Background. In developing countries with scanty resources it is very important to have reliable data to establish priorities for the health sector; e.g. to reduce maternal mortality it is necessary to determine the most important causes. The majority of deaths, however, occur without previous contact with the health system and consequently conventional analyses of death certificates are not feasible. Instead, studies have been carried out in some developing countries with various forms of post; mortem interviews, the so-called verbal autopsies (VA). Methods. We developed a structured interview with filter questions, which was applied to all deaths of women of fertile age in a cohort of 10,000 women living in 100 clusters in Guinea-Bissau and followed over a period of 6 years. The cause of death was ascertained by means of a series of diagnostic algorithms for the most common causes of maternal mortality, including postpartum haemorrhage, antepartum haemorrhage, puerperal infection, obstructed labour, eclampsia, abortion, and ectopic pregnancy. Results. Of the 350 deaths of women of fertile age, 32% were maternal and it seems unlikely that a significant proportion of maternal deaths have not been classified correctly. Using the diagnostic algorithm 70% could be given a specific diagnosis, the most important causes being postpartum haemorrhage (42% [29/69]), obstructed labour (19% [13/69]), and puerperal infection (16% [11/69]). We attempted to identify the factors that are critical for obtaining sufficient information to reach a diagnosis. In the univariate analyses, it was important whether the respondent had been present during the last illness (P = 0.04) and whether the death occurred more than one week after delivery (P = 0.04). The husband was a better respondent than a co-wife (P = 0.08), and men in general provided more specific information than women (P = 0.08). Furthermore, information appeared to be better if the woman had died in the rainy season (P = 0.08). The length of the recall period, parity, age of woman, place of death, rural/urban residence, and ethnic group were not decisive. In the multivariate analysis sex and presence of respondent and time after delivery were significantly associated with the risk of not reaching a specific diagnosis. Women are less likely to provide adequate information for a diagnosis than men (odds ratio [OR] 3.1; 95% confidence interval [CI] : 1.2-8.1). Respondents that did not reside in the village during the departed woman's illness/delivery carried equal risk of not reaching a conclusion (OR 3.1; CI : 1.1-9.1). Deaths occurring more than one week after delivery were also less likely to be classified (OR 6.1: CI: 1.7-22.0). Conclusions. The VA described in the present paper left 30% of the maternal deaths unclassified without a specific diagnosis. Had all interviews been with husbands, only 14% would have remained unclassified. If we had only asked people who were present during the terminal phase of the victim's illness the proportion of classified deaths would have risen from 70% to 75%. It is likely that delayed maternal deaths have not been adequately covered by the present algorithms, but they may also simply be more difficult to describe due to the duration of the disease episode. In contrast to methods by which cause of death is established by a panel of medical experts, the present VA should be economically and technically viable in areas where health workers have only minimal training.
