Impact of insecticide-treated net ownership on all-cause child mortality in Malawi, 2006-2010

Insecticide-treated nets (ITNs) have been shown to be highly effective at reducing malaria morbidity and mortality in children. However, there are limited studies that assess the association between increasing ITN coverage and child mortality over time, at the national level, and under programmatic conditions. Two analytic approaches were used to examine this association: A retrospective cohort analysis of individual children and a district-level ecologic analysis. To evaluate the association between household ITN ownership and all-cause child mortality (ACCM) at the individual level, data from the 2010 Demographic and Health Survey (DHS) were modeled in a Cox proportional hazards framework while controlling for numerous environmental, household, and individual confounders through the use of exact matching. To evaluate population-level association between ITN ownership and ACCM between 2006 and 2010, program ITN distribution data and mortality data from the 2006 Multiple Indicator Cluster Survey and the 2010 DHS were aggregated at the district level and modeled using negative binomial regression. In the Cox model controlling for household, child and maternal health factors, children between 1 and 59 months in households owning an ITN had significantly lower mortality compared with those without an ITN (hazard ratio = 0.75, 95% confidence interval [CI] = 0.62-90). In the district-level model, higher ITN ownership was significantly associated with lower ACCM (incidence rate ratio = 0.77; 95% CI = 0.60-0.98). These findings suggest that increasing ITN ownership may have contributed to the decline in ACCM during 2006-2010 in Malawi and represent a novel use of district-level data from nationally representative surveys. Household survey data from DHS Program were primarily used for this study. Demographic and Health Surveys were conducted in Malawi in 2000, 2004, and 2010.13–15 In addition, a Multiple Indicator Cluster Survey (MICS) was conducted in 2006 under the direction of United Nations Children’s Fund.16 Each of these surveys used a two-stage stratified sample to collect nationally representative data on sociodemographic and health indicators, including child mortality, child and maternal health interventions, and individual and household characteristics. The sampling frame for the surveys was determined based on enumeration areas from national censuses. Survey primary sampling units (PSUs/clusters) were selected with a probability proportional to population size within domains and households were selected with equal probability systematic sampling within clusters. Within a selected household all women between the ages of 15 and 49 were asked to complete an interview in which they were asked about their health and the health of their children born in the past 5 years. In addition, a full birth history was completed in which interviewed women were asked to provide date of birth and age at death for all children ever born. Finally, these surveys collected information on malaria-specific indicators including ownership and use of ITNs. In addition to national survey data, this study also used data on P. falciparum prevalence estimates among children 2 to 10 years of age (PfPR2–10) for 2010 from the Malaria Atlas Project.17 Rainfall and temperature data from the Famine Early Warning System18 and Moderate Resolution Imaging Spectroradiometer,19 respectively, were also incorporated at the cluster level. An individual-level model was developed to look at the association between ITN ownership and mortality in children 1–59 months of age using data from the 2010 DHS. Although it would be preferable to use data on ITN use by individual children as the exposure variable, the available survey data does not contain information on retrospective ITN use. However, the relationship between ITN use and ITN ownership is fairly linear; nonuse of ITNs is largely determined by lack of access.20,21 This analysis used birth histories of women interviewed for the 2010 Malawi DHS to construct a retrospective cohort of children for a 3-year period before the survey. The primary outcome measure was deaths in children less than 5 years of age during this period. The primary exposure of interest was ITN ownership which was derived from questions in the household questionnaire on bednet ownership, type of net, insecticide treatment of net, and duration of ownership. To identify an individual child’s exposure to an ITN, data on the duration of ownership of ITNs and the time of retreatment of nets (if any) was used to construct a time-varying indicator of ITN ownership for 3 years before the survey. Other time-varying covariates such as high malaria transmission season (December–May), child’s age, mother’s age, P. falciparum prevalence estimates among children 2–10 years of age (PfPR2–10), rainfall and minimum temperature were constructed. The age variables were constructed retrospectively using the child’s and mother’s reported ages from the time of the interview. PfPR2–10, rainfall and temperature lagged by 2 months were linked to DHS cluster locations using global positioning system (GPS) coordinates at cluster centroids. Other covariates of interest were included as static estimates from the time of interview. These include residence (urban or rural), household wealth quintile (based on a principal component analysis [PCA] of household assets), parity, mother’s education, and household water source. Cluster-level DPT3 immunization coverage (the percent of children 12–23 months of age with at least three doses of DPT immunization), cluster-level diarrhea prevalence (percent of children under five with diarrhea in the 2 weeks before the survey), and cluster-level coverage of skilled birth attendance (the percent of births in the past 5 years that were attended by a skilled health professional) were constructed for inclusion in models. If successful, matching mitigates confounding within nonexperimental study designs to allow causal inference to be drawn between exposure (ITN ownership) and an expected outcome (child deaths).22 Eligible children living in households with ITNs and those without were matched into strata based on key confounding factors. This preprocessing of data was done at the group level, and not by individuals, so no observations were dropped. Exact matching was conducted on collapsed attributes related to the child including household wealth (above/below median PCA score), urban/rural household residence, cluster-level PfPR2–10 (≥ 40% versus < 40%), DPT3 coverage at cluster-level (above/below median), and mother’s education (secondary+, primary, none) to reduce confounding, as has been recommended previously8 (Table 1). Exact matching was conducted using the MatchIt package (version 2.4-21) in the statistical package R.23 Basic information on variables included in the Cox Proportional Hazards model FEWS = Famine Early Warning System; MAP = Malaria Atlas Project; MODIS = Moderate Resolution Imaging Spectroradiometer; PfPR2–10 = Plasmodium falciparum prevalence rate in children age 2–10 years. The relationship between household ITN ownership and child mortality (deaths of children age 1–59 months) over the 36 months preceding the survey was assessed with Cox proportional hazards models using the matched data. The analysis time was measured in months since birth and matched strata were included as a shared frailty, which is the equivalent of a random effect to account for unobserved heterogeneity. The model was further adjusted for several time-varying covariates: season (high malaria transmission season December–May), child’s age (categorical), mother’s age (categorical), rainfall (lagged 2 months) and minimum temperature (lagged 2 months); and for several other covariates including cluster-level DPT3 immunization coverage (among children 12–23 months), cluster-level diarrhea prevalence (in the 2 weeks before the survey among children less than 5 years of age), household wealth quintile, and parity (Table 1). Rainfall and minimum temperature lags were used due to previously documented lagged relationships between climate variables and clinical incidence.11,24 The household water source (improved or not) and the cluster-level coverage of skilled birth attendance were omitted from final models due to nonsignificance. Additionally, separate models were compared for rural and urban areas, respectively, the effect of ITN age was explored through a model with ITN ownership categorized by time since the ITN was obtained (greater than or less than 18 months [1.5 years]), and child age (less than or equal to 3 years of age versus older than 3 years of age) was evaluated as a moderator of the effect of ITN ownership through a model including an interaction term. Finally, the association between the number of ITNs owned and child mortality was assessed in a model restricted to only those households with at least one ITN. The association between ITN ownership and child mortality was also assessed using a district-level model in which data were aggregated by district and by year using data from the DHS and the Malawi 2006 MICS. These two surveys are some of the few nationally representative surveys powered to provide estimates of malaria indicators and other child health indicators at the district-level. The outcome of interest was the number of under-five deaths per district per year which was estimated retrospectively from the birth histories collected in the 2010 DHS (see Supplemental Table 1). Two models were developed with different primary exposure variables measuring ITN coverage. First, district-level average household ownership of ITNs was used as the primary exposure variable. Cross-sectional data from the 2006 MICS and the 2010 DHS were used to construct retrospective, district-specific estimates of ITN ownership for survey years. Weighted averages of district-level means for ITN ownership were constructed for interim years between surveys to produce annual estimates. Overall, ITN ownership data from 27 districts were available from 2006 and 2010. For interpretation purposes, district-year average ITN ownership values were categorized into binary values (high or low) based on whether the district-year values were above or below the grand mean. The second model used modeled ITN ownership from a combination of survey data and administrative records of numbers of ITNs distributed as the primary exposure variable. Data on annual, district-level ITN distribution were used and a decay factor was applied as per previously published methodology.25 This decay factor accounts for some loss of ITNs per year (either due to attrition or physical deterioration) and includes ITNs distributed in previous years to estimate total annual ITN ownership. ITN distribution data were also adjusted for midyear district-level population available through census data. Finally, the values were included with ITN ownership data from the 2010 DHS and the 2006 MICS in Loess regression models to create best-fit district-year estimates of ITN ownership. In addition, data on other sociodemographic and child and maternal health variables were needed for these models. Table 2 provides a description of the individual- and household-level variables that were aggregated at the district-level for modeling. The choice of variables was directed by literature on child survival and data available. As was done in estimating ITN ownership, cross-sectional data from the 2006 MICS and the 2010 DHS were used to construct retrospective, district-level coverage estimates of variables for survey years. Weighted averages of district-level means for these variables were constructed for interim years between surveys. Overall, data from 27 districts were available from 2006 and 2010. For interpretation purposes, retrospective, district-year average values for each variable were categorized into binary values (high or low) based on whether the values were above or below the grand mean. Cut-off values are presented in Table 2. Basic information on variables included in district-level models ITN = insecticide-treated net; MAP = Malaria Atlas Project; PfPR2–10 = Plasmodium falciparum prevalence rate in children age 2–10 years. In addition to these individual and household-level variables, district-level estimates of malaria transmission risk (PfPR2–10) and annual district-level rainfall anomalies were linked to DHS cluster locations using GPS coordinates of the clusters centroids for inclusion in the dataset. As was done with the other variables, the values were dichotomized into high and low categories based on the average district-level value. Cut-off values are presented in Table 2. Generalized linear models with negative binomial distributions were developed to assess associations between aggregated estimates of covariates at the district-year-level and numbers of deaths in children less than 5 years of age per district per year. These models were adjusted for robust clustering at the district-level to account for temporal correlation of the data. Model fit was assessed using the Bayesian Information Criterion. Poisson models were over-dispersed thus multivariable generalized linear models with negative binomial distributions and a person-month offset were used. Stata 13 (Stata Corporation, College Station, TX) was used for all analyses.