Invited Commentary: On the Mathematization of Epidemiology as a Socially Engaged Quantitative Science

Abstract

This article is linked to "Optimally Choosing Medication Type for Patients With Opioid Use Disorder" and "Rudolph et al. Respond to “Mathematization of Epidemiology”" (https://doi.org/10.1093/aje/kwac217 and https://doi.org/10.1093/aje/kwad023).

 
• MOUD

  medication for opioid use disorder

 
• OUD

  opioid use disorder

 
• OTP

  opioid treatment program

 
• RROU

  return to regular opioid use

Editor’s note: The opinions expressed in this article are those of the authors and do not necessarily reflect the views of the American Journal of Epidemiology.

Rudolph et al., in “Optimally choosing medication type for patients with opioid use disorder” (1), have provided a welcome and novel contribution to the literature on the comparative effectiveness of medication for opioid use disorder (MOUD) treatments. A massive unmet MOUD treatment need persists in the United States. Although 3 MOUD treatments are approved by the Food and Drug Administration (buprenorphine, methadone, and naltrexone), with demonstrated effectiveness at reducing adverse health outcomes (2–5) and improving quality of life (6, 7), recent estimates indicate that nearly 90% of people with opioid use disorder (OUD) do not receive treatment with medication (8). Stigma and lack of provider awareness and education about these treatments is a major component of this treatment gap (9), and Rudolph et al. (1) make a critical contribution to the urgent task of educating and supporting clinical providers in choosing appropriate MOUD options for their patients. Their paper is also methodologically advanced, leveraging cutting-edge, data-adaptive statistical techniques to “learn” optimal, individualized treatment-decision rules from data derived from 3 high-quality randomized controlled trials of MOUD treatment effectiveness.

The strong analysis presented in Rudolph et al. (1) also provides an excellent opportunity to highlight a central tension between practical public health demands and methodological sophistication in epidemiologic research. Although this is an increasingly common feature of many epidemiologic disciplines, not just substance use epidemiology, we will use their paper as an entry point to a wider exploration of this tension in the context of MOUD access. Specifically, we investigate how this tension is constructed and negotiated via the mathematization of epidemiologic research. Ole Skovsmose (10) defines mathematization as the “formatting” of a given activity “by means of mathematical insight and techniques” (10, p. 1). This may seem trivial—the vast majority of epidemiologic research is quantitative—but mathematization involves more than mere quantitation. Sterner and Lidgard (11) describe mathematization as a “normative” scientific activity, and further argue that this normative process has a profound impact on the conduct and evaluation of science, particularly in (re)defining “what counts as scientific success” (11, p. 53). This process of mathematization is per se neutral, and in fact, many of the consequences of mathematization are clearly beneficial. The strong quantitative evidence base for the effectiveness of these MOUD treatments, for example, is itself a result of highly mathematized epidemiologic research. Even when the research is unimpeachably strong, however, the normative aspects of the mathematization process can have subtle yet profound (and possibly unintended) consequences for how epidemiologic research is formulated, conducted, interpreted, and communicated.

A comprehensive discussion of issues surrounding MOUD must begin from the fraught reality of MOUD access in the United States, beset by substantial structural barriers. The authors allude to these barriers, noting that whether and which type of MOUD treatment is received depends almost entirely on “where a patient seeks treatment” (1, p. 2). This is a consequence of the thicket of strict federal regulations surrounding the dispensation of methadone and buprenorphine. Methadone is available only from specialized opioid treatment programs (OTPs), and regulations on buprenorphine—which can be prescribed in an office-based setting—nevertheless require providers to obtain special training and a Drug Enforcement Administration waiver (called an X-waiver). This regulatory regime interacts with an even more fragmented and complex system of health finance (with different barriers across and within payer type) to produce a bumpy terrain of differential access along geographic, socioeconomic, and racial/ethnic lines. At the individual level, any single person thus faces a highly constrained choice when seeking MOUD, or—in the case of Wyoming, which has no OTPs and therefore no methadone availability (12)—no choice at all. The result of this complex access regime at the population level is to produce and reinforce well-documented inequalities in access. For example, Black patients are less likely to initiate MOUD than White patients (13), and some research has demonstrated that buprenorphine and methadone treatment rates are associated with area-level racial segregation and income composition, with buprenorphine treatment rates higher in areas with more White and high-income residents (14).

Important questions about MOUD effectiveness across heterogeneous patient populations—questions Rudolph et al. (1) tackle in this timely study—arise against this background of unmet need, structural barriers, and demonstrated inequalities in access. The authors used high-quality data from 3 randomized controlled trials of MOUD effectiveness to “learn” individualized treatment-decision rules that minimize risk of return to regular opioid use (RROU). However, this analysis has an intractable limitation: The data derive from 3 different trial environments. Each trial environment, in turn, represents a highly constrained treatment access scenario rather than a truly optimal arrangement where all 3 modalities are available with minimal or no barriers to access. The data from which the models learned the optimal treatment rules thus literally embody the structural barriers to treatment access. One trial, in an OTP setting, randomized patients to buprenorphine or methadone; another, in an office-based setting, randomized patients to buprenorphine and standard medical management or buprenorphine and standard medical management plus individual counseling; and the third, in a short-term inpatient setting, randomized patients to buprenorphine or naltrexone. Further exacerbating this limitation of the data is a mathematical requirement for valid causal inference: equality of the outcome (i.e., RROU) risk across trials, conditional on measured covariates (i.e., the transportability assumption). Because this equality of risk—which the authors appropriately tested empirically where possible—was not observed for patients randomized to buprenorphine in an OTP setting, these patients were excluded from the analytical data.

This exclusion, methodologically valid and necessary, actually compounds the inherent limitations of the data. Practically, it is of great value to know whether buprenorphine provision (associated with less onerous prescribing schedules and greater patient “freedom” (15)) improves outcomes relative to methadone in an OTP setting. Methodologically, the data addressing this direct comparison was not available to the data-adaptive models to learn from because its inclusion would have compromised the validity of the mathematical inference. Here, the tension between methodological sophistication and practical considerations emerges clearly.

Quantitation is the process of formulating a research question in quantitative terms, a precondition for almost all successful epidemiology. Mathematization, a step further, subtly affects the replacement of a public health objective (addressing structural barriers in MOUD access) with a mathematical one (minimizing a loss criterion). The normative process of mathematization thus tends to resolve the central tension of epidemiologic research in favor of technical precision. A common result of this resolution is to obscure the social context that makes a public health question salient to begin with.

The normative dimension of the inescapable mathematization of epidemiology can thus have unintended and potentially undesirable consequences. In this case, the suppression of important context via mathematization risks casting structural access inequalities as individual “clinical” characteristics, potentially amenable to physician intervention, when in fact they represent policy failures. Another, more downstream unintended consequence might be to cement the centrality of medical authority in the process of addiction treatment. The reasonable and appropriate choice of RROU as an endpoint is a fairly prosaic example of this. Incorporating this endpoint into the mathematized process of statistical learning defines the terms of treatment success according to a medicalized expert view of the goals of MOUD. Again, this is a reasonable endpoint, and many patients do indeed consider permanent cessation of regular opioid use a treatment success (16). However, there are many possible patient-centered outcomes among people receiving MOUD, including improved quality of life and treatment of other OUD-related sequelae. Therefore, an individualized treatment rule trained on one particular endpoint may not be optimal for such patients. The collection of patient-centered perspectives that may vary across populations and contexts is rarely included in quantitative epidemiologic analysis, reinforcing the artificial division between qualitative and quantitative research. Bridging this divide should be a focus of future work.

For the individual optimal treatment rules “learned” in Rudolph et al. to be maximally beneficial, changes to both federal policy (to ensure universal access to all 3 MOUD modalities) and clinical practice (to center patients’ experiences, needs, and goals as paramount) are needed. The authors understand this and articulate many of the access concerns that have been rehashed here. Beyond its value as a contribution to the epidemiologic literature, this paper provides a rare opportunity to reflect on mathematization and its tendency to push standards for epidemiologic research in ever more technical directions. Understanding this (and understanding the immense value of quantitative epidemiology broadly), we can both advance rigor and advocate for more holistic and engaged epidemiologic research practice. Critical methods and statistical pedagogy should form part of this practice, as should collaboration with qualitative researchers and scholars across disciplines. Most importantly, epidemiologic research practice should be equally grounded in technical rigor and community engagement. A critical lesson from the coronavirus disease 2019 (COVID-19) pandemic is that excellent methods are necessary but not sufficient: Epidemiologic research serves an important social function in a complex, imperfect, and highly politicized world, in situations with life-and-death stakes. The challenge for the field in the coming decades will be to resolve the tension between methodological sophistication and practical demands in an equitable and harmonious way.

ACKNOWLEDGMENTS

Author affiliations: Department of Epidemiology, Brown University School of Public Health, Providence, Rhode Island, United States (Abigail R. Cartus, Brandon D. L. Marshall).

This work was supported by the National Institute on Drug Abuse (grant R01-DA046620).

Conflict of interest: none declared.

REFERENCES