Study design and setting
The study used quantitative baseline data from the AQCESS (Access to Quality Care through Extending and Strengthening Health Systems) and IMPACT (Improving Access to Reproductive, Maternal and Newborn Health in Mwanza) projects, both executed by the Aga Khan Foundation Canada with funding from the Government of Canada, being implemented in Kenya and Tanzania, respectively. The projects aim to contribute to the reduction of maternal and neonatal mortality in Kisii (Bomachoge Borabu sub-county) and Kilifi (Kaloleni/Rabai sub-counties) counties in Kenya and Mwanza region (Illemela, Nyamagana, Buchosa, Sengerema, Ukerewe, Misungwi, Kwimba and Magu districts) in Tanzania. Bomachoge Borabu sub-county is one of the nine sub-counties in Kisii County with a population of 129,617 people. In 2015, the county had 53.3% of health facilities deliveries and 76.6% deliveries assisted by skilled providers . Kaloleni and Rabai are coastal sub-counties in Kilifi County with a population of 304,778; 52.6% health facility deliveries and 52.3% skilled birth attendance . Mwanza region lies in the northern part of Tanzania and has a population of 2,772,509 people; 57.2% of the women attending at least four ANC visits and 75.3% hospital delivery . The two countries and the regions within the countries were chosen due to their high maternal mortality [14, 15].
Sample size and sampling technique
The sample size for the households was calculated to detect a 10% difference in skilled birth attendance between the projects’ baseline and end line. The sample size was calculated using the proportion of deliveries assisted by a skilled provider for each of the study areas (61.8% for Kenya and 63.7% for Tanzania), design effect of 2, level of significance of 95%, a margin of error of 5% and non-response rate of 10%. The total number of households required for the survey were 960 in Kisii, 1,100 in Kilifi, and 1,676 in Mwanza. Out of which 518, 664, and 1176 women were eligible and interviewed in Kisii, Kilifi, and Mwanza, respectively. A community-based multi-stage cluster design was used. A total of 30 villages each in Kenya and Tanzania were selected based on the number of households in the first stage followed by a random selection of households from lists of households within the villages in the second stage. At the households, all women of reproductive age and who consented were interviewed.
Data were collected in August 2016 and August 2017 for AQCESS and IMPACT projects, respectively using a pretested questionnaire with questions about BPCR adopted from the monitoring BPCR tools for maternal and newborn health was used for data collection . The English questionnaire was translated into Swahili, Ekegusii and Kigiriama; the common languages among the study participants. Trained data collectors entered data to the Open Data Kits platform which had electronic versions of the questionnaires with in-built data validation and quality checks. Data were stored onto a secure cloud server after a completeness check by a supervisor. All the selected households were included in the interviews; in case there was no eligible respondent available at the time of data collection, three revisits attempts were made before the households were declared unavailable.
BPCR, the main outcome variable, was assessed by asking women if a member of their family or herself prepared the following on the last birth: “1) discuss the place of delivery, 2) discuss who will perform the delivery, 3) set aside funds for the delivery, 4) arrange transport, and 5) identify a blood donor.” A woman was considered to be well-prepared for birth and its complications if she mentioned at least three out of five key components of BPCR [6, 9, 13], and less prepared if mentioned less than three and not prepared, if she mentioned none. Similarly, a woman was considered to have good knowledge about danger signs if she spontaneously mentioned at least three danger signs during pregnancy, labour and childbirth and postpartum [2, 31]; poor knowledge, if she mentioned less than three, and no knowledge, if she mentioned none. A list of all danger signs in each continuum of care is included in additional file 1.
Independent variables included maternal age, level of education (none, primary, secondary+), place of delivery (home, health facility), number of ANC visits (none, 1–3 visits, 4+ visits), and knowledge of danger signs during pregnancy, labour and childbirth, and postpartum.
Categorical data were described using frequencies and percentages and continuous data using median and interquartile ranges (IQR). A univariate model was fitted to examine associations between each variable and the ordered categories of BPCR. Variables with p-value <0.25 in the univariate model were fitted in the multivariable Proportional Odds regression model to determine their association with the dependent variable (ordered categories of BPCR) while controlling for the confounding effect of the explanatory variables .
Due to the ordinal nature of the outcome, Proportional Odds model  with a logit link function was used in both univariate and multivariable regression analysis to determine the association between the explanatory variables and the outcome. For the three categories of the outcome, the response is equivalent to two binary responses; (i) well-prepared versus less prepared or not prepared and (ii) well-prepared or less prepared versus not prepared. In this case, there is a cut-off point (threshold) at well-prepared the first logit and another at less prepared to form the second logit. The model can be defined in its simplest form as follows:
Where, aj are separate intercept parameters, j is the level of an ordered category with 3 levels, B' different sets of regression parameters for each logit and x are a set of explanatory variables. The model thresholds and coefficients are estimated simultaneously using maximum likelihood procedure. Each cumulative logit has intercept, which increases with the categories of the outcome. The model assumes the same effects of B for each of the two dependent variables [33, 34].
Crude and adjusted odds ratio with their 95% confidence intervals were calculated to determine the strength and presence of associations. We used “svy” set command in Stata to adjust for clustering effect due to the complex sampling design of the study at the village level. We test the proportionality of odds for the outcomes using likelihood ratio and Brant tests. All the analysis were done using Stata version 15 . The STROBE guidelines for cross-sectional studies informed the design and reporting of this study .