Introduction

One premise in health promotion and disease prevention is that exposure to illness and injury is not predetermined; exposure can be minimized through people’s actions. Furthermore, once people are exposed to illness or injury, their prognosis, quality of life, and life span can be shaped by their actions. In this chapter, health campaigns are broadly defined as communication strategies intentionally designed to encourage people to engage in the actions that prevent illness and injury and promote wellbeing. Health campaigns may include face-to-face or mediated conversations as well as verbal and nonverbal messages. The campaigns may focus on encouraging the behavior directly, such as compliance messages attempting to encourage parents to have their children vaccinated, or indirectly, such as raising public awareness about the emerging health threat of antibiotic resistant bacterial infections (which can be slowed through vaccination). Campaigns often have explicit goals for outcomes, such as achieving 95% coverage of complete vaccination or increasing the length of cancer survivorship. The population of people associated with the target outcome may be small (e.g., the neighborhood experiencing an outbreak) or large (e.g., a continent or the globe). Regardless of the size, no two people are exactly alike with respect to their risk for illness and injury or their reactions to a campaign attempting to lower their risk.

These variations between people are important for health messaging. Some people may need to keep doing what they are doing, some need to start doing something new, and others may need to stop doing something harmful. To achieve those aims, people may need to form new beliefs to motivate them to action, to change a contrary belief, or to strengthen an existing, motivating belief. Different persuasive strategies may be needed to achieve maintenance, behavior initiation, and cessation; different theories of persuasion are more effective at forming, changing, or strengthening beliefs. To further complicate the situation, different people, based on their personalities, may respond to the same messages in different ways. Effective campaigns provide a target audience with the right persuasive strategy to inspire change based on their initial state and psychosocial predictors for change.

It is often financially and logistically unreasonable to create campaigns for each individual within a population. It may not even be necessary when people exist in similar states and share psychosocial predictors for change. On the other hand, a meta-analysis (Noar, Benac, & Harris, 2007) shows that designing one, general message is less effective in gaining behavior change than crafting messages targeting groups of people or tailoring them to a particular individual. To be clear, targeting and tailoring both refer to using information about an audience to craft messages; however, they focus on different audiences: targeting focuses on groups, while tailoring focuses on individuals (Hawkins, Kreuter, Resnicow, Fishbein, & Dijkstra, 2008).

Targeted messages are more likely to be read, understood, recalled, perceived as credible, and rated highly than one, generalized communication (e.g., Kreuter, Farrell, Olevitch, & Brennan, 2000; Noar et al., 2007). A challenging problem for health campaigns is to define those who need to be reached, and then intelligently group people based on a complex set of variables in order to identify groups with similar needs who will respond similarly to a particular persuasive strategy (Slater, 1995; Slater, Kelly, & Thackeray, 2006). The purpose of segmenting a population into audiences is to make communication efforts more effective and efficient (Slater, 1996). It is also a fundamental mechanism of audience-oriented health communication (Maibach, Maxfield, Ladin, & Slater, 1996).

Research is needed to make informed decisions. Without research, campaign designers must make assumptions about subgroups within a population. For example, imagine that a state health department has the goal of immunizing 95% of children aged 5 and younger who reside in their jurisdiction. The designers may assume that parents who are not getting their children immunized fall into three groups: those who do not know the risks of not immunizing, those with religious objections to immunization, and those who fear the unsubstantiated side effects of vaccines. With research, the designers may discover that parents of unvaccinated children may not be complying due to lack of resources (time or money), which is not addressed in the previous groups. Furthermore, parents who are uninformed about the risks of not immunizing may be the same parents with the mistaken beliefs about the side effects of vaccination; this may be one, not two, groups. The premise of this chapter is that segmentation at its best is a systematic and explicit process of research (Slater, 1996) to make informed decisions about how many audiences to consider, why the audience is doing what they are doing, and how to reach that audience effectively.

The goal of this chapter is to provide an overview of the history, concepts, and methods of audience segmentation as a research process. While personalizing communication involves both segmentation and customization (Hawkins et al., 2008), this chapter does not review how to design customized persuasive messages targeting a group or tailored to a person. Segmentation is a necessary but not sufficient condition for campaign success: decisions about message design and implementation are also critical (Slater, 1995). This chapter focuses on a primary decision for communication design: definition of the target audience (Slater, 1996).

Segmentation: Definition and History

Segmentation is a systematic and explicit process of identifying groups of people within a larger population (who need to be reached) who exist in similar states regarding the campaign’s goals and share psychosocial predictors for change (Maibach et al., 1996; Slater, 1995, 1996). The process involves finding people who are similar in their patterns of responses on a set of variables that predict an outcome aligned with a campaign’s goals (Slater, 1996).

The activity of identifying and grouping objects based on shared characteristics is not new. In ancient Greece, early biologists grouped organisms into species (i.e., taxonomy, Slater, 1996). In the study of public opinion, Dewey (1927) introduced the notions of publics as subgroups who shared similar values or interests about a given issue. In 1956, Wendell Smith introduced the concept of market segmentation in marketing research. Smith argued for developing new products for groups of consumers with similar needs and desires. To increase market share by product differentiation, Smith argued that marketers should categorize diverse consumer populations into homogenous, mutually exclusive subgroups based on shared qualities, such as product needs, and then design campaigns to best match a particular segment with the goal of influencing their purchasing behavior. This segmentation process was adopted by marketers to make advertisements more effective and efficient and, ultimately, more cost-effective for producing sales (Hines, Reser, Morrison, Phillips, Nunn, & Cooksey, 2014; Slater, 1996). In the digital age, vast amounts of behavioral data are available on consumers, from a record of previous purchases to online profiles on social media sites. Marketing firms work with Google, Facebook, and mobile phone companies, for example, to segment audiences based on online profiles, social media activities, and search activities in order to provide targeted and even tailored advertisements (Hines et al., 2014).

Health campaign researchers in various disciplines of social marketing and health communication adopted the practice of audience segmentation (Hines et al., 2014). Articles, outside of those appearing in marketing journals, appeared as early as 1975 on market segments related to media consumption (e.g., Fletcher, 1975). In 1985, Meyers and Seibold explored demographic and psychological predictors of people’s utilization of health services. In 1992, an article appeared in the journal Health Communication; the article (Hofstetter, Schultze, & Mulvihill, 1992) explored how research into grouping people based on their patterns of media exposure, and cognitive predictors of their media exposure grouping, could be used to understand exposure to health information in a multimedia environment. The first issue of the first volume of the Journal of Health Communication included an article on segmentation modeling in marketing, and its utility for health campaigns (Albrecht & Bryant, 1996), to be followed by Slater’s (1996) classic review of theory and methods in issue 4 of that same first volume.

Audience segmentation has a different purpose for health campaigns than it has for marketing. The goal of health campaigns is to influence outcomes that benefit the at-risk population and society at large (Hines et al., 2014). The segmentation solutions provide decision-makers (e.g., health agencies, nonprofit organizations, and government agencies) an empirical basis for strategic planning: making decisions about how to allocate often limited resources (human and financial) to achieve optimal outcomes (Dibb, 1999; Maibach, Leiserowitz, Roser-Renouf, & Mertz, 2011). Indeed, segmentation research may be even more important for health campaigns than for marketing, because health behaviors may be more resistant to persuasive messages than consumer purchasing behavior (Maibach et al., 1996).

Segmentation Analysis

Audience segmentation involves dividing a larger population that need to be reached into groups that exhibit similar patterns of responses across a complex set of variables, and differences from other groups on those same variables (Grunig, 1989; Slater, 1996). The aim of segmentation is to minimize within-group differences and to maximize between-group differences. After groups are selected, each group is analyzed based on their values on the variables. The assumption is that subgroups may need different strategies to achieve beneficial outcomes (Slater & Flora, 1991; Storey, Saffitz, & Rimon, 2008). Best practices for audience-influenced design involve creating messages in response to a group’s profile and then pilot-testing those messages with members of the target audience before dissemination occurs (Slater & Flora, 1991).

Two common approaches for defining groups are demographic separation and empirical clustering. Demographic separation involves dividing a larger population into groups based on shared demographic characteristics, such as race, gender, ethnicity, income, age, education, and economic status (Slater, 1995). It is not uncommon for demographic separation to be conducted with cross-tabulated tables of demographic characteristic and outcomes (Hines et al., 2014; Slater, 1995, 1996). Grouping by demographics is useful for campaign design to the degree to which the demographics are correlated with the campaign’s goal outcomes or determinants of those outcomes. For example, demographics may represent or create shared life circumstances that lead to shared motivations to enact or constraints on enacting a health behavior. Demographic variables, in this example, are distal predictors related to the health behavior through more proximal predictors of behavior (e.g., motivations and constraints). One advantage to demographic variables is that they are readily available through large, national, epidemiological surveys. The disadvantage is that proximal variables are better predictors than distal ones. Furthermore, some proximal predictors of behavior transcend demographic boundaries (Maibach et al., 1996). In this case, demographic variables may not minimize the within-group differences and maximize between-group differences on critical predictors. Also, theories of message effects typically focus on the proximal predictors, making it difficult to translate demographic-based findings into persuasive strategies in a campaign (Slater & Flora, 1991). A more productive strategy is to group people into segments based on shared psychosocial predictors.

Empirical clustering refers to statistical procedures for collapsing complex data into meaningful and interpretable answers (Slater, 1996). While empirical clustering can be done with demographic variables, it is typically performed on proximal, psychosocial predictors. Clustering procedures fall into two broad types: hierarchical and nonhierarchical (Hines et al., 2014). Hierarchical clustering involves selecting a statistic that quantifies how far apart (or similar) two people are, and then selecting a method for forming groups. It is possible to have as many groups as there are people (leading to individual tailoring). This technique involves iterations of combining and dividing data and looking at the solutions (e.g., with dendrograms or agglomeration schedules) until the researcher determines how many groups are needed to represent the data (Moutinho, 2011). With nonhierarchical clustering, the researcher specifies the number of groups, and then the statistical procedures allocate people into that number of groups. The statistical procedures include k-means clustering (Hartigan & Wong, 1979; Moutinho, 2011), Q methodology (McKeown & Thomas, 2013), and latent class analysis (LCA; Collins & Lanza, 2010). LCA has a few advantages over other techniques: it is less sensitive to missing data, it can include categorical and interval variables, and it provides fit indices to guide decisions to determine which number of groups best fits the data. For these reasons, it has been considered the optimal technique for audience segmentation (Maibach et al., 2011). The rest of this essay discusses empirical decisions with LCA in mind.

Latent Class Analysis

In brief, latent class analysis (LCA) is used to fit models in which a population is divided into mutually exclusive and exhaustive subgroups (Collins & Lanza, 2010). It is conceptually similar to other latent variable models (e.g., factor analysis); it is used when the latent variable is categorical, such as audience segments. The model presumes that the grouping variable cannot be observed, but can be inferred from multiple, measured variables. The general research question for an LCA is this: is there a latent class structure that represents the heterogeneity in the targeted population?

For LCA, the researcher measures a set of variables for the sample, and the measures are dichotomized. The researcher specifies the number of groups to consider in a model. The procedure then estimates two parameters: (1) the likelihood of members in a latent class (often presented as frequencies) and (2) the likelihood of providing a particular response on a measured variable conditional on the set of classes. The procedures provide many goodness-of-fit indices for the models, which are used to select the best number of classes (e.g., four-class or five-class model); lower scores are better (Collins & Lanza, 2010). LCA’s primary aim is not to identify a person’s group membership. Each person has some probability of belonging to each latent class, thus any person’s actual group members is uncertain (Collins & Lanza, 2010). Instead, LCA is used to identify coherent subgroups based on the measured indices and to describe a group’s profile based on their responses to the measured indices and to determine how much of the population is likely to fall within each group.

For example, according to the risk perception attitude (RPA; Rimal & Real, 2003) framework, there should be four groups based on people’s risk and efficacy assessments related to a health action. The researcher would then measure people’s risk perceptions about a health condition, efficacy perceptions to engage in recommended actions, and intentions to do so. The responses would be dichotomized to indicate, for example, feeling at risk for the health condition or not feeling at risk. The researchers could use PROC LCA (Lanza, Collins, Lemmon, & Schafer, 2007) to calculate fit indices for two- to seven-class models. PROC LCA provides a variety of goodness-of-fit indices for the models, such as Akaike information criterion (AIC) and Bayesian information criterion (BIC). In the best of circumstances, the fit indices all converge on one model; to provide support for RPA, the four-class model should fit best, with one group characterized by high risk and high efficacy, a second by high risk and low efficacy, a third by low risk and high efficacy, and a fourth by low risk and low efficacy. Furthermore, people intending to engage in the recommended action should appear in group one, but not group four. Although the RPA does not predict how many people to expect in any segment, an LCA framed by RPA about genetic health beliefs showed the most people in the high-risk/low-efficacy (34%) and low-risk/low-efficacy groups (33%; Smith, Greenberg, & Parrott, 2014).

Segmentation Process

Debates and Future Directions

Not everyone favors audience segmentation. One question is whether it is really cost-effective. Resources—time, finances, and human capital—are required to properly perform audience segmentation (Hines et al., 2014). Audience segmentation has also been described as a means to create social division (Corner & Randall, 2011; Gandy, 2001). Critics argue that audience segmentation emphasized the value of difference over similarity (e.g., Gandy, 2001). Gandy (2001) argues that audience segmentation leads to mobilizing groups based on shared in-group values, instead of mobilizing them as members of a larger, diverse community, and is fundamentally discriminatory because it treats people differently based on group membership. Gandy (2001) writes that “the effect, if not the primary purpose, of segmentation and targeting is the exclusion of participants who are deemed unlikely to support the preferred view” (italics in original, p. 142). If targeting or tailoring messages entails excluding people from the flow of information, then it threatens full participation in the public sphere (Gandy, 2001), and may lead to marginalization of certain groups (Corner & Randall, 2011; Gandy, 2001). In response, Hines et al. (2014) argue that audience segmentation does not have to lead to division and marginalization; it can provide information to craft persuasive strategies that promote community cohesion and bring groups closer together.

Philosophically, there is debate about what segments identified through statistical techniques represent. One perspective is that distinct, homogenous groups exist and these statistical techniques discover them. A contrasting perspective is that identified segments are empirical constructions made by researchers (e.g., Gandy, 2001); empirically identified segments do not exist independently from the research process (Hines et al., 2014). Multiple valid solutions exist, and some solutions may be more useful than others (Hines et al., 2014). Although this does not address whether segments are discovered or made by researchers, it is important for researchers to be systematic and explicit about their decisions, including noting what was not included in the analysis and its limitations. For example, Maibach et al. (2011) noted that they included motivations, behaviors, and policy preferences as indicators to identify different audience segments related to climate change, but they did not include structural and contextual factors. In their words, “this segmentation system is optimized for efforts to educate or engage the public about global warming per se, and less optimized for campaigns intended to promote changes in energy use behavior” (p. 7).

A third topic is the assumption that audience segmentation should increase the efficiency and efficacy of campaigns (Hines et al., 2014). Few studies test the effect of potential campaign messages on audience segments (Hines et al., 2014). A next extension for person-based methods as modes of research is to analyze how well persuasive strategies, in fact, move people from one segment to another. Latent transition analysis is available to analyze transitions among profiles over time (Collins & Lanza, 2010). Such analysis would reveal whether the campaign message moves the targeted group as intended, as well as its effects on other segments. A related concern is that audience segmentation typically occurs at only one time: during the design phase (Hines et al., 2014). Few studies track the changes in the size of segments over time (see the work on tracking the “six Americas” as an exception at http://climatecommunication.yale.edu/. It is also reasonable that if campaigns are effective, some segments should disappear and/or new ones may emerge. Audience segmentation, as a part of larger program evaluation efforts, may be conducted at multiple times to design effective campaigns and to understand their effects.

The next frontiers for audience segmentation arise with the massive amounts of intensive, longitudinal data available on people’s behaviors and the rise in consumer savvy about personalized campaigns. Information available by tracking search engine, social media, and phone activities provides unprecedented access to observation of human interaction patterns and behavior. Research is underway to determine the best ways to make use of this information. Furthermore, with advances in customization, it may soon be as cheap and easy to tailor a message to a person instead of a group. The public is aware of these activities. In a survey (Turow, Delli Carpini, Draper, & Howard-Williams, 2012), 86% of Americans were opposed to political advertisements tailored to their interests. It is unclear how much of the discomfort noted over personalized political advertising also appears with health campaigns.

Conclusion

Effective campaigns include formative research to identify important variations in the public, groups of people in similar states, and persuasive strategies to move those groups based in sound theories of social influence. Audience segmentation can help campaign designers understand the diversity of public understanding and responses to health threat, and provide evidence from which to develop effective persuasive strategies to increase health prevention and promotion. Audience segmentation provides critical information for planning effective campaigns that use resources strategically (Maibach et al., 2011). It can provide insights into who is doing what, and why groups are doing different things (Maibach et al., 1996) so that public health messages can reach the right people in the right way to maximize improvements in public health.