Effectiveness of conditional cash transfers (Afya credits incentive) to retain women in the continuum of care during pregnancy, birth and the postnatal period in Kenya: A cluster-randomised trial

Objectives Given high maternal and child mortality rates, we assessed the impact of conditional cash transfers (CCTs) to retain women in the continuum of care (antenatal care (ANC), delivery at facility, postnatal care (PNC) and child immunisation). Design We conducted an unblinded 1:1 cluster-randomised controlled trial. Setting 48 health facilities in Siaya County, Kenya were randomised. The trial ran from May 2017 to December 2019. Participants 2922 women were recruited to the control and 2522 to the intervention arm. Interventions An electronic system recorded attendance and triggered payments to the participant’s mobile for the intervention arm (US$4.5), and phone credit for the control arm (US$0.5). Eligibility criteria were resident in the catchment area and access to a mobile phone. Primary outcomes Primary outcomes were any ANC, delivery, any PNC between 4 and 12 months after delivery, childhood immunisation and referral attendance to other facilities for ANC or PNC. Given problems with the electronic system, primary outcomes were obtained from maternal clinic books if participants brought them to data extraction meetings (1257 (50%) of intervention and 1053 (36%) control arm participants). Attendance at referrals to other facilities is not reported because of limited data. Results We found a significantly higher proportion of appointments attended for ANC (67% vs 60%, adjusted OR (aOR) 1.90; 95% CI 1.36 to 2.66) and child immunisation (88% vs 85%; aOR 1.74; 95% CI 1.10 to 2.77) in intervention than control arm. No intervention effect was seen considering delivery at the facility (90% vs 92%; aOR 0.58; 95% CI 0.25 to 1.33) and any PNC attendance (82% vs 81%; aOR 1.25; 95% CI 0.74 to 2.10) separately. The pooled OR across all attendance types was 1.64 (1.28 to 2.10). Conclusions Demand-side financing incentives, such as CCTs, can improve attendance for appointments. However, attention needs to be paid to the technology, the barriers that remain for delivery at facility and PNC visits and encouraging women to attend ANC visits within the recommended WHO timeframe. Trial registration NCT03021070. We conducted a cluster randomised controlled trial, with equal allocation to intervention and control arms, in Siaya county, Kenya. The units of randomisation were level 2 or 3 health facilities (Dispensaries and Health Centres, respectively). The randomisation of facilities was stratified by the six subcounties and ensured equal allocation to study arms within each stratum without any overlap of catchment areas, as described in detail in the trial protocol.25 In summary, at a public forum with the county government early 2016, the implementing partner wrote the names of 60 shortlisted facilities on pieces of paper and folded them to hide the names, then included them in transparent boxes, one for each subcounty. Each subcounty had an (even) number of facilities to recruit to the trial proportional to subcounty size. The health management teams from each subcounty selected the pieces of paper, one by one. The first was allocated to intervention, second to control. For each selected facility, county officials from the selected subcounty mapped the location and catchment area of the facility on a large map of the county. If a subsequently selected facility had an overlapping catchment area with a previously selected facility, the newly selected facility was rejected, and another drawn to take its place. This process continued until 48 facilities were selected and allocated for the trial. Health facility staff determined whether a pregnant woman met the study eligibility criteria by administering screening questions at the end of her first ANC visit, with the screening questions provided in the trial protocol.25 All women meeting the criteria were eligible for recruitment during the study recruitment period. Criteria for enrolment were women attending their first ANC visit; long-term resident of the catchment area served by the health facility (living in the area for at least 6 months); access to a mobile phone that belongs either to themselves or to a member of their household or person whom they trust. The criterium on residence provided additional assurance that women went to the facility within their catchment area, thereby reducing contamination with other facilities. Oral informed consent was asked in the local language, and then written down on the participant’s enrolment form. Refusals were recorded. The intervention was a CCT payment for each facility appointment attended for ANC, delivery, PNC and childhood immunisation. Detailed definitions can be found in online supplemental appendix S1. For each scheduled health visit made following enrolment, women in the intervention arm received a cash transfer of KSH 450 (US$4.5) on their mobile phones. Women at the control clinics were granted KSH 50 (US$0.5) mobile phone airtime for each scheduled visit to encourage them to bring their clinic booklet to appointments. In both trial arms, women were issued with a trial card at recruitment and at all facilities there was a card reader, which provided the connection between the trial card and an online portal which stored participants’ data on visits and payments. Payments to the women were triggered by tapping the card on a card reader, which also logged the visit in an online portal.26 In the event of problems with the card reader, or if the woman did not bring her card to the appointment, payments could alternatively be processed manually by contacting the implementing partner: once the visit was verified with the facility, the implementing partner entered the visit data in the portal, which would then trigger a payment as well nurses were given KSH 400 (US$4) per woman enrolled during the trial, and an additional KSH 100 (US$1) per woman enrolled at the end of the trial for their collaboration in the trial. These payments were transferred to the nurses electronically. Details of the intervention design are presented in the protocol.25 bmjopen-2021-055921supp001.pdf Patients or the public were not involved in the design, conduct, reporting or dissemination plans of the research. The assessment of the burden of the intervention on patients is reported in Dickin et al.27 The primary outcomes were: (1) attendance or missed attendance at each eligible ANC appointment after recruitment; (2) delivery at a health facility; (3) attendance for at least one PNC appointment between 4 and 12 months after delivery; (4) attendance or missed attendance at each expected child immunisation appointment; (5) attendance at referrals to other facilities for ANC, PNC or child immunisation. We define ‘eligible visits’ for the purpose of statistical analysis of the impact of the intervention within a clearly defined framework, but all scheduled prenatal and postnatal clinic visits should have triggered a payment. The PNC attendance outcome described in the protocol was the ‘proportion of required postnatal visits honoured after recruitment into the study’. However, we have used a simplified outcome here because the required appointment schedule was not recorded for each patient. We also restricted attendances to the period 4–12 months after delivery because on blinded review of the available data, prior to writing the statistical analysis plan (SAP), very few visits prior to 4 months postdelivery were coded as PNC. The vast majority of visits prior to 4 months were recorded as vaccination appointments. though it is likely some women also received PNC. Vaccinations over the 12 months after delivery were recorded. The details of vaccinations given were not collected, and so the vaccination outcome is based purely on the number of recorded visits from an expected number of four. The following secondary outcomes are also reported and analysed according to the trial arm: (1) attendance of all eligible maternal, newborn and child healthcare visits, both prenatal and postnatal, for each woman; (2) the count of attended ANC and child immunisation clinic visits eligible for the primary outcome variables for each woman; (3) the total number of ANC, child immunisation and PNC clinic visits (without applying any eligibility criteria) for each woman; (4) gestational age (GA) at first ANC visit (and enrolment to study). The secondary outcomes of clinic visit counts and GA at first ANC visit were not listed in the trial protocol but were prespecified in the SAP to aid interpretation of the study results. The primary outcome of attendance at referrals to other facilities for ANC, PNC or child immunisation is not reported because of very limited data. The following planned secondary outcomes are reported without formal statistical analysis because of low levels of data completeness: maternal and neonatal mortality, self-rated wellness, exclusive breast feeding and contraceptive use. The following secondary outcomes were dropped because of lack of available data: timeliness of health visits (recorded visits could not consistently be matched up to scheduled dates), and infection screening. These changes were specified in the SAP. Online supplemental appendix S1 contains further details of changes made to the data collection and analysis compared with the protocol, arising from data collection challenges. Data were collected throughout the trial from an electronic card reading system, and baseline and follow-up surveys. The electronic card reading system captured the enrolled women’s phone number, expected delivery date, parity, the clinic she enrolled at, the visits she attended and the payments she received. The baseline survey, carried out by telephone following enrolment, collected sociodemographic data. We initially planned to conduct follow-up interviews with 50% of enrolled women at 6 and 12 months postdelivery to collect secondary outcomes. However, due to limited resources and lower than anticipated response rates after some 6-month interviews had been conducted, we adopted a pragmatic approach and conducted one follow-up survey at around 12 months after delivery. Problems with the implementation of the technical system and periods of disengagement from clinic staff resulted in a large proportion of visits not being registered on the system. Therefore, data on women’s visits were manually extracted from clinic health records after the trial was completed. Trial staff visited each facility, after arranging for trial participants to be invited to attend the facility with their maternal clinic book. Data on the primary outcomes of ANC, child immunisation and PNC visits and delivery at a healthcare facility were extracted from maternal clinic books. The books very rarely contained records of any referral visits. For women who did not attend the data extraction, data on the primary outcome of ANC visits were extracted from health facility registers, but data on other visits were not available from this source. Any missing payments were transferred to the participating women following the manual data extraction at the end of the trial. Data on payments made to trial participants, whether triggered by the card reader or manually, were extracted from the trial portal. These data are reported in the process evaluation paper, along with details of the challenges with the electronic system.27 We analysed all primary outcomes jointly, to maximise power, aid interpretation and minimise testing (see the Statistical analysis section). However, for our sample size calculation, we considered the power to detect an effect of the intervention on one primary outcome, which we also assumed to be a binary indicator that all attendances were made, as this is simple and conservative in the power achieved. The expected prevalence of these indicators in the control arm ranged between 30% and 80%. In the absence of specific information on the likely intra-cluster correlation (ICC) we considered a range between 0.005 (low) and 0.025 (moderate). Our planned sample size was 48 clusters covering the catchment areas of selected level 2 and 3 health facilities (24 per arm) and an average cluster size of 150 participants. At a low ICC, the design effect (DE) would be 1.745 and hence the effective sample size (ESS) would be 2063 participants per study arm. At a moderate ICC, the DE would be 4.725 and hence the ESS 762 per arm. Power to detect absolute differences is lowest when the prevalence is 50% and highest when the prevalence is either high (towards 100%) or low (towards 0%). Here, we considered the prevalence in the control arm to range between 50% (‘worst-case scenario’) and 80% (‘best-case scenario’). We considered the standard 5% significance level. If the prevalence of the outcome is 50% in the control arm, the sample size provides 80% power to detect an improvement to 54.5% in the intervention arm if the ICC is low and 57.5% if the ICC is moderate. If the prevalence of the outcome is 80% in the control arm, the sample size provides 80% power to detect an improvement to 83.5% if the ICC is low and 85.5% if the ICC is moderate. For the purpose of statistical analysis, eligible ANC visits were defined based on the recorded ‘next scheduled visit date’ and noting that this was typically 4 weeks later each time. For each recorded visit starting with enrolment, we evaluated whether the next observed ANC visit was within ±2 weeks of the ‘next scheduled visit date’: recording a successful ANC visit if yes, but a missed visit if not. However, we did not count the visit as either successful or missed if the next observed visit was more than 2 weeks early compared with the next scheduled visit or delivery occurred within 2 weeks of a ‘missed’ scheduled visit. Whether that next visit is early, on time, or late, we assessed subsequent visit attendances in the same way. We created hypothetical scheduled visits every 4 weeks for any gaps in observations, judged according to the same criteria. The ‘successful’ and ‘missed’ appointments were then summed over all scheduled and hypothetical appointments for each woman. For the primary outcomes of PNC visits, we defined a binary indicator of one or more PNC visit 4–12 months post partum. This was used to provide a simple indicator of engagement with PNC for each woman, given that the appropriate number of PNC visits may differ between women. We defined child immunisation visits as the total number of visits recorded post partum (excluding vaccination at delivery, but without other time restrictions) truncated at a maximum of four, since that is the typical number required for full immunisation. Retention in the full continuum of ANC, perinatal and PNC (a secondary outcome) was defined as a binary indicator of attendance of all eligible ANC, child immunisation and PNC visits and delivery at a healthcare facility for each woman and is available for those women who brought their clinic book for data extraction. The primary outcomes of attendance of eligible ANC and child immunisation appointments comprise repeat binary observations for each individual woman, while the primary outcomes of delivery at a health facility and PNC attendance between 4 and 12 months are each single binary variables for each woman. A summary OR is presented as the main effect measure for the trial, estimated from a model that assumes the OR is the same across the four primary outcomes. We also report separate effect estimates for each primary outcome. A mixed effects logistic regression model was used to jointly analyse the observed primary outcomes. This approach assumes that, given their ANC attendances (recorded for nearly all women), whether a woman did or didn’t bring their clinic book for extraction of the other outcomes was unrelated to the values of those outcomes. At the level of each woman correlated random effects were specified for (1) attendance of ANC clinic visits and (2) grouped outcomes of delivery at a healthcare facility, attendance of vaccination visits and attendance of at least one PNC clinic visit at 4–12 months post partum. Clinic-level random effect terms were defined for (1) ANC visits, (2) delivery at a healthcare facility and (3) PNC and child immunisation visits, with unrestricted correlations between these. The secondary outcomes of counts of ANC clinic, PNC clinic and child immunisation visits were analysed using a multivariate Poisson mixed effects model, with random intercept terms at patient and the clinic levels. Marginal mean differences in counts between intervention groups were estimated. The secondary outcome of GA at enrolment to the trial was analysed using a linear mixed effects model, with random intercept term at clinic level. Retention in the full continuum of care was analysed using a logistic regression mixed effects model, with random intercept term at clinic level. As the completion rate of the follow-up survey was lower than expected, the secondary outcome data obtained is reported in a descriptive summary but not compared between trial arms. Our main analyses are conducted as randomised (ie, intention to treat) but a ‘per-protocol’ style sensitivity analysis of the primary outcomes was also prespecified. As there was not a clear division between clinics that did and did not achieve the intended payment schedule, the sensitivity analysis included the 12 intervention clinics and the 12 control clinics with the highest proportion of enrolled women with payment within 31 days of first ANC clinic visit. All analyses were adjusted by the baseline maternal parity (‘0’ vs ‘≥1’), by the presence of any maternal medical conditions leading to classification of the pregnancy as high-risk (HIV with or without antiretroviral therapy (ART), diabetes, hypertension, malaria, each coded with separate indicator variables) and by the clinic-level variable of subcounty. Adjusted effect measures are considered the primary effect measures though unadjusted effect estimates are also reported. The analysis followed a prespecified SAP, which was finalised after data collection but prior to any unblinded analysis. Selection of maternal characteristics as adjustment variables was based on their inclusion in the core enrolment dataset, the associated absence of missing data for these items and their potential to predict the outcomes of interest. Subcounty was included as an adjustment variable because of its use as a stratification factor in the randomisation process for the study. As recommended by the MRC framework for the evaluation of complex interventions,28 a detailed process and economic evaluation will be published in two forthcoming papers.27 29