Socioeconomic, demographic and environmental factors inform intervention prioritization in urban Nigeria

Nigeria is one of three countries projected to have the largest absolute increase in the size of its urban population and this could intensify malaria transmission in cities. Accelerated urban population growth is out-pacing the availability of affordable housing and basic services and resulting in living conditions that foster vector breeding and heterogeneous malaria transmission. Understanding community determinants of malaria transmission in urban areas informs the targeting of interventions to populations at greatest risk. In this study, we analyzed cluster-level data from the Demographic and Health Surveys (DHS) and the Malaria Indicator Survey (MIS) as well as geospatial covariates to describe malaria burden and its determinants in areas administratively defined as urban in Nigeria. Overall, we found low malaria test positivity across urban areas. We observed declines in test positivity rate over time and identified the percentage of individuals with post-primary education, the percentage of individuals in the rich wealth quintiles, the percentage of individuals living in improved housing in 2015, all age population density, median age, the percentage of children under the age of five that sought medical treatment for fever, total precipitation and enhanced vegetation index as key community predictors of malaria transmission intensity. The unrepresentativeness of the DHS and MIS in urban settings at the state and geopolitical zonal level, regional differences in malaria seasonality across Nigeria, and information detection bias were among likely factors that limited our ability to compare malaria burden across geographic space and ultimately drove model uncertainty. Nevertheless, study findings provide a starting point for informing decisions on intervention prioritization within urban spaces and underscore the need for improved regionally-focused surveillance systems in Nigeria.

CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) Introduction

61
Nigeria accounts for 27 percent of all global malaria cases and deaths, respectively, making it the 62 greatest contributor to the global malaria burden [1] Underlying Nigeria's malaria burden are spatial 63 and temporal differences in malaria risk driven by variations in ecological and climatic factors, 64 intervention histories, health system factors, land-use, and urbanization [2-4]. Nigeria is one of three 65 countries expected to account for nearly one-third of global urban population growth between 2018 and 66 2050 [5]. Half of Nigeria's roughly 200 million population lived in urban areas in 2018 and the proportion 67 of urban residents is projected to increase to 70 percent by 2050 [5]. Although urbanization is expected 68 to decrease malaria transmission by reducing natural vector breeding sites and biting risk [6][7][8], rapid 69 and unplanned urbanization characterized by the siting of farms within urban neighborhoods and the 70 development and expansion of informal settlements with inadequate basic services and poor sanitary 71 conditions support the creation of natural and artficial habitats for larval development [9][10][11]. In 72 addition, urban settlers, particularly those of lower socio-economic status, maintain rural residences and 73 connections and the resultant urban-rural mobility could introduce parasites from rural villages, 74 sustaining transmission in urban spaces [12][13][14]. The impact of environmental and anthropogenic 75 factors are often reflected in city-level heterogeneities in malaria risk [15], prompting the need to 76 prioritize interventions in areas of greatest need as well as deprioritize interventions in areas of minimal 77 risk. Intervention targeting would ideally be achieved by stratifying localities using epidemiological, 78 ecological, health system and socio-economic determinants following the World Health Organization's 79 recommendations [15], but information on how these determinants relate to malaria transmission, to 80 facilitate locality grouping in urban Nigeria is lacking. 81 82 Community determinants of malaria transmission in urban Nigeria are not well-researched. Most studies 83 examine only individual level factors [16][17][18][19], are limited to specific urban settlement archetypes [17,19] 84 or are hospital-based [17,20]. In a study of patients receiving care at two hospitals in the city of Ibadan, 85 Awosolu and colleagues found that distance from streams within 1km and travel to a rural area in the 86 last months were statistically significant risk factors for malaria infection, however, participants' place of 87 residence was not accounted for, increasing the likelihood of inaccurate results [20]. Findings from a 88 cross-sectional survey in an urban town in Nigeria's South-West region indicated that the types of 89 windows and environmental hygiene significantly predicted malaria prevalence within households but 90 this study provided no information on categorization of these covariates and their measures of 91 association [19]. Insufficient information and the methodological gaps in the extant literature motivate 92 the need for additional studies that elucidate major drivers of malaria risk in urban Nigeria. 93 94 Georeferenced survey data, such as the Malaria Indicator Surveys (MIS) and Demographic and Health 95 surveys (DHS), and modeled geospatial data can be utilized to understand urban malaria transmission 96 risk. Unlike routine surveillance systems that collect solely information on malaria infection status and 97 are biased towards individuals that live close to health facilities or seek care in public institutions 98 [21,22], the MIS and DHS uses a cluster sampling methodology to collect data on individual level 99 infection status and risk factors that can be aggregated for urban clusters, and when supplemented by 100 geospatial rasters facilitate examination of risk factor associations. Moreover, definitions of urban areas 101 within the DHS and MIS survey are aligned with those of local administrators increasing the likelihood 102 that analysis results will be accepted by policymakers. Although DHS and MIS data is too sparse to 103 characterize malaria risk within cities, understanding transmission risk between different urban settings 104 can provide a level of insight. 105 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted March 18, 2022. ;https://doi.org/10.1101https://doi.org/10. /2022 106 Therefore, in light of data limitations and the need to understand how to appropriately group 107 geographic areas in Nigeria for intervention targeting, this study analyzed DHS, MIS and geospatial data 108 to 1) quantify spatio-temporal variation in malaria test positivity rate among children under the age of 109 five years (U5), 2) identify predictors of U5 malaria test positivity rate, and 3) generate model effect 110 plots to describe associations between covariates and U5 malaria test positivity rate. Predicted effects 111 observed in unadjusted effect plots represent correlations between dependent and independent 112 variables that could be leveraged to choose potential thresholds for intervention prioritization. Adjusted 113 effects provide insight into the potential causal impact of independent variables, and are essential for 114 understanding the public health impact of interventions. 115 116 Methods

117
Data 118 Cluster-level data from the 2010 and 2015 MIS, 2018 DHS and publicly available geospatial malaria 119 covariate data were used for this study (see Table 1 for references where each of the are vectors of coefficients multiplying their associated vector natural spline basis 165 functions ℎ and 0 represents the model intercept; .. are covariate values; ( ) is a cluster specific 166 stationary AR(1) process (type of autoregressive model) for modeling temporal dependence by month 167 and year of survey, , for each study cluster ; ( ) is spatial Matern process for modeling spatial random 168 effects using each cluster coordinate, ; NSZ is the event non-structural zero; represents the offset 169 term which is the number of children, 6 -59, years tested for malaria and the zero components were 170 modeled with the equation in (2) with representing the probability of observing zero counts. Due to 171 the large computational power required to generate cubic plots from complex models, unadjusted and 172 adjusted effect plots for the final model covariates were produced and described using linear splines. All 173 code written in support of this manuscript is available via this doi: 10.5281/zenodo.6350331 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted March 18, 2022. . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)

189
Describing spatial-temporal variation in malaria test positivity in urban areas The majority of clusters in Lagos (90%), Borno (84%) and Akwa Ibom (83%) had a zero test positivity rate 213 (Figure 2c). However, the DHS 2018 report states that 11 LGAs in Borno were dropped during initial 214 sampling due to insecurity [45], implying that the distribution of test positivity rate may not be 215 representative. At the regional level, 66% of clusters in the North East geopolitical region had a zero test 216 positivity rate, 63% in the South-South, 62% in the North-Central, 50% in the South-East, 49% in the 217 South-West and 41% in the North-West (Figure 2d).
. CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

252
The final multivariable prediction model of malaria test positivity was a poisson model with the lowest 253 AIC values was chosen from among a series of 21 models constructed using a combination of model 254 covariates. The model QQ plot showed that model predictions closely approximated the poisson . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted March 18, 2022. ;https://doi.org/10.1101https://doi.org/10. /2022 distribution (Supplementary figure 20a). The final model covariates were percentage of individuals with 256 post-primary education, percentage of individuals in the rich wealth quintiles, percentage of individuals 257 living in improved housing in 2015, all age population density, median age, percentage of children under 258 the age of five that sought medical treatment for fever, total precipitation and enhanced vegetation 259 index.

260
Clusters with the lowest educational attainment and wealth were at highest risk for malaria 261 Socioeconomic variables were negatively associated with malaria transmission intensity, although the 262 effect was less pronounced and more uncertain in the multivariate analysis (Figure 4). Malaria test 263 posivity rate declined with increases in the percentage of individuals with post-primary education. 264 However, the effect of educational attainment flattens out beyond 20% in the unadjusted analysis 265 (Figure 4a). Declines in malaria test positivity rate were also observed with increases in the percentage 266 of individuals in the rich wealth quintiles, particularly between 0 and 50%, and 80 -100%. In both the 267 adjusted and unadjusted analysis, clusters in the lowest socioeconomic status were at highest risk for 268 malaria. is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

280
High population density and younger median age correlated with higher malaria transmission intensity

281
In the unadjusted analysis, malaria test positivity rate declined with increases in all age population 282 density, up to 20,000 persons per square kilometer, before flattening out and going back up after 30,000 283 persons per square kilometer (Figure 5a). In the adjusted analysis, malaria test positivity rate remained 284 mostly flat up to 20,000 persons per square kilometer after which increases in malaria test positivity was 285 observed, albeit with high levels of uncertainity (Figure 5b). Declines in malaria test positivity rate 286 followed increases in median age, particularly beyond median ages of 15 years old (Figure 5c -d).

298
Higher enhanced vegetation index was positively associated with U5 malaria test positivity rate

299
In the unadjusted analysis, increasing malaria test positivity rate was correlated with increases in 300 enhanced vegetation index, which characterizes vegetation cover and growth status. The greatest 301 reductions in malaria test positivity rate, while highly uncertain, was observed at around vegetation 302 indicies of 0.5 and 0.7 (Figure 6a). However, the effect of enhanced vegetation index was significantly 303 diminished in the adjusted analysis (Figure 6b). of five years that sought medical treatment for fever and malaria test positivity rate was nearly flat in 318 both the adjusted and unadjusted analysis, the highest test positivity rate were observed at roughly 319 between 20 and 50% (Figurer 7c -d). With regards to total precipitation during the month of the survey, 320 the highest malaria test positivity rate were observed at 200 meters in the unadjusted analysis whereas 321 the adjusted analysis, a clear negative relationship was observed (Figure 7e -f). Collectively, these 322 findings suggests observed relationships may be due to the presence of information detection bias 323 and/or residual confounding . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted March 18, 2022. ;https://doi.org/10.1101https://doi.org/10. /2022  CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted March 18, 2022. ;https://doi.org/10.1101/2022 across survey years, clusters sampled within the same months were compared but due to small sample 353 sizes, we were unable to account for geopolitical zone-related differences in malaria seasonality. 354 355 The distribution of covariates obtained from the DHS/MIS are also likely unrepresentative of urban 356 settings. For instance, educational attainment, as measured by the percentage of individuals with post-357 primary education (Secondary or college education) was low in majority of the sampled clusters whereas 358 a greater proportion of clusters fell in the rich wealth quintiles or had improved housing infrastructure. 359 Assuming that individuals correctly reported their educational attainment, it is improbable that most 360 clusters will fall in the rich wealth quintiles or have improved housing infrastructure given the well-known 361 positive correlation between educational attainment and these factors [46,47]. We attempt to overcome 362 these limitations through the use of modeled covariates, where possible. However, we also provide 363 visualizations of study covariates in the supplement to inform further study or data collection efforts. 364 365 The fitted lines depicting the relationship between U5 malaria test positivity rate and indicators of 366 educational attainment, wealth, age distribution and vegetation cover and growth are aligned with 367 findings from the existing literature [9,48]. In their summary of factors related to malaria transmission in 368 urban areas across sub-Saharan Africa, Silvia and Marshall cited the literature showing that access to piped 369 water, better refuse collection and improved access to prevention methods and treatment are among the 370 numerous factors that reduce malaria risk among those of higher socioeconomic status in urban areas [9]. 371 The higher risk of malaria infections among children has been previously described and be may be 372 attributed to the higher malaria burden observed in clusters with more children [48]. Areas with high 373 vegetation cover such as urban farms are well-known vector breeding sites [9]. However, the fitted line 374 for the medical treatment seeking for fever were in discordance with the research literature [50], and 375 suggest the presence of residual confounding or information detection bias. Additional investigations is 376 essential to understand the sources of confounding and bias related to treatment seeking. 377 378 Despite the aforementioned data gaps, the methods from this study can be applied to local data to inform 379 stratification interventions in urban areas. The multivariable model can be used to predict areas with high . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)

380
U5 malaria burden and explore the causal impact of various factors. The methods used for generating the 381 functional forms of the bivariate association with predictors can be helpful in selecting thresholds for 382 intervention prioritization and deprioritization. For instance, an effect plot that shows that urban localities 383 where < 20% of residents have post-primary education are at high malaria risk could be used to inform 384 prioritization of vector control tools to these areas, if transmission is believed to be local or for the 385 provision of prophylaxis if transmission is heavily influenced by mobility patterns. Improved data that 386 adequately captures malaria burden and related sociodemographic and environmental factors is required 387 for the application of the study methologies to meaningfully inform intervention prioritization and 388 deprioritization decisions. 389 390 The bivariate analysis provides several insights that require further exploration. Among them is the 391 observed regional differences in bednet use. We found that U5 malaria test positivity rate was highest 392 among clusters with the lowest bednet use in the Northern geopolitical regions whereas in the Southern 393 regions, malaria test positivity rate appeared to be uniform in the South West region irrespective of 394 bednet use or higher among those that use nets as in the South-East region (Supplementary figure 18a). 395 Further study is needed to clarify why the effects of net use would vary by geopolitical region. Additionally, 396 further research is needed to understand the role of employment type on community malaria 397 transmission. Using the 2018 DHS, which was the only survey year with employment type, we found that 398 increases in the cluster proportions of male partners who are agricultural workers correlated with small 399 increases in the number of malaria positives in children (Supplementary figure 12a). Routine surveillance 400 data has higher temporal resolution compared to survey data and associated data collection systems can 401 be used to collect data on how the distribution of different parental employment types impact malaria 402 transmission among children residing in urban settings. Understanding regional variance in bednet use 403 and the role that employment is important for the design of contextually relevant malaria interventions. 404 405 This study complements the existing research literature by describing spatial-temporal variations in U5 406 malaria test positivity and associated risk factor relationships in urban areas within Nigeria. Efficiency in 407 intervention distribution could potentially lead to reductions in malaria burden and when locally 408 representative data is available for individual cities, study methods can be applied to inform intervention 409 stratification strategies. With increasing urbanization, there is growing interest to address malaria in 410 urban areas as evidenced by the recent convening of experts by the WHO as part of a technical 411 consultation on the burden and response to malaria in urban settings. By highlighting the data gaps 412 inherent in the Nigerian DHS/MIS, we hope that this encourages additional investments to improve its 413 design and its utility, as well as in routine surveillance systems. 414 415 Acknowledgements: None . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted March 18, 2022. ;https://doi.org/10.1101https://doi.org/10. /2022