What You Cannot See Can Still Kill You: On the Use of Latent Constructs

Research by Vachon, Given, Given, and Dunn (2019), published as an online exclusive with this issue of Oncology Nursing Forum, highlights the unique challenges faced by patients with cancer who are prescribed oral oncolytic agents (OOAs), including temporary stoppages of treatment that can have serious health consequences. The study analyzed associations between the occurrence and frequency of OOA temporary stoppages and concepts, such as burden of treatment (BOT) and multimorbidity. BOT is conceptually defined as a combination of the actual workload associated with treatment and the patient’s perception of this workload (Vachon et al., 2019). Multimorbidity is not explicitly conceptually defined, but operationally defined as a count of the number of conditions requiring medication management at the outset of OOA therapy, abstracted from study participants’ clinical records (Vachon et al., 2019). In the research, BOT and multimorbidity were included as predictor variables in multiple logistic regression models of factors associated with temporary stoppages of OOA therapy. After examining these associations within a secondary analysis of data collected from 272 people with cancer who were newly prescribed OOAs across six National Cancer Institute (NCI)–designated cancer centers, Vachon et al. (2019) concluded that BOT and multimorbidity had no statistically significant effect on the occurrence or frequency of OOA temporary stoppages. The sample of participants for this study was highly educated, entirely insured, aged an average of 61 years, and identified primarily as Caucasian.

To interpret and draw conclusions from the research conducted by Vachon et al. (2019), measurement choices and assumptions associated with their statistical models must be critically analyzed. These include variables representing the concepts of BOT and multimorbidity, both of which could be considered latent constructs. Although latent constructs can be defined in multiple ways, one of the simplest definitions is that a latent construct represents a phenomenon that exists within a model but does not appear directly in a dataset (Bollen, 2002). In other words, the concept cannot be directly observed and captured empirically. Therefore, the researcher must make decisions about what observable indicators can be measured to allow for testing hypotheses involving the latent construct. An example of a latent construct might be the concept of symptom burden. One might say that symptom burden is impossible to observe directly, at least not in the same way that a cancerous tumor cell may be directly observed under a microscope. However, oncology nurse researchers often include variables, such as symptom burden, in their analyses by hypothesizing which observable indicators are associated with symptom burden and designing self-report assessments to capture those observable indicators in the form of responses to items on a symptom burden scale. Although the researchers have not directly observed and measured symptom burden, the scale score represents symptom burden in the model. When evaluating the validity of the scale designed using classical test theory, researchers assume that associations between individual items within the scale are driven by the underlying, unobservable latent construct of symptom burden. Remove the symptom burden, and systematic associations between the individual items on the scale are assumed to fall away.

Defining Latent Constructs

Variables representing latent constructs are used practically everywhere in research for the health sciences. However, there is no single, universal definition for a latent variable. Rather, latent variable definitions are largely model-dependent (Bollen, 2002). One of the most common is known as the local independence definition (Bollen, 2002). This definition assumes that a latent construct explains statistical associations observed between indicator variables (also known as observed variables) that represent the presence of the latent construct. When the value of the latent construct is held constant, indicator variables are expected to function independently. The local independence definition has several implications: (a) errors associated with measuring each indicator variable are independent, (b) indicator variables have no direct effect on each other, (c) there is more than one indicator variable, and (d) each latent construct in the model has a direct effect on one or more observable indicator variables. The definition also implies that indicator variables represent effect indicators (e.g., that their value is a direct result of the latent construct) rather than causal indicators (phenomena directly impacting or causing the latent construct). This final point is a vitally important conceptual distinction that profoundly affects measurement choices.

Measurement tools that capture observable indicators of latent constructs should be valid. Assessment of construct validity includes evaluating (a) the degree to which operational definitions of indicator variables representing the construct flow directly from conceptual definitions and theory, and (b) whether tools designed to measure indicator variables are valid and reliable (Waltz, Strickland, & Lenz, 2017). If indicator variables function as effect indicators, a single indicator variable may adequately represent a unidimensional latent construct, with a certain amount of measurement error reflecting the difference between the observed and the true value. If the conceptual definition of a construct includes more than one dimension, as is often the case with constructs commonly used in nursing research such as quality of life, indicator variables representing each dimension of the construct may be required to fully capture its function. Failure to specify indicators representing every dimension of the construct can result in additional measurement errors and underrepresentation of the true value of the construct in the model.

Defining Burden of Treatment

Although Vachon et al. (2019) never state that they treated BOT as a latent construct, their methods and review of the literature provide a textbook example. They conceptually define BOT as “the combination of a patient’s workload and his or her perspective of the condition and workload” (p. E135). Recent qualitative and quantitative research on BOT across various chronic illness populations is briefly summarized, including operational definitions of BOT in quantitative research designs using indicator variables, such as treatment complexity, number of medications, number of interactions with healthcare providers, and difficulty managing treatment, to represent the otherwise invisible construct. Vachon et al. (2019) explain that prior research used these variables “in place of the actual variable or concept, because it may be more feasible to collect the indicator variable rather than a BOT-specific measure” (p. E136). The authors do not state whether these variables represent causal or effect indicators. Elsewhere in the literature, BOT has been described as a dynamically changing phenomenon that can be objectively measured in terms of the time and energy required to manage workload associated with treatment, and subjectively measured by assessing the perceived effect of workload and capacity to manage the workload through self-report (Austin, Schoonhoven, Kaira, & May, 2019). Theorists cited by Vachon et al. (2019) have also defined BOT as “the impact of everything [people living with an illness] have to do to care for themselves on their quality of life and well-being” (Tran, Messou, & Mama Djima, 2019, p. 266), suggesting that indicators representing BOT might be best imagined as effect, rather than causal, indicators.

Although Vachon et al. (2019) present BOT as a distinct, emergent concept in health research, the act of conceptualizing illness as work is decidedly not. In their classic study of chronic illness self-management, Unending Work and Care: Managing Chronic Illness at Home, grounded theorists Corbin and Strauss (1988) described the work of illness as having multiple dimensions, including medical work associated with treatment (such as attending clinical appointments and managing medications). The definition of medical work by Corbin and Strauss (1988) overlaps almost perfectly with the emergent definition of BOT described by Vachon et al. (2019). In addition to medical work, the research by Corbin and Strauss (1988) described emotion work (such as managing distress associated with a cancer diagnosis) and biographical work (such as adjusting life roles, occupation, and lifestyle to accommodate symptoms and late effects of cancer therapy). The research depicted the overlapping and reciprocally interactive ways in which tasks associated with cancer treatment that might seem isolated and distinct, such as taking prescribed oral medications like OOAs, directly influence and are affected by life context and other forms of labor performed by patients with cancer. This foundational research helped launch an entire subfield of nursing inquiry dedicated to understanding the complexity of the work of chronic illness self-management, including an extensive and growing body of research specifically describing BOT associated with cancer therapy.

Measuring Burden of Treatment

The two indicator variables used to operationalize the latent construct of BOT for the purposes of Vachon et al.’s (2019) analysis included participant workload, measured with an adapted version of George, Phun, Bailey, Kong, and Stewart’s (2004) Medication Regimen Complexity Index, and the BOT-patient perspective, measured as scores on the Cancer Symptom Experience Inventory, capturing symptom interference with functioning at the outset of OOA therapy. Before discussing the psychometric properties of these two measurement tools, the match between BOT as conceptually defined and the two indicator variables should be examined. Indeed, compatible with Vachon et al.’s (2019) conceptual definition of BOT, the two indicator variables represent a combination of an objective measure of workload (medication regimen complexity, as abstracted from clinical records) and self-report (perceived symptom interference with functioning). However, BOT theorists have also defined the phenomenon as a series of dynamic states (Austin et al., 2019). As cross-sectional measures, these indicator variables do not capture changes in BOT during the course of treatment. It is unclear from the description of their methods whether Vachon et al. (2019) imagined their two BOT indicator variables operating as effects or causes of BOT (or neither). The indicator variable participant workload is discussed somewhat causally in the text, whereas BOT patient perspective (on symptom interference) could easily be imagined as an effect of BOT, rather than a cause. In the authors’ visual depiction of their model, BOT (as a variable) has arrows flowing away from it, reflecting causal directionality (rather than effect), but the authors do not claim to be testing causal pathways per se.

Although medication regimen is an aspect of workload associated with cancer therapy, it represents only one aspect of a much more complicated picture. To reduce representation of BOT workload to this single indicator variable is to exclude dimensions of work beyond self-management of medication, such as the time and energy required to travel to appointments, dealing with insurance, and palliating symptoms associated with treatment. Resources lost because of cancer therapy, such as reduced time at work and lost wages, are also not accounted for by this definition. In addition, this operationalization assumes that said workload is equal for individuals with equal medication regimen complexity scores who may, in actuality, vary greatly in the degree of time and energy required to manage a particular regimen. If researchers rely on theory to extend conceptual and operational definitions of BOT workload beyond the strict boundaries of quantifying medications, they must also include dimensions of everyday life work that may be increased in the context of cancer treatment, all of which could potentially paint a distinctly different picture of BOT workload from one individual to the next, even among those with equal regimen complexity scores.

Vachon et al. (2019) acknowledge limitations associated with their choice of measurement tools to capture BOT workload. The Medication Regimen Complexity Index was developed to capture regimen complexity among people with chronic obstructive pulmonary disease, not cancer. The scoring of the tool was altered by the authors to add 1 point in cases where individuals were receiving IV chemotherapy. This adaptation raises questions about the validity and interpretation of scores. The authors acknowledged these limitations, although without further psychometric evaluation of the adapted instrument’s construct validity, additional conclusions cannot be drawn about its ability to accurately capture the burden of cancer therapy and, more specifically, burden associated with OOAs.

Defining and Measuring Multimorbidity

Multimorbidity was also a central component of the hypotheses tested in Vachon et al.’s (2019) analysis, serving as both a predictor variable and an interaction term in their model of factors potentially affecting temporary stoppages of OOA therapy. As a construct, there is no single universally accepted definition of multimorbidity. In their literature review, Vachon et al. (2019) discuss multimorbidity in terms of non-cancer diseases co-occurring in an individual with a cancer diagnosis. Elsewhere in the literature, multimorbidity has been defined as multiple chronic or acute diseases and medical conditions that co-occur in a single person (Bayliss, Edwards, Steiner, & Main, 2008). Operational definitions for multimorbidity range from qualitative and categorical descriptions of the existence, patterns, and/or severity of multiple acute or chronic conditions to the measurement of multimorbidity as a single continuous variable (count of conditions), with the implied assumption that each additional condition has equal weight as the others. Vachon et al.’s (2019) secondary analysis operationalized multimorbidity in terms of the latter, measuring it as number of conditions requiring medication management, abstracted from clinical records at the outset of the study. Measuring multimorbidity in terms of conditions that require medication documented in the health record prior to OOA treatment excludes conditions which may increase BOT (e.g., arthritis, dementia) and other daily work that do not require medication, as well as conditions that might emerge or be exacerbated over the course (and possibly as a direct result) of OOA therapy. The authors acknowledge the limitations of measuring multimorbidity this way. Critiques on the state of the science of measuring concepts, such as multimorbidity and multiple chronic conditions, have exhorted scientists to move beyond characterization of multimorbidity in terms of disease counts, to facilitate a better understanding of the impact of specific combinations and temporal patterns of comorbidities on individuals’ unique and complex systems (Baker et al., 2017; Melis, Gijzel, & Olde Rikkert, 2017).

Identifying Other Latent Constructs Potentially Present in the Model

Latent constructs occasionally operate within models without being named or accounted for in statistical analyses. Participants for this secondary analysis of data (Vachon et al., 2019) were gathered from an NCI-sponsored (R01) randomized controlled trial (RCT) of an OOA self-management intervention. The sample was 89% Caucasian, 8% African American, and the remainder classified as other. Most participants were highly educated, and all had some form of insurance. Although participants’ gender identities and sexual orientations were not reported, marital status was reported and the variable sex was dichotomized as either male or female, with equal proportions of participants in both categories. The authors of the present analysis (Vachon et al., 2019), as well as the original RCT (Sikorskii et al., 2018), have acknowledged some limitations associated with this fairly racially and socioeconomically homogenous sample.

Analyses of persistent cancer-related health disparities that disproportionately affect people who identify as Black, indigenous, and people of color (BIPOC); ethnic minorities, including Latinx; immigrants and refugees; sexual and gender minorities (SGM), including agender, non-binary, and transgender individuals; and the poor or un-/underinsured have led to continued calls for greater representation (oversampling) of these populations in cancer clinical trials (Asare, Flannery, & Kamen, 2017; Watkins Bruner, Pugh, Yeager, Bruner, & Curran, 2015). In light of this imperative, oncology nurse scientists must continually question the extent to which latent constructs that are not specified in statistical models, reflecting unnamed structural and social determinants of health, such as economic oppression, institutionalized racism, and heteronormativity, nonetheless continue to operate (Pallok, De Maio, & Ansell, 2019).

Adopting a Structural Lens on Health

Scholars of structural competency, distinct from concepts such as cultural competency, point out the following:

Health disparities emerge and persist through complex mechanisms that include socioeconomic, environmental, and system-level factors. To accelerate the reduction of health disparities and yield enduring health outcomes requires broader approaches that intervene upon these structural determinants. (Brown et al., 2019, p. S72)

Application of the structural lens to data generated from cancer clinical trials generates wholly new lines of inquiry, and different ideas as to what types of interventions will most effectively address present health disparities. The structural lens forces the researcher to ask new questions about the environments in which health is constructed and where research is conducted, such as the following:

•  To what extent do racially homogeneous (e.g., overwhelmingly non-Hispanic Caucasian) participant samples reflect legacies of institutionalized racism and biases related to who receives cancer screening, who is prescribed oral chemotherapy, and who is offered and subsequently able to participate in cancer clinical trials? Is institutionalized racism not a potential latent construct in the data (unnamed and unaccounted for in statistical analyses)?

•  Is the lack of representation of participants who are poor, un-/underinsured, or who identify as SGM, BIPOC, and/or Latinx potentially driven by structural and social determinants? Are disparities in access to OOAs another takeaway from Vachon et al.’s (2019) trial?

•  How does the exclusion of variables such as gender identity or sexual orientation, and narrow structuring of categories of sex (male/female) and marital status (versus intimate partner status, co-habitation, or presence of a caregiver) within health records and clinical trials create deficits in understanding the experiences and mechanisms driving health disparities, particularly among people who identify as SGM?

Implications for Research and Practice

Latent variables were born out of the necessity of accounting for phenomena that have a real effect on health, healing, and well-being, but that cannot be observed and measured directly. Usually, choices about conceptual and operational definitions for latent constructs are driven by theory and prior empirical work, allowing for the construction of measurement tools and subsequent interpretation of their functions and effects. When measuring phenomena that cannot be observed directly, researchers’ decisions about how to define a construct and capture it via indicators that can be observed, such as responses to self-report questionnaire items, directly affect study rigor. All measurement choices entail the risk of measurement error that can never be completely eliminated, particularly when faced with the practical constraints of a secondary analysis. However, when a concept such as BOT is incompletely defined, or when observable indicators do not match a latent construct’s conceptual definition, validity suffers, and results of the research may be difficult—if not impossible—to interpret and generalize. Mitigation of this tension is as much about maintaining a constant humility about all the unknown ways a variable that cannot yet be measured directly may be operating, as it is about vigilance and skill.

Latent constructs and the variables used to represent them stand in as symbolic mathematical representations of what are hypothesized to be real forces in the world. Some of these phenomena, such as inner strength or supportive care, can serve to heal and protect. Other phenomena, such as structural racism that bars access to quality care or participation in clinical trials, can literally kill. Although it may seem like a purely academic exercise, naming, defining, and interpreting how these latent constructs operate in the world is far from a benign task. These measurement choices about seemingly abstract concepts have real consequences.

Research constructs a certain empirical reality. A version of reality that is, at best, always slightly out of focus. Research that incorrectly, incompletely, or inaccurately captures how latent constructs operate in the world (which is all research involving latent constructs, as all research entails some error of measurement and bias), constructs an empirical reality that is flawed. This flawed empirical reality then serves as a basis for generalization of information to other settings and care. As nurse researchers conceptualize and operationalize latent constructs for use in research, they must keep these stakes, and the limitations of all evidence, in mind.

If all evidence is flawed, should researchers abandon all attempts to interpret and translate evidence to practice? Rather than throw their hands up in despair, researchers should try to proceed as thoughtfully as possible and maintain an empirical humility that holds space for the unknown, the possibility of multiple explanations and competing hypotheses, and drawing new conclusions as the science evolves. This mindset is critically important as health care enters an era where endeavors, such as precision medicine and the use of artificial intelligence in health care, are increasingly driven and automated by data.

Conclusion

Vachon et al.’s (2019) novel study raises the visibility of important challenges faced by patients with cancer prescribed OOAs, particularly the necessity of providing high-quality nursing care that assesses and addresses factors that may increase risks of temporary stoppages in treatment. Disparities in OOA stoppages observed in both women and among patients prescribed specific forms of OOA associated with potentially greater symptom burden, such as kinase inhibitors, merit further investigation. As more and more patients with multiple chronic conditions requiring medication management are prescribed complex oral chemotherapy regimens that must be taken at home, the need for nursing care and interventions to support OOA therapy self-management will continue to grow in urgency.

Oncology nurses apply their expertise to support the health, healing, and well-being of people affected by cancer. Study designs and measurement strategies that describe the constellation of factors that influence and embody latent constructs, such as health, healing, and well-being, are indispensable tools in the pursuit of knowledge to guide the practice of oncology nursing. Measurement error and the lack of a structural lens on health results in bias that affects the validity and generalizability of research. As researchers move forward with the science, they should remain reflexive and vigilant as to how they define and interpret latent constructs present in the models—those that are named and those that are not.

About the Author(s)

Rachel Walker, PhD, RN, is an assistant professor in the College of Nursing at the University of Massachusetts in Amherst; Anna L. Paskausky, PhD, MS, RN, FNP-BC, is an advance practitioner supervisor and lead trainer at Baystate Health in South Hadley, MA; and Shonta Chambers, MSW, is the executive vice president of health equity and community engagement at the Patient Advocate Foundation in Hampton, VA. This article was sponsored, in part, by a Career Catalyst Research grant from the Susan G. Komen Foundation for Breast Cancer Research and the National Institute of Nursing Research of the National Institutes of Health (P20 NR16599). The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health or the Susan G. Komen Foundation. Walker has also previously received an honorarium from Johnson & Johnson Nursing. Walker can be reached at r.walker@umass.edu, with copy to ONFEditor@ons.org.