This study confirms that despite of substantial decline in the past 30 years, south rural Guinea Bissau still has high levels of child mortality. The under five mortality dropped about 44% in the last 30 years (from 1977-1986 period), with neonatal and infant mortality decreasing 43% and 37% respectively. In the last time period (2002-2007) under our investigation, neonatal death still accounts for 40% of all child mortality with another 38% deaths occurring during the post-neonatal stage below the age of one year, thus future interventions need to focus on these periods of life.
An increase in neonatal, infant and child mortality is observed in the interval of 1997-2001, and we hypothesize that this related to the armed conflict of 1998-1999. Although most of the fighting took place in the capital, up to a third of the country's population was displaced (estimated in 350 000) . No refugee camps were established, but international aid agencies were providing food only to the internally displaced people. Like Aaby et al. , our results suggest that the hosting population was also very affected by the 1998-1999 war.
The actual detrimental effect of the armed conflict could be even higher if results were presented only for interval of two years that the war lasted. Nielsen et al.  demonstrated a steep rising in child mortality in a population that fled from Bissau during the war, showing a peek of under-five mortality between June and November 1998, when it was 2.07 times higher than expected. In the same area, the hosting population showed higher levels of malnourishment and child mortality than refugees .
For the most recent birth period (between 2002 and 2007) the estimated U5MR was 135 per 1,000 live births (95% CI: 127, 143) (Table 5). This figure is slightly lower than the published estimate for mid 2006 in rural Guinea Bissau that is 179 per 1,000 births . The difference could be explained by the fact that rural data in MICS  refers to the whole country and it has been estimated that the mortality level in southern regions is lower than in other areas (east and north) .
Many studies have shown that ethnicity affects child mortality [12–17]. Yet, it is difficult to identify the mechanisms that explain observed differences, and these are not necessarily the same for different populations. Several aspects linked to ethnicity may underlie the differences in mortality. Ethnicity may define dissimilarities in socioeconomic characteristics, child care, use of medicines, and health seeking behaviour , all aspects that have proven to play crucial roles as determinants of child mortality. Balantas, Fulas and Beafadas have similar economies, based on small-scale agriculture with some cattle husbandry, and it is not clear whether there is a socioeconomic disadvantage by any ethnic group. These ethnic groups, however, have different religions, rituals and settlement patterns. Balantas are Animists, whereas Fulas and Beafadas are Muslims. Their beliefs regarding death are very different, but it is not clear if this results in more or less pragmatic use of western medicines. Although all ethnic groups use traditional doctors and medicines, it has been suggested elsewhere  that the use of health services is lower among Balantas.
Overall our data show that Balantas have higher child mortality rate at all ages under five years compared to Fula and Beafada (Table 6). However, as shown in Table 7 this pattern is not consistent over birth periods, with Fula showing higher child mortality in the most recent period. There is no evident reason for why we observe this temporal pattern.
The published literature is also inconclusive about whether Balantas have higher mortality compared to other groups. A longitudinal (1990-1995) study that followed children from rural villages in other regions of Guinea Bissau (Bafatá, Biombo, Cacheu, Gabú and Oio) suggests that Balantas have higher neonatal mortality than other ethnic groups due to lower vaccination coverage and antenatal care . However, a different study showed that in the 1983 measles epidemic, Balantas had a lower risk of dying of measles compared to other ethnic groups [19, 20]. The authors suggest that less overcrowding in Balanta houses could be an important factor to explain their observation . In contrary, there is evidence in our study that Balanta women have more co-wives than Fula and Beafada and often live with their co-wives. This would lead to the opposite hypothesis, i.e. Balanta children live in houses with more children. However, with the present data it is not possible to confirm whether there is a significant difference in crowding of children per house. In the capital, Bissau, an ongoing demographic surveillance suggests that Pepel group has higher mortality compared to other ethnic groups . On the other hand, national data from MICS  suggests that U5MR are higher among Balanta than Beafada, Fula and other groups.
More research is needed to explain the ethnic differences in mortality, addressing specifically whether these are related to differences in health seeking behaviour rather than just to the physical distance to a health centre, and to investigate the role of household crowding.
Our study showed a slight effect of the distance to the nearest health centre (walking time) on child mortality. Although it is reasonable to expect this association, this could simply indicate the 'level of isolation' of these villages rather than being a measure of the importance of these health centres. A study carried out in Bissau showed that in spite of good health seeking behaviour, the low quality of health services in health centres, especially in recognizing the severity of cases, largely contributed to infant and child mortality . Thus, it is important to assess the quality of services in health centres as well as the care-seeking behaviour. Child mortality is usually lower in urban centres, especially between the ages of one and five years . Even though being close to an urban centre is linked to having a health centre nearer, there are other aspects like schools, quality of housing, socioeconomic status that need to be considered as they might be underlying the effect of health centres proximity.
The fact that we cannot control for maternal education and socioeconomic status might have influenced our results. It has been shown in a national survey that there is a difference in the number of literate women per ethnicity. According to MICS 2006 , a higher percentage of Balanta women (23%) were literate compared to Fula and Mandinga women (17%). Socioeconomic status could be also related to 'distance to the health centre' as wealthier families are more likely to live close to facilities.
Child mortality in Africa is still poorly studied, and empirical data during war are often of poor quality or not available. To achieve substantial decline in child mortality it is essential to understand local patterns. This study adds to our knowledge on child mortality in Guinea Bissau, and confirms the increase in child mortality during the civil war period in the southern rural area of the country. It also shows a difference in under-five mortality by the different ethnic groups, although the difference is changing by the period of births.