This study is the analysis of the stillbirth component of the systematic review of maternal mortality and morbidity undertaken by the UNDP/UNFPA/WHO/World Bank Special Programme of Research, Development and Research Training in Human Reproduction (HRP), Department of Reproductive Health and Research at the WHO. The objective of the systematic review was to obtain prevalence/incidence data on maternal mortality and a range of conditions including stillbirths. The detailed methodology which followed a pre-defined protocol has been described elsewhere .
Identification of the articles
The search for articles involved bibliographic databases (Medline, EMBASE, SocioFile, CAB Abstracts, Econlit, Cinahl, LILACS, Popline, BIOSIS, PAIS), WHO regional databases (African Index Medicus, Index Medicus for the Eastern Mediterranean Region), internet, reference lists, contacting experts in the field, and hand-searching of relevant documentation in the WHO Library. We developed specific search strategies for electronic databases according to their structured thesaurus terms or using appropriate keywords in collaboration with two librarians from the WHO and Cochrane Pregnancy and Childbirth Group. Detailed strategies for electronic databases have been previously reported and are available from the authors . The search was limited to articles dated from 1997 to 2002. The decision for this was arbitrary. There were no language restrictions.
Assessment for inclusion
Two reviewers evaluated titles and abstracts of the identified citations for potential inclusion in the review. Prior to this initial evaluation, we assessed inter-observer agreement using the kappa statistics (0.60 95% CI 0.52 to 0.69) which showed moderate to substantial agreement . We discussed and resolved points of disagreement. In case of doubt, we obtained full text articles of citations. We assessed full-texts of the articles deemed to be potentially relevant at the initial stage. Studies in all languages were eligible for inclusion if they reported data relevant to outcomes of interest, specified dates for data collection period, included data from 1990 onwards, and had sample sizes of greater than 200.
Data extraction and quality assessment
We developed and used a data extraction instrument including 48 items distributed in five modules three of which were relevant to this analysis. Modules were designed to collect information on (i) study level characteristics (sampling design, population, setting, completeness of data/response rate, reference period), (ii) outcome measures, and (iii) definitions and identification procedures for outcomes. We defined four key criteria for the quality assessment of the articles. These were: sampling schemes conducted as either random or consecutive, adequate description of population characteristics, definition of both the numerator and the denominator of the reported rate, and response rate/completeness of information in the data sets exceeding 75%. We considered the overall quality as adequate if a study fulfilled at least three of the four criteria. We did not exclude studies on the basis of inadequate quality, but accounted for this in the statistical analysis.
Selection of studies
Prior to the analysis, we developed a protocol that defined inclusion criteria and specified the approach to the analysis. Cross-sectional studies reporting stillbirth rates with representative sampling schemes were eligible for inclusion. For studies reporting information relevant to the same population for more than one year, we included data only from the most recent year. In order to prevent a woman's appearance more than once in a data set, and because the durations of studies extending beyond 12 months were highly variable, we limited analysis to studies with reference periods of 12 months. For studies where no definition for stillbirth was reported we assumed the conventional definition of more than 28 weeks of gestation . If a study reported results separately for different definitions, we used data referring to the conventional definition.
We calculated the pooled prevalence estimates for various subgroup categories weighted by the sample size of individual studies. A meta-regression was conducted to identify significant sources of heterogeneity .
The independent study-level variables included in the meta-regression were as follows: development status of the country where the study was conducted (developed versus less/least developed according to the United Nations classification system , definition of numerator of stillbirth rate (late stillbirths – more than 28 weeks gestation or more than 1000 g birth weight versus all stillbirths – other categories involving earlier gestational ages starting from more than 20 weeks or birth weight more than 500 g), definition of denominator of stillbirth rate (live births versus pregnancies/deliveries), overall quality of the study (adequate versus inadequate), scope of study (national versus sub-national), source (journal versus non-journal) and language of the article (English versus non-English).
For the purposes of statistical inference, the prevalence rates were transformed using the empirical logistic transformation  given by
where ai is the numerator of the prevalence rate, and ni is the denominator. This transformation is used to help normalize the distribution of the dependent variable in preparation for the subsequent regression analyzes. The estimated inverse variance was used as weight in these analyzes, where the variance is given as:
For studies using a multistage design, this variance was estimated as :
where deff is the estimated design effect  for neonatal mortality.
The SAS Procedure REG, was used to conduct the weighted least squares regression . The option BACKWARD was specified to allow selection of the subset of independent variables that best predict the dependent variable. This procedure first fits a model with all the candidate variables included, followed by the deletion of variables in a stepwise fashion. The level of significance for stepwise removal from the model was set at 0.10.