Study Design and Oversight
In this prospective cohort study, we followed children from early infancy to adulthood. The study was approved by relevant ethics review boards in Tanzania, Switzerland, and the United Kingdom. Parents or guardians of young children gave oral consent for data collected from 1998 through 2003. In 2019, we obtained written consent from participants interviewed in person and oral consent from those interviewed by telephone. The first and last authors voucher for the completeness and accuracy of the data.
This study was conducted within the Ifakara Rural Health and Demographic Surveillance Site (HDSS) in the Kilombero and Ulanga Districts, Tanzania.13 The study area originally comprised 18 villages, which were later split into 25 (Fig. S1 in the Supplementary Appendix, available with the full text of this article at NEJM.org). During household visits conducted every 4 months from May 1998 through April 2003, all the children born to an HDSS resident between January 1, 1998, and August 30, 2000, were enrolled in the longitudinal cohort study. From 1998 through 2003, participants were followed by means of HDSS visits every 4 months (Fig. S2). From 2004 through 2015, the survival status of participants known to be residing in the area was recorded at routine HDSS visits. In 2019, we conducted a follow-up survey through community outreach and mobile telephones, verifying the survival status of all the participants, independent of residency and HDSS records. This survey relied on household information provided at the time of enrollment. We created search lists for each HDSS village, which showed the first and last names of all former household members for each participant, as well as the date of birth and the community leader responsible for the household at the time of enrollment. In meetings with local community leaders, this list was reviewed, and additional community members were identified to help with the tracking.
Social Marketing of Treated Nets
With support from the Swiss Agency for Development and Cooperation and the government of the United Republic of Tanzania, a program for implementation research on treated nets was set up in the study area in 1995.14 In 1997, a social marketing initiative to distribute, promote, and recover part of the cost of nets and net treatment was introduced.15 A nested case–control study showed that treated nets were associated with a 27% increase in survival among children 1 month to 4 years of age (95% confidence interval [CI]3 to 45).15
Outcomes and Variables
The primary outcome was survival that was verified during household visits. For participants who had died, the age and year of death were obtained from parents or other family members. The primary exposure variable was the use of nets between birth and 5 years of age (“early-life net use”). We analyzed both individual use and community-level availability of nets. For individual net use, the child’s mother or caregiver was asked during each household visit during the 1998-2003 period whether the child had slept under a net the previous night and, if so, whether and when the net had been insecticide-treated or washed . We summarized each child’s early-life exposure to treated nets as the percentage of visits in which children were reported to be sleeping under a treated net. For village-level ownership of treated nets, we combined all household records collected from 1998 through 2003 to compute the proportion of households in each village, and according to year, owning at least one treated net.
Data on malaria parasitemia were collected in 2000 as part of the Integrated Monitoring Project for Antimalarial Combination Therapy.16 Parasitemia was measured by means of thick-film microscopy among all household members 6 months of age or older in a representative sample of HDSS households in May through July of 2000, 2001, 2002, 2004, 2005, and 2006.16
Data quality and completeness
To maximize data quality and completeness of follow-up in 2019, we recruited and trained a team of experienced interviewers who already possessed extensive local knowledge. For some families, information on caregiver education, household income, and time to reach a health facility was not available. Multiple imputation with chained equations was used to address missing covariate data in our main results. All the variables that are listed in table 1 were included as predictors for these imputations. Additional complete-case analysis was conducted to ensure that the results were not sensitive to the imputation method chosen.
Initial descriptive statistics included mean person-time of follow-up and mortality according to sex, year of birth, caregiver education, and household income categories. Mortality was estimated as the number of deaths per 1000 person-years.
We present data on how net coverage changed over time. To illustrate the relation between household ownership of treated nets at the village level and local malaria transmission, we created scatter plots of village-level treated net coverage against village-level parasitemia prevalence in 2000.
To estimate the associations between net use and long-term survival, we first estimated unadjusted standard Kaplan-Meier survival curves comparing survival outcomes among children reported to be sleeping under a treated net during at least 50% of the early-life visits with those among children reported to be sleeping under a treated net during less than 50% of the early-life visits. The 50% cutoff was chosen to match a simple “majority of the time” definition. To ensure that results were not driven by this arbitrary cutoff, we also estimated unadjusted standard Kaplan-Meier survival curves comparing survival outcomes among children always reported to be sleeping under a treated net with those among children never reported to be sleeping under a treated net. We estimated unadjusted Kaplan–Meier curves for these contrasts both for the full period (0 to 20 years of age) and for the post–early-childhood period (5 to 20 years of age). All the survival analyzes were restricted to the time between the first survey interview and the last survey interview, which resulted in both left truncation and right censoring.
We used Cox proportional-hazards models to estimate three primary contrasts of interest conditional on observable confounders — first, the association between survival and the percentage of visits in which children were reported to be sleeping under a treated net; second, survival differences between children using treated nets at half the visits or more and those using treated nets at less than half the visits; and third, differences in survival between children always reported to be sleeping under a treated net at their early-life visits and children never reported to be sleeping under a treated net at these visits. For the first association, the percentage of visits was analyzed as a linear term. Martingale residual analysis was conducted to confirm the appropriateness of this linearity assumption. Schoenfeld residual analysis17 was used to verify the proportional-hazards assumption. To address confounding concerns, all multivariable estimates for these first three contrasts were adjusted for household income category, time to the nearest health facility, the caregiver’s education category, the child’s sex, and the child’s year of birth. All multivariable models also included 25 village-specific intercepts, which allowed us to rule out systematic differences in unobservable village-level factors as potential confounders. To ensure the robustness of the presented results with respect to the empirical model chosen, we also estimated the two binary contrasts using kernel, caliper, and exact matching algorithms.
Given that early-life use of treated nets could be explained by unobserved household or caregiver traits such as health knowledge or individual ability to access health services, we also estimated a village-level model as a fourth contrast. For this contrast, we used village-level mean household ownership of treated nets in the first 3 years that the child was observed (entered as a linear term) as our primary exposure variable. Village-level exposure has the advantage of being less dependent on individual or household-level covariates and should thus be less subject to confounding. Conceptually, increasing village-level coverage should yield larger protective effects than increasing individual coverage owing to larger effects on mosquito populations and malaria transmission.18
To account for village-level net treatment as well as village-level correlations more generally, standard errors were calculated with the use of Huber’s cluster-robust variance estimator. Results are reported as point estimates with 95% confidence intervals. The widths of the confidence intervals were not adjusted for multiplicity, so the intervals should not be used to infer definitive associations. Our primary analysis was not prespecified; therefore, no P values were reported. The statistical analysis was conducted with the use of Stata SE software (StataCorp), version 16.0.19