Hidden patient groups revealed: when a popular statistics tool helps and when it misleads

Latent profile analysis has quietly become one of the most fashionable statistical tools in supportive cancer care research, and a new commentary in the journal Supportive Care in Cancer argues that the field needs to slow down and think more carefully about how it is being used. Written by Derek K. Smith of the University of Iowa College of Dentistry and Loren E. Smith of the University of Iowa College of Medicine, the commentary draws a sharp distinction between two fundamentally different ways of applying the technique, and warns that conflating them can lead researchers to draw conclusions that their data cannot actually support. The piece arrives at a moment when journals in oncology and psychosocial research are publishing latent profile studies at a remarkable pace, covering everything from financial toxicity to sleep disturbance to information needs among patients with blood cancers.

To understand the argument, it helps to understand what latent profile analysis actually does. The method belongs to a family of person-centered statistical approaches, meaning that instead of examining how variables relate to one another across a sample, it looks for subgroups of individuals who share similar response patterns. Technically, the procedure is a finite mixture model: it assumes that the observed population is a mixture of a small number of unobserved, or latent, subpopulations, each with its own characteristic distribution on a set of measured indicators. The analyst specifies a candidate number of classes, the model estimates the probability that each participant belongs to each class, and a series of fit statistics and theoretical judgments determine how many classes best describe the data. When done well, the result is an empirically derived typology, a set of patient profiles that might otherwise remain invisible in averages and correlation matrices.

The appeal for supportive care researchers is obvious. Cancer care is increasingly organized around the recognition that patients are heterogeneous, and that a one-size-fits-all approach to symptom management, psychological support, or rehabilitation leaves many people underserved. Latent profile analysis promises to turn that heterogeneity into something concrete: named subgroups, each with a prevalence estimate and a set of associated risk factors, that can in principle guide screening and targeted intervention. Recent studies cited in the commentary illustrate the breadth of this ambition. Researchers have used the method to identify profiles of reproductive concerns in young women with cervical cancer, to classify the severity of financial toxicity in young and middle-aged patients, to characterize multidimensional dyspnea in advanced lung cancer, and to map patterns of information needs among patients with hematological malignancies in China.

But the Smiths’ central point is that not all of these applications are doing the same kind of work, and the difference matters. They distinguish between mono-construct applications, in which the indicators fed into the model are all facets of a single underlying construct, and multi-construct applications, in which the indicators come from several conceptually distinct domains. In a mono-construct analysis, the latent profiles represent different levels or configurations of one thing. A study of dyspnea, for example, might profile patients according to multiple dimensions of the same breathlessness experience, such as its intensity, its emotional impact, and its functional consequences. The resulting classes are, in effect, a severity typology, and interpreting them as such is straightforward.

Multi-construct applications are a different animal entirely. When a researcher combines indicators of, say, sleep quality, fatigue, and emotional distress into a single latent profile model, the resulting classes are not levels of any one construct. They are instead complex syndromes, constellations of co-occurring problems that may or may not share a common cause. A class characterized by poor sleep and high fatigue and elevated distress is a clinically meaningful pattern, but it is not a point on a single dimension, and the language used to describe it should reflect that. The commentary suggests that much of the confusion in the literature stems from applying mono-construct interpretive habits, such as ordering classes from low to high, to multi-construct solutions where no such ordering exists.

The stakes of this distinction become clearer when one considers how latent profile results are typically translated into clinical recommendations. A mono-construct severity typology invites a fairly direct response: patients in the highest-severity class need the most intensive intervention, and screening tools can be calibrated to detect them. A multi-construct syndrome typology demands more nuanced thinking, because membership in a class does not by itself indicate which problem should be treated first or whether the co-occurrence of symptoms reflects a shared mechanism, a cascade of consequences, or mere statistical coincidence. Treating a syndrome class as if it were a severity group risks designing interventions that target the wrong thing, or that assume a causal structure the analysis never tested.

Several of the recent supportive care studies highlighted in the commentary show how varied the practice has become. One team examined the co-occurrence of severe pain and sleep disturbance in oncology outpatients receiving chemotherapy, a question that sits naturally at the boundary between mono- and multi-construct thinking, since pain and sleep are distinct constructs whose interaction is itself the object of study. Another group profiled sleep hygiene behaviors and dysfunctional sleep beliefs in lung cancer patients undergoing chemotherapy, and examined their impact on cancer-related fatigue. A third applied latent profile analysis to meaning in life among young and middle-aged cancer patients, while a fourth combined latent profiles of emotion regulation strategies with mediation analysis to explore psychological flexibility and subjective well-being in breast cancer patients. Each of these choices about which indicators to combine carries interpretive consequences.

The commentary also touches on the temporal dimension of person-centered modeling. Latent transition analysis, a longitudinal extension of the same modeling framework, was used in one cited study to track how symptom patterns changed among patients undergoing surgery for esophageal cancer. Extending profile analysis over time adds another layer of interpretive complexity, because researchers must now decide whether the classes themselves are stable entities that patients move between, or whether the underlying structure changes as recovery progresses. The Smiths’ framework, by distinguishing what the indicators represent before any modeling begins, offers a way to keep these questions grounded: the meaning of a transition between classes depends on whether those classes describe levels of one construct or configurations of several.

What practical guidance emerges from all of this? The commentary’s implicit message is that rigor in latent profile analysis begins before the software is opened, with a clear articulation of the construct or constructs the indicators are meant to capture. In mono-construct applications, researchers should be able to defend the claim that their indicators are facets of a single dimension, and should interpret class differences along that dimension. In multi-construct applications, they should resist the temptation to rank classes or to describe them as mild and severe, and should instead characterize them as distinct patterns of co-occurrence whose clinical significance requires separate justification. Fit indices and statistical tests, while necessary, cannot settle these questions, because they are questions of meaning rather than of model fit.

For a field that increasingly relies on person-centered methods to justify personalized care, the message is timely. Supportive care in oncology is built on the premise that understanding individual differences improves outcomes, and latent profile analysis is a powerful tool for that purpose when it is matched to the right kind of question. The Smiths’ commentary, published as Volume 34, article 1019 of Supportive Care in Cancer, does not call for the method to be abandoned; it calls for it to be used judiciously, with researchers stating plainly whether their profiles describe degrees of one thing or combinations of many. As the volume of latent profile studies in cancer supportive care continues to grow, that simple act of clarification may prove to be one of the most important methodological habits the field can adopt.

Subject of Research: Methodological guidance on latent profile analysis in supportive cancer care research

Article Title: Judicious use of latent profile analysis in supportive care: the distinction between mono- and multi-construct applications

Article References: Smith, D. K., & Smith, L. E. (2026). Judicious use of latent profile analysis in supportive care: the distinction between mono- and multi-construct applications. Supportive Care in Cancer, 34(10), Article 1019. https://doi.org/10.1007/s00520-026-11226-4

Image Credits: AI Generated

DOI: 10.1007/s00520-026-11226-4

Keywords: latent profile analysis, supportive care, cancer, statistics, person-centered methods, methodology, symptom clusters, oncology, psychosocial research, multivariate analysis, patient subgroups, research methods