Overall, the Agincourt study shows that in rural South Africa, and potentially Southern African settings more widely, pregnancy is not a risk factor for most diseases other than obstetrical conditions. Indeed, the selection in favor of healthy pregnant women was so strong that, for most causes of death, mortality during the maternal risk period was about half of that outside the MRP. This effect remains poorly documented elsewhere, making it difficult to compare the Agincourt findings with other situations.
This finding raises questions on the definition and validity of “indirect cause” concept 
. If there is no increased risk, and therefore no attributable risk arising from the pregnancy, is it possible for other conditions to be an “indirect cause” of maternal death? This issue has major implications for the measurement of maternal mortality in situations where HIV/AIDS is highly prevalent. If most causes other than obstetrical causes have little to do with the pregnancy, why include them when computing maternal mortality? This serves only to inflate the maternal mortality ratio (MMR), is misleading when comparing data over time and space, and can seriously bias reported trends in MMR. Would it not be a better reflection of reality to restrict the definition of maternal mortality to obstetrical causes of death?
There are a few conditions classified among the indirect causes known to be associated with pregnancy. Among infectious diseases, this is the case for instance of listeria, a rare disease for statistical purposes. Among chronic diseases, this is the case for instance of rheumatic heart disease, which may be exacerbated by pregnancy. More important, there are some violent deaths (homicide or suicide) which may be the consequence of pregnancy. These cases are important, but they require in-depth specific studies, and are unlikely to be properly captured in causes of death statistics of indirect causes. Categorizing these cases according to their ICD code, and coding independently the pregnancy status will be more informative for further statistical analysis. Analyzing them in specific studies, such as ‘Confidential Inquiries’ will also provide more useful information.
A similar argument could be made for HIV/AIDS and tuberculosis. Mortality from these conditions went up and down in Southern Africa, because of increasing HIV prevalence followed by widespread use of ARV treatment. Monitoring these changes is important for public health purposes, and coding separately whether women were pregnant or not is a valuable additional information. Monitoring toxic effects of ARV is also important. However, these points are beyond the classic coding of causes of death, and deserve further and specific research.
Selection in favor of healthy pregnant women can be easily understood. When women have a chronic infectious or non-communicable condition that may be terminal, they are less likely to become pregnant both for biological and behavioral reasons. As a result, women who are selected into partnership and pregnancy are less likely to die from these conditions. This adverse selection is more likely to apply to chronic conditions such as HIV/AIDS, TB, cancer, diabetes, hypertension and similar conditions. It is less likely to apply to acute conditions unless strong behavioral factors intervene, in which case the effect would probably be temporary.
The major causes of death among women 15–49 years were HIV/AIDS and TB accounting for 71% of determined causes for non-pregnant women. Selection for healthy pregnant women was therefore important in this population. We did not find comparable quantitative evidence elsewhere in the literature and this finding deserves further work. The lower risk of death from HIV/AIDS or TB remains plausible since the comparison is between women in the final stages of these diseases who die within a few months, and women who were healthy enough to become pregnant and therefore less likely to die within the maternal risk period. This does not mean that HIV/AIDS does not have any effect on the outcome of the pregnancy, whether through immunological and metabolic effects (such as anemia), or via treatment (ARVs may have an impact on hypertension and hepatic function). It simply means that the selection effect is much stronger and thus outweighs any deleterious effect.
The example of malaria deserves comment despite the small number of cases. Malaria can be an acute condition, determined by a mosquito bite and vector-borne transmission independent of the pregnancy. Here, one would expect the same mortality risk as for non-pregnant women. However, it can also be a chronic condition which can worsen because of the pregnancy by aggravating anemia. In this case one could expect an increased risk during pregnancy. More research on these effects is needed as the literature on this topic is controversial and the health outcome depends also on the local malaria profile (parasite strain, pattern of transmission), on treatments, and on parity. In Agincourt, the four malaria deaths among pregnant women occurred at younger ages (<35 years), whereas for non-pregnant women malaria deaths were more evenly spread (18 cases below age 35 years, 11 cases above). This pattern suggests an interaction between age, pregnancy and malaria, and in particular a higher risk associated with the first pregnancy, as found elsewhere 
Selection for cancer-related death or mortality from diabetes seems to be of similar nature as that associated with HIV/AIDS. The only cancer death among pregnant women was a case of cervical cancer in a 33 year old. There was no death from diabetes among pregnant women, compared with 12 deaths among the non-pregnant. Among the five cardiovascular disease deaths in pregnant women, one could be related to the pregnancy: a 23-year old woman who died of rheumatic heart disease, a condition exacerbated in pregnancy.
Selection for accidental deaths was also strong, since no road traffic accident and no other accident was noted among pregnant women, whereas these causes accounted for 2.3% (n
32) of known causes of death among non-pregnant women. This could however be an effect of sample size.
Among deaths from violent causes, there was one suicide of a 19-year old which might have been related to the pregnancy. Of the two homicides, one seemed independent, a 16-year old killed by “bodily force”; and the other case, in a 30-year old woman, was unclear.
Returning to the coding of cause of death, we conclude that keeping focus on readily identifiable obstetrical causes as the underlying cause is very likely the most accurate and meaningful strategy for characterizing maternal deaths, especially in high HIV/AIDS environments. This will permit more valid comparisons across different settings and eliminate the need for assumptions on the proportion of HIV/AIDS deaths during the maternal risk period that contribute to indirect causes. Other causes are better coded according to their ICD-10 category, whether infectious, non-communicable or external, and with their pregnancy status coded separately. Further investigation will then be needed to assess whether these non-obstetric deaths occurring during the maternal risk period are potentially associated with the pregnancy.
This paper raises questions on the definition and validity of the concept of “indirect cause” and argues that including causes other than obstetrical causes may inflate the maternal mortality ratio and bias comparisons over time and space. Our analysis suggests that limiting the current definition of maternal mortality to the more specific obstetric causes would provide a better focus on Safe Motherhood, defined by the risk associated with pregnancy.
Our perspective was focused on medical statistics and epidemiology. A clinician’s perspective will, of course, be different, as they treat any condition experienced by pregnant or delivering women, irrespective of alternate international classification systems of direct or indirect ‘disease or cause of death’.