Search and study selection

This paper was prepared in accordance with the meta-analysis of observational studies in epidemiology (MOOSE) and PRISMA guidelines [Reference Stroup8,Reference Liberati9]. The literature searches were performed by one author (LVC) in January 2017 and updated in September 2017. PubMed, Web of Science, the Cochrane Library and the grey literature were searched for papers reporting carriage of N. meningitidis in defined age groups in locations within the African meningitis belt, using the following search terms: ‘(‘Neisseria meningitidis’ OR ‘N meningitidis’ OR meningitis OR meningococcal OR meningococci OR meningococcus OR meningitidis) AND (carriage OR carrier OR carriers) AND (Africa OR ‘meningitis belt’ OR Gambia OR Senegal OR Guinea-Bissau OR Mauritania OR Guinea OR Ghana OR ‘Burkina Faso’ OR ‘Upper Volta’ OR ‘Cote d'Ivoire’ OR ‘Ivory Coast’ OR Togo OR Benin OR Nigeria OR Niger OR Chad OR Tchad OR Cameroon OR ‘Central African Republic’ OR Sudan OR ‘South Sudan’ OR Eritrea OR Ethiopia OR Uganda OR Kenya OR ‘Democratic Republic of Congo OR Zaire)’.

Studies were eligible for inclusion if they reported pharyngeal carriage of all meningococcal serogroups in defined age groups in generalisable populations in the African meningitis belt. Longitudinal, cross-sectional and serial cross-sectional study designs were included. Studies reporting carriage rates among household contacts of Hajjis or other known carriers were excluded. Studies in both English and French were included in the systematic review. Two studies in Russian were excluded because abstracts could not be obtained. The reference lists of included papers were hand-searched to identify any papers that were missed by the electronic search.

Data extraction and classification

We contacted authors for additional information if studies reported age groups wider than 10 years and the median age was not reported or if the time period of the study was not reported. Where carriage prevalence was reported in age groups of range 10 years or less, we used the midpoint age for analysis. For open-ended age groups, we calculated a midpoint with an upper age limit of 60 years. Using additional data from the African Meningococcal Carriage Consortium cross-sectional studies, we divided the oldest age group previously reported as 30 plus years into two groups, 30–44 years and 45 plus years.

For papers that included results from several populations or studies, data were extracted only for observations meeting the inclusion criteria. Extracted data included the study location, time period, design, any mention of an outbreak or epidemic of meningitis coinciding with the study period, whether the study occurred before or after the MenAfriVac mass campaigns in the study region, use of random sampling, laboratory used to culture the pharyngeal swabs, time between sampling and plating, age range of the individuals, median age of individuals in each age group, number of individuals swabbed and the number of individuals positive for capsulated meningococcal carriage. We excluded individuals carrying capsule-null isolates from our definition of capsulated meningococcal carriage but isolates characterised as non-determinate or non-groupable were included.

Studies were assigned to the dry or rainy season on the basis of the month in which the study took place, defining the dry season as December–June and the rainy season as July–November. Additional data from the studies by Kristiansen and colleagues in Burkina Faso and Ethiopia were used to divide prevalence estimates from these studies into dry and rainy season measures.

Studies were classified as having used random sampling if this was specifically reported; otherwise ‘no’ was recorded.

Data were extracted independently by LVC and AK and differences were checked by CLT.

Data analysis

For the meta-analysis, we excluded data from age bands wider than 20 years. For longitudinal studies, only the first observation in the time series was used to avoid over-representing a particular population in the meta-analysis.

A natural cubic spline (a piecewise cubic polynomial function with linear tails) was used to model group median age [Reference Christensen2,Reference Marsh and Cormier10]. The basis matrix for the spline was generated using the ‘ns()’ function in the R core package ‘splines’. Boundary knots (points in the spline below and above which the function is defined to be linear) were placed at 0 and 30 years because of the low density of sampling in older age groups. The number of internal knots was selected by comparing the AIC of models with two and three internal knots, placing knots at appropriate quantiles (inner tertiles for two knots, inner quartiles for three knots. Optimal placement of these knots was determined by evaluating the AIC of models with every possible combination of knots drawn from the set of ages equidistant from every consecutive observed age between 0 and 30 years. Internal knots placed at 9.25, 9.75 and 28.5 years of age gave the best model fit.

We modelled season and outbreak status as a three-tiered fixed effect and a nested interaction term of location and year of swabbing as random effects. We used fixed effects for factors that had a constant effect in all studies and random effects for factors for which only a subset of all possible levels had been observed. We tried using the physical laboratory as a proxy for differences in ambient temperature, growth medium, swab type and other unreported factors which might affect the likelihood of isolating meningococci from a sample, but this did not improve model fit. Other factors that were considered but did not improve model fit included mass group A conjugate vaccination, country, decade, latitude, time to plating and study.

We started with a simple logistic regression because this is the traditional model for analysing prevalence data. However, when we found that a low proportion of observations fell within the bootstrapped 95% CI, we investigated three additional models which are commonly used to capture overdispersion: an observation-level random effects model, a beta-binomial or hierarchical model and a zero-inflated binomial model [Reference Harrison11]. We found that the simple logistic regression captured more of the variability in the data than the observation-level random effects model and provided a better fit than the beta-binomial and zero-inflated binomial models, so we used a logistic structure for our final model.

All candidate models allow overall carriage prevalence to vary but constrain the distribution of prevalence by age to be the same across all years and locations.

Models were fit using the lme4, spaMM and glmmADMB packages in R [Reference Bates12–Reference Rousset and Ferdy15]. To generate confidence intervals for the model predictions and random effects, 10 000 iterations of bootstrapping with replacement were performed and bias-corrected 95% confidence intervals were calculated [Reference Efron16]. Profile confidence intervals for fixed effects parameters were calculated using the lme4 package [Reference Bates12]. We assessed model fit by AIC and by performing leave-one-out cross-validation, whereby the model is refit on the full dataset excluding a single observation and this new model is used to predict prevalence in the excluded observation. This prediction is then compared to the true value. To check the results of the spline model, we identified the age group with the highest carriage prevalence for each study measuring carriage in more than one age group and performed a one-way proportion test between this peak value and prevalence in all other age groups. To adjust for repeated significance testing, we used an upper P-value cut-off of 0.0005, yielding an overall confidence of >95%.

We checked our season definition against average monthly rainfall anomalies in each site using publicly available data from the National Oceanic and Atmospheric Administration (Fig. S1). In a sensitivity analysis, we excluded data from four sites with non-characteristic rainfall patterns. These sites all fell on the edges of the meningitis belt: Butajira, Ethiopia (2 sub-sites), Arba Minch, Ethiopia (4 sub-sites) and Kpalkpalgbeni, Bring-Ahafo District, Ghana.

For the visual presentation of the fixed-effects portion of the model, we adjusted the observed carriage prevalence (P _observed) to account for the random effects in the full model as follows: P _adjusted = exp(log (P_observed)-σ_Location-σ_{Location-year}), where σ _Location and σ _{Location -year} are the random effects intercept for location and year for a given observation.

To estimate the proportion of carriers that would be targeted in a mass vaccination campaign covering four age groups (0–15 years, 0–17 years, 0–19 years and 0–29 years) we assumed a population structure consistent with that of Niger in 2017 and used the fitted estimate of carriage prevalence for the rainy season (although results are consistent across seasons) [17]. We accounted for uncertainty in this estimate by simulating 100 000 draws from a uniform distribution with boundaries at the 2.5^th and 97.5^th percentile of the fitted prevalence, using these draws as the rates for a random binomial draw with population size corresponding to each single year age cohort in Niger 2017 and then taking the 2.5^th and 97.5^th percentiles of the proportion covered to yield a 95% confidence interval.

Assessment of study quality and heterogeneity

We assessed the role of study design by comparing a model with and without sampling procedure (random vs. non-random) and with and without time to plating of swabs as fixed-effect variables.

Role of the funding source

The sponsor of the study had no role in study design, data collection, data analysis, data interpretation, or writing of the report. The corresponding author had full access to all the data in the study and had final responsibility for the decision to submit for publication.