Latent Class Analysis: A Statistical Approach for Advancing Research on Race and Culture

An emerging trend among social scientists is to contextualize research designs, analyses, and findings (Goings et al. 2023; Moffat 2015; Wilkinson et al. 2017). One type of statistical analysis that can support this endeavor is latent class analysis (LCA). This method is used to detect latent (or unobserved) heterogeneity by identifying subgroups among an apparently homogenous group (Hagenaars and McCutcheon 2002). For example, it has been used to detect whether latent subgroups of social determinants of health exist among Black youth (Weller et al. 2020). Because it identifies similarities within and differences between groups, findings from LCA can challenge stereotypes, monolithic ideologies, and biases. It can also be used to challenge systemic racism and power structures by centering analyses on race and culture to deconstruct assumptions about people, which is consistent with an antiracist research perspective (Goings et al. 2023). Further, by identifying subgroups, results from LCA can be used to identify specific interventions, practices, and policies relevant to the identified subgroups. Despite the potential for LCA to challenge current paradigms, researchers and graduate programs/curricula offer little guidance on how to use this statistical method to promote social and racial justice in counseling. Therefore, this article provides suggestions on how LCA can be used to advance multicultural counseling research on race and culture. Before describing how LCA can be employed, though, I first provide a nonstatistical description of this approach. LCA uses quantitative data to identify latent (unobserved) groups or classes (Hagenaars and McCutcheon 2002). Like other structural equation modeling approaches (e.g., Watkins et al. 2018), in LCA, a latent phenomenon (in this case, classes or groups) drives participants’ responses to survey items (Muthén and Muthén 2000). This foundational concept of LCA is often complex to grasp because researchers may assume scores on a measure or survey item drive the latent group, but it is membership to a group that drives participant responses. Therein lies the power of LCA—it can uncover unobserved groups of people with shared characteristics or traits that cannot be directly measured. For readers interested in learning more about LCA or wanting a statistical definition of this approach, I encourage them to review existing resources (e.g., Muthén 2002; Nylund-Gibson and Choi 2018; Sinha et al. 2021; Weller et al. 2020). The above examples would use cross-sectional data; however, researchers can also use longitudinal data to examine whether latent classes change over time. For example, researchers can ask the question: How do Latinx youth's ability to use behavioral modification skills change while participating in multicultural counseling? To answer this question, researchers would conduct Latent Transition Analysis, which is like LCA but examines changes over time (see Abarda et al. 2020 for details on Latent Transition Analysis). With a working understanding of LCA, I now turn to sharing my recommendations on how to use it to advance multicultural counseling research on race and culture. These recommendations are based on my experience conducting LCA (e.g., Weller et al. 2013), publishing best practices on LCA (Weller et al. 2020), providing national workshops on how to conduct LCA, coauthoring a study that systematically reviewed the presence of LCA in the literature (Killian et al. 2019), and my experience reviewing manuscripts and being an associate editor. Therefore, I provide recommendations as a user, an educator, and a reviewer of LCA. Although I used LCA as a foundation for these recommendations, they also apply to other similar types of analyses (e.g., latent profile analysis and latent transition analysis). My first recommendation for using LCA is to have a literature-informed rationale for conducting the study. It is critical that researchers do not rely on the claim that it hasn't been done before or it hasn't been done in this population. Reviewers usually consider these claims to be insufficient to justify the necessity of the study. Further, manuscripts that rely on these claims are often the main reason reviewers suggest rejecting a manuscript. Therefore, like Kliewer (2005), I recommend that authors provide one of two reasons for the study: “The study…is the logical next step in a line of investigation, or prior studies have been somehow deficient in some way that the current study addresses” (p. 593). By using one of these two reasons (and literature to support the reason) and recommendations on writing a manuscript, such as those provided by Bowen (2003), researchers can point to gaps in the prior research and show how their study advances the professional discourse and clinical practice. In addition to having a literature-informed rationale for the study, my second recommendation is for researchers to identify the practice, educational, and/or policy implications of the potential LCA findings. This process occurs as researchers begin to conceptualize their study. In fact, as scholars embark on using LCA, it is important to consider questions such as the following: Why does knowing whether the presence of latent classes matter? What will this information contribute to clinical practice and/or education? If providers and policymakers do not have this information, what will happen? Can these classes help providers select which intervention to provide? Answering such questions is critical because it will guide the utility of the findings and make it possible for the results to have implications. Without a study clearly describing the implication of the findings, I have consistently seen reviewers mention one of two comments to authors that can result in a study being rejected: the researchers did not provide an answer to the “So what?” question, or the authors appeared to do fancy statistics, but there is no reason for the study. To avoid such feedback, when researchers provide thoughtful implications of their findings, reviewers often support publishing the manuscript. To decrease the likelihood of such feedback from reviewers, it is helpful when authors briefly mention the implications of LCA findings in the introduction section and then provide details in the discussion section of their manuscript. My third recommendation is to use a theory and/or framework to guide the LCA study. Alas, I have seen many manuscripts and published articles that are atheoretical or have an underdeveloped theoretical rationale guiding the study. When this occurs, findings may not advance the scholarly literature or may have limited clinical utility. Further, if the theory/framework does not apply to the population of interest, findings may not be inclusive of race or culture. For instance, when conducting an LCA study that focuses on identifying latent groups of Black youth identity, it would be important to use a theory of identity development that applies to Black youth. Therefore, I caution researchers away from general frameworks that could result in classes labeled as a low risk group, a medium risk group, or a high risk group. Instead, researchers need to focus on creating conceptually rigorous groups. The theory also informs the measures or survey items used to detect the classes, which are referred to as indicator variables in LCA models. For example, if the study is about identifying latent groups of racial identity, then having indicators that assess various aspects of identity would be needed. The theory also can guide how researchers code the survey items. For example, in addition to statistical criteria, theory can help researchers decide whether the indicator variable will be categorical, which will call for the use of LCA, or continuous, which will use Latent Profile Analysis (an approach like LCA but differs because the indicator variables are continuous; see Williams and Kibowski 2016 for a review of latent profile analysis vs. LCA.). Further, theory informs researchers in selecting the final class solution (i.e., the final number of latent classes). The process of selecting the final class solution can be perhaps the most complex aspect of conducting LCA (see Nylund-Gibson and Choi 2018 and Weller et al. 2020 for details on selecting a final class solution). Like others (e.g., Muthén and Muthén 2000; Nylund et al. 2007; Stringaris et al. 2013), I contend that in addition to using statistical criteria to determine the final class solution, researchers need to ensure that the classes make theoretical or conceptual sense based on the probabilities of each indicator within the classes. To illustrate this point, I provide an example. A researcher conducted LCA with the purpose of identifying whether classes of multicultural counselors exist. The study was guided by the multicultural competency model (Sue 2001), and the researcher used indicators to assess providers’ multicultural awareness of attitudes and beliefs, knowledge, and skills across multiple foci and dimensions. The results suggest a two-class solution: one class generally had high probabilities of indicators of multiculturalism, and the other class had generally low probabilities of indicators of multiculturalism. The researcher would use theory to determine whether having two classes makes sense and whether the high (or low) probabilities of the indicators make sense for each class. Embedded in the recommendation of using theory to select the final class solution is the practice of labeling or naming the classes. Theory can help researchers avoid a common criticism from reviewers: the “naming fallacy” of classes, wherein the name or label given to a class is not an accurate representation of the class. Theory is an important tool for mitigating this risk because it can provide a conceptually sound label for the class. In the example above, labeling the two classes as the multicultural counselor class and monocultural counselor class likely has more utility to the profession than using labels such as high multiculturalism class and low multiculturalism class. Theory, and existing literature, can also assist researchers in deciding whether to include covariates in the LCA model. Covariates can broadly be conceptualized as variables included in the analysis to determine whether classes are statistically more likely to be comprised of specific demographics. For example, researchers can examine whether the multicultural counselor class is statistically comprised of more women, men, or nonbinary individuals. In such an example, including covariates can support an intersectional perspective (Crenshaw 1989). However, as a word of caution, researchers need to be thoughtful about existing research, their sample, and sampling frame before including covariates. For instance, I encourage researchers to provide a strong rationale for including multiple racial groups in the analyses and then provide the reason (with citations) that specific racial groups may be more likely to be one group versus another group. Further, like other researchers (Johfre and Freese 2021), I suggest researchers not use White as a reference group in their analyses because it reinforces research centered on Whiteness. Additionally, when including covariates, authors should ideally include a research question related to the covariates (see Weller et al. 2013 for example). Researchers should also select a theory that has been used to explain the specific outcomes of interest. Once the classes have been identified, researchers examine whether class membership is associated with a specific outcome. This approach both helps validate the class findings and provides additional clinical utility for the results. Case in point, with the multicultural example, it would be important to determine whether class membership in the multicultural counselor class or monocultural counselor class was associated with client retention. Returning to the broader recommendation of using theory to select the final class solution, as a word of caution, I have seen researchers attempt to claim that even though the statistics did not support the final class solution, the researcher chose a specific class solution because of theory. However, I have rarely seen a manuscript that was able to provide a plausible enough explanation for the statistics not to support the final class solution. Therefore, I posit that it is critical that the statistics also support the theoretical rationale for the classes in studies on race and culture. In the absence of this information, it lends the study to criticism and does not validate the findings. Moreover, it can decrease the utility of the findings for clinical care. However, finding a final class solution that is supported by theory and statistics makes it difficult for reviewers and others to negate the importance of the findings. Thus, giving researchers a powerful way to advocate and potentially influence clinical care relevant to race and culture. My fourth recommendation for using LCA for multicultural research is to discuss its limitations. Of particular note is the limitation that this approach is based on probability—the probability of individuals being a member of a class (Muthén and Muthén 2000). Therefore, individuals in the sample rarely completely belong to one latent group. That is, individuals may not have the probability of 1.0 of being in one group. Rather, an individual may have a probability of 0.78 of being in one class and 0.22 in another class; thus, individual cases in the sample will be assigned to the class with the greatest probability. Therefore, it is important when discussing the findings that the researcher acknowledges this limitation. My final recommendation for using LCA for multicultural research is to consider whether the study could be part of a mixed methods project. For instance, researchers could conduct a sequential mixed methods study (Creswell et al. 2011) that first identified latent classes and then employed qualitative methods to contextualize the class solutions. Or researchers could use qualitative data to guide the conceptualization or utility of identifying latent classes and then use those findings to guide the development of the latent class study (from selecting indicator variables to labeling the classes). Although using mixed methods is not possible for all research projects because of various constraints (e.g., time and funding limitations), studies using qualitative methods to further understand, describe, or validate the class findings might provide important contextual information that results in clinically relevant findings. In conclusion, LCA is a valuable statistical method for advancing multicultural counseling research on race and culture. This article is the first to describe how LCA can be used to contextualize research in this way, which contributes to the growing number of articles and resources that describe how to conduct LCA (e.g., Nylund-Gibson and Choi 2018; Weller et al. 2020). In addition to sparking discussion on LCA, I hope this commentary provides researchers with ideas on how this method can be used to conduct anti-racist research and advance multicultural counseling research on race and culture. The author declares no conflicts of interest.