TORTINI

For your delectation and delight, desultory dicta on the law of delicts.

Links, Ties, and Other Hook Ups in Risk Factor Epidemiology

July 5th, 2023

Many journalists struggle with reporting the results from risk factor epidemiology. Recently, JAMA Network Open published an epidemiologic study (“Williams study”) that explored whether exposure to Agent Orange amoby ng United States military veterans was associated with bladder cancer.[1] The reported study found little to no association, but lay and scientific journalists described the study as finding a “link,”[2] or a “tie,”[3] thus suggesting causality. One web-based media report stated, without qualification, that Agent Orange “increases bladder cancer risk.”[4]

 

Even the authors of the Williams study described the results inconsistently and hyperbolically. Within the four corners of the published article, the authors described their having found a “modestly increased risk of bladder cancer,” and then, on the same page, they reported that “the association was very slight (hazard ratio, 1.04; 95% C.I.,1.02-1.06).”

In one place, the Williams study states it looked at a cohort of 868,912 veterans with exposure to Agent Orange, and evaluated bladder cancer outcomes against outcomes in 2,427,677 matched controls. Elsewhere, they report different numbers, which are hard to reconcile. In any event, the authors had a very large sample size, which had the power to detect theoretically small differences as “statistically significant” (p < 0.05). Indeed, the study was so large that even a very slight disparity in rates between the exposed and unexposed cohort members could be “statistically significantly” different, notwithstanding that systematic error certainly played a much larger role in the results than random error. In terms of absolute numbers, the researchers found 50,781 bladder cancer diagnoses, on follow-up of 28,672,655 person-years. There were overall 2.1% bladder cancers among the exposed servicemen, and 2.0% among the unexposed. Calling this putative disparity a “modest association” is a gross overstatement, and it is difficult to square the authors’ pronouncement of a “modest association” with a “very slight increased risk.”

The authors also reported that there was no association between Agent Orange exposure and aggressiveness of bladder cancer, with bladder wall muscle invasion taken to be the marker of aggressiveness. Given that the authors were willing to proclaim a hazard ratio of 1.04 as an association, this report of no association with aggressiveness is manifestly false. The Williams study found a decreased odds of a diagnosis of muscle-invasive bladder cancer among the exposed cases, with an odds ratio of 0.91, 95% CI 0.85-0.98 (p = 0.009). The study thus did not find an absence of an association, but rather an inverse association.

Causality

Under the heading of “Meaning,” the authors wrote that “[t]hese findings suggest an association between exposure to Agent Orange and bladder cancer, although the clinical relevance of this was unclear.” Despite disclaiming a causal interpretation of their results, Williams and colleagues wrote that their results “support prior investigations and further support bladder cancer to be designated as an Agent Orange-associated disease.”

Williams and colleagues note that the Institute of Medicine had suggested that the association between Agent Orange exposure and bladder cancer outcomes required further research.[5] Requiring additional research was apparently sufficient for the Department of Veterans Affairs, in 2021, to assume facts not in evidence, and to designate “bladder cancer as a cancer caused by Agent Orange exposure.”[6]

Williams and colleagues themselves appear to disavow a causal interpretation of their results: “we cannot determine causality given the retrospective nature of our study design.” They also acknowledged their inability to “exclude potential selection bias and misclassification bias.” Although the authors did not explore the issue, exposed servicemen may well have been under greater scrutiny, creating surveillance and diagnostic biases.

The authors failed to grapple with other, perhaps more serious biases and inadequacy of methodology in their study. Although the authors claimed to have controlled for the most important confounders, they failed to include diabetes as a co-variate in their analysis, even though diabetic patients have a more than doubled increased risk for bladder cancer, even after adjustment for smoking.[7] Diabetic patients would also have been likely to have had more visits to VA centers for healthcare and more opportunity to have been diagnosed with bladder cancer.

Futhermore, with respect to the known confounding variable of smoking, the authors trichotomized smoking history as “never,” “former,” or “current” smoker. The authors were missing smoking information in about 13% of the cohort. In a univariate analysis based upon smoking status (Table 4), the authors reported the following hazard ratios for bladder cancer, by smoking status:

Smoking status at bladder cancer diagnosis

Never smoked      1   [Reference]

Current smoker   1.10 (1.00-1.21)

Former smoker    1.08 (1.00-1.18)

Unknown              1.17 (1.05-1.31)

This analysis for smoking risk points to the fragility of the Agent Orange analyses. First, the “unknown” smoking status is associated with roughly twice the risk for current or former smokers. Second, the risk ratios for muscle-invasive bladder cancer were understandably higher for current smokers (OR 1.10, 95% CI 1.00-1.21) and former smokers (OR 1.08, 95% CI 1.00-1.18) than for non-smoking veterans.

Third, the Williams’ study’s univariate analysis of smoking and bladder cancer generates risk ratios that are quite out of line with independent studies of smoking and bladder cancer risk. For instance, meta-analyses of studies of smoking and bladder cancer risk report risk ratios of 2.58 (95% C.I., 2.37–2.80) for any smoking, 3.47 (3.07–3.91) for current smoking, and 2.04 (1.85–2.25) for past smoking.[8] These smoking-related bladder cancer risks are thus order(s) of magnitude greater than the univariate analysis of smoking risk in the Williams study, as well as the multivariate analysis of Agent Orange risk reported.

Fourth, the authors engage in the common, but questionable practice of categorizing a known confounder, smoking, which should ideally be reported as a continuous variable with respect to quantity consumed, years smoked, and years since quitting.[9] The question here, given that the study is very large, is not the loss of power,[10] but bias away from the null. Peter Austin has shown, by Monte Carlo simulation, that categorizing a continuous variable in a logistic regression results in inflating the rate of finding false positive associations.[11] The type I (false-positive) error rates increases with sample size, with increasing correlation between the confounding variable and outcome of interest, and the number of categories used for the continuous variables. The large dataset used by Williams and colleagues, which they see as a plus, works against them by increasing the bias from the use of categorical variables for confounding variables.[12]

The Williams study raises serious questions not only about the quality of medical journalism, but also about how an executive agency such as the Veterans Administration evaluates scientific evidence. If the Williams study were to play a role in compensation determinations, it would seem that veterans with muscle-invasive bladder cancer would be turned away, while those veterans with less serious cancers would be compensated. But with 2.1% incidence versus 2.0%, how can compensation be rationally permitted in every case?


[1] Stephen B. Williams, Jessica L. Janes, Lauren E. Howard, Ruixin Yang, Amanda M. De Hoedt, Jacques G. Baillargeon, Yong-Fang Kuo, Douglas S. Tyler, Martha K. Terris, Stephen J. Freedland, “Exposure to Agent Orange and Risk of Bladder Cancer Among US Veterans,” 6 JAMA Network Open e2320593 (2023).

[2] Elana Gotkine, “Exposure to Agent Orange Linked to Risk of Bladder Cancer,” Buffalo News (June 28, 2023); Drew Amorosi, “Agent Orange exposure linked to increased risk for bladder cancer among Vietnam veterans,” HemOnc Today (June 27, 2023).

[3] Andrea S. Blevins Primeau, “Agent Orange Exposure Tied to Increased Risk of Bladder Cancer,” Cancer Therapy Advisor (June 30, 2023); Mike Bassett, “Agent Orange Exposure Tied to Bladder Cancer Risk in Veterans — Increased risk described as ‘modest’, and no association seen with aggressiveness of cancer,” Medpage Today (June 27, 2023).

[4] Darlene Dobkowski, “Agent Orange Exposure Modestly Increases Bladder Cancer Risk in Vietnam Veterans,” Cure Today (June 27, 2023).

[5] Institute of Medicine – Committee to Review the Health Effects in Vietnam Veterans of Exposure to Herbicides (Tenth Biennial Update), Veterans and Agent Orange: Update 2014 at 10 (2016) (upgrading previous finding of “inadequate” to “suggestive”).

[6] Williams study, citing U.S. Department of Veterans Affairs, “Agent Orange exposure and VA disability compensation.”

[7] Yeung Ng, I. Husain, N. Waterfall, “Diabetes Mellitus and Bladder Cancer – An Epidemiological Relationship?” 9 Path. Oncol. Research 30 (2003) (diabetic patients had an increased, significant odds ratio for bladder cancer compared with non diabetics even after adjustment for smoking and age [OR: 2.69 p=0.049 (95% CI 1.006-7.194)]).

[8] Marcus G. Cumberbatch, Matteo Rota, James W.F. Catto, and Carlo La Vecchia, “The Role of Tobacco Smoke in Bladder and Kidney Carcinogenesis: A Comparison of Exposures and Meta-analysis of Incidence and Mortality Risks?” 70 European Urology 458 (2016).

[9] See generally, “Confounded by Confounding in Unexpected Places” (Dec. 12, 2021).

[10] Jacob Cohen, “The cost of dichotomization,” 7 Applied Psychol. Measurement 249 (1983).

[11] Peter C. Austin & Lawrence J. Brunner, “Inflation of the type I error rate when a continuous confounding variable is categorized in logistic regression analyses,” 23 Statist. Med. 1159 (2004).

[12] See, e.g., Douglas G. Altman & Patrick Royston, “The cost of dichotomising continuous variables,” 332 Brit. Med. J. 1080 (2006); Patrick Royston, Douglas G. Altman, and Willi Sauerbrei, “Dichotomizing continuous predictors in multiple regression: a bad idea,” 25 Stat. Med. 127 (2006); Valerii Fedorov, Frank Mannino, and Rongmei Zhang, “Consequences of dichotomization,” 8 Pharmaceut. Statist. 50 (2009). See also Robert C. MacCallum, Shaobo Zhang, Kristopher J. Preacher, and Derek D. Rucker, “On the Practice of Dichotomization of Quantitative Variables,” 7 Psychological Methods 19 (2002); David L. Streiner, “Breaking Up is Hard to Do: The Heartbreak of Dichotomizing Continuous Data,” 47 Can. J. Psychiatry 262 (2002); Henian Chen, Patricia Cohen, and Sophie Chen, “Biased odds ratios from dichotomization of age,” 26 Statist. Med. 3487 (2007); Carl van Walraven & Robert G. Hart, “Leave ‘em Alone – Why Continuous Variables Should Be Analyzed as Such,” 30 Neuroepidemiology 138 (2008); O. Naggara, J. Raymond, F. Guilbert, D. Roy, A. Weill, and Douglas G. Altman, “Analysis by Categorizing or Dichotomizing Continuous Variables Is Inadvisable,” 32 Am. J. Neuroradiol. 437 (Mar 2011); Neal V. Dawson & Robert Weiss, “Dichotomizing Continuous Variables in Statistical Analysis: A Practice to Avoid,” Med. Decision Making 225 (2012); Phillippa M. Cumberland, Gabriela Czanner, Catey Bunce, Caroline J. Doré, Nick Freemantle, and Marta García-Fiñana, “Ophthalmic statistics note: the perils of dichotomising continuous variables,” 98 Brit. J. Ophthalmol. 841 (2014); Julie R. Irwin & Gary H. McClelland, “Negative Consequences of Dichotomizing Continuous Predictor Variables,” 40 J. Marketing Research 366 (2003); Stanley E. Lazic, “Four simple ways to increase power without increasing the sample size,” PeerJ Preprints (23 Oct 2017).