CUBES: A practical toolkit to measure enablers and barriers to behavior for effective intervention design

Constructing a best-practice framework of behavior

Grounding intervention design in a comprehensive and actionable behavioral framework is important. Many such models exist, with varying levels of evidence supporting their components. We surveyed models that were a) most influential and b) had an evidence base confirming that they predict behavior, rather than a comprehensive survey of all models in existence. We defined ‘influential’ as having been used to guide behavioral-intervention design across sectors, including health behaviors, and ‘evidence-based’ as the existence of original research evaluating the predictive power of individual components of a model (such as ‘perceived severity of risk’²⁴ or ‘conscientiousness’¹²) on behavior.

We began with a list of models known to the authors, then surveyed key approaches cited in these models, then searched PubMed and Google Scholar for the following terms: ‘behavior’ AND (model OR framework OR drivers OR barriers OR facilitators OR enablers), and focused on the first five pages of results in the search engines. The main search was performed between October and December 2016, with additional targeted searches until December 2018. We identified 17 models fitting the criteria, each focusing on a different set of behavioral drivers (Table 1). The drivers for which we found experimental evidence of moderate to high predictiveness on behavior were then placed into a framework. We subjected the drivers included in the framework to critical review by four experts in behavioral science, health psychology, and the development sector, focusing on the drivers’ comprehensiveness and applicability to global development. Finally, we applied the framework to design research in our own large-scale programs (see sample case study in this paper) to test for actionability and to further refine its components.

Assessing and developing a curated set of research methods to measure drivers

To help programs select appropriate research tools to capture enablers and barriers to behavior, we surveyed methods used across disciplines, using literature research and expert conversations. Unless otherwise mentioned below, we then applied and tested these methods in various combinations in our own large-scale development programs to assess each method’s feasibility, strengths and weaknesses in insight generation. In Zambia and Zimbabwe, we investigated voluntary medical male circumcision^4,5,23, and in different areas of India we conducted programs on household behaviors relating to maternal and child health (family planning, antenatal care, institutional delivery, postnatal care), tuberculosis care-seeking behaviors, and healthcare provider and front-line worker behaviors within medical facilities and communities (unpublished reports).

Cubes: a practical framework of behavior

Following our review of influential models of behavior, we distilled their most evidence-based components into a practical behavioral framework that programs can use to evaluate existing evidence, conduct research to close evidence gaps, and ultimately design interventions to match barriers to behavior (Figure 1). The CUBES framework articulates three critical components of behavior change. First, the path toward a target behavior consists of a series of distinct stages. Second, the progression through each of these stages is influenced by a set of contextual and perceptual drivers. Third, these barriers and enablers may be transmitted to the individual through influencers (such as friends, family, or community members), either directly or through media channels. Below we outline each component and the contributions of the behavioral models surveyed (Figure 1).

Figure 1. The CUBES behavioral framework.

Contextual and perceptual drivers combine to act as enablers and barriers along an individual’s path from knowledge – encompassing awareness and skills – to intention (or motivation to act) and action and beyond. Layers of influencers can affect these drivers and reach an individual through various channels.

Stages of the behavioral change process

Both contextual and perceptual drivers influence whether individuals possess the knowledge needed for behavior change, intend to act, or are already acting. These drivers either hinder or facilitate progression along the stages of change.

In global development, many intervention programs focus on enhancing awareness (passive knowledge) and skills (active knowledge). However, it is possible to be aware of an option, and even have the skills to do something, without intending to take advantage of it. For example, in Uttar Pradesh, India, increasing nurses’ skills did not always increase their practices accordingly²⁵. Clearly, knowledge does not equal action. Therefore, understanding where people are on the pathway to behavior, and why they are not moving forward, is key to designing interventions that can move people toward action. This is an essential first step for program designers to orient themselves when evaluating behavior.

The Transtheoretical Model¹³ proposed a series of relatively rigid stages: precontemplation (being unaware), contemplation, preparation, action, maintenance, and termination, and adds a set of interventions (‘processes of change’) to help an individual progress from one stage to the next. However, the efficacy of assigning individuals to very detailed stages has been called into question in several systematic reviews^26–28. In line with the Health Action Process Approach¹⁶, the CUBES framework therefore divides the behavioral change process more simply into three stages of knowledge (ending with the necessary awareness or skills to engage in a behavior), intention, and action. While the path to action and beyond can be understood as a sequence, people can also move back and forth between already-reached stages.

Almost no models of behavior focus on repeat behaviors and habits, despite the importance of sustained change in real-life contexts. Different drivers become more or less important for repetition and habit. The intention to repeat a behavior becomes more likely when two factors converge: a positive evaluation of the previous experience (‘experienced utility’), and a revised self-efficacy in comparison to what was expected^16,29. However, merely repeating a behavior does not create habits. The creation of a habit requires the development of automaticity (the behavior is performed with low awareness, control, attention, or intention) and an association of a behavioral response with contextual cues and an experience of reward^30–32. Some behavioral drivers that can be targeted for single actions, such as intentions, goals, or beliefs, are much less important to the formation of a habit^16,30,32. For example, delivering a child in a health facility is a one-time behavior, for which intention, goals, and beliefs matter greatly, but exclusive breastfeeding after the child is born is close to a habit, which after initiation does not require a woman to form the intention from scratch every single time. Instead of targeting beliefs and intentions, then, restructuring of environmental cues and conscious inhibition of unwanted habits may have greater success in creating lasting habits^33,34.

Contextual drivers of behavior

Behavior emerges out of a complex system of interactions between individuals and the systems they act in. The Social-Ecological Model introduces the concept of ecosystems, which examines the dynamic ways that different layers of the social sphere influence each other^8,35,36. Adapting the Social-Ecological Model^8,35,36, social norms and customs influence individuals in an ecosystem in several layers, and the central tenet of a norm is that deviating from it invites judgment or other negative consequences. These norms may be explicit, but they can also emerge from an implicit layer of what Practice Theory defines as ‘shared cultural knowledge’ that expresses itself as routines and habits^7,37.

Structural factors are further contextual drivers that may shape and constrain perception and action. These aspects are usually not fleshed out in behavioral models. However, unless these constraints are removed, interventions acting on perceptual drivers will not allow for the target behavior to occur. We propose differentiating between infrastructure, policies and laws, and systems and processes, all of which vary strongly with their respective context. For example, infrastructure drivers may include availability and condition of roads leading to a health facility, equipment to perform a test, or seeds for farmers to use. Policies and laws constraining behavior can exist on several levels, from national laws to facility-level guidelines. Systems and processes could mean supervision, training, feedback, or incentive systems under which healthcare providers operate in a facility, or the tools teachers have available to plan lessons and receive feedback on their tuition. Finally, demographic factors such as age or education level act as contextual constraints on behavior.

Perceptual drivers of behavior

Multiple models recognize that behavior is also shaped by perceptual drivers that, together with contextual drivers, combine to make a target behavior more or less likely.

Biases and heuristics. Most aspects of an individual’s cognition and behavior are influenced by ‘automatic’, often unconscious, mental shortcuts or rules of thumb (heuristics) that can bias decision-making. Many biases can be difficult to change, but knowing about them makes it possible to construct environments where the best decisions are the easiest. A typical example is setting the default option of a pension program to ‘opt-out’ instead of ‘opt-in’: staying in a pension program after auto-enrollment is easier than making the effort to join in the first place^38,39. Examples of heuristics and biases are the optimism, confirmation, and availability biases, anchoring-and-adjustment, hyperbolic discounting heuristics, and the status quo and representative biases^{17,18,38–44}. MINDSPACE, an influential checklist for behavior change developed in part by UK government agencies^20,45, is an example of using biases (such as a bias to choose the default option) as tools to design interventions.

Beliefs. Beliefs are formed by learned experience, differentiating them from biases (which humans share to varying degrees as a result of how our brains evolved). In the Integrative Model of Behavioral Prediction/Reasoned Action Approach, beliefs about what outcome can be expected from a behavior (called ‘attitudes’ in that model), normative beliefs (how others will judge a behavior), and beliefs about the extent of one’s control over the behavior (self-efficacy) all influence intention, which is seen as the main driver of behavior^19,46,47. Experiments have shown that some beliefs predict behavior better than others. For example, perceived control (self-efficacy) and beliefs about a behavior’s outcome are better predictors of behavior than normative beliefs^48,49. Indeed, self-efficacy beliefs have emerged as a strong predictor of behavior across other models, such as Social Cognitive Theory, the Health Belief Model, and the Transtheoretical Model. The strong influence of self-efficacy on behavior has been shown experimentally in many studies^50–54.

Outcome expectations are another example of a belief. An individual appraises a potential behavior by weighing perceived costs and perceived benefits, which may be emotional (‘How will the outcome of the behavior make me feel?’), social (‘How will others judge me?’), or functional (‘How does this help or hurt me?’). Outcome expectations, together with the perceived severity of an outcome, the perceived susceptibility to that risk, and self-efficacy, are central to one of the most widely-used models of behavior, the Health Belief Model^14,24,55. Evidence from several systematic reviews shows that increasing perceived benefits and decreasing perceived costs to a behavior will be most likely to cause an individual to engage in the target behavior^10,24,56,57.

Emotion. The experience of emotion (affect) also drives behavior. Affect arguably colors all perception and powerfully shapes decision-making²⁰. While affect is used as a distinct driver of behavior in the MINDSPACE checklist²⁰, the COM-B framework²² includes it as a rapid, automatic component of motivation. One example of targeting affect to drive motivation to engage in a new behavior can be seen in an intervention promoting soap use in Ghana: education around the benefits of soap did little to drive up its popularity, but emphasizing the feeling of disgust from ‘dirty hands’ resulted in significantly increased soap use⁵⁸. This emphasis on automatic emotional responses, such as disgust or a desire to conform with others, is a key component of the Evo-Eco behavioral model⁵⁹.

Personality. Personality traits are not often included in models of behavior—not even in the Theoretical Domains Framework, which includes the perhaps most comprehensive list of barriers and facilitators⁶⁰ —but they can strongly influence an individual’s propensity to engage in and maintain health behaviors^61,62. Current applications of personality models to behavior prediction focus on aggregates of behavior, or behavioral patterns of behavior such as going to check-ups and regular physical activity (preventive health behaviors). Such models often do not attempt to predict single instances of behavior, which has been shown to be much less reliable⁶¹. Currently, the dominant personality model with a large evidence base behind it is the so-called Five-Factor Model^12,62–64 with the five broad traits of openness to experience, conscientiousness, extraversion, agreeableness, and neuroticism (‘OCEAN’). Conscientiousness appears to be an especially strong predictor of behavior patterns, such as sticking to preventive health routines^61,62.

Influencers and channels

Following from the Social-Ecological Model^8,35,36, influencers surround an individual in layers. Family and friends or peer groups are the closest layer to the individual. The next layer usually consists of relationships in the community, for example in workplaces, schools, and neighborhoods. The most distant layer is the larger social context. Influencers can reach individuals either directly or at scale via various media channels, which is important information to determine how to deliver interventions. For example, female self-help groups can serve as a channel for rural women in India to reinforce or change social norms relating to a certain target behavior. For an intervention to work, the content, the type of influencer, and the channel through which they reach individuals must be identified as relevant to the target individuals.

Interaction between the ‘building blocks’ of CUBES

Depending on the behavior, all the drivers mentioned above will be relevant to varying degrees to any one individual, and their combination will result in a larger or smaller tendency to act. The elements of CUBES influence each other deeply. For instance, punitive supervision (systems and processes) by medical-officers-in-charge (influencers) in a health facility might lead nurses on the receiving end to experience high anxiety (emotions), and to beliefs that trying hard will not result in any benefit to them (outcome expectation belief). It is important for programs to disentangle these components, even when they influence each other, because this determines what kind of intervention is best placed at what level.

Therefore, the enablers and barriers in the different ‘building blocks’ of CUBES can be seen as a checklist for programs that they can utilize to design effective interventions.

Models of behavior that simplify to the point of ‘motivation’ or ‘ability’ (such as COM-B or the Fogg model) are simpler, but also less actionable for programs on the ground. For example, if a study finds a lack of motivation to go to the doctor despite having symptoms indicative of tuberculosis, this alone is not actionable by an intervention. Instead, programs need to know where potential patients are on the knowledge–intention spectrum (they have clearly not yet taken action), whether they perceive their symptoms and the disease as a danger to their and others’ health (risk perceptions), whether they think going for a check-up will actually alleviate symptoms (outcome expectations), whether they feel able to skip work and other responsibilities to attend an appointment (self-efficacy), whether appropriate facilities are even available, affordable, and accessible (structural factors), and whether there is stigma involved in seeking care (social norms). All these components would feed into the concepts of ‘motivation’ or ‘ability’, but require very different interventions to effect change.

Once CUBES has been used to understand and categorize existing evidence, evidence gaps can be closed in a focused way with primary research. Not all components of CUBES are best captured and intervened on in the same way. In the following sections, we introduce a method mix designed to identify a comprehensive set of drivers.

Choosing the right type of method for different stages of research

Research methods can be divided into descriptive, experimental, and simulated approaches; the last two are relatively under-used in global development. Whether qualitative or quantitative, descriptive methods such as interviews or observation aim to describe and explain behavior without testing the effects of manipulating variables systematically. While they can explore ‘how’ and ‘why’ questions (and are therefore often called ‘exploratory’), they do not have the ability to systematically relate the effect of change to outcomes (‘confirmatory’). Experimental approaches can be used to test hypotheses and find causal relationships by systematically varying variables and testing their effects on outcomes. However, they lack the ability to survey a broad spectrum of factors at once that exploratory methods provide. Finally, simulated methods can assess cause–effect relationships in a virtual environment when experimental field methods are not possible or are too complex, but they rely on many assumptions to construct the simulation.

Expanding the descriptive toolkit

Journey mapping. Developed and primarily used in market research^65,66, journey mapping systematically tracks people’s experiences and interactions with a product, service, or life event over time, as people form beliefs about the product or event and make decisions, perhaps via influencers, to interact with or avoid it. This method is especially well-suited to get a sense of stages of change. Journey mapping can also help form hypotheses about segments of customers who share distinct characteristics in order to target them with bespoke messages via different channels. In addition, it can be useful in generating hypotheses about underlying behavioral drivers that can then be tested further.

Journey mapping uses many different techniques to collect data, including one-on-one qualitative interviews, focus groups, ethnography, web analytics, customer reports via apps, and (qualitative) network mapping^65,67. Below, we show in a case study how journey mapping was successfully integrated in a program understanding decisions around voluntary medical male circumcision⁵.

Observation. Observation is a versatile and routinely used tool in global development to collect data about what people do, the context that surrounds them, and how they interact with processes, objects, or each other⁶⁸. The spectrum of observation techniques ranges from researchers interacting closely with communities (participant observation), to covert or overt ‘natural observation’ without participation, to controlled observation, where procedures are highly standardized. Measuring time and frequency of practices of medical residents in hospitals⁶⁹, or of nurse practices in hospitals in India²⁵, are typical examples of controlled observation. Such time-and-motion studies observe the time taken and actions of participants executing distinct components of a process. Observation is also a tool to measure contextual drivers such as infrastructure or processes. For example, facility and infrastructure audits commonly combine observations with interviews of key stakeholders to track characteristics such as hospital staff coverage, equipment availability, or communication tools⁷⁰.

To get a holistic sense of contextual drivers, observation can track more than behaviors or supplies. Instead, immersive observations in the participants’ natural environment could be structured to assess the set of contextual enablers and barriers outlined in the CUBES framework. We call this approach ‘structured immersive observation’. For example, in a recent observational study on nurses in healthcare facilities, we measured key contextual dimensions as follows. First, we assessed facility infrastructure available to nurses: whether equipment and drugs required for routine tests were available at the time of testing, staff coverage throughout the period of observation, availability of beds, water, and electricity, and transport options for patients to be referred. Second, we assessed systems and practices, such as interactions with and feedback from other staff, time spent with patients and tests performed, documentation systems and job aids, and training records. Third, we assessed community norms by observing community interactions and communications with the nurse.

Enhanced quantitative surveys. Quantitative surveys are a critical tool to obtain insights on many enablers and barriers to behavior simultaneously and at scale, and consequently are a mainstay of global development research. Survey design is a broad field with a large array of approaches. Here, we recommend three key techniques that can enhance the design of quantitative surveys to measure potential behavioral drivers. Survey questions can be structured to account for as many components of CUBES as possible. In sensitive contexts, surveys can also be enhanced to counteract respondent biases. Finally, while not within the scope of this article, programs would benefit from leveraging quantitative data for insight beyond descriptive analyses: for example, population segments can be found for targeted intervention design⁵.

Driver-structured surveys. Programs can use the CUBES framework as a checklist to assess whether a survey captures the range of potential contextual and perceptual enablers and barriers, influencers and channels, and stages of change relating to behaviors of interest. We often find surveys only measure a narrow set of drivers, which presents the opportunity to generate a more holistic view of behaviors of interest and the system. This can also be enhanced by adapting existing validated tools, such as scales testing personality or self-efficacy (see below).

Standardized scales. A simple and high-yield modification to quantitative surveys is the inclusion of previously validated and standardized rating scales relating to CUBES perceptual drivers. Standardized scales that test specific cognitive processes include the Risk Propensity Scale⁷¹ and the ten-item General Self-Efficacy Scale⁷². Types of emotions and their felt strength have also been widely measured with graphical rating scales, such as the Self-Assessment Manikin⁷³. All these scales can be flexibly adapted to specific contexts. For example, in a study investigating women’s propensity to engage in breast cancer prevention, self-efficacy was asked in two items: ‘the extent to which participants were confident that they could conduct breast self-exams every month; and when they conducted a breast self-exam, how confident they were in their ability to identify a “lump that needs medical attention.”’⁷⁴.

Personality tests, widely used in the private sector, also use standardized rating scales. Many studies show some predictive value of the OCEAN model’s ‘Big Five’ personality traits on health behaviors⁶¹, especially conscientiousness⁶². Questionnaire designers can tap a large number of validated instruments to test OCEAN components, such as the public-domain International Personality Item Pool⁷⁵.

All standardized scales have the advantage of using previously validated instruments, and that individual differences can be captured with relatively little effort. A limitation is that, like all self-reports, such tests are susceptible to reporting bias, since participants can deduce or guess socially desired responses (those that the respondent thinks will make them appear in a favorable light). Responses may therefore be compatible with the participants’ sense of self rather than their actual behavior.

Informal confidential voting interviews and polling-booth surveys. Methods that stress anonymity and confidentiality, such as polling-booth surveys (PBS) and the Informal Confidential Voting Interview (ICVI) approach, can be used to probe sensitive topics, as they counteract social desirability bias. The ICVI consists of a one-on-one interview followed by self-completion methods⁷⁶. Similarly, PBS collects feedback from a group of people who respond anonymously and independently through a ballot box. Comparison of one-on-one interviews with PBS⁷⁷ and with ICVI⁷⁸ on sexual risk behavior in Indian men and women demonstrated their value, as more risky behaviors were reported with each method.

Standardized patients. Standardized patients (SPs) are people trained to play the role of a patient with certain medical and personal characteristics, who interact with healthcare providers in a realistic setting. SPs are comparable to ‘mystery shoppers’ in consumer research⁷⁹, in that the healthcare professional does not know the patient is not real⁸⁰. In other scenarios, both parties know about the setup⁸¹. The SP method has been applied to investigate healthcare provider behavior in many contexts, such as prescription practices of pharmacists⁷⁹. It has also been used to assess how doctors communicate with their patients, such as how surgeons disclose medical errors to patients⁸¹. The SP approach can compare expected with actual behaviors, and analyze communication, such as the content of advice given to patients (80). While the method on its own is very well suited to capture practices and contextual drivers such as infrastructure and processes (of patient interaction), other drivers, such as beliefs or biases, are less accessible to investigation.

Social network analysis. Social network analysis (SNA) maps relationships between people, organisms, groups, or devices⁸². When analyzing behavior, SNA can be an excellent tool to describe which influencers and which channels are most important to transmit certain norms and information. SNA can be both qualitative or quantitative. Data can be generated from surveys, ethnography, or observation, or mined from existing resources such as GPS coordinates or twitter messages^83,84. Many field studies have used SNA to focus on where in the system to intervene. For example, in Uganda, researchers mapped the process of obtaining a diagnosis for tuberculosis through provider and patient networks, and the steps where delays were most common could be identified⁸⁵. Ultimately, SNA is a flexible and versatile method, but specifically focuses on identifying centers of influence.

Leveraging ‘passive’ datasets. As in the SNA example, insight can also be generated from leveraging ‘passive’ datasets, generated for a different original purpose, without direct interaction with or observation of respondents. Examples are information obtained from GPS, satellites, and sensor systems, as well as other databases. To investigate contextual drivers, satellite images can map physical conditions of the built environment, which can then be related to behaviors and drivers from other datasets⁸⁶. For some audiences, social media data can be an appropriate source of aggregate estimates of positive or negative sentiment⁸⁷. More analog data sources can also help generate insight, as in an analysis of Kenyan newspaper articles about voluntary medical male circumcision, which provided insight on the types of risks that were presented to readers⁸⁸.

Assessing decision drivers through ‘in vitro’ experiments

We propose that programs can benefit from ‘in vitro’ experimental methods before testing specific interventions in lengthy trials in the field. Experimental methods that track the decisions participants make in laboratory-like conditions serve several purposes: programs can systematically change and test enablers and barriers to behavior, predict behaviors in response to specific interventions, determine those features of a service or product that are most likely to align with the customers, and forecast the market size of a product or features based on predicted behaviors. All these factors narrow down the potential characteristics of an intervention to be tested in the field, and ultimately make the design of effective interventions more probable.

Discrete choice experiments. Discrete choice experiments (DCE), extensively used in market research, uncover preferences and value attribution from the choices that participants make, rather than from the participant disclosing them. They are a powerful tool to predict behaviors ‘in vitro’, assess which features of a product or message are most important to the customer, and to forecast market shares of products. DCE have been shown to be predictive of health behaviors⁸⁹. For feature selection, participants are typically shown multiple iterations of sets of products with varying features. In each trial, the participant picks one option. For example, a discrete choice experiment in South Africa evaluated which characteristics of HIV prevention products, such as the method of use, or the protection against diseases other than HIV, would be most valued by participants⁹⁰. From participant choices, a model can be built showing which level and which combinations of a product’s features predominantly drive decisions, where the tipping points of certain preferences lie, and forecasting product market share. DCEs can also be used to test various ‘what-if’ scenarios, and results can then be used as a funnel to select the most promising attributes for a field intervention. Simulated test marketing is a related concept, in which the consumer is asked to make choices in a realistic environment, with similar systematic manipulation of test variables.

Decision games. Purely quantitative DCE approaches are mostly used to cycle through permutations of features and types of products or interventions. However, a related ‘decision game’ method mixing quantitative with qualitative elements can help investigate which behavioral drivers most influence choices made by participants. In a recent study with healthcare providers in Uttar Pradesh, India, we used such a group ‘decision game’: participants were given a set of scenarios, each with a set of response options, and were asked to choose the option they thought other participants would select (unpublished reports). Response options coded for different behavioral drivers, and participants were later qualitatively probed on their choices. For example, nurses were asked which nurse in a scenario was likely to be most stressed: the one working in an understaffed facility (coding for infrastructure-staffing), the one dealing with demands from patient family members (influencers-community), or the one facing constant scrutiny and accountability from supervisors (systems and processes – supervision).

Implicit attitude tests. Few research methods are suitable to measure enablers and barriers that respondents cannot or do not want to report. Over the last two decades, experimental psychology has developed a battery of experimental approaches to measure ‘implicit’ biases, or biases that are inaccessible to conscious awareness and self-report, but nevertheless influence behavior. The underlying concept of implicit attitude tests is that our brains perform unconscious evaluations of concepts, people, and objects, which have arisen from past experiences and cannot be measured by explicit questioning.

The most widely-used implicit attitude test is the Implicit Association Test (IAT). This test is based on the concept that participants can perform a task more quickly when they see two concepts as related than when they do not associate them with each other⁹¹. IATs have been used to measure a plethora of social stereotypes, such as gender and racial biasesⁱⁱ, for instance in the Democratic Republic of Congo⁹². In market research, the IAT has been used to gauge consumer attitudes toward different products⁹³. The predictive value of implicit attitude tests on behavior is still under debate^94,95. For this reason, and because IATs can only test a small set of associations within a test that requires training participants, we only recommend implicit tests when a behavior is likely to be influenced by a specific, deep-seated bias that respondents are unlikely to report.

Using simulations to model ‘what-if’ scenarios

Simulations have the unique advantage that complex ‘what-if’ scenarios can be explored at the push of a button, which can be used to supplement and inform data collection or to optimize interventions. Through the construction of ‘virtual worlds’, mathematical models can simulate the impact of implementing certain interventions, or of targeting interventions to specific sub-groups. They can also generate hypotheses on what a likely driver for behavior might be. As an example, agent-based models have been used to simulate the large-scale effects that emerge from the actions of many single agents, such as the spread of disease or of social norms and beliefs^96,97. Similarly, Bayesian cognitive mapping builds probabilistic models of the likelihood that agents make certain decisions⁹⁸. However, simulations are only as good as the model and assumptions that underlie them. The relevant and correct starting parameters must be chosen with caution, which includes a degree of subjectivity⁹⁹, and generalizations from a model based on specific assumptions are limited. Table 2 summarizes the approaches discussed, and their strengths and weaknesses.

Choosing the right method at the right time for the right purpose

The goal of any program is to implement successful interventions in the field. Figure 2 depicts the process of setting the research agenda, generating insights, and designing and optimizing interventions that programs can use, depending on the knowledge level at the start of the research process. First, when programs define a target behavior to be changed, they can evaluate existing evidence against the components of the CUBES framework. This can be done either from existing literature or analyzed from datasets, or both. To directly choose an intervention that works, programs must already have narrowed down specific drivers to intervene on. This may be the case in a data-rich environment; in other cases, political or resource constraints limit what is testable in the field. In such cases, primary research may not be required or appropriate. On the other end of the spectrum, a program might know what it wants to change—for example, to increase the uptake of modern methods of contraception—but have little systematic knowledge of the types of drivers that may be involved. In this case, exploratory research is warranted (‘insight generation’ in Figure 2). We have found that a quantitative survey often provides the best practical balance between assessing many components of CUBES at scale. Either before a survey (to inform its design) and/or after (to dive deeper into specific findings), specific descriptive or experimental methods offer particular strengths assessing specific CUBES components and can supplement a survey (Table 2).

Figure 2. Decision aid for choosing the right research approach at the right time, for the right purpose.

Descriptive methods can be qualitative or quantitative, and many can take both forms. To choose a qualitative or quantitative focus, programs can consider whether the freedom to explore limited aspects in depth is most important (which means a greater focus on qualitative research), or representativeness at scale and the relative likely impact of each driver (which points to quantitative methods). Often, programs use preliminary qualitative research to the inform the design of quantitative research, especially in field settings where quantitative research is conducted in person and therefore is expensive and time-consuming, whereas small-sample qualitative research is less resource-intensive. The freer structure of qualitative research also typically allows for follow-up questions and clarification. However, we recommend that this order not be followed by default, but rather examined on a case-by-case basis. In our experience, qualitative research can sometimes divert resources from quantitative research, which can be wasteful if results from a smaller sample cannot be generalized. Instead, qualitative and quantitative research can also be run in parallel, with a complementary focus on different drivers; pockets of qualitative components can be mixed into qualitative research; or qualitative back-checks can be conducted after quantitative research.

In global development, most programs tend to focus on descriptive methods for insight generation, followed by field implementation, where randomized controlled trials assess the effectiveness of an intervention. In addition to preliminary evidence synthesis and using a flexible methods toolkit for specific deep-dives, we argue that this approach misses a key step, namely ‘in vitro’ experimental methods. These methods can narrow down the many potential hypotheses emerging from exploratory research, so that only the enablers and barriers likely to be most impactful are ultimately tested in the field. This optimization can be used to choose between different types of interventions (such as monetary incentives versus more information on risks and benefits), as well as the components of a specific interventions (such as the magnitude of incentives likely to be most effective). If rich descriptive data is already available, and so a limited set of specific hypotheses around limited drivers can be formed from the outset, programs can directly skip to this step. This step can also be done in parallel with exploratory research, if evidence is strong on specific drivers but weak on others and a holistic picture is desired. As detailed above, purely simulation-based methods can be of use here to model the effect of many different changes. As a result of this step, field testing will be based on much stronger evidence.

Designing interventions to increase the uptake of voluntary medical male circumcision (VMMC)

In Figure 3, we briefly outline the methodological choices made to investigate enablers and barriers to uptake of voluntary medical male circumcision (VMMC) in Zambia and Zimbabwe. VMMC is a highly cost-effective intervention for preventing HIV acquisition that is being scaled up in eastern and southern Africa^5,100. The achievement of the program’s ambitious targets necessitated shifting the behavior of many men in the community who either did not consider circumcision, or if they did, did not take action. Therefore, the program needed to understand the multiple interacting factors that facilitate or inhibit men’s decision to get circumcised, and to test interventions to address those factors. A synthesis of previous studies revealed a variety of existing insights on many behavioral drivers, such as concerns around pain, complications, or cost, as well as patterns of influencers¹⁰¹. Analysis of a large quantitative survey further showed a relatively small awareness-intention gap, but a large drop from intention to action¹⁰² A total of 64% of men intended to get circumcised, but only 11% did. However, a holistic view assessing the prevalence of each of these, and other, drivers and their relative strength was lacking, and most research was either small-scale or qualitative⁵. The existing studies could not answer why there was a strong intention-to-action gap. Also, the research did not examine or reveal the heterogeneity among men—the fact that a given enabler or barrier may be important to one man, while not as relevant to another.

Figure 3. Process of evidence evaluation, insight generation, and intervention design and optimization in a VMMC program.

A broad set of CUBES drivers therefore needed to be captured at scale to assess the relative importance of each driver in a single dataset. However, as the stages of change appeared to be of primary importance, journey mapping and qualitative decision games were first used to understand the stages of change and associated beliefs and influencers at each stage in more detail⁵. In summary, qualitative research pointed to a much more nuanced picture. Men develop positive as well as negative beliefs, influenced by individuals around them, as they move through the various stages of change. These competing beliefs and associated emotions move men towards or away from the decision of getting circumcised in distinct stages of change. For example, beliefs in the early stages include “VMMC protects myself and my partner from STIs” (positive) and “the procedure is painful”. As men move to consider undergoing the procedure, there emerges a strong conflict between the positive beliefs and negative ones, such as circumcision threatening self-identity, leading to distrust between the man and his female partner, and the perceived long healing time. The conflict between emotions such as shame, distrust, and fear and perceived potential benefits move men into a state of cognitive dissonance and stall them from taking-action. However, it was also clear that not all men held all beliefs equally strongly, or were equally subject to the same type and strength of influencers. These insights informed the design of a large-scale quantitative survey investigating CUBES components comprehensively. Having this dataset available then allowed us to segment men on the enablers and barriers to VMMC, so that messaging interventions could be designed targeting the most important drivers for each segment. For instance, among the six segments found in Zimbabwe, Embarrassed Rejecters had mostly negative beliefs about VMMC, as well as fears and concerns regarding the procedure, and had little social support⁵. However, they did not lack the knowledge that strongly characterized another segment, Neophytes. Accordingly, segment-specific interventions, including messaging through front-line workers or media campaigns, could be developed. In addition, the roll-out of circumcision devices as an intervention to improve the uptake of circumcision was an important consideration for programs. However, it was unclear what the potential demand and market share of these devices would be. The demand for different devices to carry out the VMMC procedure was forecast using simulated test marketing, a technique related to discrete choice experiments, so that the right devices could be marketed with the right message to the right people¹⁰⁰. Interventions designed based on this research are currently being piloted at national scale in Zambia and Zimbabwe.