Acid Suppression and Esophageal Adenocarcinoma: A Critique of Causal Inference

1. Introduction

The annual incidence of esophageal adenocarcinoma (EAC) increased in the United States by more than 750% between 1973 and 2017, the highest growth rate of any epithelial malignancy during that period; melanoma ranked second at approximately 275% [1]. EAC is thought to arise from chronic exposure of the distal esophagus to refluxed gastric and duodenal contents (e.g. acid, bile), and gastroesophageal reflux disease (GERD) is an established risk factor in EAC etiology [2]. Approximately 10% of those with reflux symptoms who undergo upper endoscopy are found to have Barrett’s esophagus (BE), the precursor lesion to EAC [3], and 0.1-0.5% of those with BE will progress to EAC in any given year, with a higher progression rate among those with dysplasia [4]. Though trends in obesity are thought to partially explain the increased prevalence of GERD [5], it is generally unclear why certain GERD patients develop BE/EAC while others do not [3].

Symptoms of GERD are commonly treated with acid-suppressive therapy (AS), traditionally through the use of proton pump inhibitors (PPIs) and H2-receptor antagonists (H2RAs) [6]. Despite evidence that acid can damage DNA [7], the experimental literature has proposed several hypotheses by which the use of AS may nevertheless increase the risk of EAC. These hypotheses suggest that by increasing gastric pH, AS modulates the composition [8,9,10] and solubility [8,11,12,13] of specific reflux constituents implicated in EAC pathogenesis [9,11,12,14], and separately promotes the progression to EAC by enhancing proliferation in reflux-conditioned tissue [15,16,17,18]. There is moreover a cluster of in vivo studies reporting higher rates of EAC, dysplasia, or BE under less acidic conditions [9,11,15,19,20,21], and although the results of these studies are heterogeneous [22,23], reviews of the animal literature have suggested that the use of AS may explain the rise in EAC incidence [10,12].

Observational studies often report a positive association between EAC and the use of AS [24,25,26,27,28], but clinical guidelines have long maintained that this association reflects confounding by GERD, whereby acidic reflux promotes BE/EAC and the use of AS is a benign marker of risk [29,30]. This interpretation [29,30] appears to be largely informed by the findings of a small collection of studies that estimate either the AS–EAC association among those without reflux-related indications [26] or the common treatment odds ratio (OR) after adjusting for predictors of reflux burden (e.g. GERD diagnosis, hiatal hernia, BMI) as main-effect covariates [25,26,31]. Though these estimates are elevated in general, their statistical nonsignificance appears to have been widely read as reassuring evidence that the AS–EAC association is non-causal [29,30,32], including in current clinical guidelines on the management of GERD [30].

This paper does not argue that AS contributes to EAC risk in human populations but rather that the findings of these observational studies have been interpreted within an inferential framework that is misaligned with the experimental literature. The statistical non-significance of the estimates reported in these studies may be read [29,30,32] as evidence against a GERD-independent causal effect under a loose evidentiary standard. However, the hypotheses of the experimental literature do not posit the existence of such an effect: they presuppose that the patient’s underlying reflux burden determines the duration of esophageal exposure to refluxate that is modified by AS, or that it determines the cellular context (e.g. mutational burden, receptor expression) in which the hormonal effects of AS are exerted. In this setting, the findings of these observational studies do not inherently distinguish confounding by GERD from a causal signal, and they cannot be read as reassuring evidence against causality.

2. Reflux Burden as an Effect Modifier

In a 2015 review of rat reflux models, Greene et al. [12] conclude that the suppression of gastric acidity in rats appears to “potentiate the development of esophageal adenocarcinoma" [12] and point to two mechanisms articulated in the reviewed studies. The first concerns the pH-dependent toxicity of bile [9,11], a reflux constituent that is suspected to be a central driver in BE/EAC pathogenesis [33,34]: while many bile acids precipitate irreversibly in normal gastric acidity [35,36], weakly acidic conditions maintain many bile species in a soluble state from which they can passively diffuse through cell membranes [13,37,38]. The second mechanism concerns the bacterial reduction of nitrate to nitrite, which Fein et al. [9] had speculated to facilitate the formation of carcinogenic N-nitroso compounds.

These hypothesized mechanisms are illustrated in Figure 1. Under these mechanisms, the use of AS alters the carcinogenicity of refluxate, but (treated) reflux burden determines the magnitude of esophageal exposure to that refluxate. As such, no causal effect exists in the absence of reflux, and the magnitude of this effect on risk generally increases with (treated) reflux burden. Several other proposed mechanisms can be represented similarly [8,10,39].

Next, consider the subsequent mouse model of Lee et al. [15], in which the authors examine the hypothesis that PPI-induced hypergastrinemia could promote the progression from BE to EAC [17,18]. The use of AS promotes secondary hypergastrinemia [40], and gastrin can function as a pro-oncogenic stimulus [16,17,41]. Separately, reflux-related inflammation damages DNA [42] and appears to enhance tissue responsiveness to gastrin through upregulation of the receptor CCK2R [15,18]. Consistent with an interactive effect, Lee et al. observe that PPI-induced hypergastrinemia promotes the development and dysplastic progression of Barrett’s-like metaplasia in chronically inflamed mouse esophagi, but do not observe any metaplastic alteration in mouse esophagi that lack the inflammatory stimulus [15].

This hypothesized mechanism is illustrated in Figure 2. Under this mechanism, the use of AS promotes hypergastrinemia, but the carcinogenic susceptibility to this pro-oncogenic stimulus depends on the prior accumulation of reflux-related cellular changes, namely CCK2R overexpression and mutational burden. Accordingly, the effect on EAC risk generally increases with (historical) reflux burden, and unless there is a widespread alternative source of esophageal injury that mimics the cellular context produced by chronic reflux—a source that does not appear to exist [12,43]—then virtually no effect on risk manifests in patients who do not experience reflux.

3. Epidemiologic Findings and Their Interpretations

The causal hypotheses proposed in the experimental literature presuppose that the effect of AS on EAC risk would depend on reflux burden, yet the source of the AS–EAC association seems to have been adjudicated within a framework that omits this possibility. In what appears to be regarded as the most probative observational study on the source of the AS–EAC association [29,30,32], Garcia Rodriguez et al. [26] make this framework explicit: “Long term pharmacological gastric acid suppression is a marker of increased risk... However these associations are most likely explained by the underlying treatment indications being a risk factor for the cancer rather than an independent harmful effect of these agents per se." Within this binary, GERD-independent causation is the only form of causation, and it would therefore suffice to observe evidence for the absence of a GERD-independent causal effect in order to infer the absence of causality in general. In practice, this implies that a near-null AS–EAC association among those without GERD would be sufficient to determine that an elevated association among those with GERD is non-causal.

Garcia Rodriguez et al. [26] distinguish long-term (3+ years) AS users with esophageal indications (e.g. reflux symptoms, esophagitis, BE) from long-term users with non-esophageal indications (e.g. peptic ulcer disease, gastritis) and estimate separate ORs for each group computed against a common referent of non-users. The authors observe that the OR is statistically significant for the esophageal group (OR: 5.42; 95% CI: [3.13, 9.39]) and statistically non-significant for the non-esophageal group (OR: 1.74; 95% CI: [0.90, 3.34]), and Spechler et al. [29] present this finding in a technical review for the American Gastroenterological Association (AGA):

PPI therapy also has effects that, conceivably, might promote the development of cancer in Barrett’s esophagus. For example, use of PPIs often is associated with an increase in the serum levels of gastrin, a hormone that has been shown to increase proliferation in Barrett’s epithelium... Garcia Rodriguez et al. found that patients who were treated with acid suppression for an ‘esophageal indication’ such as GERD had a significantly increased risk of developing esophageal adenocarcinoma… In contrast, for patients who were treated with acid suppression for a ‘gastroduodenal indication’ such as peptic ulcer disease, there was no significantly increased risk of adenocarcinoma... The lack of an association with cancer in patients taking PPIs for gastroduodenal disease suggests that the positive association in the patients with esophageal disease resulted from confounding by indication. In other words, it was likely the GERD, not the GERD treatment, that increased the incidence of cancer.

Spechler et al. appear to recognize what they consider to be evidence against an independent causal effect and subsequently infer the presence of confounding by GERD. However, this inference does not follow from the premise that the authors themselves articulate: if AS-induced hypergastrinemia promotes the progression from BE to EAC by increasing proliferation in BE [18,41]—a tissue context characterized by heightened responsiveness to gastrin [15,41] and a substantial mutational burden [44]—then one would only expect to observe an effect among those with BE or perhaps the GERD that causes it [3]—both are esophageal indications [26].

More generally, the stratification performed by Garcia Rodriguez et al. creates a contrast between those who would tend to have a pathological reflux burden and those who would not. Because the non-esophageal group’s OR (1.74; 95% CI: [0.90, 3.34]) is perfectly compatible with hypotheses that predict a relatively small effect on risk in those with small reflux burdens (or no effect in those without large reflux burdens), it would not by itself weigh against the causal hypotheses of the in vivo literature. Moreover, the group ORs are structured so that any of the causal hypotheses would predict a greater OR for the esophageal group than for the non-esophageal group. Let $O_r$ be the EAC odds for the common untreated reference group, let $p_e$ and $p_n$ be the baseline risks for the esophageal and non-esophageal groups, respectively, and let ${\delta }_e$ and ${\delta }_n$ be the change in absolute risk attributable to treatment. The OR for group i is given by:

$$O{R}_i={\displaystyle \frac{O_i}{O_r}}={\displaystyle \frac{p_i+{\delta }_i}{(1-p_i-{\delta }_i)O_r}}.$$

As GERD is an established risk factor for EAC [2], we can assume that $p_e>p_n$, and because the typical reflux burden is larger in the esophageal group than in the non-esophageal group, each of the mechanisms through which AS has been hypothesized to contribute to EAC risk would increase risk such that ${\delta }_e>{\delta }_n$. It then follows straightforwardly that each of these hypothesized mechanisms would be expected to produce a causal signal where $O{R}_e>O{R}_n$, and we cannot therefore infer from the non-esophageal group’s OR that the esophageal group’s OR reflects confounding by GERD as opposed to a causal effect.

A similar interpretive constraint applies in the context of main-effect regression adjustment [24,25,26,28,45]. As an illustrative example, consider Chow et al. [25], in which the authors report that the OR for 4 or more H2RA prescriptions attenuates from 4 (95% CI: [1.3, 12.4]) to 1.5 (95% CI: [0.4, 5.4]) after adjusting for the presence of reflux-related conditions as main-effect covariates, while also reporting “no excess risk” among those without said conditions. Both the authors and the AGA technical review quoted above interpret these findings as affirmative evidence against causality, but this could only be the case if a causal signal were expected in those without reflux-related conditions and the unreported treatment OR among those without said conditions were incompatible with even a small effect on risk.

Alternatively, the common GERD-adjusted treatment OR would not itself weigh against a causal effect as—in addition to being elevated—it may still understate the treatment OR among those with reflux-related conditions, i.e. those for whom AS would be predicted to have the greatest effect on risk. If we take the authors’ words to imply that the treatment OR among those without reflux-related conditions is smaller than 1.5, it would at least be suggestive of a larger treatment OR within some subset of individuals with reflux-related conditions, and the authors do not present any affirmative evidence suggesting that this elevated treatment OR would be non-causal. This argument also applies to the GERD-adjusted treatment ORs presented by Garcia Rodriguez et al. [26], as well as to the estimates presented in subsequent studies [45,46,47].

4. Discussion

In a context where positive but statistically non-significant associations are read as null associations [25,29,30], a causal signal may only be distinguishable among those with pathological reflux. That said, the interpretability of a treatment OR in some high-reflux stratum is still limited by the available proxy for reflux burden. A GERD diagnosis can itself correspond to a wide range of underlying reflux burdens [48] such that a positive treatment OR within the GERD-positive stratum may still reflect confounding by GERD. Self-reported symptom severity can instead be stratified finely [31], but it is inherently influenced by treatment [49] such that the group with the greatest underlying reflux burden may not be identifiable.

A distinct but closely related subliterature examines whether PPIs may be chemopreventive among those who have already progressed to BE, consisting primarily of observational studies that compare PPI users with non-users (including H2RA users) and a randomized controlled trial (AspECT) [50] that compares high and low doses of esomeprazole [51,52]. Whether epidemiologic evidence concerning PPI chemoprevention among those with BE can adjudicate the source of the AS–EAC association observed in the general population requires distinguishing three questions: first, whether the evidence genuinely supports a chemopreventive effect of PPIs among those with BE; second, whether any such protectiveness would generalize to those who have not yet developed BE; and third, whether such evidence bears on the safety or protectiveness of H2RAs.

Regarding the first question, a recent meta-analysis reported an adjusted OR of 0.46 (95% CI: [0.25, 0.86]) for PPI use among those with BE, but it also reported significant heterogeneity ($I^2=92.15\%$) and determined that the results of each observational study were at risk of either serious or critical bias (ROBINS-I) [51]. While the AspECT trial observed a trend towards superior EAC-related outcomes within the high-dose group [50], this would only necessarily constitute evidence of chemoprevention (and safety) if the safety of the lower dose is established elsewhere or if the PPI dose-risk curve is monotonic. This monotonicity is not guaranteed because the toxicity of bile acids is non-monotonic across the acidic pH spectrum [37,38] and PPIs possess pleiotropic effects [22,53,54,55]. PPI pleiotropy has been discussed within the in vivo literature [12,22] and could plausibly explain why the AspECT trial’s primary endpoint was substantially driven by non-EAC mortality, which was lower in the high-dose group [50].

As for the second question, if PPIs were to exert a protective effect in BE via increased gastric pH [29] or reduced refluxate volume [11], then there would not be an obvious reason as to why this protectiveness would be stage specific. If this protectiveness instead reflected a pleiotropic effect that manifests at a later stage of carcinogenesis [54,55], then protectiveness among those with BE would still be compatible with PPIs contributing to both BE pathogenesis and marginal EAC risk. To the third and final question, the protectiveness of PPIs would not speak to the safety of H2RAs insofar as this protective effect could reflect PPI-specific pleiotropy [22,53,54], a reduction in refluxate volume that is not achieved by H2RAs [56], or a more complete and prolonged suppression of gastric acidity [57] that maintains certain bile species in an ionized state [8,37] for a greater duration than do H2RAs.

To conclude, this paper does not argue that AS contributes to EAC risk in human populations but that the presumption to the contrary may be unwarranted. The in vivo literature establishes the biological plausibility of a causal role for AS in EAC etiology [11,12,15,19], and the epidemiologic findings that had been thought to best adjudicate the causal question do not do so [25,26]. Until the risk associated with untreated reflux has been adequately separated from the treatment that it precipitates, the AS–EAC association should not be dismissed as spurious.