Effects of the EQUIP quasi-experimental study testing a collaborative quality improvement approach for maternal and newborn health care in Tanzania and Uganda

We used a quasi-experimental, plausibility design to evaluate EQUIP [29] (Fig. 1) and compared two purposefully selected districts in each of Southern Tanzania and Eastern Uganda; a detailed description of the intervention and study design is presented elsewhere [23, 30]. Briefly, the two pairs of districts were selected from areas where the research team had established research collaborations and (1) the districts were rural, so results would be relevant to other rural districts in the two countries and (2) the districts were of comparable size with similar health infrastructure. However, in Uganda, the government split our comparison district 3 months after our decision was made in December 2010. Of the two new districts, we selected the newly created district Namayingo, as this was the most similar to the intervention district in terms of geographical features (both border Lake Victoria). This meant that the comparison district in Uganda lacked a hospital, had a smaller population than the intervention district and had a less developed health infrastructure (Table 1).

The intervention districts of Tandahimba and Mayuge mainland and the two comparison districts of Newala and Namayingo, in Tanzania and Uganda, respectively, were predominantly rural. In Tandahimba, Tanzania, 32 public and faith-based health facilities offered maternal and newborn care for a population of 227,514. In Mayuge, Uganda, 33 public and faith-based facilities served a population of 412,500 (a ratio of 1.4 and 0.8 facilities, respectively, per 10,000 population). We did not include private for-profit facilities, as the few facilities operating in the study areas did not provide maternal and childbirth services (Additional file 1: webannex 1 maps).

In each country, we compared time trends of coverage and health care quality indicators and mothers’ knowledge of danger signs between intervention and comparison districts using independent continuous facility and household surveys. The plausibility design also included the regular assessment of contextual factors likely to affect maternal and newborn health other than the study intervention, as recommended by Victora et al. [31]. Community members and health staff were not masked to the intervention. The survey team worked independently and was trained to be as neutral as possible but was not masked to the intervention area. However, the survey team was unaware of the primary and secondary outcomes that were chosen for evaluation.

We based the QI approach on the collaborative model of improvement [11], which is a short-term, rapid-learning method to improve quality in a focused area based on PDSA cycles and clearly defined and agreed upon indicators for monitoring [32]. We described our methodology in more details in our protocol paper and in the annexes of our protocol paper [23, 33].

In intervention districts, we implemented two strategy components (1) collaborative QI with (a) district health managers, (b) health facility staff and (c) community health workers and (2) continuous household and health facility surveys, with results communicated to district health managers, health facilities and communities using report cards once every 4 months [26]. In comparison districts, we implemented continuous household and health facility surveys for evaluation, with results communicated to district health managers using a written report once per year [30].

For the QI strategy, the main health facility participants were health staff working in the area of maternal and newborn health. Community participants were volunteers, either selected by the community, often on grounds of prior experience as community volunteers (Tanzania) or chosen from active village health teams (in Uganda). The district QI team was composed of the council health management team (Tanzania) and the district health team (Uganda).

Every 3 to 4 months, we invited both health facility and community members to participate in learning sessions to review progress. Learning sessions were either joint or separate depending on the chosen improvement topic. These sessions, which typically lasted 1 day, introduced and reminded participants about QI techniques, including the PDSA cycles, and new topics for improvement and also provided a platform to review progress and allow teams to learn from each other. Separate learning sessions were held with district managers (Additional file 1: webannex 2 table). During action periods, which were times between learning sessions when the teams implemented the improvement work, the QI teams were mentored regularly by EQUIP project staff and district managers. In Tanzania, health facility and community QI teams were mentored on average two to three times each quarter, half of the time in the form of “cluster meetings”, where health facility and community QI teams met together locally. Similarly, in Uganda, two to three coaching and mentoring visits for health facility and community QI teams were done per quarter. The learning sessions were supported with feedback from the survey results presented in the form of report cards covering selected indicators (Additional file 1: webannex 5 report cards). The EQUIP team met with district health managers 11 and 12 times over the 30 months of EQUIP implementation in Tanzania and Uganda, respectively.

A piloting period took place in both intervention districts from June to October 2011. In Tanzania, full implementation, after the pilot and roll-out periods between June 2011 and February 2012, was achieved in 31 primary health facilities, one hospital and 157 villages (including each roughly 250–500 households each) from March 2012 to April 2014. In Uganda, gradual implementation was done from November 2011 to February 2013, and QI teams were active in 31 health facilities and 72 parishes (each comprising almost 1000 households) from February 2013 to April 2014 (Additional file 1: webannex 3 timeline of assessment and implementation).

QI teams selected improvement topics from a QI charter, a planning tool where key areas where improvement is needed are outlined (Additional file 1: webannex 4, improvement charter) which used the WHO guideline of recommended essential interventions, commodities and guidelines as a basis [3]. During the course of the project, prioritisation of which essential intervention to address first were frequently chosen based on discussions between project staff and district mentors and taking into account continuous survey results, which were summarised in report cards. QI teams implemented and tested various change ideas, such as new strategies for birth preparation counselling (e.g. going through birth preparation checklists) or changing implementation routines to address defined problems (e.g. having a delivery tray including oxytocin prepared at all times, see more details Additional file 1: webannex 6 vignettes). The community QI teams implemented a variety of change ideas (Table 2), including home visits to pregnant women, community discussions and the establishment of community savings funds. The improvement topics ‘facility delivery’, ‘uterotonics within one minute’ and ‘post-partum care’ were introduced early on in both countries. In Tanzania, the 1-day training, ‘Helping Babies Breathe’ was offered in March 2013 during one learning session for all intervention facilities preceding the national roll-out in the region. Health facility QI teams used job aids, timely and improved ordering of drugs and supplies, sensitization to and improved counselling of clients and better communication with district managers to improve implementation. In Tanzania, QI teams also started to access funds collected as part of the national cost-sharing strategy and from community health funds, which had been accumulating funds without being used. District QI teams worked on improving (1) the human resource situation, such as requesting new staff or staff to be transferred from the central level, (2) the drug supply from the Medical Stores Department to facilities and (3) the supervision of district managers of primary care facilities and communities to improve quality of care. The change ideas were implemented and tested during a 1–2-month period using locally generated data (Additional file 1: webannex 7 run charts) and widely implemented if internal monitoring suggested improvements.

In intervention districts, two external medical and social science experts were employed to facilitate learning sessions and mentoring and coaching. Additionally, in each district, three staff members from the health and community sectors, and a varying number of government-salaried staff working at the lower level of the district, were involved in mentoring and coaching and were remunerated through daily allowances as per government guidelines. In Tanzania, these mentors supported one district and 32 health facility QI teams, as well as about 300 village volunteers organised in 10 cluster QI teams. In Uganda, the mentors supported one district, 30 health facility and 61 parish QI teams (see annex II of [23]).

Our primary coverage outcomes were (1) the percentage of women delivering in a health facility and (2) breastfeeding within 1 h after delivery, as assessed through the continuous household survey using reports from women of reproductive age with a live birth in the 12 months before the survey (Table 3).

The primary quality outcome indicator was the proportion of births in which an uterotonic drug was administered within 1 min after delivery. This indicator was constructed by multiplying women’s reports of the place of birth assessed through household surveys with reported health workers practicing on the usage of uterotonics in surveyed facilities. The health worker’s report was based on a last event module that asked staff providers about their practices in a narrative and non-threatening way [30]. As this measurement mode is not validated in low-resource settings, we did observations of delivery practices in selected facilities to validate health worker reports of implementation practices [34].

The primary knowledge indicator was the proportion of women who knew danger signs both in pregnancy and for newborns, as measured amongst mothers who gave birth in the year before the continuous household survey. We used open-ended questions to assess knowledge of danger signs, which was defined as recalling all three critical signs of severe vaginal bleeding, oedema of the face/hands and blurred vision/headache in pregnancy and all four critical signs in newborns (convulsions, difficult breathing, lethargy/unconsciousness and very small baby) [23]. The secondary outcomes were seven indicators that were constructed to reflect the improvement work such as post-partum care and clean birth kits (Tables 2 and 3).

Assuming a design effect of 1.4 and 10% refusals, the size of the survey was calculated to provide at least 80% power to detect small absolute increases (fewer than 10 percentage points) between the beginning and end of the intervention period for outcomes across the continuum of care, 10-percentage-point increases each year of the intervention for most outcomes (including institutional delivery and immediate breastfeeding) and larger increases more frequently. See Marchant et al. [30] for a detailed discussion.

We implemented continuous household cluster and health facility surveys in the intervention and comparison districts. The details are presented elsewhere [30]. Briefly, questionnaires were developed based on well-established sequences of questions as used in Demographic and Health Surveys and Service Provision Assessments [35] and also drew upon earlier work in the study areas [5, 36]. Data collection was organised in 5-month cycles, comprising a 4-month ‘round’ of field work and a 1-month break for aggregated analysis and planning for the next cycle.

Based on a sampling frame of the lists of sub-villages with the total number of households (Tanzania) and parish-level lists (Uganda), we sampled, each month, and in each district, 10 clusters (comprised of 300 households) using probability proportional-to-population-size sampling. We thus included no repeated sample of the same women over time but 24 independent probability samples of household clusters to represent each district each month while. We systematically selected each cluster of 30 households from a household list using a fixed fraction (total number of households in the sub-village divided by 30) [30]. After completion of 4 months of data collection, data were aggregated (1200 households and an estimated 152 women with a recent birth, per district) for analysis, and report cards were prepared (Additional file 1: webannex 5 report card). Each 4-month round also included a census of all health facilities to assess readiness and also included a last event module whereby the birth attendant for the last birth recorded in the health facility register was identified and interviewed about the care they had given during that birth.

The contextual factors were assessed using pre-defined indicators based on the health system building block framework to inform the plausibility analysis. The selection of indicators was informed by the work of Victora et al. [31]. Systematic investigation into concurrent context changes is needed to draw any conclusion when randomisation is not feasible [29, 31]. The detailed methodology of our context analysis is described elsewhere [23, 33]. In brief, we included indicators of financing, leadership, human resources, drugs and supplies, health information and service delivery [37]. Data were collected using (1) district reports, (2) interviews every fourth month with the district health managers and (3) the continuous household and health facility surveys. A structured questionnaire addressed to the district medical officer and his or her team collected information on the district planning process including participation of civil society, implementation of district plans and other aspects of leadership and governance [23]. Qualitative information on leadership and governance were triangulated with information available in district reports. Analysis used a qualitative and explanatory approach.

We used an approach adapted from the analytic methods to model data points over time (see Additional file 1: webannex 8 statistical methods). For each of the six time points of the continuous survey rounds, we calculated the difference in the indicator estimate between the intervention and comparison district [38]. We used meta-regression to fit a regression line through the resulting six data points over the 30 months of data collection and used this regression to estimate the difference-of-difference value between intervention and comparison districts from the baseline (first data collection round) to the end line (last data collection round). Estimates were adjusted for the sampling method (using svy commands in STATA). The delta method was used to estimate the variance of the difference-of-differences measure and present confidence intervals [39]. Rather than using a time-series approach based on autoregressive integrated moving average (ARIMA) models, we chose a simpler and more transparent analytical method that we thought was more appropriate for our small number of data points over time. We did, however, perform a sensitivity analysis with ARIMA models to investigate correlation over time and found no significant impact on the results for our primary outcomes. We did not adjust the models for potential confounding factors. We used information on potential confounders qualitatively [29]. Descriptive tabulation was done for the outcomes selected to present the context. Analysis was done using STATA 13 (StataCorp, Texas, USA).

Ethical clearance was obtained from the institutional review boards of Ifakara Health Institute, the Tanzania Commission for Science and Technology, the Uganda National Council of Science and Technology, Makerere University School of Public Health and the London School of Hygiene and Tropical Medicine (LSHTM). This activity underwent human subjects review process at CDC and was approved as not being engaged in human subjects research. Advocacy and sensitization meetings with district and sub-district authorities were held at the start of the project. Communities and health facilities were informed about the survey by a survey team member 1 day prior to the interview, using information sheets in the local languages. Written informed consent to participate in the surveys was obtained from household heads, women, facility in-charge and health staff interviewed.

Of the 14,400 sampled households in each country over the six survey rounds, 14,255 (99%) and 13,125 (93%) households in Tanzania and Uganda, respectively, consented to be included in the survey (Fig. 1). The household interviews listed 13,239 and 14,718 eligible resident women of reproductive age. In Tanzania, 11,835 (89%) women were interviewed and 1415 had a live birth in the 12 months before the survey. In Uganda, 12,870 (87%) were interviewed and 2993 had a live birth in the previous 12 months. Of the 378 and 360 planned facility assessments during the six rounds, 354 (94%) and 302 (84%) were completed in Tanzania and Uganda, respectively. A total of 409 and 291 last event interviews were done with health workers in Tanzania and Uganda.

The internal monitoring data shown on the run charts of the QI teams suggested improvements in facility delivery (Additional file 1: webannex 6a and 6b). The combined run charts of the 34 included facilities in Tanzania suggested an increase in the coverage levels of facility births from around 85 to 95% over the implementation period, a 10-percentage-point increase. However, results warrant careful interpretation because home births were not reliably documented in facility records.

According to the household survey data, facility delivery increased from 55 to 87% in the intervention district and 62 to 78% in the comparison district, suggesting no evidence of any association between the intervention and facility delivery in Tanzania (difference-in-difference 7%; 95% CI −7 to 21%, Table 3, Fig. 2). In Uganda, facility delivery increased from 56 to 68% in the intervention and 31 to 42% in the comparison area, giving no evidence of an association between the intervention and facility delivery (difference-in-difference −3%; 95% CI −15 to 9%).

We found some evidence that the EQUIP approach increased the proportion of mothers who reportedly received uterotonics within 1 min of childbirth. In Tanzania, we observed an increase in the proportion of women with a live birth in the year prior to the survey who received uterotonics, with a difference-in-differences of 26 percentage points (95% CI 25–28%). In Uganda, the difference was smaller at 8 percentage points (95% CI 6–9%). We found no evidence of an association between the EQUIP intervention and the primary outcomes of immediate breastfeeding or knowledge of maternal and newborn danger signs.

In Tanzania, the analysis of the secondary coverage outcomes, chosen to reflect the prioritised improvement topics, indicated some evidence of an association between the EQUIP intervention and birth preparedness through preparation of clean birth kits for home deliveries (31%; 95% CI 2–60%). No evidence of an association was observed for early postnatal care for home deliveries (17%; 95% CI −8 to 40%). There was no evidence of an effect of the intervention on increased availability of selected items for infection prevention in health facilities (21%; 95% CI −4 to 46%). We observed some association between the EQUIP approach and improvements in supervision visits by district managers to primary health facilities (14%; 95% CI 0–28%). In Uganda, we saw no evidence of effect of the intervention on any secondary coverage outcomes.

The context analysis indicated differences in key health system indicators between intervention and comparison districts and between both countries (Table 1). District reports on finances indicated that in 2013 and 2014, the Tanzanian comparison district had 12 USD per capita for health expenditure, compared to 7 USD in the intervention district, Tandahimba. For Uganda, comparable data were unavailable, but other studies suggest levels around 6 USD per capita for health [40]. In Tanzania, the situation on human resources was better in the intervention district compared to the comparison district (0.97 compared to 0.79 nurses per 1000 population). In Uganda, more human resources were available in the intervention district (0.70 compared to 0.53 nurses per 1000 population). There were few in-service-training sessions in maternal and newborn health in both districts in Tanzania other than those supported by EQUIP. In Uganda, a few training sessions on maternal and newborn health care were provided in the intervention area, supported by other international partners. Information from the comparison area was not available. Availability of drugs and supplies was better overall in Tanzania than those in Uganda. For example, magnesium sulphate was available in the intervention district in Tanzania in the first and sixth round of facility assessments in 21 and 30% of facilities, whereas the respective figures were 3 and 0% in the intervention district in Uganda.

In both countries, health planning was based on health information from the health management information system (HMIS). In Tanzania, the EQUIP data were not included in the formal planning process, as planning guidelines recommend the use of HMIS data alone. In Uganda, however, the district team started to use the EQUIP data in planning in the second year of EQUIP support.

Our study has four major strengths. Firstly, it is to our knowledge the first to conduct QI at all levels of a district health system, including district management, all primary and referral facilities and communities through a network of volunteers and community health workers. Secondly, we used a plausibility design as recommended for health systems interventions [29, 31] which provides more robust evidence than a before-and-after design. Moreover, our context analysis helped to interpret findings from a plausibility perspective [29]. Thirdly, we evaluated the QI intervention using independently generated population- and facility-based data in both intervention and comparison areas. Fourthly, the same intervention was implemented in two different countries, which provided learning in different contexts.

Our study also has limitations. Although the survey team collected data that were independent from the QI teams, they were not masked. However, we think it unlikely that interviews with mothers in intervention and comparison areas or in health facilities differed systematically, because the survey team operated independently and were unaware of the chosen outcome indicators.

We implemented QI in only one intervention district in each of the two countries, and our quasi-experimental design limits both generalisability and internal validity. Our context analysis showed differences between intervention and comparison districts. While these differences did not point to a clear advantage of either district in Tanzania, the implementation district in Uganda seemed to be better equipped in terms of human resources and drugs and supplies, biasing the results in favour of the intervention.

For the analysis over time, our household surveys were powered to detect statistically significant changes of 10 percentage points or greater, which may have contributed to the lack of evidence of associations between the EQUIP approach and improved outcomes. Also, we had a limited number of data points to be included in the analysis over time.

The methods we used to assess changes in knowledge of danger sign and breastfeeding within 1 h are subject to measurement biases, despite our use of a well-established sequence of questions [49]. Misinterpretation of breastfeeding questions, for example, has been reported [50]. For many important intrapartum quality indicators, such as uterotonics after childbirth, no standard method of measurement is available. We used last event reports of health workers; however, their reliability can be questioned. However, a small add-on observation study in the implementation district confirmed implementation levels reported by health workers. Moreover, we considered injection of an uterotonic within 1 min as the most important aspect of Active Management of the Third Stage of Labor as supported by the literature [51] and omitted two other aspects of controlled cord traction and uterine massage.

We phased short implementation period, coupled with the prolonged roll-out in a larger implementation district in Uganda than Tanzania. The issues of relatively low implementation strength and short duration were exacerbated by the use of women’s report to establish coverage. In household surveys, mothers are commonly asked to report on births that occurred within a defined period before the interview date, thus providing information about the past. We used a recall period of 12 months prior to the date of assessment for population indicators, which meant that only the last survey round covered a time period where all teams were active, but not all improvement topics were implemented for this time period (Additional file 1: webannex 3).

This design constraint might have biased out the results versus the nullhyposthesis.

For QI to have an impact, sufficient input and support is required. Many of the change ideas were only fully implemented after a 3-month piloting phase so that implementation in all facilities and communities was achieved only 6–12 months before the end of the project. Improvement ideas need to mature, and it often takes several months for an improvement topic to reach levels above 90% [15]. Moreover, our intervention might have had too little focus on the improvement areas for which we had pre-defined primary indicators as we allowed the districts to prioritise improvement topics from a much broader pre-defined list shown in the improvement charter. Finally, there was some misalignment of QI topics teams worked on and primary outcomes for breastfeeding in Tanzania.