Menu
Menu
Menu
Menu
Background. Infant mortality is an important indicator of population health in a country. It is associated with several health determinants, such as maternal health, access to high-quality health care, socioeconomic conditions, and public health policy and practices. Methods. A spatial-temporal analysis was performed to assess changes in infant mortality patterns between 1992-2007 and to identify factors associated with infant mortality risk in the Agincourt sub-district, rural northeast South Africa. Period, sex, refugee status, maternal and fertility-related factors, household mortality experience, distance to nearest primary health care facility, and socio-economic status were examined as possible risk factors. All-cause and cause-specific mortality maps were developed to identify high risk areas within the study site. The analysis was carried out by fitting Bayesian hierarchical geostatistical negative binomial autoregressive models using Markov chain Monte Carlo simulation. Simulation-based Bayesian kriging was used to produce maps of all-cause and cause-specific mortality risk. Results. Infant mortality increased significantly over the study period, largely due to the impact of the HIV epidemic. There was a high burden of neonatal mortality (especially perinatal) with several hot spots observed in close proximity to health facilities. Significant risk factors for all-cause infant mortality were mother’s death in first year (most commonly due to HIV), death of previous sibling and increasing number of household deaths. Being born to a Mozambican mother posed a significant risk for infectious and parasitic deaths, particularly acute diarrhoea and malnutrition. Conclusions. This study demonstrates the use of Bayesian geostatistical models in assessing risk factors and producing smooth maps of infant mortality risk in a health and socio-demographic surveillance system. Results showed marked geographical differences in mortality risk across a relatively small area. Prevention of vertical transmission of HIV and survival of mothers during the infants’ first year in high prevalence villages needs to be urgently addressed, including expanded antenatal testing, prevention of mother-to-child transmission, and improved access to antiretroviral therapy. There is also need to assess and improve the capacity of district hospitals for emergency obstetric and newborn care. Persisting risk factors, including inadequate provision of clean water and sanitation, are yet to be fully addressed. © 2010 Sartorius et al; licensee BioMed Central Ltd.
The Agincourt health and socio-demographic surveillance system (HDSS), established in 1992, extends over an area of about 400 km2 and consists of 21 villages with approximately 11,700 households and a population of 70,000 people at the end of 2007 (Figure (Figure1).1). A full geographic information system (GIS) covers all households within the site and is updated annually. For these analyses the study population consisted of all infants who were either born or migrated into the site between 1992 and 2007 and who either survived or died in their first year of life. Location of the Agincourt HDSS site [33], South Africa. A verbal autopsy (VA) was conducted on every death to determine its probable cause [34]. Interviews administered by trained lay fieldworkers were assessed independently by two physicians to determine probably cause-of-death. Where consensus could not be reached, a third independent medical assessment was made. The VA was first validated in the mid-1990s [35] and again in 2006 with particular reference to HIV/AIDS related mortality. International Classification of Diseases (ICD-10) was used to classify main or underlying, immediate and contributory causes of death. For this study, cause-specific analysis was limited to main causes from 1992-2006 as VA’s had not yet been assessed for 2007. Covariates included: infant demographic variables (gender, nationality); 5-year time periods; maternal factors (former refugee status, age at pregnancy, death in first year of child’s life, education); fertility factors (parity, birth intervals, sibling death); household mortality experience, socio-economic status (SES) and food security; distance to health facility; antenatal clinic attendance; and household elevation (climatic proxy). Every two years since 2001, an asset survey was conducted in all households within the HDSS [36]. Information on living conditions and assets, building materials of main dwelling, water and energy supply, ownership of modern appliances and livestock, and means of transport available were recoded (one being higher SES and zero lower status), summed to give an overall score for a household, and then used to construct wealth quintiles for SES ranked by increasing score from most to least poor. The negative binomial is an alternative for the commonly used Poisson distribution, often regarded as the default distribution for integer count data. The Poisson assumes that expected mean equals its variance. The negative binomial differs from the Poisson distribution in that it allows for the variance to exceed the mean. Since the negative binomial distribution has one more parameter than the Poisson distribution, the second parameter is used to adjust the variance independently of the mean. Our data displayed evidence of being highly overdispersed and thus the negative binomial model was chosen. A preliminary negative binomial regression analysis was carried out to assess the relationship between infant mortality and each covariate. Covariates significant at the 10% level (without substantial missing values) were then incorporated into the multivariate model. The multivariate Bayesian negative binomial model was fitted in WinBUGS to examine the association between the significant covariates and all-cause infant mortality. Observation dates were used to calculate the person-days contributed by each infant (offset). Spatial random effects were used at a village level to take into account spatial correlation. Temporal random effects were also used at yearly intervals to account for temporal correlation. Village specific random effects were modelled via a multivariate Gaussian process (multivariate Gaussian distribution with covariance matrix expressed as a parametric function of distance between pairs of village centroid points) [37]. Standard Bayesian autoregressive (AR) approaches, with priors for the AR(1) and AR(2) processes defined by Schotman [38] and Zeller [39] respectively, as well as a Poisson generalized autoregressive moving average (GARMA) approach [40], were tested to model the temporal random effects. Various order models for the AR and MA terms were assessed and the one that best fitted the data was used. MCMC simulation was employed to estimate the model parameters [41]. Further details of the statistical modelling approach are given in the appendix. Deviance Information Criterion (DIC) [42] was used as the first step in comparison of model fit and the one giving the lowest DIC was chosen. Models were then also validated by fitting the models for 1992-2006 and predicting outcomes for all infants in 2007. Credibility intervals were constructed and the model providing the best predictions (along with low DIC) were used as the final model. The negative binomial models, particularly the AR(1) and AR(2) to model the temporal random effect, provided the lowest DIC (8618.07 and 8617.34 respectively) by some margin when compared to other approaches such as GARMA. In Bayesian statistics, a credible interval is a posterior probability interval which is used for interval estimation, in contrast to point estimation (confidence intervals). In other words, the credibility interval refers to the distribution of parameter values while a confidence interval pertains to estimates of a single value. In this study the negative binomial AR(2) predicted the outcome much better than the AR(1) model based on these Bayesian credibility intervals. Thus the AR(2) process was used in the final multivariate model. A baseline model was used that included no covariates but a constant and site-specific (village centroid) random effect. All identifying features (village centroids, geographic boundaries) were removed, and the prediction area expanded irregularly (~740 km2) to double the normal size, in order to ensure confidentiality and avoid stigmatizing of villages. The HIV/TB map is not shown for this reason. Simulation-based Bayesian kriging [43] at prediction points (regular grid) within the site was used to produce maps of mortality risk for the whole HDSS site. Model estimates were exponentiated to represent incidence rate ratios (IRR). Data extraction and management was done using Microsoft SQL Server 2005. The analysis was carried out in STATA 10.0, WinBUGS and R. The predictions of the fitted spatial models were mapped in MapInfo Professional 9.5.
N/A
