Literature search

Using a predefined protocol, we carried out a systematic search of the literature following guidelines for meta-analysis of observational studies in Epidemiology (MOOSE) [29] and preferred reporting items for systematic reviews and meta-analysis (PRISMA) (see Additional file 1). The search was conducted between January and August 2013. A physician with support of a librarian conducted the search. Medical and social science databases and journal libraries included, but were not limited to, PubMed, EBSCO Host, Web of Science, ScienceDirect, Wiley, Cochran library and Google Scholar. We searched for studies conducted in SSA, involving pregnant women and reporting on risk of maternal and perinatal mortality/death by place of birth or delivery. Reference lists of selected publications were assessed in order to identify other potential papers of interest.

The key words used were “Maternal mortality”, “Maternal deaths”, “Perinatal mortality”, “Perinatal deaths”, risk, “Place of birth/delivery”, study, Africa and “sub-Saharan Africa”. In the selected publications, the following definitions were used. Maternal deaths were defined as all direct and indirect obstetric deaths during pregnancy, delivery, and the first 42 days after birth. Perinatal deaths were defined as pregnancy losses occurring after seven completed months of gestation (stillbirths), or deaths within the first seven days of delivery of a live born child (early neonatal deaths) weighing 1000 grams or more. Place of delivery was either facility-based (defined as birth or delivery in a formal health facility whether or not attended by a skilled medical attendant) or home (defined as birth or delivery outside of a formal health facility whether or not attended by a skilled traditional birth attendant). Studies reporting on perinatal mortality were considered eligible if they reported on pregnancy outcomes from 7 complete months until 7 days after birth. Studies reporting on maternal mortality were eligible if they reported on maternal deaths as soon as pregnancy was identified until 42 days after birth. In addition the studies had to be written in English and published between 1990 and 2013.

Data quality and extraction

Each study was subjected to a quality review using a modified Newcastle-Ottawa scale for cohort studies [30]. Key quality items assessed included representativeness of the population, population characteristics such as gestation age at enrollment and duration of pregnancy follow up, information about study design (population based cohort vs. demographic and health surveillance), ascertainment of exposure and proportion of home deliveries, and use of standard definitions for main outcome measures and denominators used (e.g. births, live births). We also extracted information that may have affected the outcomes such as frequency of data collection, proportion of refusals, loss to follow up and sample sizes.

We noted general study information (e.g. year of study publication and authors) and the primary health outcomes for the study: maternal and perinatal mortality ratio by place of delivery. As the outcomes of interest are ratios, we extracted information on relevant numerators and denominators that would enable independent calculations of these ratios. In cases where an appropriate denominator was not provided (e.g. number of live births at home), but the numerator and corresponding appropriate ratio was given, we estimated the denominator by simple proportion. We precluded studies that aimed to ascertain risk of a particular exposure (e.g. HIV infection or severe anaemia in pregnancy) or an intervention (e.g. Prevention of mother to child transmission of HIV) on perinatal/maternal mortality, with insufficient (<25%) assignment of outcomes of interest to place of delivery and where reported data could not completely fill a 2 × 2 table. Where possible, we contacted some authors either to confirm information or provide extra details. Two individuals independently extracted the information and resolved disparities by consensus.

Statistical analysis

Study quality was scored in two categories: high if more than 60% of the quality items were reported and low otherwise. For both primary health outcomes of interest, we estimated the crude odds ratio (OR) by place of delivery in each study and then calculated the weighted average of the OR across the studies using meta analytic procedures [31]. The OR from each study can be combined using a variety of analytical methods, which are classified as: a) fixed effect models which weight studies according to the amount of information they contain; or b) random-effects models, which incorporate an estimate of between-study variation in the weighting [32, 33]. We assessed study heterogeneity using I² statistic, which measures the percentage of variation in OR attributable to heterogeneity between studies. The fixed effects model was used when I² was low < 50% [34] otherwise we used random-effects model to calculate individual study OR and corresponding 95% confidence intervals (95% CI) and to pool the results across the studies.

The potential effect of place of delivery on maternal and perinatal mortality was estimated by attributable risk percentage reduction, defined as the portion of the incidence of an outcome in the exposed that is due to the exposure: [(Ie-Iu)/Ie]*100 where Ie and Iu are incidences in the exposed and unexposed groups, respectively [35]. In our analysis, this represents the incidence of mortality in the exposed (home delivery group) that would be prevented if the exposure (home deliveries) were eliminated. We used a Poisson method to calculate mortality ratios and their 95% CIs as Poisson method approximates distribution of these ratios better [36]. STATA version 12.0 (Stata Corp, College Station, Texas) was used for analysis. No ethical review was required for the study.

Perinatal mortality

To estimate the protective effect of place of delivery, we first calculated the odds of perinatal mortality by place of delivery. Using a fixed effects model, Figure 2 shows that the pooled crude odds of perinatal mortality is significantly higher for home compared to facility delivery (OR 1.21, 95% CI: 1.02-1.46). As there is a high between-study heterogeneity I² = 73.7, we also estimated the pooled effect with a random effects model [51]. The estimate from the more conservative random effects model is exactly the same, but is no longer significant, (OR 1.21, 95% CI: 0.79-1.84). The results of the individual studies are also mixed. Two studies are in favor of home delivery [47, 52], one is neutral [46] and three are in favor of facility-based delivery [43–45].

Maternal mortality

Figure 3 shows that the estimated pooled crude odds for maternal mortality at facility is 2.29 (95% CI: 1.58-3.31) compared to home settings. There is no variation in OR attributable to heterogeneity, I-squared = 0%. The increased odds for maternal mortality at facility are consistently high across all the three individual maternal studies and stable across both fixed and random effect models. The overall weighted maternal mortality ratio (MMR) is 727 (95% CI: 570–913) deaths per 100, 000 live births. The MMR is 599 (95% CI: 424–823) and 945 (95% CI: 658–1315) per 100,000 live births for home and facility deliveries, respectively (Table 2).

Perinatal mortality

In this study, we attempted to estimate the risk of mortality for home deliveries compared to facility-based deliveries. The study illustrates the complexity of the relationship between risk of mortality and place of delivery. At an individual level, we found evidence for increased chance of perinatal losses following home compared to facility-based deliveries. However, opposite conclusions can statistically be reached, significant versus non significant results, depending on whether we make fixed model or random effect model assumptions. It has been argued that a finding that is statistically significant with the latter, but not with the former model, should be viewed with caution [51, 53].

We estimated the overall weighted perinatal mortality ratio in SSA to be 63 (95% CI: 54–73) deaths per 1000 births. This figure compares well with current WHO estimates of PMR (56 per 1000 births) in the region [18]. Our results imply that at best, an expected 14 perinatal deaths could be averted per 1000 births if women delivered at facility instead of home. This represents a 21% (95% CI:-6,40) reduction in perinatal mortality risk for the home delivery group. In 2012, 36.8 million births were estimated to have taken place in SSA of which 18.0 million (49%) were home deliveries [54]. In view of the high number of home deliveries in the SSA region, it can be argued that a 21% reduction in perinatal mortality risk might produce important public health improvements, if only a significant portion of the women who currently give birth at home could be motivated to deliver in health facilities.

Through health worker trainings and WHO policy [5], the current practice in the region is to encourage active or self referral of pregnant women, especially those with high risk factors (e.g. twin pregnancy, pregnancy induced hypertension) to deliver at health facilities. Moreover, some of the studies included in the analysis made special arrangements for facility referral of at risk pregnancies [43, 47]. This selection was likely to have made the facility and home delivery groups different with respect to perinatal risk factors, the latter being on average less at risk than the former. The results might therefore represent an underestimation of PMR in the home delivery group. The observation that perinatal mortality is on average high in the home delivery group, in spite of its relative low risk profile, should be of concern to policy makers, program implementers and health care providers in SSA.

Maternal mortality

Regarding maternal mortality, the results show that pregnant mothers delivering in facilities have a significantly higher risk of experiencing a maternal death than women delivering at home. Several reasons could account for this surprising result. First, in settings where access to facility-based delivery is low (<40%), those seeking care at facilities may be complicated cases with higher risk of mortality [55, 56]. In such settings, high maternal mortality at facilities would be expected due to high risk selection, as already reported in the literature [57]. Only one [36] out of the three studies reporting on maternal mortality had a low facility delivery rate. Precluding this study from the analysis does not significantly change the results. This suggests that low facility delivery rate alone does not appear to explain the relatively high risk of mortality observed for facility-based deliveries in this study. Second, and related to the risk selection, the study by Høj et al. [50] shows a progressive increase in risk of maternal mortality from health center to hospital compared to home: (OR 1.49,95% CI: 0.73-2.76) and (OR 2.72, 95% CI:1.64-4.38) respectively. This suggests existence of a differential in terms of the complexity of case mix by level of care. This can be interpreted as an indication that referral for facility-based delivery is actually working with secondary or tertiary referral facilities treating women with more complex conditions [36], although probably the timeliness of care is not optimal to make a difference [58]. Thirdly, facility-based delivery as a strategy to reduce maternal mortality does not simply entail delivery at a health facility, but also access to an enabling environment, such as availability of health workers with midwifery training, diagnostic tools, drug supplies and access to blood bank for effective care. High risk of maternal mortality at facility may therefore reflect lack of requisite capacity for facilities to offer quality care to high-risk women [59]. A large study in the region has shown that most maternal deaths occurring at facilities are among women who receive substandard care [36]. Moreover, not all facilities-based deliveries are attended by health workers with midwifery training due to unfilled vacancies and staff absenteeism [60], supporting the assertion that poor quality of care at health facilities is likely to explain a significant portion of the observed high risk of facility-based maternal deaths. Fourthly, although hard evidence is lacking, the possibility that mothers might be harmed at facilities, due for example to poor infection control or other human errors, cannot be ruled out [55].

Strengths and limitations

The strength of our paper accrues from the deliberate choice to include studies with a homogenous design, using standard definitions of outcomes (perinatal and maternal mortality ratios), and that largely collected data prospectively and at regular intervals to minimize recall bias and increase validity of the outcome measures. Typically, estimates of perinatal and maternal mortality in African countries are derived from complex statistical modeling techniques [48] or from nationally representative demographic and health surveys that use sisterhood approach methods [61]. However, poor quality data, recall bias and selective under reporting tend to affect such estimates [25].

Still, our study suffers from some limitations. Population-based cohort studies are costly and therefore only few have been conducted in the SSA region. Despite the high quality of individual eligible studies, the pooled analysis is based on a limited number of studies, thus the results should be interpreted with caution. We also noted subtle operational variations in the exposures. For example, home delivery in some settings meant delivery by non-trained traditional birth attendants (TBAs) [43] while in other settings it entailed delivery by trained TBAs [52]. Despite the uniform study design, such conceptual differences can lead to clinical variations [32] and heterogeneity.

Geographical, temporal, health systems and individual patient risk profiles can also affect maternal and perinatal outcomes [58] and confound interpretation of health outcomes by place of delivery. It is known for instance that rural areas tend to have worse perinatal and maternal health outcomes than urban areas [4]. Within the SSA region, important variations in the risk of mortality exist between sub-regions [62]. At an individual level, age, parity and existence of other co-morbidities such as hypertension can influence both maternal and perinatal outcomes [47, 50]. Ideally, these factors should be controlled before making statements about the relative effectiveness of one place of delivery compared to another [55]. Inadequate number of studies and particularly the lack of patient level information from the available studies precluded the possibility of such accurate analysis in our own review.