Child survival and BCG vaccination: a community based prospective cohort study in Uganda

The study was conducted in Mbale district, 300 km North-East of Kampala with an estimated population of 720,000, at the time of study commencement in 2006 [23]. The study area is served by Mbale Regional Referral Hospital, which double as the district and regional referral hospital. The HIV prevalence among pregnant women in antenatal clinics in Mbale was approximately 5% during the study period. Most of the people living in the area were subsistence farmers. The study was conducted in the two biggest of the 7 counties in Mbale district, namely Bungokho County (rural) and Mbale Municipality (urban). Twenty four clusters were included in the study, 18 rural and 6 urban. Six urban clusters in Mbale municipality were selected from all its three municipal divisions. Most of the urban areas were large informal settlements (urban slums). Eighteen rural clusters in Bungokho County were chosen from eight out of its eleven sub-counties. Clusters were included if they neighboured the main road out from Mbale Municipality or were on the 1^st or 2^nd branch off the main road, had a population of at least 1,000 inhabitants and represented a social and administrative unit. Children in the study area receive vaccines on the Ugandan immunization schedule which includes BCG, DPT-HepB + Hib, polio and measles. BCG vaccine (BCG-Denmark) is administered intra-dermally, on the right upper arm at birth or at first contact (dosage 0.05 ml up to 11 months and 0.10mls after the 11 months). DPT-HepB + Hib is administered in three doses, of 0.5mls each at monthly intervals, starting at 6 weeks (or first contact after that age). Polio is administered orally, in four doses (2 drops each). The first dose is given at birth or within the first two weeks and the subsequent doses are monthly starting at 6 weeks of age. Measles is given as a single dose at nine months of age.

Between January 2006 and May 2008, all pregnant women in the selected clusters were identified by study recruiters. Recruiters were mostly women representatives on the local village councils in the study area. They were well respected members of the community, with good knowledge of the local geography, community members and social structures. They regularly went door to door in order to identify new pregnancies. The pregnant women were then approached by the study team and invited to participate in the trial if they resided in the study area, were seven or more months pregnant, opted to breastfeed their infants, were not intending to leave the area during the study period, and consented to participate in the study. Of the 886 pregnant women who were identified and approached, 875 (99%) accepted to participate. Of these, 12 (1.4%) women did not meet the eligibility criteria and 28 (3.2%) relocated out of the study area after recruitment, while 16 (1.8%) women experienced a stillbirth. We analysed data for the remaining 819 women.

Ethics approval was obtained from the Makerere University Research and Ethics Committee, the Uganda National Council for Science and Technology, and from the Regional Committee for Medical and Research Ethics for Western Norway (REK VEST, approval number 05/8197).

At recruitment, trained data collectors fluent in the local language administered a pre-tested structured questionnaire in the local language, Lumasaaba. They collected information on the current pregnancy as well as on socio-demographic characteristics, antenatal care attendance (ANC) and marital status. After birth, the mother/child pairs were visited at weeks 3, 6, 12 and 24, between 18–24 months of age, between 3 and 5 years of age and between 5 and 7 years of age. Data on vaccination, illness and the child’s vital status was collected at each of the visits. Information on vaccination status was obtained through maternal interviews and review of child health cards. To minimize information bias, child health card data was prioritized in this study. A standard World Health Organization (WHO) verbal autopsy questionnaire that had been validated in Uganda was used to collect information for a standard algorithm determining the likely cause of death [24]. The questionnaire had both an open-ended section for reporting verbatim and a closed-ended section with filter questions.

We categorized marital status into three categories: ‘Married’, ‘Co-habiting’ and ‘Other’. The ‘Other’ category included women that were divorced, widowed or separated. In Uganda, it is now common to find couples living together without being formally married and we classified these as ‘Co-habiting’. Place of birth was categorized into ‘home births’, which were those that took place at home with or without a traditional birth attendant or those occurring on the way to a health facility. Health facility births were those that happened at a hospital, clinic or local maternity unit. We defined parity according to the number of previous live births. Based on the Ugandan education curriculum in which primary education is scheduled for seven years, education was grouped into ‘7 years of school or less’ and ‘more than 7 years of school’. Age was categorized into four categories (less than 20 years, 20–24, 25–30 and greater than 30 years).

We created a composite index of assets (socio-economic status) using multiple correspondence analysis (MCA). Because the MCA technique allows combination and ranking of a large number of variables into fewer variables without prejudgment, it is considered a more accurate indicator of socioeconomic status (SES) than single items such as occupation or possession of particular items [25]. Also, in comparison to principal component analysis (PCA), the MCA technique is more appropriate for discrete variables. This was important in this study because several relevant variables could only be categorical. Furthermore, unlike PCA, which clusters variables together, MCA clusters the categories within these variables together [25]. We used MCA on possession of a TV, radio, mobile phone, chair, cupboard, refrigerator, type of toilet, type of house walls as well as presence of electricity and water in the home. Scores were obtained and categorized into centiles (the poorest 20%, middle 40% and the richest 40%).

Data was analysed with Stata version 9.2 (StataCorp LP, TX, U.S.). Continuous variables were summarized with means, standard deviations, medians, ranges and interquartile ranges while percentages were calculated for categorical variables. The primary outcomes were post-neonatal infant death and post-infancy death. We calculated their confidence limits with the exact method. The neonatal mortality risk was the proportion of deaths within the first 28 days of life, the infant mortality risk the proportion of deaths in the first year of life [26], all per 1,000 live births. Post-neonatal infant mortality risk was the proportion of deaths between one month and 1 year of age while post-infancy death risk was the proportion of deaths between 1 and 5 years of age. The main outcome variables were post-neonatal infant death and post-infancy death. The main exposure variable was BCG vaccination. Other variables included as potential confounders included maternal age, parity, mother’s education, place of birth, antenatal care attendance, marital status, residence and household wealth index.

In order to take into account variable follow up time, Cox proportional hazards models were used to estimate hazard ratios (HR) and 95% confidence intervals. In the models, BCG vaccinated children contributed person-time from the date of vaccination with BCG until death, migration or last follow-up visit while non-vaccinated children contributed person-time from birth till death, migration or last follow-up visit. The multivariable regression models included variables which, in a crude analysis, were associated with deaths yielding a P-value < 0.25. The final models were also adjusted for the child’s age in days at the time of vaccination and for the design effect of the cluster-randomized PROMISE-EBF trial. We also undertook a so called landmark analysis to thwart the effects of a possible survival bias induced by the so-called “immortal person-time” [27].