Menu
Menu
Menu
Menu
Background Pneumococcal conjugate vaccine (PCV) and rotavirus vaccine (RV) are key tools for reducing common causes of infant mortality. However, measurement of population-level mortality impact is lacking from sub-Saharan Africa. We evaluated mortality impact and vaccine effectiveness (VE) of PCV13 introduced in November 2011, with subsequent RV1 roll-out in October 2012, in Malawi. Methods We conducted two independent community-based birth cohort studies. Study 1, in northern Malawi (40000population), evaluated population impact using change-point analysis and negative-binomial regression of non-traumatic 14–51-week infant mortality preintroduction (1 January 2004 to 31 September 2011) and postintroduction (1 October 2011 to 1 July 2019), and against three-dose coverage. Study 2, in central Malawi (465 000 population), was recruited from 24 November 2011 to 1 June 2015. In the absence of preintroduction data, individual three-dose versus zero-dose VE was estimated using individual-level Cox survival models. In both cohorts, infants were followed with household visits to ascertain vaccination, socioeconomic and survival status. Verbal autopsies were conducted for deaths. Results Study 1 included 20 291 live births and 216 infant deaths. Mortality decreased by 28.6% (95% CI: 15.3 to 39.8) post-PCV13 introduction. A change point was identified in November 2012. Study 2 registered 50 731 live births, with 454 deaths. Infant mortality decreased from 17 to 10/1000 live births during the study period. Adjusted VE was 44.6% overall (95% CI: 23.0 to 59.1) and 48.3% (95% CI: −5.9 to 74.1) against combined acute respiratory infection, meningitis and sepsis-associated mortality. Conclusion These data provide population-level evidence of infant mortality reduction following sequential PCV13 and RV1 introduction into an established immunisation programme in Malawi. These data support increasing coverage of vaccine programmes in high-burden settings.
We conducted two prospective population birth cohort studies in the Northern (Study 1) and Central (Study 2) regions of Malawi (online supplementary eFigure 1).24 Study 1 investigated PCV13 impact, defined as the population-level reduction in infant mortality; Study 2 investigated both impact and effectiveness, with VE defined as the individual risk ratio in vaccinated versus unvaccinated infants.25 bmjgh-2020-002669supp001.pdf Study 1 was conducted at the Karonga Health and Demographic Surveillance Site (KHDSS). KHDSS was established in 2002, currently covering a population of 40 000 people in 41 villages, with one rural government hospital and four health centres.26 Demographic data are included from 1 January 2004 to 1 July 2019, and vaccine coverage data are complete up to the 1 July 2017. PCV13 was introduced on 12 November 2011, therefore all children born from 1 October 2011 onwards were 6 weeks old at introduction and eligible to receive dose-one. This study provides a pre–post impact evaluation of PCV13 introduction. Births and deaths were recorded monthly by 230 community-based volunteers. Vaccine and socioeconomic status were collected for each household on an annual basis, using a rolling recensus by trained interviewers. Verbal autopsies (VAs) were conducted by medical assistants at a median of 1 month (range: 2 weeks to 20 months) following death, using a modified version of the WHO 2012 tool.27 All data underwent double entry into a Microsoft Access database and conflicts were flagged for cleaning. Impact in Study 1 was estimated ecologically using negative binomial regression of study area-wide annual trend in non-traumatic 14–51-week infant mortality pre-PCV13 and post-PCV13 introduction, adjusted for year to account for underlying downward trends in infant mortality and RV1 introduction. The annual trend was derived using locally weighted 12-month moving averaging as follows: where Y^t and Yt are the trend estimate and the observed incidence at month t. Additionally, we used a change-point model with the full time series to determine whether PCV13 and RV1 introduction occurred before significant trend changes in infant mortality. In change-point analysis an intervention time point is not prespecified, therefore, with fewer assumptions than interrupted time series analysis it assesses whether: (1) changes in incidence have occurred; (2) identifies the most likely time for the change point.28 We used the Stata—bayesmh—function to fit a negative binomial Bayesian model for the above specified locally weighted 12-month moving averaging annual trend. We used uninformative prior for the mean and a uniform prior for month (all values are equally likely), 50 000 MCMC iterations with 10 000 burn-in period and specified one change point. The resulting change point (month), pre and post change-point mean, mean ratio and the corresponding 95% credible intervals were calculated. Postintroduction impact was ecologically estimated using negative binomial regression of study area-wide mortality versus monthly three-dose PCV13 population coverage, adjusted for year. Coverage was calculated as the cumulative number of infants who received three-doses of PCV13, divided by the cumulative total number of age-eligible infants residing in the study area and surviving to 14 weeks. Small population size precluded individual VE analysis. Study 2 was conducted in Mchinji, a rural district with a population of 465 000 in 1832 villages, based on a census we conducted in March 2012. Healthcare was provided at 1 government hospital, 11 health centres, 354 community healthcare workers and 4 rural hospitals with limited inpatient facilities that provide care for a small fee. Cohort recruitment ran from 24 November 2011 (soon after PCV introduction) to 1 June 2015, and follow-up with mortality surveillance ran from 1 March 2012 (before RV introduction) to 1 June 2016. This site does not contribute any pre-PCV13 data, precluding a pre–post impact analysis. Pregnancies, pregnancy outcome and deaths in under 5 year olds were recorded monthly by 1059 volunteers. Field enumerators conducted household visits at 4 months and 1 year of age to collect vaccine status, socioeconomic variables and verify survivorship. Under-five deaths had a VA conducted at median 14 months (range: 2–50 months) following death using the WHO 2012 tool, with vaccine status recorded by senior monitoring and evaluation officers.27 Data were single entered into a Microsoft Access database with in-built validation rules and underwent automated monthly cleaning; errors in identification were sent for field verification. A random subset of 4-month and 1-year interviews were redone quarterly and all vaccine clinics were visited to audit documentation for quality control. The sample size for individual VE was calculated for 80% power to detect a 25% reduction in non-traumatic infant mortality, assuming 14–51-week infant mortality of 15/1000 live births, 80% three-dose vaccine coverage and 15% loss to follow-up. A sample of 45 520 births surviving to 14 weeks and 552 death events were required. In the absence of pre-PCV13 data, population-level impact in Study 2 was estimated using negative binomial regression of yearly mortality versus yearly vaccine coverage, by geographical cluster, adjusted for two-dose RV1 coverage. Coverage was calculated as the number of dose-eligible infants who received one, two or three doses of PCV13 by 52 weeks of age, divided by the total number of infants residing in the cluster and surviving to 14 weeks. Geographical clusters were 354 government-defined community healthcare worker catchment areas, with a median population of 1300 people (IQR: 984–1687). Unadjusted and adjusted individual-level VE of three versus zero dose PCV13 receipt against non-traumatic mortality in infants aged 14–51 weeks was estimated using Cox regression as the primary analysis.29 30 VE was derived as: PCV13 doses received were modelled as time-dependent covariates, using date of vaccination recorded from caregiver-held health records (health passports) to split survival time into vaccinated and unvaccinated periods. Missing vaccination dates were imputed using chained equations with 10 imputations; all variables included in the primary model were used in the imputation (online supplementary eMethods 1). Infants who migrated did not contribute any survival time as vaccine status could not be determined.31 The proportional hazards assumption was tested using Schöenfeld’s residuals. Decided a priori, analyses were adjusted for a range of potential confounding factors associated with both risk of mortality and vaccine uptake. These included maternal survival, education, age and marital status, household assets, household construction, water and sanitation facilities and a binary indicator of RV1 introduction. Individual RV1 and other EPI vaccine receipt (including Haemophilus influenzae B vaccine) were not included due to collinearity with PCV13 receipt, and too few children exclusively received PCV13 to conduct a subanalysis with this group. Distance to the nearest health facility (in kilometres) and season (rainy/dry) were investigated post hoc as possible proxies of vaccine access; however, neither showed any association with vaccine uptake, or survival, and were not included in the final model. The following sensitivity analyses were conducted: Royston-Parmar flexible parametric survival models to describe time-varying vaccine effects32; using acute respiratory infection (including pneumonia), meningitis and sepsis-associated mortality and diarrhoea-related mortality as the outcome (with children who died of other causes excluded from the model); using individuals who survived to 6 and 26 weeks as the eligible population30; random effects models to account for cluster-level effects. Analyses were conducted using Stata SE V.14. We recorded receipt of zero, one, two and three doses of PCV13, with three versus zero doses as the primary exposure of interest. Vaccine date and receipt were collected during interviews and VAs from health passports, or caregiver recall if a written record was unavailable (online supplementary eMethods 2). We recorded deaths among three-dose eligible infants (ie, aged between 14 and 51 completed weeks) from a non-traumatic cause, as defined by the WHO 2012 VA guidelines (online supplementary eMethods 3). This primary outcome, although aetiologically non-specific has the advantage of being free from limitations in cause of death classification using VAs.33 Cause-specific mortality was included as a sensitivity analysis, using InterVA-4, a probabilistic Bayesian algorithm, to automate the analysis of the VAs and assign probability weighted cause of death.34 Overall, InterVA has been found to have reasonable agreement (concordance coefficient=0.81) with physician coded cause of death in infants, but was lower for acute respiratory infections in similar settings.35 Prior to the start of the study, the protocol was presented to the District Executive Committee and District Health Management teams in Mchinji and Karonga districts for input and approval. Extensive community engagement was conducted for this new data collection activity in Mchinji district, including the recruitment of village-level volunteers and meetings with traditional leaders and area development committees. Community consent was sought during study introduction. Verbal informed consent was obtained for all interviews. The study was approved by the National Health Sciences Research Ethics Committee in Malawi [#837], London School of Hygiene and Tropical Medicine [#6047] and Centres for Disease Control and Prevention [#6268].
N/A
