Whose Knowledge Counts? Considering Gender,Sexuality,and Race in Description,Prediction,and Causal Inference in Oral Epidemiology

Introduction

Epidemiology plays a central role in shaping our understanding of health and disease through 3 core tasks: describing patterns of population health and disease, predicting health-related outcomes, and examining causal relationships (Galea and Hernán 2020). While often portrayed as neutral, these processes are embedded in social contexts and shaped by the researchers’ social positioning across race, class, gender, and sexuality (Robinson and Bailey 2019). As Poirier and colleagues (2023) argue, neutrality in scientific inquiry is neither real nor desirable; the production of epidemiological knowledge is inseparable from the assumptions and power dynamics that shape it, and what counts as epidemiological knowledge is therefore historically and socially situated. This is especially relevant in social epidemiology, where attention to structural determinants of health necessarily involves grappling with the assumptions and power relations that shape the questions asked, the methodological choices made, and how findings are interpreted (Krieger and Davey Smith 2016).

These dynamics are also evident in oral epidemiology. Scholars have argued that the gap between knowledge production and mobilization reflects the adoption of research practices that have historically marginalized certain voices and worldviews. For example, our own work shows that women remain underrepresented in senior authorship roles in dental research and that co-authorship patterns often mirror systemic exclusion (Haag et al 2023). These disparities echo broader patterns across academia (Larivière et al 2013) and are amplified in dentistry, where men continue to dominate leadership roles and institutional support is unequally distributed (Lalloo 2025). Racial minorities are even less visible: population-level data show that racialized groups are significantly underrepresented in the dental workforce (Mertz et al 2017) and likely even more so in oral health research. In turn, these patterns of exclusion reflect not only pipeline disparities but the positivist assumptions on which epidemiology has long been built, including claims to neutrality, universality, and objectivity. Oral epidemiology has inherited these traditions, shaping what is studied, who is visible in the data, and who holds the authority to frame questions, define categories, and interpret results.

Underrepresentation in authorship intersects with the dominance of research methodologies that devalue co-design, participatory methods, and lived experience as valid forms of knowledge. Studies show that researchers from marginalized backgrounds are more likely to pursue equity-oriented questions and use community-engaged methods (Hoppe et al 2019; Hofstra et al 2020), yet their approaches are often dismissed as less rigorous. These findings highlight how structural exclusion from academia not only limits who participates in research but also devalues the contributions and methodologies of those who do. In the words of Bhakuni and Abimbola (2021), “knowledge systems are social systems, with their share of social prejudices and implicit biases that interfere with people’s ability to participate fully and equally in knowledge production, use, and circulation.”

In this commentary, we argue that oral epidemiology has inherited key assumptions from mainstream epidemiology, including claims to neutrality, universality, and objectivity. These assumptions have contributed to the exclusion of racialized, gendered, and sexually diverse populations from shaping the field. Such exclusions hinder the achievement of Hernán’s 3 core tasks for the discipline: description, prediction, and causal inference (Hernán et al 2019). Drawing from empirical evidence, we show how these dynamics lead to (1) narrow and selective descriptions of data, (2) unfairness in predictive algorithms, and (3) simplistic or fragmented causal models that overlook the upstream determinants of oral health inequities. We build on previous work in social epidemiology and explore how these biases can be counteracted by embedding intersectionality, inclusive governance, and researcher reflexivity across all stages of knowledge production. To our knowledge, such an explicit critique has not previously been undertaken in oral epidemiology. Drawing on intersectional and feminist population health research (Krieger 2012; Bauer 2014; Krieger 2016; Poirier et al 2023; Krieger 2024b) and decolonial and global health scholarship (Paine et al 2020; Bhakuni and Abimbola 2021), we offer a novel and challenging perspective intended to expand, rather than dismiss, the field’s core tasks.

Narrow and Selective Description of Data

Descriptive epidemiology is focused on examining patterns of health and disease in populations, helping to identify population groups in need of care (Fox et al 2022). There are persistent challenges related to the recruitment and retention of historically excluded and underrepresented groups. Practical recommendations to address this issue include elements related to the research team, such as cultural competency and the identification with the sample; meaningful engagement with communities; logistical aspects, such as innovative sampling strategies; and the return of value to the participants (Islam et al 2010; Farooqi et al 2022; Cunningham-Erves et al 2023). Here, we focus on 2 key problems in how oral health inequities are typically described: (1) narrow and selective reporting and (2) a lack of inclusive, reflexive methodologies that interrogate the framing and representation of data.

Narrow reporting may arise from a superficial understanding of how social identities are shaped and help shape structural disadvantage. Reginaldo and colleagues (2022) reviewed 53 studies on racial oral health disparities and found that most failed to justify their use of race, explain how it was classified, or clarify how data on race were collected. A similar problem applies to gender or sexuality. Despite growing evidence that sexual and gender minorities experience worse oral health outcomes (Schwartz et al 2019; Gupta et al 2023), these populations are often invisible due to limited collection of data and use of reductionist categorizations that obscure their experiences.

The lack of intersectional analysis compounds this erasure, leading to what is widely known as “intersectional invisibility” (Purdie-Vaughns and Eibach 2008). Most oral health inequity studies disaggregate data by either race, gender, or sexuality, without attending to how these categories intersect to shape the distribution of oral diseases. This results in a fragmented understanding of inequities, in which the specific experiences of, for example, Indigenous women, sexual minority men, or nonbinary asexual individuals are flattened into broader categories that hide their unique individual and group experiences. In this sense, narrow reporting in descriptive oral epidemiology is unlikely to be a methodological oversight, but it reflects broader choices about what and who counts in knowledge production.

Selective reporting is equally shaped by decisions that are often framed as neutral but inherently involve value judgments on how social groups are defined and compared. In a seminal commentary, Harper and colleagues (2010) underscored that absolute and relative measures to describe health inequities often tell conflicting stories. They demonstrated that, between 1960 and 2000, although absolute differences in mortality between US Black and White children declined, relative inequities remained constant. This highlights how choosing one measure over another frames the equity narrative differently, and the power lies in the hands of researchers to decide which narrative to promote. Transparently reporting both absolute and relative measures is therefore recommended, as this allows readers to consider the distinct and complementary perspectives each provides.

As highlighted by Reginaldo and colleagues (2022), numerous examples exist in oral health research in which arbitrary decisions were made without reflexivity or transparent disclosure. We strongly endorse the call by Poirier et al (2023) for epidemiologists to reflect on their positionality as central to equity-focused research. This would make visible the values and assumptions that shape how data are categorized and reported, reducing the risk that specific interpretations are presented as objective truths. When paired with intersectionality frameworks for health equity and community-led data governance, these practices help shift control over research priorities, category definitions, and interpretive frames (Bauer 2014). Frameworks, such as Indigenous data sovereignty (Paine et al 2020), exemplify how redistributing this authority can lead to more accurate and transparent representations of oral health inequities.

Prediction: The Challenge of Algorithm Fairness

Prediction is another core task of modern epidemiology, and the increasing use of machine learning (ML) in oral health research brings both promise and peril to predictive endeavors (Tuygunov et al 2025). Predictive models are being developed to identify individuals at risk of poor oral health, target preventative interventions, and improve health service delivery. Some of the most common sources of algorithmic bias are developing models with biased training data or with insufficiently representative data on some groups. With limited attention to data representativeness and systemic bias embedded in algorithms, these initiatives risk reinforcing existing inequities.

Algorithmic fairness, which postulates that predictive models should yield equitable outcomes across groups (Chen et al 2023), must be assessed across intersecting subgroups, not through reductive binary comparisons such as “White vs non-White,” which obscure important within-group variation. For instance, a previous study evaluating fairness in oral health models using US data found particularly poor model performance for Black adults, an issue that would have been masked by lumping all non-White groups together (Schuch et al 2023). Accordingly, a study predicting oral health services use in Brazil identified the poorest model performance for mixed-race people (Chisini et al 2025). Assessing and addressing these issues requires rethinking how race, gender, and sexuality are conceptualized and operationalized in predictive models. These constructs are often used as proxies for unmeasured or poorly understood social factors, including the effects of structural racism and sexism, despite wide variation within and between groups (Krieger 2024a). This practice can obscure the root causes of inequities and reinforce bias in algorithmic systems. As such, the dynamic nature of intersectionality calls for the development of ML approaches that go beyond simplistic categorical variables and can meaningfully account for the interdependence of multiple social identities, without reinforcing stereotypes or oversimplifying lived realities (Mhasawade et al 2021).

To address the lack of data from minority groups, federated learning offers a promising path forward. This ML approach enables models to be trained collaboratively across multiple data sources without sharing raw data (Xu et al 2021). This can help overcome the problem of underrepresentation by incorporating diverse, locally relevant data while preserving patient privacy. However, its success depends on the availability of locally relevant data and thoughtful data governance, including equitable agreements about access, control, and benefit sharing (Chen et al 2023).

Equally critical is the inclusion of people with lived experience at every stage of predictive model development and implementation. There are multiple metrics used to evaluate algorithmic bias, such as equalized odds, demographic parity, and predictive parity, but these fairness criteria often conflict and cannot be simultaneously satisfied (Chen et al 2023). As such, it is essential to clearly define which fairness metric best aligns with the ethical and practical goals of the specific prediction task. This choice should not be left solely to technical teams, as neither the model nor its developers, without domain expertise and lived experience, can adequately determine which tradeoffs are most appropriate (eg, whether to prioritize minimizing false-positives or false-negatives).

Although ML has been an innovation for dental care and education (Tuygunov et al 2025), studies have paid little to no attention to model fairness. As argued above, models should be based on high-quality and representative data, with close evaluation and transparent reporting of all steps of the model-developing pipeline (Allareddy et al 2023). Algorithmic fairness is not a checkbox; it is a continuous process of design, dialogue, and accountability. This includes engaging interest holders, including patients and health professionals, from the outset to define priority questions and determine the most appropriate point in the care pathway for model implementation. When guided by these justice-oriented principles, prediction tools have the potential to improve efficiency and center the needs of marginalized populations. As ML becomes more deeply embedded in oral health research and practice, equity must be built into every stage of development, validation, and deployment.

Causal Inference: Which Causes Are Studied, and How Are They Defined?

While causal thinking has been central to epidemiology for centuries, the past 2 decades have seen exponential growth in the use of formal methods aimed at drawing causal inferences from observational data, such as counterfactual frameworks, and propensity score matching (Krieger and Davey Smith 2016). This trend is increasingly evident in dental research. A recent review identified 85 articles using these methods between 2012 and 2024, with more than 75% published in the past 5 y (Dao et al 2025). This shift poses a reflection on important questions about how causal relationships are constructed, which determinants are rendered visible, and whose realities are excluded in oral epidemiology.

Recent debates have revealed tensions between the rise of formal causal inference methods and the long-standing epistemological commitments of social epidemiology. While proponents of counterfactual approaches, such as Galea and Hernán (2020), argued that social exposures such as income or race can and should be studied using the same formal tools applied to clinical exposures, some social epidemiologists have argued otherwise. Robinson and Bailey (2019), for instance, have stated that these frameworks often fail to capture the structural nature of social determinants. Building on critical race theory (Delgado and Stefancic 2001) and ecosocial theory (Krieger 2012; Krieger 2024a), some scholars have emphasized that processes such as racism and colonialism are not reducible to fixed or isolated variables. Rather, they are dynamic systems of power that operate through material and symbolic conditions as well as lived experiences. While useful in some contexts, attempts to emulate counterfactual experiments to assess the health effects of systems of oppression risk stripping away the complexity that makes these determinants so important for (oral) health. This complexity is essential to capturing how structural forces interact to reproduce health inequities. Oversimplified models can misrepresent causality, obscure determinants, and lead to interventions that are ineffective or even harmful. Krieger and Davey Smith (2016) cautioned against such “spurious causal inference,” especially when social constructs such as race and gender are framed as immutable biological features, and the role of racism and sexism is invisible in epidemiological modeling.

This reductionist logic is visible within the field of oral epidemiology (Baker and Gibson 2014). Most studies examine the effect of race on oral health using statistical models that include socioeconomic position as a covariate. In doing so, socioeconomic position is treated as a confounder rather than as a potential pathway through which race influences oral health. A review by Celeste and colleagues (2023) found that among articles published in 2020, fewer than 15% explicitly included racism as a causal factor in explaining racial oral health inequities. Recent work on sexual minority health further illustrates the limitations of narrow causal frameworks in oral epidemiology. Gupta et al (2023) identified heightened cost-related dental care avoidance among bisexual adults, suggesting that conventional models fail to capture the broader social and structural dimensions of these disparities. Yet such mechanisms are often flattened in epidemiologic models that treat sexual orientation as a static covariate rather than a dynamic axis of oppression intersecting with race, class, and gender. These practices exemplify what Krieger and Davey Smith (2016) referred to as “bad social science”: stripping social variables of their historical and political content and misrepresenting structural violence as statistical noise. They also reflect dentistry’s broader struggles in addressing the social determinants of health in dental education (Leadbeatter and Holden 2021) and how social justice is framed in the profession’s social contract (Holden and Quiñonez 2021).

We argue that causal inference in oral epidemiology must also incorporate intersectionality and reflexive perspectives. Intersectionality frameworks enable researchers to examine how systems of power intersect to shape exposure to oral health risks, linking structural oppression to individual and group oral health outcomes. Reflexivity can expose the normative choices underpinning causal models, challenging assumptions about neutrality. Complementing this, community-led governance structures, such as Indigenous data sovereignty, empower communities with authority and oversight of research priorities, methods, and ethics, providing mechanisms for collective control over what gets measured, how causal models are constructed, and which assumptions can (or should) be held (Doerksen et al 2024). Together, these approaches shift the focus from isolating variables to understanding the broader systems underlying oral health inequities.

Final Considerations

This brief commentary aims to foster reflection on the ways we describe population patterns of health and disease, predict oral health–related outcomes, and assess causes of oral health. Addressing the exclusions embedded in these tasks requires an epistemological shift from positivist traditions toward interpretivist approaches that foreground context, power, and justice. Following some social epidemiology scholars, we argue that reflexivity, intersectionality, and inclusive governance are not optional additions but foundational to transforming knowledge production.

There will never be an epidemiological design or method capable of capturing the full extent of researcher influence on health equity research, nor should we aspire to that. However, failing to interrogate the foundations of knowledge production constitutes a form of structural violence against racialized, gender-diverse, and sexually diverse communities. This occurs not only through underrepresentation in data but also through exclusion from shaping the questions, analytical categories, and interpretations. Redressing this requires reflexivity, intersectional frameworks, and governance that centres marginalized communities as knowledge producers.

Author Contributions

D.G. Haag, H.S. Schuch, contributed to data conception and design, drafted the manuscript; J.L. Bastos, G.H. Soares, B. Poirier, L. Jamieson, contributed to data conception and design, critically revised the manuscript. All authors gave their final approval and agreed to be accountable for all aspects of the work.