Trial Design

We conducted a matched-pair, cluster-randomized, controlled trial in government health facilities that received either the BetterBirth program, a coaching-based implementation of the WHO Safe Childbirth Checklist (60 facilities), or the existing standard of care (60 facilities). We have described the methods of the BetterBirth Trial,19 the BetterBirth intervention,20,21 and our data quality-assurance system22 elsewhere. The trial protocol and statistical analysis plan are available with the full text of this article at NEJM.org. The trial sponsor (the Bill and Melinda Gates Foundation) reviewed the trial design and sample-size calculations but was not involved in data collection, management, analysis, or interpretation; the writing of the manuscript; or the decision to submit the manuscript for publication.

Trial Setting and Participants

The most populous state in India (population of 204 million, 77% rural),23 Uttar Pradesh is a high-priority region for national and international public health organizations owing to its persistently high neonatal mortality (32 per 1000 live births) and maternal mortality (258 per 100,000 births).24,25 The government of Uttar Pradesh permitted the trial to proceed in 38 districts, in which we identified 320 eligible facilities. We considered a facility to be eligible if it was designated as a primary health center, community health center, or first referral unit; had at least 1000 deliveries annually; had at least three birth attendants with training of at least the level of an auxiliary nurse midwife; had no other concurrent quality-improvement or research programs; and had district and facility leadership willing to participate. The final trial sample included 120 facilities across 24 districts (see the Supplementary Appendix, available at NEJM.org).

We matched facilities (i.e., clusters) on the basis of the following criteria before randomization: geographic zone, functional classification (primary health center, community health center, or first referral unit), distance to a district hospital, annual birth volume, and number of birth attendants. We randomly assigned facilities to trial groups within each matched pair. After matching and randomization, we collected baseline data on practice adherence in 10 sites to confirm successful matching.

Women who were registered for labor and delivery — excluding women who delivered outside the facility, women who were referred from another facility, or women who were admitted for abortion — were eligible for the trial. At each intervention facility and its matched control site, we began enrolling patients 2 months after the initiation of the intervention. Enrollment continued until the target sample size of the site was reached or until 24 weeks after the completion of the intervention, whichever occurred first, with a 12-week minimum follow-up.

Intervention

We implemented the BetterBirth program in accordance with the Engage–Launch–Support model (Fig. S1 in the Supplementary Appendix) that was piloted at nontrial sites in Karnataka and Uttar Pradesh, India.18,21,26 Coaches (nurses) and coach team leaders (physicians or public health professionals), all of whom were unaffiliated with the facilities and comprehensively trained to apply a standard behavior-change framework, conducted site visits during the 8-month Support phase.20,21 We expected coaches to conduct 43 daylong visits to each facility, beginning twice weekly and tapering to monthly visits. Coach team leaders accompanied coaches on alternating visits (23 total visits). Each facility chose at least one staff member to serve as a childbirth quality coordinator, a local champion for use of the checklist and continued coaching.

Coaches motivated birth attendants to use the checklist and to identify, understand, and resolve barriers to providing quality care.20,21 Coach team leaders supported facility leadership in fostering team communication and addressing gaps in care at facility and district levels by accessing resources through the established health care system. Cloud-based data collection enabled rapid feedback on the progress of a facility. We provided no clinical-skills training, financial support, or clinical supplies (except paper copies of the checklist).

Data Collection and Outcomes

The primary outcome was a composite outcome of events occurring within the first 7 days after delivery, incorporating stillbirth; early neonatal death; maternal death; or self-reported maternal severe complications, including seizures, loss of consciousness for more than 1 hour, fever with foul-smelling vaginal discharge, hemorrhage, or stroke. We selected measures of complications on the basis of definitions from the WHO guidance on maternal severe complications, using questions that had previously been validated for the reporting of complications by patients.27-30 We calculated a prespecified, additional composite outcome consisting of maternal and perinatal death only.

Secondary maternal outcomes by 7 days after delivery included maternal death, maternal complications, interfacility transfer (referral), cesarean section, hysterectomy, blood transfusion, and return to the facility for a health problem. Secondary newborn outcomes included stillbirth, early neonatal death, and interfacility transfer. We assessed all outcomes from facility register information and by contacting women or close family members by telephone between 8 and 22 days after delivery. If we received no response by 22 days after delivery, a field worker conducted a home visit and attempted to follow up until 42 days after delivery.

In addition, we selected a convenience sample of 15 matched pairs of facilities in which trained nurse–data collectors directly observed birth attendants providing care during a 12-hour (daytime) shift at 2 months after the initiation of the intervention (during the highest intensity of coaching) and 12 months after the initiation of the intervention (4 months after the cessation of coaching). These independent observers measured practice adherence, including supply availability (Table S3 in the Supplementary Appendix). Intervention staff and independent observers were not present at the same facility simultaneously.

Owing to the nature of the intervention, we were unable to prevent any facility staff, most trial staff, or any investigators from being aware of the identity of intervention and control facilities. Call-center staff, who collected the majority of outcome data, were unaware of facility assignments.

Ethics Compliance

The leadership at each facility provided facility-level consent for participation and permission for trial staff to collect deidentified data on every eligible woman from facility registers. Before patient discharge, we obtained verbal informed consent and contact information from each woman or her surrogate for follow-up. Data collectors reconfirmed verbal consent at the start of the follow-up call or visit. In directly observed births, women or their surrogates provided written consent for observation.

At trial initiation, birth attendants and facility staff verbally agreed to participate. Before an independent observer collected data, the birth attendant verbally reconfirmed agreement. Electronic data were deidentified and stored in a Health Insurance Portability and Accountability Act–compliant database to ensure participant privacy.

The trial protocol was approved by ethics review boards at Community Empowerment Lab, Jawaharlal Nehru Medical College, the Harvard T.H. Chan School of Public Health, Population Services International, the WHO, and the Indian Council of Medical Research. The protocol was reviewed and reapproved on an annual basis. A data and safety monitoring board met every 6 months after the initiation of enrollment; one interim analysis was conducted when 30% of the data were collected (see the Supplementary Appendix).

Sample Size

We hypothesized a priori that the rate of the primary composite outcome would be 15% lower in the intervention group than in the control group. We estimated the intracluster (within-facility) correlation to be 0.01 and the matching effect to reduce the standard error by 45%, basing these parameters on previous studies.31 We aimed to enroll 171,964 women (85,982 per group) to detect a 15% relative difference (60 events per 1000 births in the control group vs. 51 events per 1000 births in the intervention group) with 80% power and an alpha level of 0.05. On the basis of limited data available from Uttar Pradesh, the baseline rate of the primary composite outcome may have been as high as 120 events per 1000 births. The baseline rate that was used in calculations was purposively set lower than the estimated rate owing to limited information as well as the inclusion of home-based birth events in the available data, which may have elevated mortality.

In assessing practice adherence, we assumed an intracluster (within-facility) correlation of 0.01 and a design effect of matching of 0.75. With 15 matched pairs, we had more than 80% power at an alpha level of 0.05 to detect an absolute difference of 8.5 percentage points in the rate of any birth practice between the trial groups.

Statistical Analysis

Using an intention-to-treat approach, we compared outcomes between trial groups using a Rao–Scott chi-square test, accounting for the matched-pair, cluster design.32 The main outcome was the dichotomous composite outcome that was present if any of the three main outcomes occurred (maternal death, stillbirth or early neonatal death, or maternal severe complications). This variable was then used to estimate a composite relative risk.33,34 An additional secondary composite outcome included maternal or perinatal death only. In secondary analyses, each of the main outcomes was compared across groups; a Rao–Scott test with 3 degrees of freedom was used to assess the overall causal effect. No adjustment for multiplicity of testing was made.

For the subgroup of directly observed births, we calculated adherence frequencies for each prespecified measured practice at 2 months and 12 months after the initiation of the intervention. Furthermore, we calculated an aggregate adherence score by summing the total number of 18 checklist practices that should be performed regardless of the health status of the mother–newborn dyad (see the Supplementary Appendix). We generated the mean number of practices (presented as a fraction of 18) performed in each trial group and compared the differences at each time point, using a Rao–Scott chi-square test.32 For comparison of individual practices, we used a bias-corrected logistic-regression approach that can handle zero cells and complete separation of points within strata and clusters.35 We conducted all statistical analyses using SAS software, version 9.4 (SAS Institute).