Nonsteroidal Anti-inflammatory Drugs and Risk of Esophageal and Gastric Cancer

The hypothesis that nonsteroidal anti-inflammatory drugs (NSAID) prevent gastrointestinal cancers is gaining support. Mechanisms involving NSAIDs inhibitory effect on the cyclooxygenase-2 enzyme's production of various prostaglandins, resulting in modulation of inflammation and immunoresponse, induction of cell apoptosis, and inhibition of angiogenesis have been suggested (1). The majority of epidemiologic studies have addressed colorectal cancer (2-10). The relation between NSAIDs and upper gastrointestinal (UGI) cancers has been less studied, and the results are partly conflicting. Two recent reviews conclude that, although the data are limited in several aspects, a synthesis of existing studies support an inverse association between use of aspirin or nonaspirin NSAIDs and risk of esophageal cancer (11, 12), equally strong in both main histologic types (11). Pooled estimates of seven case-control (13-19) and two cohort studies (20, 21) showed that “any” exposure to aspirin or nonaspirin NSAIDs carried a 43% reduced risk of esophageal cancer (11). All studies did not support a negative association, however (16). Furthermore, a large cohort study indicated an increased risk of esophageal carcinoma among patients with rheumatoid arthritis, who presumably used NSAIDs in a vast majority (22). Nevertheless, clinical trials testing the potential prevention of esophageal cancer by the use of NSAIDs have already been initiated (23). One randomized, double blind, placebo-controlled trial did not detect any regression in the severity of squamous cell dysplasia among users of celecoxib, a cyclooxygenase-2 enzyme inhibitor, in a population at high risk of esophageal squamous cell carcinoma (24). With regard to gastric cancer, case-control studies have reported a reduced risk in the range of 20% to 70% among users of NSAIDs (13, 14, 16, 17, 25, 26). A few cohort studies have reported results in support of a protective effect (20, 22, 27), whereas others have not (28, 29). Our aim was to test whether use of aspirin and nonaspirin NSAIDs reduces risk of esophageal or gastric cancer in a prospective, nested case-control study conducted in the General Practitioner Research Database (GPRD). This database has been used in a previous study (17) conducted between 1993 and 1995 in testing the association between NSAID and cancer. However, these two studies differ considerably in design with regard to study period, efforts of case validation and tumor classification, drug exposure definitions, and in the extent of adjustments for potential confounding factors, including UGI disorders.

The GPRD has been described in detail elsewhere (17, 30). In brief, this database consists of computer-based information from over two million patients visiting GPs in the United Kingdom. Prescriptions from participating GPs are computer linked to the GPRD. The GPs also prospectively supply information regarding the patients, including demographics, details from the visits, diagnoses from referrals to specialist doctors and hospital admissions, results from laboratory tests, and indications for new courses of drug therapy. A modification of the Oxford Medical Information System classification system is used to code diagnoses, and a dictionary based on data from the Prescription Pricing Authority is used to code drugs. Data are sent anonymously to the Medicines and Healthcare products Regulatory Agency, which organizes the information in the GPRD for research purposes. Validation studies of the GPRD have found that >90 % of all referrals are entered into the GPs computer with a diagnosis code that reflects the specialist's diagnosis (31-33). These validation studies did not address the completeness or misclassification of exposure variables, however. A recent study validated the information on tobacco smoking status in the GPRD finding that recordings of current and former use were 79% and 29%, respectively, of expected rates according to a population-based survey (34).

In the GPRD, we identified all patients ages 40 to 84 years during the study period, January 1994 through December 2001. Patients became members of the study cohort on the date when they met the criteria of at least 2 years of enrollment with the GP and at least 1 year of computerized prescription history. Patients with any cancer recorded in the GPRD before this start date were excluded.

The members of the study cohort were followed up until the earliest occurrence of any of the following end points: (a) detection of an esophageal or gastric cancer, (b) detection of any other cancer, (c) age of 85 years, (d) death, or (e) end of study period (December 31, 2001).

The computerized search of the study cohort identified 2,128 patients with a diagnosis code indicating an esophageal or gastric cancer during follow-up. One investigator (M.L.) reviewed all computerized patient profiles to verify and further classify all tumors. To avoid bias, all exposure information was removed before the start of this review. Patients were excluded from further analyses if (a) the tumor was benign; (b) the site of the primary cancer was unknown; (c) the tumor was a metastasis; (d) the patient had another, concurrent, cancer; (e) the cancer was first diagnosed before the start date in the study cohort; or (f) the histologic diagnosis was not adenocarcinoma or squamous cell carcinoma. For a sample of 1,280 (60%) consecutive case patients, identified by the computerized search, the GPs were asked to provide information on the site and histology of the tumor and to send available paper-based information (e.g., operation and pathology reports, letters from specialists) related to the diagnosis. We received complementary information on 993 (78%) of these patients, resulting in an additional exclusion of 2.7 % of the computer-detected cases. Because the effect of this additional exclusion review was negligible, we did not request complementary information from the GPs for the remaining computer-detected case subjects. Index date was set to be the date when the tumor was first recorded or when the manual review revealed an earlier date of diagnosis.

A date within the study period was generated at random for each of the members of the entire source cohort. If this random date was included in the eligible person-time for that cohort member, we used this random date as the index date and marked that person as an eligible control subject. Thereafter, a total of 10,000 controls, free of cancer, were randomly selected and frequency matched by sex, age (within 1 year), and same calendar year from the pool of eligible controls. The number of 10,000 controls was chosen as being close to five controls per case, a ratio considered to bring close to as much information as any other ratio greater than it.

Any drug exposure before index date was considered. We classified exposure for each drug into nonuse or ever use. Nonuse represented no recorded use and ever use was defined as any recorded use of that drug. Ever users were further grouped into current or past users. Current use was defined as when prescriptions for that drug were issued within the year before index date, whereas past use represents whenever the most recent prescription for that drug was issued at least 1 year before index date. Among current users, the duration of drug use was calculated adding the periods of “consecutive” prescriptions, defined as an interval of <6 months between two prescriptions of the same drug. Duration was grouped into three levels (<1, 1-3, or >3 years). The latter duration category was defined as long-term use. Daily dose of nonaspirin NSAIDs was categorized as low medium or high for each individual drug (see Appendix 1), whereas daily dose of aspirin was grouped into three levels (up to 75 mg, >75 but <300 mg, and ≥300 mg).

In cohort analyses, we calculated the incidence rates by dividing the total number of observed cases by the corresponding total person-time followed up in each age and sex specific strata. We used Poisson regression to estimate relative risk and 95% confidence interval (95% CI; ref. 35). In nested case-control analyses (i.e., the main analyses), we used unconditional logistic regression to calculate odds ratios (OR) with 95% CIs. In this study design, the OR is an unbiased estimator of the incidence rate ratio. All estimates of risk were adjusted for the following potential confounding factors: age (in 10-year categories), sex, calendar year, smoking status (categorized into nonsmokers, ex-smokers, current smokers, or unknown), alcohol use (categorized into 0-2, 3-16, 16-34, >34 units per day, or unknown), body mass index (BMI) in kg/m² (categorized into <20, 20-25, 26-30, >30, or unknown), and a history of UGI disorders (yes or no). UGI disorders was defined as any of the following events recorded up to 6 months before index date: gastroesophageal reflux disease, peptic ulcer, dyspepsia, or prescription of acid suppressing drugs (i.e., proton pump inhibitors or H₂-receptor blockers). Stratified analyses were done by sex, age (<70, or >70 years), UGI disorders (yes or no), and by use of health services during the 2 years before index date. The use of health services was defined as recorded visits to the GP, referrals, or hospitalizations. To further evaluate the influence of protopathic bias (i.e., reversed causality due to that the use of a drug might be determined by early symptoms of a, yet undetected, cancer), all main analyses were recalculated using 2 years of lag time (i.e., only including information recorded at least 2 years before index date).

The Scientific and Ethical Advisory Group in the United Kingdom approved the study.

Among 4,340,207 person-years of follow-up, we identified 2,128 patients with esophageal or gastric cancer and 10,000 control subjects. The manual review revealed 85 patients with a concurrent cancer or a cancer diagnosed before the start of the study, 38 patients with other or unknown site of their primary cancer, 28 patients with a benign esophageal or gastric tumor, and 27 patients with a tumor of other histology than adenocarcinoma or squamous cell carcinoma. After exclusion of these 178 patients (8%), 1,950 case patients remained for final analysis. The manual review rendered a change of the tumor site from esophageal to gastric in 74 (8%) patients and from gastric to esophageal in 14 (2%) patients. Thereafter, the tumor site was classified as esophageal in 909 patients, as gastric in 1,023, whereas in 18 patients it was impossible to discern whether the site was esophageal or gastric. The median time of follow-up between first drug prescription and index date was similar between cases (2,235 days) and controls (2,200 days). As presented in Table 1, the distribution of potential confounding variables was even among case patients and control subjects.

The crude incidence rates for esophageal cancer was 20.9/100,000 person-years and for gastric cancer 23.6/100,000 person-years. Age- and sex specific relative risks of esophageal and gastric cancer are presented in Table 2. The age and sex distributions were similar between patients with esophageal and gastric cancer. There was a clear male predominance and an increasing incidence by age in both these types of cancer (Table 2).

The overall analyses revealed that use of nonaspirin NSAIDs was inversely associated with risk of both esophageal and gastric cancer (Table 3). Long-term users of nonaspirin NSAIDs were at 18% reduced risk of esophageal cancer (OR, 0.82; 95% CI, 0.57-1.18) and a 35% reduced risk of gastric cancer (OR, 0.65; 95% CI, 0.44-0.94). No clear dose-response relation was found. Among long-term users who had used low-medium and high daily doses, respectively, ORs were 0.70 (95% CI, 0.42-1.18) and 0.96 (95% CI, 0.59-1.55) for esophageal cancer, and 0.53 (95% CI, 0.31-0.92) and 0.78 (95% CI, 0.47-1.27) for gastric cancer. There was no apparent difference in OR between adenocarcinoma (OR, 0.59; 95% CI, 0.29-1.22) and squamous cell carcinoma (OR, 0.59; 95% CI, 0.21-1.64) of the esophagus among long-term users, whereas in the group of esophageal cancers without a defined histologic classification in the database there was no association (OR, 1.02; 95% CI, 0.65-1.59). No important changes of the results were found when we repeated all basic analyses using duration of drug exposure as the total cumulative time of drug use (data not shown). The results were not materially altered after exclusion of persons who also had used aspirin (data not shown).

In analyses of exposure recorded at least 2 years before index date (i.e., 2 years lag time), the association between nonaspirin NSAIDs and risk of esophageal and gastric cancer were virtually unchanged among ever users (data not shown). Long-term use of nonaspirin NSAIDs was not associated with esophageal cancer (OR, 0.98; 95% CI, 0.66-1.43), whereas the inverse association remained for gastric cancer (OR, 0.59; 95% CI, 0.37-0.92).

Users of aspirin with >3 years of duration were associated with a decreased risk of esophageal cancer (OR, 0.76; 95% CI, 0.53-1.08), but no reduction in risk was found for gastric cancer (Table 3). Among persons who used daily doses of 75, 150, and ≥300 mg, ORs for esophageal cancer were 0.79 (95% CI, 0.47-1.33), 0.86 (95% CI, 0.46-1.61), and 0.60 (95% CI, 0.29-1.23), respectively. Among long-term users, there were no major differences between esophageal adenocarcinoma (OR, 0.75; 95% CI, 0.41-1.36) and squamous cell carcinoma (OR, 0.87; 95% CI, 0.35-2.16). No important changes of results were found in analyses using duration as the total cumulative time of drug use (data not shown). Furthermore, the results did not change importantly after exclusion of nonaspirin NSAIDs users (data not shown). In analyses with 2 years lag time, current use of aspirin was associated with a decreased risk of esophageal cancer (OR, 0.75; 95% CI, 0.58-0.97) but not gastric cancer (OR, 1.17; 95% CI, 0.95-1.45). Among long-term current users of aspirin, no association with esophageal cancer (OR, 0.99; 95% CI, 0.65-1.51) or gastric cancer (OR, 1.04; 95% CI, 0.71-1.54) was observed.

A history of UGI disorders increased the risk of both esophageal cancer (OR, 1.20; 95% CI, 1.04-1.39) and gastric cancer (OR, 1.69; 95% CI, 1.48-1.92) after adjustment for all other potential confounding variables listed in the Materials and Methods section. In analyses with 2 years lag time, there were still increased risks of both esophageal cancer (OR, 1.19; 95% CI, 1.02-1.38) and gastric cancer (OR, 1.27; 95% CI, 1.11-1.46). Among control subjects, UGI disorders was linked with increased use of both nonaspirin NSAIDs and aspirin (Table 4). When we excluded adjustment for UGI disorders in the model, the risk estimates only increased slightly (data not shown). When stratifying the analysis by UGI disorders, ORs for both esophageal and gastric cancer changed compared with the overall analysis (Table 5). Using data without lag time, we found that among subjects without UGI disorders, the negative association slightly approached the null among nonaspirin NSAIDs long-term users, whereas among subjects with UGI disorders, the negative association was reinforced (Table 5). This finding was more pronounced when using the 2 years lag time data and extended to encompass also the association with aspirin long-term use (Table 5). However, the differences between the two strata of UGI disorder, in their association to use of NSAIDs, never reached the level of statistical significance. Analyses of each separate UGI disorder (i.e., gastroesophageal reflux disease, peptic ulcer, dyspepsia, or prescription of acid suppressing drugs) rendered similar results as the composite variable (data not shown).

Stratification by sex and age did not reveal any clear effect modification (data not shown). The number of GP visits during the 2 years before index date revealed weakly increased risks of esophageal and moderately increased risks of gastric cancer. Using zero to two GP visits as reference, the ORs for esophageal and gastric cancer, respectively, were in three to five GP visits 1.27 (95% CI, 1.01-1.59) and 1.56 (95% CI, 1.22-1.99), and more than five GP visits 1.25 (95% CI, 1.03-1.53) and 1.86 (95% CI, 1.50-2.29). After stratification by GP visits (0-5 versus >5), a weak negative interaction, when combining use of analgesics and more frequent visits was found (data not shown).

This study identifies an inverse association between long-term use of nonaspirin NSAIDs and risk of gastric cancer, which seemed to be stronger among subjects with a history of UGI disorders. No other associations were firmly identified after adjustment for potential confounding, neither with nonaspirin NSAIDS nor aspirin.

The main advantages of our study are the large sample size and the prospectively assessed exposure data, with detailed information on dates and types of drug use and medical data, including UGI disorders. The recordings in the database are based on the GPs' routine medical care. This renders a degree of missing values and misclassification that, in some of the potential confounders, could differ between cases and controls. We found that the missing values were evenly distributed between cases and controls, however. This implies that the exposure information was equally good for cases and controls. Furthermore, because the information regarding potential confounders was recorded prospectively, before the occurrence of any cancer, any possible exposure misclassification should mainly be nondifferential and could, at the most, slightly dilute the risk estimates. A potential weakness of our study is the case ascertainment, because there was no cancer registry for identification of case patients. Previous studies of the diagnoses assessed in the GPRD have shown good validity, however (32, 33). Furthermore, our manual review of all computer-selected cases together with the review of additional information provided by the GP in a large sample of cases should have reduced tumor misclassification. Another source of error was that the GPRD does not contain adequate information regarding some potential confounding factors, including dietary habits, education, or familiar antecedents. However, it is unlikely that any of these factors could explain our results, as they will most likely be shared equally between users and nonusers of NSAIDs. We did not have access to Helicobacter pylori status, which possibly could have an influence on the use of NSAIDs. On the other hand, we had access to data regarding several other, and possibly more important variables, including UGI disorders. We were unable to capture exposure to widely available over-the-counter anti-inflammatory drugs, but the impact of this possible error has previously been reported to be negligible, especially when the exposure of interest is long term (36). Another limitation is that the computerized database started in the late 1980s and therefore lacks information before that period. Consequently, the lag-time analyses were inherently less precise when evaluating the role of long-term duration of drug use. Furthermore, our definitions of current use and long-term use are arbitrary and the data collection regarding the long-term usage (i.e., current use of at least 3 years) probably includes many subjects with <5 years of use, which might dilute the group of “true” long-term users. Yet, this should not interfere importantly with our results, because consistency has been high among previous studies of the association between NSAIDs and gastrointestinal cancers despite major heterogeneous exposure definitions (12). We also assessed, in a subanalysis, the role of total cumulative exposure of NSAIDs (aspirin and nonaspirin NSAIDs), without finding any important differences with the main analysis. Similarly, the GPRD lacks information regarding the presence of medical disorders (e.g., UGI disorders) occurring earlier than the beginning of the computerized database, which might introduce residual confounding.

Although data are somewhat conflicting, the majority of previous studies support an inverse association between use of aspirin or nonaspirin NSAIDs and risk of esophageal and gastric cancer (13, 14, 16-18, 20-22, 25-27). In most studies, personal interviews or self-administered questionnaires were employed, however (14-16, 19, 25, 26). Problems with such retrospective data collection are nondifferential misclassification, and more importantly, differential misclassification. Retrospective data collection is particularly prone to recall bias, because it relies heavily on the subject's ability to recall past drug exposure, especially when duration of use is the main interest. This ability of recalling might be influenced by the newly verified cancer among cases. Most of our risk estimates did not bear statistical significance, which could be due to insufficient numbers of subjects, but the point estimates in our study were in line with previous results. However, the protective effects weakened when restricting the analysis to persons without UGI disorders. Only three previous studies attempted to adjust their results for UGI disorders (14, 25, 26). In a U.S. population-based case-control study, there was no change in the protective effect of aspirin or NSAIDs on the risk of esophageal or gastric cancer in analyses restricted to subjects without UGI disorders (14). Two case-control studies of aspirin and gastric cancer in Russia (25), and in Sweden (26), adjusted for UGI disorders and found that the negative association was compelled to H. pylori–positive subjects, however.

Appropriate adjustment for UGI disorders should be of great importance when investigating these associations, because patients who experience early symptoms from a yet unrecognized UGI tumor may increase or decrease their use of NSAIDs (i.e., protopathic bias; refs. 11, 12, 23, 37). Furthermore, it might be crucial to consider interaction between NSAIDs and known risk factors for the occurrence of esophageal or gastric cancer (i.e., reflux in esophageal adenocarcinoma; ref. 38), or H. pylori infection and peptic ulcer disease in gastric cancer (39). Hence, the indication (or the contraindication) for treatment might act as a confounder. These types of biases are common sources of error in pharmacoepidemiology (40-42), a problem that has not been satisfactory addressed in the previous literature (11, 12, 23, 37). Our data indicate that a history of UGI disorders is an independent risk factor for esophageal and gastric cancer. Moreover, in agreement with previous studies (43-47), such disorders were linked with the use of analgesics among our controls, representing the source population. Hence, if UGI disorders occur before NSAID use, they may act as confounders (42). Therefore, we attempted to adjust for UGI disorders in the analyses by both logistic regression, stratification, and by changes of lag time. Because we could not firmly establish the temporal relation between NSAID use and UGI disorders, the divergent results after stratification are complex to interpret. There are at least four alternative explanations. First, the use of NSAIDs might reduce the progression from UGI disorders to esophageal and gastric cancers (i.e., UGI disorders are in the causal chain). This would imply a true protective effect of NSAIDs, stronger among subjects with UGI disorders. Second, patients with a yet unrecognized esophageal or gastric cancer might be less prone to use NSAIDs because of early symptoms (UGI disorders) from their cancer (i.e., reversed causality). Third, UGI disorders might be affected by prior NSAID use and therefore stratification may have introduced selection bias (48, 49). Fourth, certain UGI disorders (reflux and gastric ulcer) are well established independent risk factors of esophageal and gastric cancer, respectively. It is also well known that NSAIDs may cause several UGI side effects (e.g., gastric ulcer and dyspeptic symptoms), usually leading to contraindication for NSAIDs prescription. If contraindication was more likely for cases than controls, as cases were more likely to have a history of UGI disorders than controls, the reduced ORs among persons with UGI disorders, but not among those without UGI disorders, could be an artifact and occur even in the absence of causal association between NSAID use and cancer risk (i.e., confounding by contraindication).

In conclusion, our large, prospective nested case-control study provides further indication of a protective effect of nonaspirin NSAIDs on the risk of gastric cancer but still leaves an unresolved question on all other studied associations. Our results suggest that UGI disorders could distort the associations, although we could not show this with statistical significance. If such bias was true, the previously reported inverse associations might, at least partly, be explained by lack of appropriate adjustment for such disorders. Further studies that could analyze with validity the exact timing and interplay between NSAIDs, UGI disorders and these tumors are warranted before any firm conclusions can be drawn.

We thank all collaborators that participate in the GPRD for making this study possible; the valuable statistical and epidemiologic support from Weimin Ye at Karolinska Institutet; and Ken Rothman at Boston University, Hans-Olov Adami at Karolinska Institutet, and Miguel Hernan at Harvard University who have contributed with valuable input into the study.