Household Demographic Surveillance System, Cause-Specific Mortality 1992-2012

Cause of death data based on VA interviews were contributed by fourteen INDEPTH HDSS sites in sub-Saharan Africa and eight sites in Asia. The principles of the Network and its constituent population surveillance sites have been described elsewhere [1]. Each HDSS site is committed to long-term longitudinal surveillance of circumscribed populations, typically each covering around 50,000 to 100,000 people. Households are registered and visited regularly by lay field-workers, with a frequency varying from once per year to several times per year. All vital events are registered at each such visit, and any deaths recorded are followed up with verbal autopsy interviews, usually 147 undertaken by specially trained lay interviewers. A few sites were already operational in the 1990s, but in this dataset 95% of the person-time observed related to the period from 2000 onwards, with 58% from 2007 onwards. Two sites, in Nairobi and Ouagadougou, followed urban populations, while the remainder covered areas that were generally more rural in character, although some included local urban centres. Sites covered entire populations, although the Karonga, Malawi, site only contributed VAs for deaths of people aged 12 years and older. Because the sites were not located or designed in a systematic way to be representative of national or regional populations, it is not meaningful to aggregate results over sites.

All cause of death assignments in this dataset were made using the InterVA-4 model version 4.02 [2]. InterVA-4 uses probabilistic modelling to arrive at likely cause(s) of death for each VA case, the workings of the model being based on a combination of expert medical opinion and relevant available data. InterVA-4 is the only model currently available that processes VA data according to the WHO 2012 standard and categorises causes of death according to ICD-10. Since the VA data reported here were collected before the WHO 2012 standard was formulated, they were all retrospectively transformed into the WHO 2012 and InterVA-4 input format for processing.

The InterVA-4 model was applied to the data from each site, yielding, for each case, up to three possible causes of death or an indeterminate result. Each cause for a case is a single record in the dataset. In a minority of cases, for example where symptoms were vague, contradictory or mutually inconsistent, it was impossible for InterVA-4 to determine a cause of death, and these deaths were attributed as entirely indeterminate. For the remaining cases, one to three likely causes and their likelihoods were assigned by InterVA-4, and if the sum of their likelihoods was less than one, the residual component was then assigned as being indeterminate. This was an important process for capturing uncertainty in cause of death outcome(s) from the model at the individual level, thus avoiding over-interpretation of specific causes. As a consequence there were three sources of unattributed cause of death: deaths registered for which VAs were not successfully completed; VAs completed but where the cause was entirely indeterminate; and residual components of deaths attributed as indeterminate.

In this dataset each case has between one and four records, each with its own cause and likelihood. Cases for which VAs were not successfully completed has a single record with the cause of death recorded as “VA not completed” and a likelihood of one. Thus the overall sum of the likelihoods equated to the total number of deaths. Each record also contains a population weighting factor reflecting the ratio of the population fraction for its site, age group, sex and year to the corresponding age group and sex fraction in the standard population (see section on weighting).

In this context, all of these data are secondary datasets derived from primary data collected separately by each participating site. In all cases the primary data collection was covered by site-level ethical approvals relating to on-going demographic surveillance in those specific locations. No individual identity or household location data are included in this secondary data.

1. Sankoh O, Byass P. The INDEPTH Network: filling vital gaps in global epidemiology. International Journal of Epidemiology 2012; 41:579-588.

2. Byass P, Chandramohan D, Clark SJ, D’Ambruoso L, Fottrell E, Graham WJ, et al. Strengthening standardised interpretation of verbal autopsy data: the new InterVA-4 tool. Global Health Action 2012; 5:19281.