altitude gradient most likely contributed to the greater association between biting propensities and the environmental covariates.

Figure 6. Correlation heatmap of household biting propensities and environmental and housing covariates.

See Table 1 for descriptions of the covariates. The size of the squares and the shade of colors both illustrate the magnitude of the correlation. Red-shaded squares denote positive correlation whereas blue-shaded squares denote negative correlation.

Figure S5 (Supplementary File 1) shows the density of Anopheles mosquitoes captured by the light traps in Tororo before and after the deployment of IRS. Figure S6 (Supplementary File 1) shows the average predicted Anopheles mosquito (An. gambiae (s.l.) and An. funestus (s.l.)) density per household per night across the whole study region before and after the deployment of IRS, predicted using environmental covariates. Clearly, the environmental covariates did not account for a big portion of the heterogeneity in mosquito density, especially in the density of An. gambiae (s.l.) before the deployment of IRS (upper left panel of Figure S6).

A portion of the heterogeneity in mosquito biting may be attributed to household attractiveness caused by housing characteristics or structures, consistent with those reported previously²¹.

Among the three sites, the housing covariates had the strongest associations with ω in Tororo, followed by ω in Jinja, and ω in Kanungu (Figure 6). Table 2 summarizes the incidence rate ratios of the best fitted model (retaining significant risk factors only) selected by the Akaike information criterion. In Jinja, houses with metal walls had a high biting propensity score compared to houses with cement walls; houses with earth or sand floors were also much more likely to attract mosquitoes compared to other types of floors. In Kanungu and Tororo, houses with mud floors had a higher mosquito abundance compared to houses with cement floors; houses with open eaves also attracted more mosquitoes compared to houses with closed eaves. On the other hand, houses with screened airbricks in Tororo had a protective effect reducing the number of indoor mosquitoes which could result in a reduced malaria transmission. For all study sites combined, houses with wood walls, brick floors, screened airbricks, and metal or tiled roofs were protective against malaria vector. These findings are consistent with those reported in 21, but we found additionally that screened airbricks reduce number of indoor mosquitoes for all three sites combined.