A systematic review protocol was performed adhering to the guidelines outlined in the Meta- Analysis of Observational Studies in Epidemiology (MOOSE) for systematic reviews (24) and the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) statement for reporting systematic reviews and meta-analyses (25). The investigators wrote a protocol and registered it with the International Prospective Register of Systematic Reviews (identification number CRD42024530774) in March 2024.

The search strategy was designed to identify studies reporting stepped-wedge study design in SSA by systematically searching literature databases from January 2000 to March 2024. We searched PubMed, Web of Science, CINAHL, PsycINFO, and the Cochrane Library, alongside Google Scholar and citation searching. Searches were restricted to English-language publications due to feasibility constraints and to ensure consistent screening and extraction. Detailed search strategiesfor each database are reported in Supplementary File S1. The bibliographies of relevant reviews and eligible studies were also examined for additional sources. In addition, grey literature sources were consulted, including repositories of conference proceedings and university theses, and relevant expert sources.

In this systematic review, we focused exclusively on studies that utilized the SWD and were conducted within SSA countries. We included studies set in various environments, such as hospitals, communities, schools, and other public health-related contexts. Our review specifically considered studies that evaluated public health, clinical, or social interventions, with an emphasis on those that reported the effectiveness of these interventions and provided data on the applicability of the SWD in African settings.

To ensure the relevance and quality of the included studies, we only considered publications in English or French, covering a timeframe from the inception of each database up to the date of the final search.

We excluded studies that did not employ the SWD or used modified versions that significantly deviated from the standard SWD framework. Additionally, studies conducted outside of SSA or those focusing on non-SSA populations were not included. We also excluded studies that failed to report relevant outcomes or provided insufficient data for meaningful analysis. Furthermore, publications such as editorials, commentaries, opinion pieces, conference abstracts without full data, and non-peer-reviewed articles were excluded from this review.

We used a comprehensive strategy to identify studies on SWD in SSA countries. All records were imported into Rayyan (Rayyan Systems Inc.) for deduplication and title/abstract screening. Three reviewers (O.M.O., O.R.A, and A.Z.M) read the titles and abstracts to determine each study's eligibility for full-text review. Studies were included if they described an original research trial that used a SWD approach incorporating all the following elements: clusters, randomization, and comparison with controls. Studies retrospectively analyzed as a stepped wedge when the study was not designed initially as such were excluded. Given the focus on cluster randomized trials, trials where individuals were the unit of randomization or the stepping level were excluded. Duplicate, serial, or follow-on publications for the same study were counted as one, although data extraction may have come from several sources.

Two separate reviewers (O.R.A. and O.M.O.) independently reviewed full-texts for data extraction based on the pre-defined criteria. When further clarification information was required, authors were contacted via email. Discrepancies were resolved by discussion or involvement of a third reviewer (F.T.A.).

Three authors (A.Z.M, O.R.A., and O.M.O.) independently extracted data from published sources using a data extraction form. The extracted information included study identifiers (lead author and publication year), research area (subject areas, disease/domain, nature of intervention, setting, and country), motivation for using the stepped wedge design, general characteristics of the stepped wedge design (level of stepping, number of steps, number of clusters, total number of participants, and period between steps), methods of data analysis (primary outcome, comparisons, and statistical methods of data analysis), and selected items on quality of reporting. The data extraction form was piloted on ten articles and adjusted accordingly. Authors were contacted for complete reports for studies with only abstracts or protocol summaries. Additional information was obtained from the study authors where required. Differences in data extraction were resolved through discussion or referral to a fourth author (F.T.A).

No institutional review board approval was required for this study.

2,562 articles were retrieved from five databases (PubMed, Web of Science, CINALH, PsychINFO, and Cochrane), Google Scholar and citation Searching. The search ranges from January 2000 to March, 2024. One thousand six hundred and eighty-nine articles were removed before screening the remaining 873 were screened for titles and abstracts. Of the 873 articles, 780 were excluded based on the study exclusion criteria. Furthermore, 93 articles were retrieved for full-text screening. Finally, after a thorough full-text screening, 85 articles met the inclusion criteria (26–110) (see Figure 2).

The extracted data from the 85 studies are presented in Table 1. The studies included a total of 1,895,788 participants. The study's sample sizes range from 17 to 780,000. Of the 85 studies included, 72 (84.7%) were conducted in a facility-based (26, 27, 30–34, 36–47, 49, 50, 52–64, 66–75, 77–81, 83–96, 98–102, 104, 105, 107, 109), and 13 (15.3%) were community-based studies (28, 29, 35, 48, 51, 72, 76, 82, 97, 103, 106, 108, 110).

Sixteen (18.8%) were conducted in Uganda (26, 32, 39, 40, 46, 47, 53, 84, 85, 92–96, 100), 14 (15.3%) in Kenya (29, 35, 37, 41, 43, 48, 59, 71, 76, 88, 105, 106, 110), and 11 (12.9%) in South Africa (52, 55, 56, 63, 73, 80, 86, 90, 91, 103, 107). Other countries where SWD has been used includes 7 (8.2%) in Eswatini (36, 42, 65, 66, 87, 99) and Tanzania (49, 67, 77, 81, 82, 97, 102), 4 (4.70%) in Nigeria (33, 57, 58, 72), and Malawi (44, 54, 78, 83), 3 (3.5%) in Botswana (31, 98, 104), 2 (2.4%) in Mozambique (30, 64), Burkina Faso (28, 101), Senegal (34, 61), and Zambia (38, 79), and 1 study each in Ghana (68), Ethiopia (74), Rwanda (69), Sierra Leone (27) and Zimbabwe (70). Five (5.9%) studies were conducted in multi-countries, one study each was conducted across Bissau, Guinea-Bissau, and Gondar, Ethiopia (50); Cameroon and Kenya (75); Kenya and Zimbabwe (45); Nigeria and Tanzania (62); and one study from Senegal, Egypt, and Uganda (60).

The duration of the included studies varied widely, ranging from two months to sixty months. Notably, 4 studies (4.71%) did not report the study duration (29, 49, 72, 85). The most common study length was thirty-six months, observed in 17 (20%) studies (33–37, 41, 61, 65, 66, 68, 87, 89, 91, 101, 106, 107, 110). Additionally, 12 (14.1%) studies were conducted over twelve months (27, 28, 38, 40, 47, 48, 50, 62, 64, 77, 100, 109).

Seven (29.2%) studies each were conducted over twenty-four months (32, 42, 45, 57, 76, 97, 104) and forty-eight months (31, 67, 71, 80–82, 108) durations. Furthermore, 5 (5.9%) studies were conducted over eight months (26, 43, 69, 83, 94), and 4 (4.7%) studies spanned 60 months (44, 59, 79, 103). 2 studies each had the following durations: six months (63, 92), ten months (52, 95), fifteen months (46, 98), eighteen months (55, 60), nineteen months (56, 90), twenty-three months (30, 105), twenty-eight months (53, 88), and thirty months (39, 78). Finally, 1 study was conducted over each of the following periods: two months (51), three months (54), 11 months (73), fourteen months (96), sixteen months (74), twenty months (102), twenty-two months (75), twenty-five months (86), thirty one months (84), thirty-five months (70), thirty-seven months (99), forty-two months (93), and forty-six months (58).

Figure 3 shows the distribution of SWDs by country and health domains in SSA. A significant proportion of the studies focused on HIV/AIDS, with thirty-four (40%) studies falling within this domain (31, 36, 37, 39–43, 45, 48, 53, 55, 57, 59, 63–67, 70, 71, 78, 80, 81, 84, 87–89, 93, 99, 103, 105, 107, 109). Seventeen (20%) studies addressed maternal and child health (27, 30, 32, 33, 46, 47, 49, 60, 72, 86, 95–98, 100, 101, 110), while eleven (12.9%) studies focused on tuberculosis (26, 50, 68, 74, 75, 85, 90–92, 94, 104). Malaria was the subject of seven (8.2%) studies (28, 34, 37, 61, 76, 82, 106), and another seven studies (8.2%) were conducted in the general health domain (35, 44, 54, 62, 79, 102, 108).

Two (2.4%) studies each were carried out in the domains of neonatal health (57, 69), and sexual and reproductive health (29, 77). Additionally, one (1.2%) study was conducted in the following domains: cardiovascular health (51), respiratory health (58), and malnutrition (83). Moreover, two (2.4%) studies covered multiple domains, specifically HIV/TB (52) and pneumonia/malaria/diarrhea (58).

Table 2 provides a comprehensive overview of the characteristics of the studies included in the analysis. It details key elements such as the first author and publication year, the definitions of the clusters used, the nature of the interventions implemented, and the randomization process. The table also includes information on the rationale for choosing a stepped-wedge design, the number of steps and clusters involved, the time between steps, sample size calculations, and the primary outcomes of each study. This structured presentation allows for a clear comparison of study designs and outcomes across different interventions and settings.

Figure 4 shows the definition of clusters in SWD in SSA. Health facilities, strategically chosen as clusters in numerous studies, underscore the emphasis on healthcare delivery and interventions within controlled environments. The everyday use of hospitals and clinics as clusters is rooted in their pivotal role in healthcare delivery, enabling the implementation and evaluation of interventions to enhance patient outcomes (26, 27, 30–34, 36–47, 49, 50, 52–71, 73–75, 77–81, 83–89, 91–96, 98–102, 104, 105, 107, 109).

Some studies focused on specific units within hospitals or clinics to target interventions at a more granular level. Primary care facilities were often chosen as clusters in low-resource settings (37, 52, 55, 56, 62, 70, 91, 92, 101, 108).

Clusters defined by geographical or administrative boundaries were used to capture broader population-based interventions, particularly in rural or resource-constrained areas. Administrative units such as districts and wards allowed for evaluating public health interventions across larger populations (26–28, 30, 33, 34, 36, 38, 42, 47, 51, 52, 62, 69, 71–73, 75–79, 84–86, 91, 92, 94–97, 99, 101, 104, 108, 109).

In more localized studies, clusters were often defined at the village or community level to capture the effects of interventions on specific populations (28, 29, 39, 48, 54, 72, 78, 81, 103, 106, 108).

Some studies utilized educational institutions as clusters to implement and evaluate public health interventions for children and adolescents. Schools served as effective clusters for interventions focused on health education, disease prevention, and other child-focused outcomes (35, 110).

Religious institutions were used as clusters in studies where community engagement and trust were crucial for the success of the intervention. Churches were selected as clusters in studies that leveraged community trust and leadership to deliver health interventions (29, 48).

In some studies, social groups or communities defined clusters to target interventions at a more personal and community-driven level. Clusters based on social networks allowed for the examination of interventions within the natural groupings of individuals (28, 29, 35, 48, 51, 72, 76, 78, 81, 82, 90, 97, 103, 106, 108, 110). Community-level clusters enabled the study of interventions that required collective participation or were community-driven. Some studies employed clusters that spanned multiple settings, such as combining health facilities with community or household interventions, reflecting the complexity of specific public health interventions. Studies that integrated health facility-based interventions with community outreach often used mixed-setting clusters (33–36, 38, 48, 54, 61, 67, 72, 76, 78, 80, 92, 99, 103, 106–108).

Figure 5 shows the nature of interventions used for SWD identified in the included studies. Several studies focused on implementing health technologies and digital solutions to improve healthcare delivery and patient outcomes (26, 27, 32, 54, 62, 79, 85, 90, 91, 94, 98, 101, 104, 105). These interventions often involved using digital tools or devices designed to enhance the accuracy, efficiency, and accessibility of healthcare services. For example, a study utilized a 99DOTS-based TB treatment supervision strategy that utilizes a digital adherence tool where patients call toll-free numbers hidden in blister packs to confirm medication adherence (26). The intervention was supported by automated messaging systems and health worker monitoring to improve treatment adherence. The integration of point-of-care diagnostic tools was another common theme (41, 45, 68, 101). Emma Sacks (2020) implemented routine POC early infant diagnosis (EID) of HIV to provide timely on-site testing for infants, improving diagnosis and treatment initiation (45). Similarly, Eduard J. Sanders (2021) used point-of-care HIV-1 nucleic acid testing (NAAT) to detect acute HIV infections in symptomatic adults (41). Scott Dryden-Peterson (2015) introduced an automated SMS-based CD4 result platform that wirelessly distributed CD4 results to clinics, streamlining the process of HIV patient management in antenatal clinics (98).

Many of the interventions involved distributing and administering pharmaceutical treatments to prevent or manage diseases (28, 34, 36, 38–40, 42, 61, 90, 91, 93, 96, 98). Badara Cisse (2016) and J. L. Ndiaye (2016) implemented SMC with sulfadoxine-pyrimethamine (SP) plus amodiaquine (AQ) for children under ten years during the malaria transmission season, reducing the incidence of clinical malaria (34, 61). Dorothy Thomas (2022) and Elizabeth M. Irungu (2021) integrated PrEP delivery into existing HIV care services, offering it to HIV-negative partners of serodifferent couples, alongside ongoing technical assistance and training for healthcare providers (40, 43). The integration of ART services was a crucial focus in studies such as Elaine J. Abrams (2019), which transitioned from CD4+ guided ART eligibility to universal ART for all HIV-positive pregnant and breastfeeding women (42), and Osondu Ogbuoji (2019), which implemented the Early Access to Antiretroviral Therapy for All (EAAA) strategy in public-sector clinics (87).

Many studies involved interventions centered around training healthcare workers or providing health education to patients and communities (29, 32, 33, 48, 49, 62, 70, 78, 81, 86, 88, 89). Anne Antonia Cornelia van Tetering (2021) focused on a one-day simulation-based acute obstetric training for healthcare providers, emphasizing medical technical skills and teamwork in managing obstetric emergencies (32). Pascal Geldsetzer (2020) provided training for healthcare providers on PrEP delivery as part of HIV prevention efforts (89). Ali M. Giusto (2017) implemented a family- and church-based intervention aimed at preventing HIV risk behaviors among adolescents by strengthening family relationships through community-based education (29). Anne Cockcroft (2022) provided universal home visits by trained health visitors to educate pregnant women and their spouses on local maternal and child health risk factors (33).

Behavioral and psychosocial interventions aimed at altering behavior and improving mental health outcomes were a crucial focus in several studies (33, 48, 49, 78, 80, 81, 88, 93, 102, 105, 107, 110). Margaret W. Gichane (2020) introduced the Women's Health CoOp (WHC), a gender-focused intervention combining HIV and alcohol risk reduction with behavioral skills training, focusing on sexual empowerment and communication skills (80). Wendee M. Wechsberg (2021) also utilized the WHC intervention in a different context, aiming to reduce HIV-related risks among women who use alcohol and drugs (107). Stella-Maria Paddick (2017) implemented Cognitive Stimulation Therapy (CST), a psychosocial intervention to improve cognitive function in individuals with dementia (102). Eva Wodeya Wanyenze (2022) focused on providing orientation sessions for birth companions to support labor through emotional and physical support techniques (47).

Some interventions aimed to strengthen healthcare infrastructure and systems to improve service delivery and patient outcomes (27, 49, 52, 54, 56–58, 60, 62, 65, 66, 68, 70, 72, 74, 77, 79, 84–88, 90, 91, 93–96). Hamish R.Graham (2019, 2021) focused on implementing improved oxygen systems in hospitals, including the introduction of pulse oximetry and full oxygen systems for children in Nigeria, aiming to reduce mortality from hypoxemia (57, 58). Wilbroad Mutale (2023) introduced the Better Health through Mentoring and Assessment (BHOMA) project, which involved intensive clinical training, quality improvement measures, and improved community outreach to enhance healthcare service delivery at the primary care level (108).

Integration care models were also implemented in some studies combining multiple health services or interventions into a cohesive approach (44, 77, 83, 97). Emily B. Wroe (2021) expanded an existing community health worker (CHW) program for HIV and TB into a broader household-based model that included non-communicable diseases (NCDs), malnutrition screening, and family planning services (44). Leigh Senderowicz (2022) focused on integrating postpartum intrauterine device (PPIUD) insertion into routine antenatal family planning counselling (77). Monica P. Patel (2005) provided supplemental feeding with Ready-to-Use Therapeutic Food (RUTF) for children at risk of malnutrition, integrating nutritional support into routine maternal and child health services (83). Samson Winani (2007) combined the use of clean delivery kits with health education on the "six cleans" to prevent maternal and newborn infections (97).

Simple randomization, often using computer-generated sequences, was a standard method employed to determine the order in which clusters received the intervention (26, 30, 32, 35, 38, 47, 58, 65). This approach ensures that each cluster has an equal chance of being assigned to any given sequence, reducing selection bias. Adithya Cattamanchi (2021) used a simple, unrestricted two-stage process where health facilities were randomly assigned to blocks and the order of intervention was determined through a random process (26). Similarly, Ryan R. Thompson (2022) used a random process to determine which clinics would receive the 99DOTS intervention each month (94). In the study by Hamish R.Graham (2021), hospitals were allocated to receive an oxygen system based on a computer-generated random sequence, ensuring a phased introduction of the intervention (58). This method was similarly employed by Eva Wodeya Wanyenze (2023), where a simple random technique determined the order of hospital crossover to the intervention period (47).

Stratified randomization was used in several studies to ensure balance across key variables such as geographic location, facility size, or patient volume before random assignment (40, 43, 60, 64–66, 70, 71, 88, 99, 105). This technique helps control potential confounding factors by ensuring that these variables are equally distributed across the intervention groups. Janina Isabel Steinert (2021) grouped health facilities into pairs based on geographic proximity and catchment size before randomizing them into different sequences for intervention rollout (65). This approach ensured that facilities with similar characteristics were evenly distributed across the study arms. Elizabeth M. Irungu (2021) utilized stratification by region before randomly assigning clinics to the order in which they would start receiving the intervention (43). This method was also used by Pamela K. Kohler, who stratified randomization based on region and facility size to ensure a balanced distribution of intervention timing across different facility types (88). Block randomization was employed to ensure that equal numbers of clusters received the intervention at each stage, particularly in studies with phased rollouts. This method is useful in maintaining balance throughout the intervention period. Dorothy Thomas (2022) randomized facilities into groups of four, with each group beginning the intervention six months apart (40). This ensured that the implementation was evenly distributed over time while controlling for potential confounders related to the timing of intervention delivery. Katherine R. Sabourin (2022) used block randomization to balance the number of urban and rural sites, ensuring that the timing of intervention rollout was similar across different settings (70). This method ensured that both urban and rural sites had comparable exposure to the intervention phases (see Figure 6).

In some studies, randomization was performed using a lottery or public draw method, which added an element of transparency and community involvement in the randomization process (29, 43, 78, 106). This method was particularly used in community-based interventions. Ali M. Giusto (2017) employed a lottery system conducted by community leaders to randomly select churches for participation (29). This approach was similarly used by Lily C. Kumbani, where communities were randomly assigned to intervention sequences by drawing names from a basket (78). In the study by Elizabeth M. Irungu (2021), randomization was conducted at stakeholder events where facility representatives picked numbered opaque balls from a bag, determining the sequence in which their facilities would receive the intervention. This method increased transparency and stakeholder engagement in the research process (43).

Some studies utilized systematic or quasi-randomized approaches when true randomization was not feasible, often due to logistical constraints or the need to align with existing organizational structure (35, 83, 106). Calum Davey (2015) employed a quasi-randomized method where schools were systematically allocated into three groups using an Excel spreadsheet, with each group receiving the intervention at different times (35). This approach was necessary due to the school system's structure and logistical considerations. Tobias Homan (2016) used the traveling salesman algorithm to assign households to clusters, ensuring an efficient rollout of the intervention while maintaining a randomized element (106). This innovative approach was particularly suited for interventions requiring a geographically coherent implementation plan (see Figure 7).

In several studies, randomization was combined with masking to reduce bias during the intervention phase (43, 56). Masking ensures that participants and sometimes even the investigators are unaware of the sequence of intervention, which can help prevent conscious or unconscious influences on the study outcomes. H. Manisha Yapa (2020) randomized clusters to receive the intervention on pre-specified dates, with investigators and clusters being blinded to the randomization sequence until two weeks before each step (56). This method was also applied in Janina Isabel Steinert (2020), where the randomization sequence was concealed to minimize potential biases (66) (see Figure 8).

One of the primary reasons for choosing the SWD was the ethical imperative to ensure that all study participants eventually received the intervention, especially when the intervention was expected to be beneficial (44, 46, 47, 52, 55, 57, 69, 96, 97, 106, 110). Emily B Wroe (2021) and Hamish R. Graham (2021) selected the SWD to ensure that all clusters eventually received the potentially beneficial interventions while allowing for rigorous evaluation (44, 58). Wit Wichaidit (2019) also used SWD to avoid the ethical dilemma of withholding a beneficial intervention (handwashing stations) from some schools, emphasizing that the intervention posed no harm and provided clear benefits (110). Samson Winani (2007) and Wendee M. Wechsberg (2021) employed the SWD to address ethical issues associated with traditional cluster-randomized trials where control clusters might not receive an efficacious intervention (97, 107). The SWD allowed for the gradual introduction of the intervention across all clusters, ensuring ethical treatment of all participants.

The SWD was often chosen to accommodate logistical and operational challenges, particularly in studies where simultaneous implementation across all clusters was impractical (26, 30, 33, 34, 36, 37, 40, 43, 45, 49, 51, 52, 58, 61, 62, 65, 66, 68, 81, 86, 91, 94, 101, 110). Dorothy Thomas (2022) and Slawa Rokicki (2021) selected the SWD to manage logistical constraints, allowing for a phased rollout of the intervention (40, 100). This design was practical in settings where resources were limited, enabling a more manageable and effective implementation process (41, 45, 50, 51, 62, 77, 83, 91, 100–102). Sophie Sarrassat (2021) and Tobias Homan (2016) used the SWD to facilitate the gradual rollout of interventions, which was more practical for implementing agencies and allowed for the effective management of resources (101). Jackie Sturt (2023) also cited capacity constraints and the need for phased implementation during the COVID-19 pandemic as reasons for choosing the SWD (62).

Quite a number of these studies selected SWD to account for time-related confounders, ensuring that temporal trends did not bias the results. Calum Davey (2015) and Badara Cisse (2016) employed the SWD to adjust for time effects, given evolving programs and external conditions that could influence the study outcomes (34, 35). By implementing the intervention in a phased manner, these studies were able to control for potential confounders related to the timing of the intervention. CE Boeke (2020) and Clara A. Agutu (2022) chose the SWD to allow for a systematic rollout of the intervention across all facilities over time, controlling for time-related confounders while ensuring that each facility eventually served as its own control (36, 37).

The SWD was frequently used to ensure that all clusters or study sites eventually received the intervention, particularly in public health studies where broad coverage was essential (28, 33–35, 37, 43–50, 52, 54–58, 60–62, 68, 70, 71, 74, 77–79, 81, 82, 85–90, 92–94, 96, 98, 99, 103, 105, 106, 108, 109). Elizabeth M. Irungu (2021) and David B. Meya (2019) opted for the SWD to facilitate a staggered approach to implementation, ensuring that all clinics or health facilities eventually received the intervention (39, 43). This was particularly important in studies integrated into broader national or regional health programs. Emma Sacks (2020) and Kenneth K. Mugwanya (2023) selected the SWD to maximize the impact of the intervention by ensuring its systematic rollout across all participating sites, enabling a thorough evaluation while maintaining operational efficiency (45, 71).

Practical considerations, including the need for a flexible and adaptive study design, were key drivers for selecting the SWD in many studies. Janina Isabel Steinert (2021) and James T. Pfeiffer (2017) used the SWD to accommodate logistical and practical constraints over extended study periods (64, 65). The phased implementation allowed for adjustments based on real-world conditions, making the design more adaptable and feasible. Geofrey Musinguzi (2020) and F. Mosha (2005) highlighted the SWD's efficiency in resource-constrained settings, where a staggered rollout allowed for better management of limited resources while still providing a rigorous evaluation of the intervention's effectiveness (49, 51).

The SWD was chosen in some studies because it allowed the intervention to be integrated into existing health programs, facilitating a smoother transition and broader acceptance. Andrew F. Auld (2020) and Tefera Agizew (2019) selected the SWD because the intervention was part of a national rollout (31, 104). The design allowed for the phased implementation of the intervention across multiple sites, ensuring alignment with existing health initiatives while enabling a robust evaluation. Alexandra E. Ridout (2023) and Holly A. Anger used the SWD to embed the intervention into routine clinical pathways, ensuring that all study sites could eventually incorporate the intervention into their standard practices while maintaining a focus on evaluating its effectiveness over time (27, 60).

The number of steps in a SWD is crucial as it influences the study's duration, complexity, and ability to detect time-related changes. The number of steps in the reviewed studies varied significantly, reflecting the flexibility of the SWD to adapt to different research needs and contexts. The fewest steps used in an SWD study was 1, as seen in Stella-Maria Paddick (2017) (102). This study likely opted for a single step due to resource limitations or the simplicity of the intervention being tested. The highest number of steps observed was 27, used in Tobias Homan (2016) (106). This extensive number of steps was likely necessary to facilitate a highly granular and phased rollout of the intervention, perhaps due to logistical challenges or the need to closely monitor the impact at multiple stages (see Figure 9).

Several studies opted for a moderate number of steps, with 3 to 4 steps being the most frequently chosen configuration. This range provides a balance between logistical feasibility and the ability to rigorously evaluate the intervention over time.

A significant number of studies employed 3 steps, such as in Andrew F. Auld (2020), Dorothy Thomas (2022), and Renee Heffron (2022) (31, 40, 93). This choice often reflects a phased approach where the intervention can be rolled out in manageable increments while maintaining sufficient control over time-related variables.

Another common choice was 4 steps, seen in studies like Emma Sacks (2020), Anne Cockcroft (2022), and Shaukat Khan (2020) (33, 45, 99). This configuration allows for a more detailed evaluation compared to fewer steps while still being relatively straightforward to manage.

The rationale behind choosing a specific number of steps is often related to the intervention's logistical demands, the study's complexity, and the need to control for temporal effects. In studies with extensive logistical demands or resource constraints, fewer steps were often chosen to simplify the implementation process. For example, Alexandra E. Ridout (2023) and Calum Davey (2015) both utilized two steps to streamline the rollout of their interventions (27, 35).

Studies requiring detailed monitoring or involving complex interventions tended to use more steps. For instance, Hamish R.Graham (2019) used 12 steps to introduce an oxygen system in hospitals, which allowed for careful monitoring and adjustment at each stage of the rollout (57).

Some studies balanced complexity and feasibility by choosing a moderate number of steps. Anne Antonia Cornelia van Tetering (2021) used 7 steps to implement a training program for healthcare workers, which allowed for phased implementation and comprehensive evaluation without overwhelming the study team (32).

Several studies did not explicitly define the number of steps or used a variable number of steps based on the study's progression.

Studies like David B. Meya (2019) and Eduard J. Sanders (2021) involved multiple steps without a fixed number, with facilities transitioning from control to intervention phases at different intervals (39, 41). This approach provided flexibility to adapt the study design based on real-time developments and operational constraints.

In some cases, the exact number of steps was not clearly specified in the study documentation, as seen in CE Boeke (2020) and Noah Kiwanuka (2023) (36, 85). This could indicate a more fluid approach to study design, where the number of steps was determined by external factors such as funding, participant availability, or logistical support (see Figure 10).

The number of clusters in the studies varied widely, reflecting the SWD's flexibility in accommodating different research contexts and logistical constraints.

The study with the fewest clusters used only 3 clusters, as seen in Johanna Åhsberg (2023) and Jeffrey Rewley (2020) (67, 68). These studies likely involved small-scale interventions or were conducted in highly controlled environments where limited clusters were sufficient for achieving the study objectives. The study with the most clusters included 142 clusters, as seen in Pren Naidoo (2017) and Nynke van den Broek (2019) (86, 90). These studies were typically large-scale public health interventions that required a significant number of clusters to cover diverse populations and settings, allowing for a robust evaluation of the intervention across various contexts (see Figure 11).

Several studies employed a moderate number of clusters, with 6 to 10 clusters being the most frequently chosen configuration. This range provided a balance between statistical power and logistical feasibility.

A significant number of studies used 6 clusters, such as Leigh Senderowiez (2022), Emily B Wroe (2021), and Frauke Rudolf (2021) (44, 50, 77). This number of clusters allowed for a manageable study size while still providing enough variation to detect differences across clusters.

Ten clusters were also commonly used, as seen in studies like F. Mosha (2005) and Ana Pilar Betrán (2018) (30, 49). This configuration offered a good balance between complexity and the ability to generalize the findings across different settings.

The rationale behind choosing the number of clusters is often related to the study's logistical constraints, the need for statistical power, and the geographical spread of the intervention.

Studies often selected a moderate number of clusters based on logistical feasibility. For example, Badara Cisse (2016) used 54 clusters (health posts) due to the need to implement a phased intervention in different regions over time (34). Similarly, Katherine R. Sabourin (2022) used 8 clusters to ensure a balanced implementation across both rural and urban healthcare settings (70).

The number of clusters was frequently determined by the need to achieve sufficient statistical power. Andrew F. Auld (2020) used 21 clusters to ensure adequate power to detect a reduction in ART mortality rates (31), while Anne Antonia Cornelia van Tetering (2021) used 7 clusters to provide enough data points for a statistically robust analysis (32).

In some studies, the number of clusters was influenced by the geographical distribution of the intervention sites. Leigh Senderowicz (2022) used 6 clusters across different regions in Tanzania to ensure that the intervention could be evaluated across diverse healthcare settings (77).

A few studies required an exceptionally high number of clusters to accommodate large-scale interventions or to account for a wide range of settings.

Nynke van den Broek (2019) involved 127 clusters across 11 districts in South Africa, reflecting the study's national scope and the need to cover a broad range of healthcare facilities (86). Pren Naidoo (2017) similarly used 142 clusters to evaluate a tuberculosis diagnostic intervention across multiple primary healthcare clinics, demonstrating the scale required to assess such a widespread public health intervention (90).

The time between steps in the SWD studies varied widely, reflecting the design's adaptability to different research needs and logistical constraints.

The shortest time between steps was two weeks, observed by Josee Uwamariya (2021) and Tefera Agizew (2019) (69, 104). This brief interval was likely chosen to allow for rapid rollout and assessment of the intervention in settings where timely results were critical.

The longest time between steps was 121 months (approximately 10 years), seen in Pascal Geldsetzer (2020) (89)This extended interval was likely due to the intervention's long-term nature and the need to observe outcomes over a prolonged period.

Several studies employed a moderate time interval between steps, with 1 to 6 months being the most frequently chosen configuration (26–28, 30–32, 36, 38–46, 48, 49, 52, 55–58, 62–64, 67, 70, 73, 74, 76, 80–82, 86, 87, 91, 92, 94, 96, 98–101, 105). This range provides a balance between allowing sufficient time for the intervention to take effect and maintaining the momentum of the study.

A significant number of studies used 1 month between steps, such as William P. Killam (2010), Lea Multerer (2021), and Ryan R. Thompson (2022) (94, 109). This relatively short interval allowed for a quick assessment of the intervention's impact while keeping the study timeline manageable.

Six months was another common choice, as seen in Adithya Cattamanchi (2021), Dorothy Thomas (2022), and Wendee M. Wechsberg (2021) (26, 40, 107). This duration provided ample time for the intervention to be fully implemented and for its effects to be observed before moving to the next cluster.

The rationale behind selecting specific time intervals between steps is often related to the nature of the intervention, the study's logistical constraints, and the need to control for time-related confounders.

The complexity and expected impact of the intervention often dictated the time between steps. For instance, Eduard J. Sanders (2021) used a 6-month interval to allow sufficient time for the intervention (HIV prevention services) to take hold and produce measurable outcomes (41). In contrast, Holly A. Anger (2020) used just 1 week between steps, likely due to the rapid nature of the intervention and the need for quick implementation (60).

Logistical considerations, such as the availability of resources and the need to coordinate with multiple sites, influenced the time between steps. Hamish R.Graham (2019) chose 4 months between steps to allow for the installation of oxygen systems in hospitals, which required careful planning and execution (57).

In studies where time-related confounders were a concern, longer intervals were chosen to ensure that any effects observed were attributable to the intervention rather than external factors. Anne Cockcroft (2022) and Badara Cisse (2016) both used 1 year between steps to account for seasonal variations and other time-related factors that could influence the outcomes (33, 34).

Some studies did not have a fixed time between steps or employed variable intervals based on the study's specific needs.

Wit Wichaidit (2019) used variable intervals ranging from 3 to 24 weeks depending on the group, reflecting the need to adapt the study timeline to different settings or phases of the intervention (110). Katherine R. Sabourin (2022) used 7–8 weeks between steps, providing some flexibility in the study design (70).

In certain studies, the time between steps was not explicitly defined, allowing for flexibility in the study design. Lily D. Yan (2017) and Margaret W. Gichane (2020) had undefined intervals, likely due to the need to adapt to changing circumstances or logistical challenges during the study (34, 79).

The sample size calculation in SWD studies varied significantly, reflecting the diversity in study designs, outcomes of interest, and the complexity of the interventions.

Some studies provided detailed sample size calculations, incorporating specific assumptions about expected outcomes, intracluster correlation coefficients (ICC), and power. For example, Adithya Cattamanchi (2021) calculated a sample size with 89% power to detect a 10% increase in treatment success based on 18 clusters and an ICC of 0.001 (26). Similarly, Alexandra E. Ridout (2023) aimed for nearly 100% power, factoring in within-period intra-cluster correlations and cluster autocorrelation coefficients (27).

Some studies had to account for multiple outcomes in their sample size calculations, making the process more complex. Anne Antonia Cornelia van Tetering (2021) calculated the sample size based on the combined maternal and neonatal mortality ratios, with adjustments for cluster design effects and varying cluster sizes (32).

In contrast, several studies did not provide detailed sample size calculations, either due to the complexity of the design or because the primary focus was on feasibility rather than precise power estimates. For instance, Dorothy Thomas (2022) and Eduard J. Sanders (2021) did not detail their sample size calculations, possibly due to challenges in estimating the required sample size in real-world settings (40, 41).

The number of clusters expected effect size, ICC, and study design were key factors influencing sample size calculations in SWD studies.

The number of clusters and the ICC were critical factors in sample size determination. For example, Ana Pilar Betrán (2018) assumed a baseline frequency of 30% for health practices and calculated the sample size with an ICC of 0.05 (30). Similarly, Andrew F. Auld (2020) used a coefficient of variation of 0.2 and a 24-month enrollment period to estimate the required sample size for detecting a reduction in ART mortality rates (31).

The desired effect size and the power to detect this effect significantly impacted the sample size. For instance, Emma Sacks (2020) aimed to detect a 50% increase in caregivers receiving results, using a design effect of 2 across 18 clusters (45). Geofrey Musinguzi (2020) calculated a sample size to detect changes in cardiovascular disease (CVD) risk factors with adequate power (51).

Calculating sample size for SWD studies is particularly challenging due to several inherent complexities:

The SWD's phased implementation and the need to account for within-cluster and between-cluster variations make sample size calculations more complex than traditional designs. This complexity often requires sophisticated statistical methods and software, as noted in Frauke Rudolf (2021), who used the "steppedwedge" function in Stata for calculations (50).

Estimating the ICC is crucial but challenging, as it directly affects the power and sample size required. An inaccurate ICC estimate can lead to underpowered studies or unnecessarily large sample sizes, as highlighted in Sophie Sarrassat (2021), who assumed a design effect 2 and an ICC of 0.3 for sample size estimation.

In real-world settings, the availability of data, logistical constraints, and variability in cluster sizes can complicate sample size calculations (101). Wilbroad Mutale (2023) had to adjust the sample size during the study due to lower-than-expected mortality rates, illustrating the need for flexibility in sample size planning.

Many studies did not provide detailed sample size calculations, often due to the abovementioned challenges. Sometimes, sample sizes were based on pragmatic considerations rather than formal calculations.

Studies like Dorothy Thomas (2022) and Eduard J. Sanders (2021) likely prioritized feasibility and real-world applicability over precise sample size calculations, focusing on enrolling as many participants as possible within the available resources (40, 41).

Some studies may have based their sample size on previous similar studies or pilot data rather than conducting detailed new calculations. This approach was seen in studies like Lily C. Kumbani, where simulations were used to estimate the sample size needed to ensure 80% power (78).

Many SWD studies focused on health outcomes and managing specific diseases, reflecting the importance of improving clinical practice and patient health in various settings.

A significant number of studies targeted outcomes related to HIV and TB. For example, Adithya Cattamanchi (2021) focused on the proportion of patients with treatment success for TB, while Dorothy Thomas (2022) measured HIV viral suppression among partners living with HIV (26, 40, 41). Similarly, Eduard J. Sanders (2021) aimed to assess the number of new HIV diagnoses, combining chronic and acute cases.

Malaria-related outcomes were also common, as seen in Alfred B Tiono (2014), where the incidence of clinical episodes of malaria among children was a primary outcome (28). Lea Multerer (2021) measured clinical malaria defined as fever plus a positive RDT result for Plasmodium falciparum (76).

Alexandra E. Ridout (2023) evaluated maternal mortality, major maternal morbidity, and fetal losses as a composite primary outcome (27). Anne Cockcroft (2022) assessed male spouses' knowledge and attitudes related to maternal and child health, including the knowledge of danger signs during pregnancy and childbirth (33).

Several studies aimed to assess changes in behavior, psychosocial factors, and knowledge as primary outcomes, often in the context of public health interventions.

Ali M. Giusto (2017) explored the associations between fathers' and sons' sex-related beliefs and behaviors, moderated by parenting characteristics (29). Wit Wichaidit (2019) focused on handwashing behavior among schoolchildren, assessing the availability of handwashing materials and observed handwashing behavior after toilet use (110).

Eve S. Puffer (2016) measured family communication, HIV risk knowledge, and self-efficacy as primary outcomes (48). Stephanie R. Psaki (2022) targeted the increased use of HIV prevention and treatment services, reduction in sexual risk behaviors, and reduction in sexual and gender-based violence as primary outcomes (103).

A number of studies focused on improving clinical and diagnostic processes, with primary outcomes centered around their effectiveness.

Frauke Rudolf (2021) measured the diagnostic yield for smear-positive TB as the primary outcome, with secondary outcomes including treatment success and overall mortality (50). Priya B. Shete (2023) focused on the number of people initiating treatment for microbiologically confirmed TB within 2 weeks of presenting to the health center (92).

Hamish R. Graham (2019, 2021) evaluated child mortality and mortality from acute lower respiratory infections (ALRI) as primary outcomes, reflecting the impact of interventions like oxygen systems in hospitals.

Some studies aimed to assess the impact of interventions on patients' quality of life and satisfaction with healthcare services (57, 58).

Stella-Maria Paddick (2017) used the Brief WHO Quality of Life (WHOQOL-Bref) scale to assess changes in quality of life as the primary outcome. This focus highlights the importance of evaluating not just clinical outcomes but also the broader impacts of healthcare interventions on patient well-being (102).

Osondu Ogbuoji (2019) measured patient satisfaction as the primary outcome, including overall satisfaction and satisfaction with specific domains such as wait time, consultation time, involvement in treatment decisions, and respectful treatment (87).

Some studies focused on the performance of health systems and the effectiveness of service delivery as primary outcomes.

Geoffrey A. Jobson (2021) assessed the proportion of HIV-infected adults screened for TB at ART initiation and the proportion of adults retained in care as primary outcomes (52). Pamela K. Kohler measured youth retention in HIV care, particularly during early engagement, as a key outcome (88).

Janina Isabel Steinert (2021) evaluated total private healthcare expenditures as the primary outcome, assessing the financial impact of healthcare interventions on households (65). Wilbroad Mutale (2023) focused on age-standardized mortality among individuals to measure health system performance.

Several studies employed composite or multifaceted primary outcomes, reflecting the complexity of the interventions and the need to assess multiple dimensions of impact.

Alexandra E. Ridout (2023) used a composite rate of maternal mortality, significant maternal morbidity, and fetal losses as the primary outcome. This approach allowed for a comprehensive assessment of the intervention's impact on maternal health (27).

Margaret W. Gichane (2020) assessed the adoption, acceptability, appropriateness, cost, and fidelity of the intervention as primary outcomes, demonstrating the importance of evaluating multiple aspects of program implementation and effectiveness (80).

The quality of reporting in cluster-randomized controlled trials (RCTs) utilizing the stepped wedge design was evaluated across Eighty-five study reports, using headings and subheadings adapted from the Consolidated Standards of Reporting Trials (CONSORT) (26–110) (see Supplementary File S3). Eighty-four studies specified the method used to allocate clusters to different intervention implementation periods, as indicated in the title and/or abstract (26–93, 95–110). Eighty-one studies provided a diagrammatic representation of the design (26–31, 33–39, 41–48, 50–70, 72–110), which was particularly helpful for understanding the general characteristics of the design and correctly classifying the studies as employing a stepped wedge design. The interventions were adequately described in eighty-four studies (26–78, 80–110).

Seventy-one studies reported sample size calculations (37, 44, 49, 52, 63, 66, 70, 71, 79, 91, 94, 97, 107, 109), though only Fifty-nine considered the design effect, including the intracluster correlation coefficient (26, 27, 29–36, 38, 39, 41–45, 48, 51, 53–62, 64, 65, 67, 68, 70, 72–74, 77–82, 84, 85, 88, 89, 92, 93, 96, 99–101, 103–106, 108, 109). Eighty studies described sequence generation (26–48, 50, 52–70, 72–82, 84–90, 92–110), while Forty-seven reported allocation concealment (27–31, 33, 36–39, 42, 43, 46, 47, 52–60, 62, 65, 69, 73, 74, 77–81, 85, 86, 88, 89, 97–99, 101, 102, 104, 106, 107, 109, 110). Additionally, sixty-one studies documented restricted randomization (27, 29–31, 33–35, 38, 39, 41–43, 45, 48, 50, 51, 54, 56–62, 64, 65, 67, 68, 70, 72–78, 80–85, 87–90, 92, 93, 95, 96, 98–106, 108, 109), and twenty-nine reported employing blinding (56–60, 62, 65, 69, 73, 74, 77–80, 97–99, 101, 102, 104, 106, 107, 109).

The use of intention-to-treat principles was reported in sixty-nine studies (26, 27, 29–31, 33–39, 41–43, 45, 47, 48, 50, 53–65, 67–70, 72–75, 77, 78, 80–90, 92–96, 98–106, 108, 109). A participant flow diagram was included in eighty-one studies (26–31, 33–48, 50–75, 77–93, 95–110), with all studies indicating the number of clusters and individuals included in the analysis (26–110). Baseline data and summaries of primary outcomes were provided in all the studies (26–110). Eighty-two studies included an effect size estimate (26–34, 36–48, 50–70, 72–110), while Eighty-three reported precision estimates (26–48, 50–70, 72–110). Forty-one studies provided information on adverse events (27, 28, 35, 36, 38, 39, 46, 48, 53, 55, 60–63, 66, 67, 69, 70, 74, 75, 78, 81, 84–86, 88, 90, 92, 93, 95–97, 99, 102–105, 107–110). Furthermore, all studies discussed the generalizability of the results and acknowledged their respective limitations (26–110).