Abstract

Earlier studies suggest a protective association between vitamin K antagonist (VKA) anticoagulants and the incidence of cancer. The authors examined the associations between VKA therapy and incidence of 24 site-specific cancers with a Danish population-based cohort study, using heart valve replacement as an instrumental variable. The authors enrolled 9,727 Danish residents who received a replacement heart valve between 1989 and 2006. The heart valve recipients were matched with 95,481 unexposed individuals on age and sex. The authors used the heart valve replacement instrument to estimate rate ratios associating VKA therapy with incidence of the 24 site-specific cancers using Poisson regression models. Direct associations between VKA therapy and incidence of the 24 cancers were estimated in a prescription validation subset. The instrumental variable associations were plotted according to the inverse normal of rank percentile and subjected to semi-Bayes shrinkage adjustment for multiple comparisons. The pattern of associations was consistent with a null-centered Gaussian distribution. No individual cancer site showed a substantial positive or negative association with VKA therapy in the prescription validation subset, the instrumental variable analysis, or the analysis with semi-Bayes adjustment. These results do not support the existing hypothesis that VKA therapy is associated with reduced cancer risk.

Recent advances in cancer biology have associated tumor growth with a hypercoagulable state (1, 2), motivating the hypothesis that antithrombotic agents could influence cancer development and progression (3–6). Vitamin K antagonists (VKAs) (e.g., dicoumarol, warfarin, and phenprocoumon) are oral anticoagulants used therapeutically since the 1950s (7). VKAs are indicated for primary prevention of stroke in patients with replacement heart valves (7, 8), chronic atrial fibrillation, or history of myocardial infarction, as well as for the treatment and prevention of deep venous thrombosis and pulmonary embolism (7).

Evidence of an association between VKAs and cancer outcomes first appeared in the 1980s with 2 reports of increased survival among patients with small cell lung carcinoma who were randomized to receive standard therapy plus warfarin, compared with patients who received standard therapy alone (9, 10). More recently, Schulman and Lindmarker (11) compared site-specific cancer incidence among clinical trial participants with a first incident deep venous thrombosis in the legs and/or pulmonary embolism, who were randomized to either 6 weeks or 6 months of warfarin therapy. After a mean of 8.1 years of follow-up, they observed a higher urogenital cancer risk in the shorter-duration arm, compared with the longer-duration arm (for a combined outcome of kidney, bladder, prostate, ovary, or uterine cancer: odds ratio (OR) = 2.5, 95% confidence interval (CI): 1.3, 5.0). Schulman and Lindmarker contended that their results supported the hypothesis that VKA drugs have antineoplastic activity, but their study was criticized for a number of unaddressed limitations (12–16).

Other studies have also estimated associations between VKA therapy and cancer incidence. Taliani et al. (17) observed a null association between 12 and 3 months of VKA therapy and the incidence of any cancer after a mean of 3.6 years, but they included only 19 exposed cancer cases. Blumentals et al. (18) observed a modest positive association between warfarin therapy and bladder cancer incidence (OR = 1.3, 95% CI: 0.85, 1.9). Tagalakis et al. (19) reported a reduced risk of prostate cancer among warfarin users compared with nonusers (OR = 0.80, 95% CI: 0.65, 0.99), but they observed null associations with bladder, kidney, uterine, and ovarian cancer incidence. Most recently, Pengo et al. (20) reported a protective effect of VKA therapy on prostate cancer incidence (hazard ratio = 0.69, 95% CI: 0.50, 0.97). A review of this topic concluded that current evidence for an association between VKA therapy and cancer outcomes is inconclusive (21).

To address limitations in earlier literature, we conducted a nationwide Danish cohort study with a relatively long follow-up period, using heart valve replacement as an instrumental variable for the associations between VKA therapy and the incidence of 24 different cancers.

MATERIALS AND METHODS

This study was approved by the Danish Registry Board.

Study population and data collection

The source population for this study was the population of Denmark between January 1, 1989, and December 31, 2006. National prescription data maintained in the Danish Registry of Medicinal Product Statistics have been available only since 1995. If we had limited our study to 1995 and beyond, the frequency of site-specific cancers would not likely yield improved precision in estimates of association compared with earlier studies. However, the Danish National Registry of Patients has electronically recorded all inpatient diagnoses and surgical procedures at nonpsychiatric hospitals since 1977 and was expanded in 1995 to include diagnoses made at outpatient clinics and emergency departments. We capitalized on this longer data collection period to extend follow-up by using heart valve replacement surgery as a proxy for VKA treatment. We argue below that heart valve replacement satisfies the conditions necessary to serve as an instrumental variable for the associations between VKA therapy and site-specific cancer incidence. Use of this variable not only extends the study period but also permits estimation of unconfounded associations between VKA therapy and cancer incidence, assuming the instrumental variable conditions are satisfied (22).

We identified all heart valve replacement surgeries in Denmark during the study period by searching the Danish National Registry of Patients for relevant procedure codes. Heart valve replacement patients should receive VKA therapy immediately and continuously after surgery (23). To allow a reasonable induction period, we began follow-up on an index date, defined as the date of valve replacement surgery plus 1 year. Replacements were for the mitral, aortic, tricuspid, or pulmonic valve and included biologic and artificial valves. We defined each subject’s index date as the date of valve replacement surgery and excluded any subject with a cancer history before, or within 1 year after, his/her index date. This roster comprised our exposed cohort. For each exposed subject, we matched up to 10 unexposed subjects from the general population, sampled randomly and without replacement, within strata of birth year and sex. Unexposed subjects were assigned the same index date as their exposed counterparts and were required to have no history of heart valve replacement or cancer on, or within 1 year after, that date. No subject appeared in the cohort more than once, so no matched unexposed subjects received a heart valve replacement over the course of follow-up. This enrollment criterion was independent of cancer incidence, and no unexposed subject’s person-time was immune or immortal (24).

Using each subject’s unique Civil Personal Registry number, we searched the Danish Cancer Registry for diagnoses of 49 site-specific cancers (25). We used the Danish National Registry of Patients to determine prevalent medical conditions on the index date, corresponding to the Charlson Comorbidity Index (26), and to document history of 2 major indications for VKA therapy, venous thromboembolism and atrial fibrillation.

We conducted a validation substudy among subjects whose residence and index dates fell within the geographic and temporal coverage of 4 Danish county-specific prescription registries (Aarhus County, data since 1996; North Jutland County, data since 1989; Viborg and Ringkøbing counties, data since 1998). Patients prescribed a VKA must present to a pharmacy and pay the entire medication cost. Once paid, the medication is dispensed and the transaction is automatically logged in the prescription registry, whereupon the Danish government reimburses a portion of the cost. The county prescription registries encode drugs according to the Anatomical Therapeutic Chemical (ATC) Classification System (27) and record dispense dates, fill quantities, and the patient’s Civil Personal Registry number. These registries allowed us to determine whether members of the validation subset received a VKA prescription during follow-up.

Definitions of analytical variables

Cancers diagnosed during follow-up were identified in the Danish Cancer Registry by using International Classification of Diseases (ICD), Tenth Revision, codes. The Danish Cancer Registry has translated past records, entered under earlier ICD revisions, into the ICD, Tenth Revision, to standardize case ascertainment. All cancers with 5 or more cases among valve recipients were considered as separate outcomes, which reduced the number of cancer sites included in the analyses from 49 to 24.

For subjects without a cancer diagnosis, we characterized end of follow-up by linkage with the Danish Civil Registry, which updates residential address and vital status for all Danish residents on a daily basis. Each subject contributed person-time from 1 year after their index date until the first of 1) cancer diagnosis, 2) emigration from Denmark, 3) death from any cause, or 4) December 31, 2006.

Age was defined as the number of complete years between the birth date and index date. On the basis of diagnosis history as of the index date, we calculated the Charlson Comorbidity Index according to the published method (26). In addition to the diagnoses included in the Charlson Comorbidity Index, we assessed history of atrial fibrillation, superficial or deep venous thrombosis, and pulmonary embolism using ICD, Eighth Revision, and ICD, Tenth Revision, codes to search the Danish National Registry of Patients. We defined a positive history of venous thromboembolism as having been diagnosed with superficial or deep venous thrombosis and/or pulmonary embolism before the index date.

For subjects in the validation subset, we searched county prescription registries for VKA prescriptions filled after index dates. These were identified by searching for Anatomical Therapeutic Chemical Classification System codes beginning with “B01AA.” Heart valve recipients are almost always placed on life-long VKA therapy after surgery (28), so determination of “ever exposure” to a VKA after the index date is expected to indicate enduring use.

Statistical analysis

We calculated the frequency of subjects and the sum and proportion of person-time within heart valve replacement groups according to age category, sex, comorbidity, and history of venous thromboembolism and atrial fibrillation. We used the validation data to calculate the positive and negative predictive values for the classification of VKA exposure by heart valve replacement status. We then calculated incidence rate ratios and 95% confidence intervals associating known VKA prescription status with the incidence of the 24 site-specific cancers.

Instrumental variable analysis

In our age- and sex-matched cohort, heart valve replacement appears to satisfy the criteria to be an instrumental variable for the associations between VKA therapy and site-specific cancer incidence (22, 29). Figure 1 is a directed acyclic graph depicting hypothesized relations among heart valve replacement, VKA prescription, and cancer incidence. For valve replacement to be a valid instrumental variable for the VKA-cancer associations, it must have no direct causal effect on the incidence of the cancers we studied (i.e., no plausible arrow “b” or equivalent open direct path not passing through VKA therapy); its effect on cancer incidence must be mediated by VKA exposure (presence of arrow “a”); and there must be no unblocked backdoor path (30) between valve replacement and cancer incidence (absence of, or adequate conditioning on, node U2). If these assumptions hold, then residual confounding of the VKA-cancer associations (node U1 in Figure 1) is negated at the expense of nondifferential misclassification of VKA exposure by the instrument (22, 29).

Figure 1.

Directed acyclic graph depicting the conditions necessary for heart valve replacement to serve as an instrumental variable (Z) for the associations between vitamin K antagonist (VKA) therapy (X; the target exposure) and site-specific cancer incidence (Y; the outcome), Denmark, 1989–2006. U1 and U2 represent vectors of hypothetical confounders.

Figure 1.

Directed acyclic graph depicting the conditions necessary for heart valve replacement to serve as an instrumental variable (Z) for the associations between vitamin K antagonist (VKA) therapy (X; the target exposure) and site-specific cancer incidence (Y; the outcome), Denmark, 1989–2006. U1 and U2 represent vectors of hypothetical confounders.

We contend that the instrumental variable assumptions hold because 1) there is no evidence to support a causal relation between surgical installation of a heart valve and the incidence of any cancer, nor is there a biologic pathway to suggest such an association (i.e., no arrow “b” or its equivalent); 2) heart valve replacement essentially mandates lifelong VKA therapy (28), while the prevalence of VKA treatment is low in the general population (i.e., there exists a strong causal association depicted by arrow “a”); and 3) after matching valve recipients to nonrecipients on age and sex, we expect no important uncontrolled common causes of valve replacement and cancer incidence (i.e., no other variables that satisfy the “U2” node for any of the cancers we examined).

Instrumental variable analyses commonly use 2-stage linear regression to estimate associations. In the primary stage, the outcome is regressed on the instrument, yielding an estimate of the exposure-outcome association that is unconfounded but distorted by misclassification (31, 32). In the secondary stage, the instrument-outcome association is scaled by the instrument-exposure association, yielding an unconfounded estimate of the exposure-outcome risk difference, adjusted for misclassification of the target exposure (31). This approach requires complete data for both the instrument and the target exposure. In our study, all subjects had data on the instrument (heart valve replacement), but only about 25% of those subjects had data on the target exposure (VKA prescription). Under a traditional 2-stage instrumental variable analysis framework, we could use our data to estimate theoretically unconfounded instrument-outcome associations. However, because not all subjects had VKA prescription data, we could not implement the secondary regression to correct the inherent exposure misclassification. Nevertheless, the strong positive and negative predictive values for the classification of VKA therapy by the heart valve replacement instrument yield an expectation of little misclassification. Therefore, our analytical approach was to make no further adjustment to the primary instrumental variable analysis and to compare the instrumental variable associations with those estimated from the VKA validation subset to judge the effect of misclassification.

We estimated incidence rate ratios and 95% confidence intervals associating heart valve replacement or measured VKA prescription with site-specific cancer incidence in separate Poisson regression models. VKA status was the sole independent variable in each of these models, and the logarithm of person-years at risk served as the offset variable. To accommodate overdispersion of Poisson parameters, we used generalized estimating equations, with covariance matrices initially presumed to be exchangeable.

We plotted the summarized incidence rate ratio estimates against the inverse normal of rank percentile (INRP). This plot was overlaid with a line depicting modeled incidence rate ratios, calculated from the inverse-variance weighted linear regression of the observed log incidence rate ratio estimates on their INRP (33). Associations drawn from a null-centered Gaussian distribution should lie along the modeled incidence rate ratio line, which would intersect the plot origin, where INRP and the log incidence rate ratio both equal 0 (33).

Semi-Bayes adjustment of instrumental variable associations

Our estimation of 24 different associations would generate nonnull observations by chance alone. To reduce expected random error, we subjected the vector of instrumental variable estimates to empirical Bayes shrinkage analysis (34). The empirical Bayes method can be applied to groups of associations from a common theoretical population of such associations, and it “shrinks” each estimate toward the overall mean, in proportion to the ratio of its variance to the empirically estimated population variance. Associations with larger relative variances are displaced further toward the mean than associations with smaller relative variances. Empirical Bayes analysis therefore leads to a deemphasis of large-magnitude associations that are measured with poor precision, which are disproportionately influenced by random error.

Our vector of instrumental variable estimates was incompatible with the constraints imposed by the algorithm for empirical population variance estimation (34). Instead, we implemented semi-Bayes adjustment using variances derived from estimated dispersion ranges for the theoretical population of all site-specific cancer associations (34). We first presumed that the population of associations was dispersed within a 2-fold range (population variance = 0.031). We then presumed that the population of associations was dispersed within a 10-fold range (population variance = 0.345). The smaller presumed population variance yields larger attenuation of the instrumental variable estimates under semi-Bayes adjustment, and it likely underestimates the true population variance because it was derived from a smaller dispersion range than what we observed for our 24 associations. The larger presumed population variance yields smaller attenuation of the instrumental variable estimates under semi-Bayes adjustment and likely overestimates the true population variance. Thus, taken together, adjusted associations from these 2 semi-Bayes analyses likely bracket the associations that would have been observed under a fully empirical adjustment.

RESULTS

We identified 9,727 subjects who received a replacement heart valve during the study period without a cancer history. These subjects were matched with 95,481 subjects from the general population with no history of heart valve replacement or cancer diagnosis. Heart valve recipients contributed 52,510 person-years of follow-up, and matched nonrecipients contributed 556,021 person-years of follow-up.

A comparison of baseline characteristics between full 1:10 matched sets of heart valve recipients and nonrecipients appears in Table 1. Proportions of person-time were similar according to these characteristics in the entire cohort, despite variation in the matching ratio across all strata (data not shown). Heart valve recipients had a higher burden of comorbidity than the matched unexposed individuals. Heart valve recipients were about 8 times more likely to have a history of atrial fibrillation than were nonrecipients (17% vs. 2.1% of person-time positive, respectively), and they were approximately 2 times more likely to have a history of venous thromboembolism (2.0% vs. 1.1% of person-time positive, respectively). The imbalance of venous thromboembolism history between exposure groups led us to recalculate associations in the cohort subset without a venous thromboembolism history, as primary venous thromboembolism increases clinical suspicion of previously undetected cancer (35–37). Results from this restricted analysis (data not shown) were not substantially different from results based on the entire cohort.

Table 1.

Baseline Characteristics of Heart Valve Recipients and Matched Nonrecipients, Denmark, 1989–2006a

Variable Heart Valve Recipients (n = 8,724)
 
Matched Unexposed (n = 87,240)
 
No. Person-Years No. Person-Years 
Sex       
    Male 5,331 29,766 61 53,310 316,671 60 
    Female 3,393 19,095 39 33,930 208,486 40 
Age category on index date, years       
    Under 18 180 1,215 2.5 1,801 12,408 2.4 
    18–24 95 819 1.7 949 8,344 1.6 
    25–34 267 2,047 4.2 2,632 21,421 4.1 
    35–44 540 3,812 7.8 5,390 41,572 7.9 
    45–54 1,175 7,905 16 11,868 87,955 17 
    55–64 2,253 13,078 27 22,562 144,491 28 
    65–74 2,901 14,877 30 29,043 159,491 30 
    75–84 1,265 4,967 10 12,513 48,253 9.2 
    85 or more 48 141 0.3 482 1,222 0.2 
Charlson Comorbidity Index       
    0 4,517 28,079 57 69,715 448,808 85 
    1 3,527 18,250 37 15,464 70,025 13 
    2 574 2,257 4.6 1,773 5,667 1.1 
    3 106 275 0.6 288 655 0.1 
History of venous thromboembolismb       
    Positive 201 958 2.0 1,394 5,944 1.1 
    Negative 8,523 47,903 98 85,846 519,214 99 
History of atrial fibrillation/flutterc       
    Positive 1,518 8,148 17 2,723 10,770 2.1 
    Negative 7,206 40,713 83 84,517 514,387 98 
Variable Heart Valve Recipients (n = 8,724)
 
Matched Unexposed (n = 87,240)
 
No. Person-Years No. Person-Years 
Sex       
    Male 5,331 29,766 61 53,310 316,671 60 
    Female 3,393 19,095 39 33,930 208,486 40 
Age category on index date, years       
    Under 18 180 1,215 2.5 1,801 12,408 2.4 
    18–24 95 819 1.7 949 8,344 1.6 
    25–34 267 2,047 4.2 2,632 21,421 4.1 
    35–44 540 3,812 7.8 5,390 41,572 7.9 
    45–54 1,175 7,905 16 11,868 87,955 17 
    55–64 2,253 13,078 27 22,562 144,491 28 
    65–74 2,901 14,877 30 29,043 159,491 30 
    75–84 1,265 4,967 10 12,513 48,253 9.2 
    85 or more 48 141 0.3 482 1,222 0.2 
Charlson Comorbidity Index       
    0 4,517 28,079 57 69,715 448,808 85 
    1 3,527 18,250 37 15,464 70,025 13 
    2 574 2,257 4.6 1,773 5,667 1.1 
    3 106 275 0.6 288 655 0.1 
History of venous thromboembolismb       
    Positive 201 958 2.0 1,394 5,944 1.1 
    Negative 8,523 47,903 98 85,846 519,214 99 
History of atrial fibrillation/flutterc       
    Positive 1,518 8,148 17 2,723 10,770 2.1 
    Negative 7,206 40,713 83 84,517 514,387 98 
a

Restricted to strata with 10 unexposed subjects matched to each exposed subject.

b

Diagnosed with superficial or deep venous thrombosis and/or pulmonary embolism before the index date.

c

The International Classification of Diseases, Eighth Revision, contained separate codes for atrial fibrillation (code 427.93) and atrial flutter (code 427.94). The International Classification of Diseases, Tenth Revision, combined these 2 diagnoses into a single code (code I48).

Heart valve recipients had a higher mortality rate than did nonrecipients over follow-up (5.1 vs. 3.2 deaths per 100 person-years, respectively). To address whether differential mortality generated informative loss to follow-up, we created a secondary data set in which follow-up of unexposed subjects was terminated on the last follow-up date of matched exposed subjects who died. We observed no material difference in the associations between heart valve replacement and site-specific cancer incidence in the original and secondary data sets (data not shown).

For subjects in the validation subset, heart valve replacement classified VKA prescription with a positive predicted value of 97% and a negative predictive value of 91% (Table 2). These validation parameters did not differ substantially when calculated in the strata of cancer diagnosis status and calendar year of the validation period (data not shown).

Table 2.

Validation of the Heart Valve Replacement Instrument as a Proxy Variable for Vitamin K Antagonist Anticoagulant Exposure, Denmark, 1989–2006a

 Prescribed a Vitamin K Antagonist Not Prescribed a Vitamin K Antagonist 
Received a replacement heart valve 2,460 75 
Did not receive a replacement heart valve 1,927 20,185 
Positive predictive value 2,460/(2,460 + 75) = 0.970 
Negative predictive value 20,185/(20,185 + 1,927) = 0.913 
 Prescribed a Vitamin K Antagonist Not Prescribed a Vitamin K Antagonist 
Received a replacement heart valve 2,460 75 
Did not receive a replacement heart valve 1,927 20,185 
Positive predictive value 2,460/(2,460 + 75) = 0.970 
Negative predictive value 20,185/(20,185 + 1,927) = 0.913 
a

Positive and negative predictive values were calculated in the subset of the cohort with prescription drug exposure data (n = 24,647).

Table 3 shows the estimated associations between VKA prescription and incidence of 24 site-specific cancers in both the instrumental variable and validation subset analyses. Incidence rate ratio point estimates from the validation subset ranged from 0.46 to 4.6. Five of the sites showed modestly elevated incidence rates among VKA-exposed subjects, counter to the protective effect hypothesized by earlier studies. Four of these associations were measured with good precision (prostate cancer: incidence rate ratio (IRR) = 1.3, 95% CI: 1.0, 1.7; basal cell skin cancer: IRR = 1.4, 95% CI: 1.0, 1.8; bladder cancer: IRR = 1.7, 95% CI: 1.2, 2.5; and colon cancer: IRR = 1.7, 95% CI: 1.2, 2.4). One association had few cases and thus poor precision (myelodysplastic syndromes: IRR = 4.6, 95% CI: 1.3, 16).

Table 3.

Associations Between Vitamin K Antagonist Therapy and Site-specific Cancer Incidence, Estimated by Using Heart Valve Replacement as an Instrumental Variable and Using the Prescription Validation Subset, Denmark, 1989–2006a

Cancer Site Valve Recipients
 
Nonrecipients
 
Instrumental Variableb
 
Validation Subsetc
 
IRR Rankd 
Cases, no. Person-Years Cases, no. Person-Years IRR 95% CI IRR 95% CI 
Leukemia, lymphocytic 47,785 126 506,584 0.59 0.28, 1.3 1.8 0.72, 4.8 
   Ovary 18,575 126 200,489 0.69 0.34, 1.4 1.3 0.58, 3.1 
   Larynx 47,798 77 506,268 0.69 0.28, 1.7 1.4 0.46, 4.4 
   Pancreas 18 47,835 271 507,197 0.70 0.44, 1.1 1.1 0.58, 2.2 
   Brain 11 47,808 158 506,707 0.74 0.40, 1.4 0.74 0.26, 2.1 
   Uterus 11 18,585 155 200,631 0.77 0.42, 1.4 1.4 0.60, 3.2 
   Lung 125 48,291 1,291 512,263 1.0 0.86, 1.2 1.1 0.80, 1.5 
   Gallbladder 47,783 49 506,228 1.1 0.43, 2.7 1.2 0.13, 10 
Myelodysplastic syndrome 47,781 49 506,225 1.1 0.43, 2.7 4.6 1.3, 16 
   Rectum 47 48,005 455 508,269 1.1 0.81, 1.5 1.1 0.67, 1.8 10 
   Kidney 16 47,801 153 506,762 1.1 0.66, 1.9 1.4 0.60, 3.3 11 
   Prostate 120 29,801 1,109 311,762 1.1 0.94, 1.4 1.3 1.0, 1.7 12 
   Colon 95 48,228 882 510,254 1.1 0.92, 1.4 1.7 1.2, 2.4 13 
   Skin, basal 166 48,446 1,514 513,280 1.2 0.99, 1.4 1.4 1.0, 1.8 14 
   Esophagus 17 47,832 154 506,672 1.2 0.71, 1.9 0.46 0.11, 2.0 15 
   Breast (female) 72 18,881 653 202,920 1.2 0.93, 1.5 0.92 0.56, 1.5 16 
Non-Hodgkin’s lymphoma 38 47,898 328 507,578 1.2 0.88, 1.7 0.87 0.44, 1.7 17 
   Bladder 80 48,123 676 509,130 1.3 0.99, 1.6 1.7 1.2, 2.5 18 
   Skin, nonbasal 40 47,987 322 507,744 1.3 0.95, 1.8 0.80 0.39, 1.6 19 
   Liver 11 47,800 86 506,388 1.4 0.72, 2.5 –e  20 
Leukemia, myeloid 10 47,781 76 506,384 1.4 0.72, 2.7 1.2 0.49, 3.0 21 
   Stomach 26 47,879 188 506,975 1.5 0.97, 2.2 1.6 0.77, 3.3 22 
   Cervix 18,554 55 200,121 1.6 0.75, 3.3 2.4 0.72, 7.9 23 
   Oral cavity 47,786 49 506,226 1.7 0.82, 3.7 1.5 0.50, 4.8 24 
Cancer Site Valve Recipients
 
Nonrecipients
 
Instrumental Variableb
 
Validation Subsetc
 
IRR Rankd 
Cases, no. Person-Years Cases, no. Person-Years IRR 95% CI IRR 95% CI 
Leukemia, lymphocytic 47,785 126 506,584 0.59 0.28, 1.3 1.8 0.72, 4.8 
   Ovary 18,575 126 200,489 0.69 0.34, 1.4 1.3 0.58, 3.1 
   Larynx 47,798 77 506,268 0.69 0.28, 1.7 1.4 0.46, 4.4 
   Pancreas 18 47,835 271 507,197 0.70 0.44, 1.1 1.1 0.58, 2.2 
   Brain 11 47,808 158 506,707 0.74 0.40, 1.4 0.74 0.26, 2.1 
   Uterus 11 18,585 155 200,631 0.77 0.42, 1.4 1.4 0.60, 3.2 
   Lung 125 48,291 1,291 512,263 1.0 0.86, 1.2 1.1 0.80, 1.5 
   Gallbladder 47,783 49 506,228 1.1 0.43, 2.7 1.2 0.13, 10 
Myelodysplastic syndrome 47,781 49 506,225 1.1 0.43, 2.7 4.6 1.3, 16 
   Rectum 47 48,005 455 508,269 1.1 0.81, 1.5 1.1 0.67, 1.8 10 
   Kidney 16 47,801 153 506,762 1.1 0.66, 1.9 1.4 0.60, 3.3 11 
   Prostate 120 29,801 1,109 311,762 1.1 0.94, 1.4 1.3 1.0, 1.7 12 
   Colon 95 48,228 882 510,254 1.1 0.92, 1.4 1.7 1.2, 2.4 13 
   Skin, basal 166 48,446 1,514 513,280 1.2 0.99, 1.4 1.4 1.0, 1.8 14 
   Esophagus 17 47,832 154 506,672 1.2 0.71, 1.9 0.46 0.11, 2.0 15 
   Breast (female) 72 18,881 653 202,920 1.2 0.93, 1.5 0.92 0.56, 1.5 16 
Non-Hodgkin’s lymphoma 38 47,898 328 507,578 1.2 0.88, 1.7 0.87 0.44, 1.7 17 
   Bladder 80 48,123 676 509,130 1.3 0.99, 1.6 1.7 1.2, 2.5 18 
   Skin, nonbasal 40 47,987 322 507,744 1.3 0.95, 1.8 0.80 0.39, 1.6 19 
   Liver 11 47,800 86 506,388 1.4 0.72, 2.5 –e  20 
Leukemia, myeloid 10 47,781 76 506,384 1.4 0.72, 2.7 1.2 0.49, 3.0 21 
   Stomach 26 47,879 188 506,975 1.5 0.97, 2.2 1.6 0.77, 3.3 22 
   Cervix 18,554 55 200,121 1.6 0.75, 3.3 2.4 0.72, 7.9 23 
   Oral cavity 47,786 49 506,226 1.7 0.82, 3.7 1.5 0.50, 4.8 24 

Abbreviations: CI, confidence interval; IRR, incidence rate ratio; VKA, vitamin K antagonist.

a

Incidence rate ratios and 95% confidence intervals from Poisson regression models with empirical variance estimation.

b

Comparing VKA-exposed with VKA-unexposed subjects by using the heart valve replacement as an instrumental variable.

c

Comparing VKA-exposed with VKA-unexposed subjects by using prescription status ascertained in the validation subset.

d

Rank is based on incidence rate ratios estimated from the instrumental variable analysis.

e

–, no exposed case.

For the instrumental variable approach, incidence rate ratio point estimates ranged from 0.59 to 1.7. All associations appeared consistent with a null effect of VKA therapy on incidence of the 24 specific cancers. The INRP analysis presented in Figure 2 supports this interpretation. In this plot, the individual instrumental variable estimates fall almost perfectly along the line of incidence rate ratio values predicted under a Gaussian distribution model, and this line approximately intersects the coordinate defined by a null incidence rate ratio and the center of the INRP scale. These alignments indicate that the vector of instrumental variable estimates is consistent with associations drawn at random from an approximately null-centered normal distribution (33). Induction period analyses, in which we stratified models by time between index date and cancer diagnosis (<5 or ≥5 years), showed similar null-centered patterns of associations in both strata (data not shown).

Figure 2.

Plot of site-specific cancer incidence rate ratios, estimated by using heart valve replacement as an instrumental variable, according to the inverse normal of rank percentile, Denmark, 1989–2006. Points are ordered from left to right according to the order of cancer sites listed in Table 3. Error bars depict 95% confidence intervals.

Figure 2.

Plot of site-specific cancer incidence rate ratios, estimated by using heart valve replacement as an instrumental variable, according to the inverse normal of rank percentile, Denmark, 1989–2006. Points are ordered from left to right according to the order of cancer sites listed in Table 3. Error bars depict 95% confidence intervals.

The semi-Bayes analyses are also consistent with a null interpretation of the cancer associations (Figure 3). The semi-Bayes estimates are particularly important for the 5 cancer sites that showed modestly elevated associations (prostate, basal cell skin cancer, bladder, colon, and myelodysplastic syndromes). For all of these sites, both of the presumed population variances yielded attenuations in the incidence rate ratios that made them appear more null centered.

Figure 3.

Semi-Bayes analysis of the instrumental variable associations, Denmark, 1989–2006. Shrunken estimates are presented for 2 derived population variances: 1) when the population of associations was presumed to cover a 2-fold range (variance = 0.031); and 2) when the population of associations was presumed to cover a 10-fold range (variance = 0.345). Trios of site-specific associations are ordered from left to right according to the order of cancer sites listed in Table 3. Error bars depict 95% confidence intervals. IV, instrumental variable; Var, variance.

Figure 3.

Semi-Bayes analysis of the instrumental variable associations, Denmark, 1989–2006. Shrunken estimates are presented for 2 derived population variances: 1) when the population of associations was presumed to cover a 2-fold range (variance = 0.031); and 2) when the population of associations was presumed to cover a 10-fold range (variance = 0.345). Trios of site-specific associations are ordered from left to right according to the order of cancer sites listed in Table 3. Error bars depict 95% confidence intervals. IV, instrumental variable; Var, variance.

DISCUSSION

Our results are consistent with the proposition that no protective association exists between long-term VKA therapy and incidence of the 24 site-specific cancers we examined. The null associations that we observed with the instrumental variable analyses were also apparent in the validation subset and semi-Bayes analyses. This concordance reassures that the instrumental variable estimates do not show null associations simply as a result of misclassification stemming from use of the heart valve replacement instrument as a proxy for VKA therapy.

The possibilities for selection bias were limited, because we used national, population-based, prospective registries and placed no restriction on potential enrollees that would be a common effect of VKA treatment and a future cancer diagnosis (38).

Nondifferential and independent measurement errors in heart valve replacement, VKA exposure, and cancer incidence would lead to an expectation of attenuated association estimates, which could explain our null findings (38). Heart valve replacement, however, requires admission to a hospital and receipt of a surgical procedure, leaving little room for ambiguous classification in the Danish National Registry of Patients. VKA prescription was our target exposure, and imperfect classification of VKA exposure by the prescription registries would lead to an expectation of nondifferential exposure misclassification in the associations measured in the validation subset. The prescription registries log a pharmacy transaction only after a patient has presented a prescription and has paid in full for the drug, after which the patient is reimbursed by the Danish government for a portion of the cost. This procedure creates a strong expectation that an entry in the registry means that a patient took possession of the medication. Although possession does not guarantee use of the medication, the consequences are potentially dire if heart valve recipients fail to comply with their anticoagulant prescription, so we expect high adherence among valve recipients. Poor adherence among the nonrecipients who were prescribed VKAs for other indications would only serve to strengthen the contrast between exposed and unexposed.

Cancer incidence should be classified with few false positives in the Danish Cancer Registry, and sensitivity is expected to be high (39). If sensitivity were compromised, the error would almost certainly be nondifferential and independent because the prescription data and heart valve data were prospectively logged in registries independent of the Danish Cancer Registry. Outcome misclassification via imperfect yet nondifferential and independent sensitivity, though with near-perfect specificity, is not expected to bias our estimated incidence rate ratios since our cancer outcomes are all rare (38). Differential outcome misclassification due to greater medical scrutiny of valve recipients, compared with nonrecipients, could have obscured truly protective associations with VKA therapy. However, such a bias would be expected to positively displace all associations, whereas we observed a vector of estimates compatible with a null-centered normal distribution.

Our null-centered results conflict with earlier reports of protective associations between VKA therapy and cancer incidence (11, 19, 20). Several sources of bias have been suggested as alternative explanations for the protective urogenital cancer association observed by Schulman and Lindmarker (11), including questionable sensitivity of outcome classification, lack of data on potentially confounding comedications such as nonsteroidal antiinflammatory drugs, selection bias arising from differential losses to follow-up, possible differential cancer surveillance in the 2 warfarin groups, and lack of a plausible biologic mechanism for VKAs to specifically modify risk for the urogenital cancers (12–15). As noted above, Tagalakis et al. (19) reported a reduced rate of prostate cancer among warfarin users compared with nonusers (OR = 0.80, 95% CI: 0.65, 0.99) but null associations similar to our instrumental variable estimates between VKA therapy and the incidence of bladder, kidney, uterine, and ovarian cancers. Pengo et al. (20) reported a reduced rate of cancer overall, but the protective association was dominated by a reduced rate of prostate cancer (hazard ratio = 0.69, 95% CI: 0.50, 0.97). Our instrumental variable result for prostate cancer equaled 1.1 (95% CI: 0.94, 1.4), and the association also was not protective in the validation subset. The association between VKA therapy and prostate cancer has thus been heterogeneously measured in these 3 large, population-based studies.

In summary, we noted modest elevations in the incidence rates for 5 specific cancers among subjects prescribed a VKA in a validation substudy, which counters the prevailing hypothesis of a cancer-protective effect. However, our instrumental variable estimates were substantially more precise than were estimates from the validation substudy, and they are also theoretically unconfounded. These estimates, in aggregate, indicate no association between VKA therapy and incidence of the 24 specific cancers we examined.

Abbreviations

    Abbreviations
  • CI

    confidence interval

  • ICD

    International Classification of Diseases

  • INRP

    inverse normal of rank percentile

  • IRR

    incidence rate ratio

  • OR

    odds ratio

  • VKA

    vitamin K antagonist

Author affiliations: Channing Laboratory, Department of Medicine, Brigham and Women’s Hospital and Harvard Medical School, Boston, Massachusetts (Thomas P. Ahern); Department of Clinical Epidemiology, Aarhus University Hospital, Aarhus, Denmark (Lars Pedersen, Claus Sværke, Henrik Toft Sørensen, Timothy L. Lash); Department of Epidemiology, School of Public Health, Boston University, Boston, Massachusetts (Thomas P. Ahern, Kenneth J. Rothman, Timothy L. Lash); and RTI Health Solutions, RTI International, Research Triangle Park, North Carolina (Kenneth J. Rothman).

This work was supported by the US National Cancer Institute at the National Institutes of Health (grant R01 CA118708), the Danish Research Foundation (grant 10-084581), and the Karen Elise Jensen Foundation. T. P. A. was supported in part by a Congressionally Directed Medical Research Programs predoctoral award (grant BC073012) and in part by a National Institutes of Health postdoctoral training award (grant NIH 5 T32 CA009001).

Conflict of interest: none declared.

References

1.
Donati
MB
Falanga
A
Pathogenetic mechanisms of thrombosis in malignancy
Acta Haematol
 , 
2001
, vol. 
106
 
1-2
(pg. 
18
-
24
)
2.
Rickles
FR
Edwards
RL
Activation of blood coagulation in cancer: Trousseau’s syndrome revisited
Blood
 , 
1983
, vol. 
62
 
1
(pg. 
14
-
31
)
3.
Agnelli
G
Taliani
MR
Prandoni
P
More on: vitamin K antagonists and cancer. WODIT Investigators
J Thromb Haemost
 , 
2008
, vol. 
6
 
8
(pg. 
1441
-
1442
author reply 1442–1443
4.
Di Nisio
M
Buller
HR
Porreca
E
Do low-molecular-weight heparins improve the survival of cancer patients?
Nat Clin Pract Oncol
 , 
2005
, vol. 
2
 
12
(pg. 
612
-
613
)
5.
Klerk
CP
Smorenburg
SM
Otten
HM
, et al.  . 
The effect of low molecular weight heparin on survival in patients with advanced malignancy
J Clin Oncol
 , 
2005
, vol. 
23
 
10
(pg. 
2130
-
2135
)
6.
Piccioli
A
Falanga
A
Prandoni
P
Anticoagulants and cancer survival
Semin Thromb Hemost
 , 
2006
, vol. 
32
 
8
(pg. 
810
-
813
)
7.
Goodman
LS
Gilman
A
Brunton
LL
, et al.  . 
Goodman & Gilman’s the Pharmacological Basis of Therapeutics
 , 
2006
11th ed
New York, NY
McGraw-Hill
8.
Katzung
BG
Basic and Clinical Pharmacology
 , 
2004
9th ed
New York, NY
Lange Medical Books/McGraw Hill
9.
Zacharski
LR
Henderson
WG
Rickles
FR
, et al.  . 
Effect of warfarin on survival in small cell carcinoma of the lung. Veterans Administration Study No. 75
JAMA
 , 
1981
, vol. 
245
 
8
(pg. 
831
-
835
)
10.
Zacharski
LR
Henderson
WG
Rickles
FR
, et al.  . 
Effect of warfarin anticoagulation on survival in carcinoma of the lung, colon, head and neck, and prostate. Final report of VA Cooperative Study #75
Cancer
 , 
1984
, vol. 
53
 
10
(pg. 
2046
-
2052
)
11.
Schulman
S
Lindmarker
P
Incidence of cancer after prophylaxis with warfarin against recurrent venous thromboembolism. Duration of Anticoagulation Trial
N Engl J Med
 , 
2000
, vol. 
342
 
26
(pg. 
1953
-
1958
)
12.
Eikelboom
JW
Mehta
SR
Venous thromboembolism and cancer [letter]
N Engl J Med
 , 
2000
, vol. 
343
 
18
pg. 
1337
  
author reply 1338
13.
Hughes-Davies
TH
Venous thromboembolism and cancer [letter]
N Engl J Med
 , 
2000
, vol. 
343
 
18
pg. 
1337
  
author reply 1338
14.
Roychowdhury
D
Venous thromboembolism and cancer [letter]
N Engl J Med
 , 
2000
, vol. 
343
 
18
(pg. 
1337
-
1338
)
15.
Zacharski
LR
Ornstein
DL
Venous thromboembolism and cancer [letter]
N Engl J Med
 , 
2000
, vol. 
343
 
18
pg. 
1338
 
16.
Zielinski
CC
Hejna
M
Warfarin for cancer prevention
N Engl J Med
 , 
2000
, vol. 
342
 
26
(pg. 
1991
-
1993
)
17.
Taliani
MR
Agnelli
G
Prandoni
P
, et al.  . 
Incidence of cancer after a first episode of idiopathic venous thromboembolism treated with 3 months or 1 year of oral anticoagulation. Warfarin Optimal Duration Italian Trial (WODIT) Investigators
J Thromb Haemost
 , 
2003
, vol. 
1
 
8
(pg. 
1730
-
1733
)
18.
Blumentals
WA
Foulis
PR
Schwartz
SW
, et al.  . 
Does warfarin therapy influence the risk of bladder cancer?
Thromb Haemost
 , 
2004
, vol. 
91
 
4
(pg. 
801
-
805
)
19.
Tagalakis
V
Tamim
H
Blostein
M
, et al.  . 
Use of warfarin and risk of urogenital cancer: a population-based, nested case-control study
Lancet Oncol
 , 
2007
, vol. 
8
 
5
(pg. 
395
-
402
)
20.
Pengo
V
Noventa
F
Denas
G
, et al.  . 
Long-term use of vitamin K antagonists and incidence of cancer: a population-based study
Blood
 , 
2011
, vol. 
117
 
5
(pg. 
1707
-
1709
)
21.
Pengo
V
Denas
G
Jose
SP
, et al.  . 
Cancer prevention and vitamin K antagonists: an overview
Thromb Res
 , 
2010
, vol. 
125
 
suppl 2
(pg. 
S103
-
S105
)
22.
Hernán
MA
Robins
JM
Instruments for causal inference: an epidemiologist’s dream?
Epidemiology
 , 
2006
, vol. 
17
 
4
(pg. 
360
-
372
)
23.
Bonow
RO
Carabello
BA
Kanu
C
, et al.  . 
ACC/AHA 2006 guidelines for the management of patients with valvular heart disease: a report of the American College of Cardiology/American Heart Association Task Force on Practice Guidelines (writing committee to revise the 1998 Guidelines for the Management of Patients With Valvular Heart Disease): developed in collaboration with the Society of Cardiovascular Anesthesiologists: endorsed by the Society for Cardiovascular Angiography and Interventions and the Society of Thoracic Surgeons
Circulation
 , 
2006
, vol. 
114
 
5
(pg. 
e84
-
e231
)
24.
Lash
TL
Cole
SR
Immortal person-time in studies of cancer outcomes
J Clin Oncol
 , 
2009
, vol. 
27
 
23
(pg. 
e55
-
e56
)
25.
Storm
HH
Michelsen
EV
Clemmensen
IH
, et al.  . 
The Danish Cancer Registry—history, content, quality and use
Dan Med Bull
 , 
1997
, vol. 
44
 
5
(pg. 
535
-
539
)
26.
Charlson
ME
Pompei
P
Ales
KL
, et al.  . 
A new method of classifying prognostic comorbidity in longitudinal studies: development and validation
J Chronic Dis
 , 
1987
, vol. 
40
 
5
(pg. 
373
-
383
)
27.
 
WHO Collaborating Centre for Drug Statistics Methodology. ATC/DDD index 2011. Oslo, Norway: Norwegian Institute of Public Health; 2008. (http://www.whocc.no/atcddd/). (Accessed 2008)
28.
Aurigemma
GP
Gaasch
WH
Otto
CM
Antithrombotic therapy in patients with prosthetic heart valves
UpToDate
 , 
2010
Waltham, MA
UpToDate
29.
Greenland
S
An introduction to instrumental variables for epidemiologists
Int J Epidemiol
 , 
2000
, vol. 
29
 
4
(pg. 
722
-
729
)
30.
Greenland
S
Pearl
J
Robins
JM
Causal diagrams for epidemiologic research
Epidemiology
 , 
1999
, vol. 
10
 
1
(pg. 
37
-
48
)
31.
Brookhart
MA
Wang
PS
Solomon
DH
, et al.  . 
Instrumental variable analysis of secondary pharmacoepidemiologic data
Epidemiology
 , 
2006
, vol. 
17
 
4
(pg. 
373
-
374
)
32.
Brookhart
MA
Wang
PS
Solomon
DH
, et al.  . 
Evaluating short-term drug effects using a physician-specific prescribing preference as an instrumental variable
Epidemiology
 , 
2006
, vol. 
17
 
3
(pg. 
268
-
275
)
33.
Cunnane
C
Unbiased plotting positions—a review
J Hydrol
 , 
1978
, vol. 
37
 (pg. 
205
-
222
)
34.
Steenland
K
Bray
I
Greenland
S
, et al.  . 
Empirical Bayes adjustments for multiple results in hypothesis-generating or surveillance studies
Cancer Epidemiol Biomarkers Prev
 , 
2000
, vol. 
9
 
9
(pg. 
895
-
903
)
35.
Sørensen
HT
Johnsen
SP
Nørgård
B
, et al.  . 
Cancer and venous thromboembolism: a multidisciplinary approach
Clin Lab
 , 
2003
, vol. 
49
 
11-12
(pg. 
615
-
623
)
36.
Baron
JA
Gridley
G
Weiderpass
E
, et al.  . 
Venous thromboembolism and cancer
Lancet
 , 
1998
, vol. 
351
 
9109
(pg. 
1077
-
1080
)
37.
Sørensen
HT
Mellemkjaer
L
Steffensen
FH
, et al.  . 
The risk of a diagnosis of cancer after primary deep venous thrombosis or pulmonary embolism
N Engl J Med
 , 
1998
, vol. 
338
 
17
(pg. 
1169
-
1173
)
38.
Rothman
KJ
Greenland
S
Lash
TL
Validity in epidemiologic studies
Modern Epidemiology
 , 
2008
Philadelphia, PA
Lippincott Williams & Wilkins
(pg. 
128
-
147
)
39.
Storm
HH
Pihl
J
Michelsen
E
, et al.  . 
Cancer Incidence in Denmark, 1993
 , 
1996
Copenhagen, Denmark
Danish Cancer Society