In contrast, a previous report from the Physicians' Health Study [18] found no association between aspirin and colorectal cancer after 5 years. In this study, assignment to aspirin was randomized; this avoided bias due to confounding, including confounding from factors influencing self-selection for aspirin consumption. In addition to the low dose of aspirin used, another limitation in the interpretation was the relatively short duration of the trial, which resulted from early termination of the aspirin arm. This was especially intriguing because a possible but nonsignificant trend (P = 0.09) from higher incidence of colon cancer at the beginning of the trial toward lower rates at the end of the randomized phase of the trial was seen in the aspirin arm.

Our current analysis extends the follow-up period to a mean of 12 years and reports on 341 incident cases of colorectal cancer, including 118 previously reported cases [18]. Our primary goal was to examine whether 4 to 6 years of regular aspirin use reduced the incidence of colorectal cancer over a 12-year period. A second goal was to examine factors influencing the choice to take aspirin regularly after the end of the randomized phase of the trial; if these factors were possible risk factors for colorectal cancer alone or in combination and were not properly controlled for, they might be sources of bias in observational studies. This may shed light on the discrepancies between findings from observational studies and those from randomized trials. A third goal was to examine 1) whether self-selected regular aspirin use during the year before the 7-year questionnaire was administered reduced the incidence of colorectal cancer over the 5-year period after the questionnaire was administered and 2) whether this effect was more pronounced in persons who had been randomly assigned to aspirin (because they received aspirin for about 5 years longer than did persons randomly assigned to placebo).

Methods

Top Methods Results Discussion Author & Article Info References

Study Sample

The Physicians' Health Study was designed to test two hypotheses: 1) that 325 mg of aspiring taken on alternate days reduces risk for cardiovascular disease and 2) that 50 mg of ß-carotene taken on alternate days decreases the incidence of cancer. The population and methods of this study are described in detail elsewhere [19, 20]. In brief, U.S. male physicians 40 to 84 years of age who did not regularly use aspirin or other NSAIDs and did not have a history of myocardial infarction, stroke, cancer, liver or renal disease, gout, peptic ulcer, or contraindications to aspirin were eligible for the study and were enrolled in a 3-month run-in phase during which they received aspirin. During the run-in phase, 22 071 men (66%) followed the study regimen, remained willing and eligible to participate, and were randomly assigned to receive 325 mg of aspirin or placebo and 50 mg of ß-carotene or placebo every other day in 1982. These men formed the study sample. The aspirin arm of the study was terminated early (in January 1988), after an average follow-up period of 5 years, primarily because a 44% reduction in risk for a first myocardial infarction was seen in the aspirin group. The ß-carotene arm of the study was continued until its scheduled completion in December 1995; after the end of the aspirin arm of the study, participants were free to choose between receiving aspirin or receiving placebo in their study calender packs along with their randomly assigned ß-carotene treatment.

Assessment

Twice in the first year and once yearly thereafter, participants were sent monthly calendar packs that contained their study medication. They also completed a short questionnaire that asked about compliance and the occurrence of relevant health outcomes. From 1982 to 1995, 99.7% of participants provided data on morbidity; mortality follow-up was 100% [20]. All reports of cancer (except nonmelanoma skin cancer) were followed up by a request to review medical records, including pathology reports. Medical records were reviewed by a committee of physicians who were blinded to treatment assignments. For this analysis, all confirmed incident cases of colorectal cancer and all cases of colorectal cancer reported until 25 October 1995 were used. The pathology reports for all colorectal cancer cases were reviewed by one of the authors, and the stage of disease at the time of diagnosis was assessed according to the modified Dukes classification [21].

Because the aspirin arm of the randomized trial ended on 25 January 1988, several definitions of aspirin exposure and follow-up are possible (Figure 1). First, we used the initial aspirin and placebo groups as assigned by randomization and compared the incidence of colorectal cancer over a 12-year follow-up period in an intention-to-treat analysis. Second, we used information on the frequency of self-selected aspirin use between year 6 and year 7 (after the end of the aspirin arm of the randomized trial) in a cohort of 20 396 physicians who were alive and free of cancer at the time of the year 7 questionnaire and looked at factors influencing the choice to use aspirin more frequently. Finally, we used an observational analysis to compare the incidence of colorectal cancer according to frequency and duration of aspirin use in this cohort in the following 5 years.

View larger version (14K): [in this window] [in a new window]   Figure 1. Design of study of aspirin use and colorectal cancer: post-trial follow-up data from the Physicians' Health Study.

Aspirin use from year 6 to year 7 was calculated by using information given on the year 7 questionnaire, which asked about the number of pills not taken from the calendar packs and the days of aspirin use outside of the study protocol during the previous 12 months. The overall number of days of aspirin use was calculated from the number of days on which participants took pills from their calendar packs, the contents of the calendar packs (aspirin or placebo), and the number of days on which participants reported taking additional aspirin or medication containing aspirin. One hundred fifteen participants (0.6%) were excluded from these analyses because their frequency of aspirin use could not be calculated. Aspirin use per year was categorized as less than 1 day per week (<52 days), 1 to less than 3 days per week (52 to <150 days), and 3 or more days per week ( 150 days). To obtain a frequency of use similar to that during the previous randomized treatment period, the first two categories were collapsed. This resulted in a dichotomy: Patients used aspirin on 3 or more days per week or on less than 3 days per week. To assess the effect of the duration of aspirin use, participants who chose to use aspirin more frequently and those who chose to use aspirin less frequently between year 6 and year 7 were grouped according to whether they had been randomly assigned to receive aspirin in the previous 5 years. Aspirin use after the year 7 questionnaire was not considered.

For the prediction of self-selected aspirin use, reports of gastrointestinal symptoms (such as those suggestive of gastritis, peptic ulcer, nausea, constipation, and diarrhea), other symptoms (such as headache), and use of medications (such as warfarin) were taken from the year 6 questionnaire to ensure that symptom reports preceded assessment of aspirin use. For the same reason, reporting of gastrointestinal diagnoses and procedures (such as peptic ulcer, gastrointestinal bleeding, and cholecystectomy) and other diagnoses (such as coronary artery disease [International Classification of Diseases (ICD) codes 410-414], ischemic stroke [ICD code 434 or 436], hemorrhagic stroke [ICD code 430 or 431], and transient ischemic attack [ICD code 435]) was restricted to the period before the assessment of aspirin use. Alcohol use was categorized as daily, weekly but less than daily, and less than weekly use. Hypertension was defined as a systolic blood pressure of 160 mm Hg or more, a diastolic blood pressure of 95 mm Hg or more, or current use of antihypertensive agents. Smoking status was derived from information on the year 5 questionnaire and was categorized as current, past, or never. Information about the usual frequency of vigorous exercise was obtained from the year 3 questionnaire or, if it was missing from the year 3 questionnaire, from baseline. Body mass index (body weight in kilograms divided by squared height in meters) was calculated from baseline information on weight and height.

Statistical Analysis

Person-time was censored at the time of an unrefuted report of cancer, death, or the last reported case of colorectal cancer (22 August 1995), whichever was earliest, for all participants. We used Cox proportional-hazards models to calculate incidence rate ratios and 95% CIs for colorectal cancer according to whether participants had been randomly assigned to receive aspirin for the total follow-up period and for consecutive 2-year periods. An interaction term with length of follow-up was used to test for a trend of the rate ratio over time and to evaluate the adequacy of the proportional hazards assumption over time. To compare the experience of the study group with that of the general U.S. population, standardized morbidity ratios were calculated for the overall follow-up period by using Surveillance, Epidemiology, and End Results (SEER) data for white men from 1983 to 1987, adjusting for age in 5-year categories [22]. We also used Cox proportional-hazards models for the observational analyses after the year 7 questionnaire, controlling for age (continuous), body mass index (continuous), smoking (current, past, or never), and alcohol consumption (daily, weekly, or less than weekly). Variables were selected by running individual, age-adjusted models.

Variables with a P value less than 0.2 were combined in a single model and then removed one by one so that their influence on the aspirin effect on incidence of colorectal cancer could be seen [23]. Multiple logistic regression was used to look at the propensity of participants to choose aspirin after the end of the aspirin arm of the randomized trial; this was defined as the odds of taking aspirin on 3 or more days per week. The influence of variables on the frequency of aspirin use was checked by running individual, age-adjusted models. Variables with a P value less than 0.2 were combined in a single model, and insignificant predictors were dropped. Because gastrointestinal symptoms and diagnoses are correlated, we put them into the model one by one (while adjusting for all other selected variables) and jointly to assess their independent effects. To test for differences in stage of colorectal cancer at the time of diagnosis in the aspirin and the placebo groups, we used a trend test, assuming the Dukes stages to be ordinal with identical intervals (Dukes A = 0, Dukes B1 = 1, Dukes B2 = 2, Dukes C1 = 3, Dukes C2 = 4, and Dukes D = 5).

Role of the Funding Sources

The funding sources had no role in the collection, analysis, or interpretation of the data or in the decision to submit the manuscript for publication.

Results

Top Methods Results Discussion Author & Article Info References

Table 1 shows the baseline characteristics of the study sample according to randomly assigned aspirin treatment. The distribution of all measured variables is almost identical, which is strong evidence for the assumption that the treatment groups are comparable. Only 11% of the participants were current smokers at baseline, and another 40% had smoked in the past. One fourth consumed alcoholic beverages at least once a day, and one fourth consumed alcoholic beverages less than once a week.

The intention-to-treat analysis is based on a mean follow-up period of more than 12 years, 258 503 person-years of follow-up, and 341 incident cases of colorectal cancer. Ninety-three percent of cases were confirmed by 25 October 1995. One hundred seventy-three occurred in the aspirin group, and 168 occurred in the placebo group. Compared with U.S. white men of the same age, study participants had an 18% reduction in incidence of colorectal cancer over the total follow-up period (standardized morbidity ratio, 0.82 [95% CI, 0.75 to 0.90]). Colorectal cancer is usually diagnosed (and surgically removed) before it has spread to local or regional lymph nodes (Dukes stages A to B2, 59%). Nevertheless, invasion through the muscle layer into the subserosa (Dukes stage B2, 26%) is slightly more common than containment within the muscle layer (Dukes stage B1, 24%). Containment within the muscle layer is much less common when positive lymph nodes are present (Dukes stage C1, 5%; Dukes stage C2, 18%). Eighteen percent of cases of colorectal cancer have already metastasized to other organs (Dukes stage D) at the time of diagnosis. When the stage at diagnosis of participants randomly assigned to receive aspirin and those randomly assigned to receive placebo is compared, there seems to be a slight excess of cancers confined to the mucosa in the aspirin group (Dukes stage A, 12%) compared with the placebo group (7%). Nevertheless, no statistically significant difference is seen in the overall distribution of stage at diagnosis in the treatment groups (P = 0.11).

Figure 2 shows the number of events and the incidence rate ratios and their 95% CIs for random assignment to aspirin and colorectal cancer in consecutive 2-year periods and over the total follow-up period. The 2-year rate ratios show no clear trend over time, and their CIs always include 1. The overall rate ratio for aspirin assignment compared with placebo assignment is 1.03 (CI, 0.83 to 1.28).

View larger version (12K): [in this window] [in a new window]   Figure 2. Relative risk for colorectal cancer in the aspirin group compared with the placebo group. Bars represent 95% CIs.

After the end of the aspirin arm of the randomized trial, 71% of participants chose to take aspirin regularly on 3 or more days per week and 29% chose not to do so. Most of the latter group (84%) used aspirin on less than 1 day per week. Predictors of self-selected regular aspirin use after the end of the aspirin arm of the study are shown in Table 2 and Table 3. Random assignment to aspirin, older age, vigorous exercise at least once a week, higher body mass index, regular alcohol consumption, hypertension, history of coronary artery disease, and headache were all independently associated with more frequent use of aspirin. On the other hand, gastrointestinal symptoms and diagnoses, history of ischemic or hemorrhagic stroke, and history of treatment with warfarin were independent predictors of less frequent use of aspirin.

Over a mean follow-up period of more than 5 years after the end of the aspirin arm of the study, we found 159 incident cases of colorectal cancer in a total of 107 810 person-years. The adjusted incidence rate ratio for more frequent use of aspirin and colorectal cancer, based on 157 cases and 106 757 person-years, was 1.07 (CI, 0.75 to 1.53) (Table 4). In addition, Table 4 shows the relative risk for colorectal cancer according to chosen aspirin use, stratified for randomized treatment assignment in the years before the chosen use. This stratification allows examination of the effect of the duration of aspirin use on the incidence of colorectal cancer. Compared with persons not regularly exposed to aspirin in the preceding years, neither those discontinuing regular use (odds ratio, 0.95) nor those starting regular use after the end of the randomized phase of the trial (odds ratio, 1.02) nor those exposed during the entire study period (odds ratio, 1.07) have a reduced risk for colorectal cancer.

Discussion

Top Methods Results Discussion Author & Article Info References

We did not find an effect of 5 years of randomized aspirin treatment on the incidence of colorectal cancer over 12 years, and the lower bound of the 95% CI precludes a substantial protective effect. When offered the choice to regularly take aspirin after the end of the aspirin arm of the study in 1988, 29% of the participants chose not to take aspirin regularly. These men were much more likely to have a history of gastrointestinal symptoms or diagnoses. We found no association between self-selected aspirin use and incidence of colorectal cancer over a 5-year follow-up period, regardless of the previous duration of aspirin use.

The Physicians' Health Study is the only published randomized trial on aspirin and incidence of colorectal cancer. In contrast to most observational studies that have reported an inverse association between aspirin and incidence of colorectal cancer, it may be limited by a relatively short duration and low-dose aspirin treatment. However, because of randomization, it has the advantage of less propensity for bias.

The duration of exposure to aspirin necessary to prevent colorectal cancer is unclear. Giovannucci and colleagues [15] found an increasing protective effect with an increasing duration of regular aspirin use. Although this reduction became significant only after 20 years of use, the relative risk for colorectal cancer began to steadily decline after 5 to 9 years of regular use. Rosenberg and coworkers [8] found a full protective effect when regular use was started in the year before diagnosis and found a weak effect of duration of use when any regular use was compared with regular use over 5 or more years. Therefore, an effect should have been at least emerging in the Physicians' Health Study.

The inhibitory effects of NSAIDs on colorectal cancer have been shown, in some cases, to be dose dependent in the rodent model [24]. Such a dose dependency has not been shown for aspirin, which, in contrast to the other NSAIDs, inhibits cyclooxygenase irreversibly. This may not be the relevant mechanism, however. In the only study that could directly compare dose and duration, Peleg and associates [11] found the association between cumulative dose and colorectal cancer to be weaker than the association between cumulative days of exposure and colorectal cancer. As for the frequency of use, reductions in colon cancer mortality or incidence have been reported for less-than-daily use ( 4 times per week [8], 4 tablets per week [15], 2 tablets per week [13], and less than monthly or 1 to 15 times per month [16]), although most of the regular users might have used aspirin more frequently.

Constipation and the resulting increase in gastrointestinal transit time [25, 26] and cholecystectomy leading to an accumulation of secondary bile acids in the colon [27, 28] have been discussed as risk factors for colorectal cancer [29]. Substantial alcohol consumption, when combined with inadequate intake of folate and methionine, may increase risk for colon cancer [30]. Moderate alcohol use was a strong predictor of aspirin use in the Physicians' Health Study. Therefore, uncontrolled confounding by alcohol use might attenuate a protective effect of aspirin on colorectal cancer, at least in moderate drinkers. Cigarette smoking has been shown to be a risk factor for colorectal cancer after an induction period of at least 35 years in men [31] and in women [32]. Current or past cigarette smoking was not independently associated with the frequency of aspirin use in our study.

Analgesic use has been noted to be differential over a 10-year period [33]. Thun and colleagues [34] found a statistically significant increase in risk for death from rectal cancer (but not risk for death from colorectal cancer overall) with the frequent use of acetaminophen in persons who do not use aspirin (relative risk, 3.08), although this finding was based on only four cases [34]. Baron and Adami [35] state that long-term aspirin use is likely to be associated with a clinically significant long-term disorder that may itself be related to cancer. Kune and coworkers [7] found a statistically significant deficit of chronic diseases (hypertension, heart disease, and chronic arthritis) among patients with colorectal cancer.

We could not demonstrate confounding by any of the above-mentioned variables in our observational analysis, but this does not necessarily exclude uncontrolled or uncontrollable confounding as plausible alternative explanations for the findings in other observational studies. First, the run-in phase of our study excluded persons who were unable to take aspirin regularly, and those persons might very well be the ones with the highest risk for colorectal cancer. Second, distribution of the variables may vary among different populations, leading to a different magnitude of the overall effect of confounding.

Laboratory methods [36] were used to assess compliance with random assignment to aspirin in a sample of participants, and this compliance was found to be very high throughout the trial [37]. If recent use rather than long-term use is important (effect on late stage of carcinogenesis), as Rosenberg and associates suggest [8, 38], both the intention-to-treat analysis and the observational analyses might be misleading because they do not consider aspirin use after the end of the aspirin arm of the trial and the year 7 questionnaire, respectively. On the other hand, if these assumptions are true, we should have seen an effect during the randomized phase of the trial.

Outcome was assessed by self-report, and persons using aspirin regularly may be more health conscious and may report cancer earlier. By restricting the observational analysis to 1 year before the last reported case of colorectal cancer, we reduced the rate ratio for colorectal cancer in frequent users of aspirin to 0.91, providing some evidence for the hypothesis without affecting the overall conclusions. Aspirin use could cause bleeding and lead to earlier diagnosis, or it could mask early symptoms and lead to later diagnosis. Either effect would influence the stage of cancer at the time of diagnosis, although in opposite directions. We found no significant differences in the stage of colorectal cancer at the time of diagnosis in the aspirin and placebo groups, making a differential diagnostic lag time unlikely.

Despite the limitation imposed by the relatively short treatment period, our findings indicate that a causal role of aspirin in the prevention of colorectal cancer is less likely [20, 39-41]. Clinicians should be cautioned about using salicylates (or other NSAIDs) for the primary prevention of colorectal cancer. Awareness that the role of factors influencing the choice to regularly use medication in observational studies (confounding by indication) is more important than previously suspected seems to be growing [20]. The smaller propensity for confounding and bias in randomized trials compared with observational studies should be weighed against limitations that result from shorter duration of treatment, especially for cancer outcomes.

Drs. Glynn, Lee, Manson, Buring, and Hennekens: Division of Preventive Medicine, Department of Medicine, Brigham and Women's Hospital, Harvard Medical School, 900 Commonwealth Avenue East, Boston, MA 02215.