Uncovering hypothesis-generating associations between prescription drugs and digestive cancer risk: a case-control drug-wide association study (DWAS)

Abstract

Background:

Prescription drug use among older adults is more common than ever, yet traditional trials are not designed to assess long-term outcomes. Thus, the potential impact of widely used medications on cancer risk remains unclear. This is particularly a concern for digestive cancers for several reasons, most notably their key role in drug absorption and metabolism. We therefore leverage large-scale data to conduct a multi-stage study, aimed at identifying potential associations between oral prescription drugs and digestive cancer risk.

Methods:

We used the Surveillance, Epidemiology and End Results (SEER)-Medicare data for a case-control drug-wide association study (DWAS), followed by targeted analyses to further address confounding. Subsequent analyses conducted include incorporation of additional adjustment for comorbidities, use of active comparators as well as propensity scores. Associations between prescription drugs and digestive cancer were assessed using adjusted odds ratios (OR), with significance assessed using Benjamini-Yekutieli (BY)-adjusted p-values.

Results:

In initial scan, 218 of 442 oral prescription drugs were associated with digestive cancer (BY-adjusted p<0.05). After our multi-stage targeted analyses, 38 drugs of interest (vs. their active comparators) remained associated with digestive cancer. Specifically, 10 drugs were positively associated (OR>1) and 28 were inversely associated (OR<1).

Conclusions:

Our study confirms some previously reported associations and identifies additional potential, novel associations between individual oral prescription drugs and digestive cancer that merit additional follow-up.

Impact:

These findings support the utility of our approach for hypothesis generation and warrant targeted confirmation in prospective studies as well as investigation of specific cancer types driving these associations.

Introduction

Older adults in the United States (U.S.) are using more prescription drugs than ever before,(1,2) with a recent survey of U.S. adults showing that nearly 90% of adults aged 65+ reported use of any prescription drugs in 2017–2020 and 43% reporting use of 5 or more prescription drugs.(2) Similar findings have been reported in Canada(3) and European countries.(4)

Despite the high prevalence of prescription drug use, little is known about the potential role of prescription drugs in long-term cancer risk.(5,6) This is particularly important for cancers of the digestive system, as the digestive tract and accessory organs play a key role in drug absorption and metabolism.(7) Furthermore, many drugs act on inflammatory and metabolic pathways, which are thought to play a role in digestive cancer development.(8–11) To this end, evidence suggests the digestive system may be a frequent target of unintended drug effects in older populations.(12–14) Yet, digestive cancers are not well suited to detection in clinical trials and post-market surveillance studies due to their long latency,(15) and recent widely-publicized recalls or withdrawals typically occurred years after approval.(16,17)

Malignancies of the digestive system constitute a leading cause of cancer death in the U.S.(18) and globally,(19) and identifying ways to mitigate their risk is crucial. Prior epidemiological studies have provided important insights on prescription drug use in relation to risk of digestive cancer; however, these studies have been focused on broad classes of drugs (e.g., statins, oral contraceptives) rather than individual drugs and cancer risk.(20–22) This is an important limitation, given that even within a drug class, individual medications may have heterogeneous properties (e.g., lipophilic or hydrophilic statins(23)) that may differentially impact risk.

Inspired by genome-wide association studies (GWAS), a drug-wide association study (DWAS) provides a broad, hypothesis-generating approach to detect unintended consequences of prescription drugs. Four DWAS on drug-cancer associations have been conducted,(5,6,24,25) providing evidence of novel, biologically plausible drug associations with cancer and demonstrating the utility of such an approach. However, limitations of these studies, such as scanning associations without controlling for drug indication (which itself can be associated with cancer), demonstrate that additional follow-up steps are needed to refine findings. To address this important gap, we leveraged large-scale real-world data to conduct a novel tiered-based approach to DWAS analysis. First, we sought to uncover associations between oral prescription drugs and risk of digestive cancers in a case-control DWAS, then conducted a series of multi-stage targeted analyses to refine drug-cancer associations to those that are robust and warrant more extensive follow-up in future studies.

Materials and Methods

Ethics Statement

This study was reviewed by the Memorial Sloan Kettering Cancer Center Institutional Review Board (IRB) and determined to be exempt from formal IRB review (#X16–030). Patients or the public were not involved in the design, or conduct, or reporting, or dissemination plans of our research.

Data Source

The Surveillance, Epidemiology and End Results (SEER)-Medicare dataset comprises cancer registry data linked to Medicare enrollment and claims. The dataset also includes a random 5% sample of Medicare beneficiaries without cancer who reside in the SEER areas and serve as a comparison group.(26,27) The SEER program collects data on cancer patient demographics, and primary tumor site, among other data. Its population-based cancer registries currently cover 45.9 percent of the U.S. population.(28) Medicare is a federally funded health insurance program that insures most US adults 65 years of age and older.(27) Its drug coverage program (Part D) went into effect in 2006 and enrolled 74% of Medicare beneficiaries in 2019.(29)

Identification of Cases and Controls

Eligible cases were participants who were diagnosed between January 1, 2009 and December 31, 2019 with any incident digestive system cancer. This included cancers of the oral cavity and pharynx, esophagus, stomach, small intestine, colorectum, liver and bile duct, pancreas, gallbladder and other biliary, and other digestive system (see ICD-O-3 codes in Supplemental Methods). To capture exposure prior to the onset of disease, we defined the index date as one year prior to the date of cancer diagnosis. Figure 1 and Supplemental Methods detail the participant selection criteria. Briefly, we excluded individuals with any of the following: prior cancer diagnosis except non-melanoma skin cancer, <2 years of Medicare enrollment prior to cancer diagnosis, <1 year of continuous Part D prescription drug coverage since first enrollment of Part D, age<66 on index date, ≥3-months Part D enrollment gap until the index date, or no socioeconomic status (SES) information. As individuals with end stage renal disease or long-term disability are eligible for Medicare before age 65, leading to their claims before age 65 being recorded by Medicare, we truncated their Medicare records to only include claims after they turned 65.

Figure 1.

Flowchart of study population assembly.

We selected potential control individuals from the random 5% sample of non-cancer Medicare beneficiaries (Figure 1; Supplemental Methods). For beneficiaries enrolled prior to age 65, we truncated their records to only include the period after they turned 65. We next excluded non-cancer Medicare beneficiaries with <1 year of continuous Part D prescription drug coverage since initial enrollment. We then matched eligible non-cancer individuals to each case by reference year of cancer diagnosis, age, sex, cumulative months of Part D enrollment, and year of first Part D enrollment. After applying remaining exclusions similar to those in our case population (Figure 1), we randomly sampled 2 to 4 eligible cancer-free controls for each case. For each control, we defined the index date as one year prior to the date of diagnosis of the matched case.

Exposures

We assessed prescription drugs administered orally—including oromucosally and sublingually—and accompanied by a valid dose, with a dispensation recorded in claims between the date of first Part D enrollment and the index date. Only drugs that are administered orally were included, given that we were interested in drugs that are absorbed and/or metabolized by the digestive tract/organs and also have direct contact with the digestive tract. We excluded drugs that would be prescribed for a colonoscopy (Supplemental Methods),(30) given the potential for reverse causality, as well as over-the-counter medications, given that these are rarely captured in Medicare claims data. We extracted ingredients from combination drugs and collapsed different salts (e.g., “doxycycline calcium” and “doxycycline monohydrate” were collapsed into “doxycycline”). We used the World Health Organization’s anatomical, therapeutic, and chemical (ATC) classification to categorize drugs.(31)

Covariates

We measured numerous covariates to ensure comparability between cases and controls. Sociodemographics included age at event date (years), sex (female; male), combined race and ethnicity (non-Hispanic White; Black/African American; Asian/Pacific Islander; Hispanic; American Indians/Alaska Natives; other; unknown), cumulative months of Part D enrollment to event date (months), year of first enrollment of Part D, census tract-level per capita income (PCI) at event date (US dollars), census tract-level percentage of residents living below poverty at event date (%), census tract-level percentage of persons aged 25+ with at least 4 years of college at event date (%).

Overall health status variables included NCI comorbidity index in 1 year prior to index date (diagnosis codes and algorithm in Supplemental Methods)(32) and healthcare utilization in the 1 year prior to index date (any use vs none).

Digestive cancer-related comorbidities diagnosed prior to index date included HIV/AIDS, diabetes, inflammatory bowel disease (IBD), peptic ulcer, alcohol use disorder, and liver disease. Additional variables included hepatitis B, gallstones, and tobacco use (diagnosis codes in Supplemental Methods).

Statistical Analysis

We used a multi-stage design that integrated a drug-wide association scan followed by multiple steps to address potential confounding (Figure 2 and Supplemental Methods). We used the Benjamini-Yekutieli (BY) procedure, a type of False Discovery Rate (FDR) correction, to account for multiple testing.(33) We used adjusted p-values to assess the presence of association at α=0.05. All analyses were conducted in SAS and R 4.3.2 GUI. Specific R packages used in our study included survival (Version 3.5–7), multtest (Version 2.56.0), ggplot2 (Version 3.4.2), dbarts (Version 0.9–30), WeightIt (Version 1.4.0), survey (Version 4.4–2), and tableone (Version 0.13.2).

Figure 2.

Multi-stage design to discover and follow-up drug associations with digestive cancer risk.

Stage I – Initial Scan

In the initial drug-wide association scan, we used multivariable conditional logistic regression to estimate odds ratios (ORs) and 95% confidence intervals (CIs) for each drug-cancer association. We conditioned on matched pairs and further adjusted for race and ethnicity, NCI comorbidity index, healthcare utilization, and census tract-level SES indicators including income, poverty, and education. We chose a minimal set of covariates since this step was used for hypothesis-generation across a range of drug exposures. We conservatively restricted analyses to include drugs used by ≥275 participants, given a priori power calculations suggested this sample size provided >80% power to detect an odds ratio of 1.5. Of note, we could still detect OR<1.5 if a large sample size provided sufficient statistical power.

Stage II – Adjustment for Comorbidities

We assessed the magnitude of statistically significant drug-cancer associations from the initial scan. Those associations with a >10% magnitude of association, reflecting potential clinical relevance, were carried forward to additionally adjust for comorbidities using multivariable conditional logistic regression. Given the range of exposures included, these comorbidities (HIV/AIDS, diabetes, IBD, peptic ulcer, alcohol use disorder, and liver disease) were selected given their potential association with risk of digestive cancer after consulting a pharmacist and a gastroenterologist.

Stage III – Active Comparator Analysis

After Stage II, we conducted active comparator analyses to control for confounding by drug indication(34).

For drugs with a statistically significant positive association in Stage II, appropriate comparators for each drug were independently selected by a pharmacist and a gastroenterologist, with differences in selected comparator resolved through discussion and consensus among the study team. When no appropriate active comparator was available, alternative strategies were evaluated, including requiring indication diagnoses in comparators that were not in the drug of interest.

For drugs with statistically significant inverse associations in Stage II, we used statins as the active comparator. This approach has been used to control for the possibility that healthier individuals may be more likely to initiate a chronic medication.(35)

In these analyses, only participants who mutually exclusively used the drug of interest or comparator were included. For drugs mainly used by females or males, we restricted analyses by sex. The drugs of interest, comparators, and covariates are listed in Supplemental Methods.

Stage IV – Active Comparator Analysis with Propensity Score Adjustment

To address the potential for residual confounding, we next used propensity score (PS) adjustment to further adjust for residual confounding in active comparator analyses. For statistically significant drugs and their comparators from Stage III, propensity scores were estimated using Bayesian Additive Regression Trees (BART) with 200 trees, 1000 iterations, and drug of interest (vs. comparator) as outcome.(36–38) We examined associations between drugs of interest (vs. their comparators) and cancer using logistic regression with overlap weights of propensity scores.(39,40) We excluded drugs used by <300 individuals in total. We used standardized mean difference (SMD) to assess covariate balance in the weighted population, and covariates that were still unbalanced (SMD≥0.1) were included in final logistic regression model.(41)

A total of 43 covariates were used for this stage, including demographics, calendar year of index date, Part D enrollment information, comorbidities/lifestyle factors, diabetic complications, insulin use, liver disease, IBD, healthcare utilization, and preventive care utilization. Details are presented in Supplemental Methods.

Sensitivity Analyses

Two sensitivity analyses were performed to assess the robustness of estimates from the final stage IV analyses. First, we evaluated statistically significant drugs of interest and their comparators using a long-term use analysis (≥180 days use of drug versus ≥180 days use of comparator) with a 12-month washout period (i.e., the period without any use of drugs of interest or comparators to better capture drug initiation) for drugs used for chronic diseases; BART with overlap weight was used to adjust for potential confounders. This approximates a dose-response analysis, which could not be directly performed in our data as we did not have drug dispensing information before age 65 for most participants. Second, for anti-dementia and anti-Parkinson drugs, we restricted analyses to individuals with at least a 5-year washout period before first use of drugs of interest or statins to further control for potential healthy user bias.

Data Availability

The conditions of the data use agreement with the National Cancer Institute prohibit direct data sharing; however, investigators may apply directly to the National Cancer Institute to receive the data used for the study.

Results

Description of Cases and Controls

A total of 288,497 cases and 1,150,361 controls were eligible for this study. Cases and controls had a similar sociodemographic characteristics and overall health status, although cases were more likely to be Black or African American (10.2% vs 8.2%, Table 1). Compared to controls, cases had higher prevalence of some cancer-related comorbidities, such as diabetes, alcohol use disorder, and liver disease.

Table 1.

Descriptive characteristics of study population by case-control status

	Overall N = 1438858	Digestive Cancer Patients N = 288497	Control Individuals N = 1150361
	N (%) / Median (IQR)	N (%) / Median (IQR)	N (%) / Median (IQR)
Sociodemographics
Age (years)	77 (72, 83)	77 (72, 83)	77 (72, 83)
Sex
Female	745405 (51.8)	149367 (51.8)	596038 (51.8)
Male	693453 (48.2)	139130 (48.2)	554323 (48.2)
Race and ethnicity
American Indian/Alaska Native	4067 (0.3)	801 (0.3)	3266 (0.3)
Asian/Pacific Islander	106595 (7.4)	22016 (7.6)	84579 (7.4)
Black/African American	123525 (8.6)	29376 (10.2)	94149 (8.2)
Hispanic	193439 (13.4)	38480 (13.3)	154959 (13.5)
Non-Hispanic White	983877 (68.4)	192688 (66.8)	791189 (68.8)
Other	17859 (1.2)	3497 (1.2)	14362 (1.2)
Unknown	9496 (0.7)	1639 (0.6)	7857 (0.7)
Months of Medicare Part D enrollment	61 (37, 89)	61 (37, 89)	61 (37, 89)
Per capita incomea	26703 (20953, 35621)	26309 (20511, 35096)	26812 (21039, 35811)
Percent of residents living below povertya	13 (7, 20)	13 (8, 20)	12 (7, 20)
Percent of residents with ≥4 years of collegea	26 (16, 39)	25 (16, 38)	26 (16, 40)
Overall health status
NCI comorbidity indexb	0.00 (0.00, 0.52)	0.00 (0.00, 0.57)	0.00 (0.00, 0.52)
Healthcare utilizationb	819519 (57.0)	164507 (57.0)	655012 (56.9)
Specific cancer-related comorbidities c
HIV/AIDS	2756 (0.2)	671 (0.2)	2085 (0.2)
Diabetes	459355 (31.9)	102106 (35.4)	357249 (31.1)
Inflammatory bowel disease	17581 (1.2)	3593 (1.2)	13988 (1.2)
Peptic ulcer	55675 (3.9)	12120 (4.2)	43555 (3.8)
Alcohol use disorder	113589 (7.9)	30219 (10.5)	83370 (7.2)
Liver disease	117006 (8.1)	30247 (10.5)	86759 (7.5)

^aVariables were measured in census tract-level. The unit of per capita income is US dollar. The unit of other variables is %.

^bVariables were measured in 1 year prior to index date.

^cVariables were measured prior to index date.

Initial Scan and Adjustment for Additional Comorbidities

Of the 442 prescription drugs included in the initial drug-cancer association scan, 218 were statistically significantly associated with digestive cancer risk (BY-adjusted p<0.05, Figure 2 and Supplementary Table 1). Of these 218 drugs, 67 were positively associated (OR>1), and 151 inversely associated (OR<1) (Supplementary Figure 1).

Out of 218 associated drugs, 129 met the a priori >10% threshold for a clinically relevant magnitude of association. We additionally adjusted for cancer-related comorbidities, and 127 associations remained statistically significant, of which 42 were positively associated and 85 were inversely associated (Figure 3 and Supplementary Table 2).

Figure 3.

Associations between prescription drug and any digestive cancer risk*

* Restricted to associations with clinically relevant magnitude (>10%) from initial scan and further adjusted for cancer-related comorbidities. Specifically, accounted for matching factors (reference year, age, sex, months of Part D enrollment, and year of Part D first enrollment) and further adjusted for race/ethnicity, census tract-level income, census tract-level poverty, census tract-level education, NCI comorbidity index, healthcare utilization, and cancer-related comorbidities (diabetes, liver disease, peptic ulcers, inflammatory bowel disease, HIV/AIDS, alcohol use disorder). P-value was Benjamini-Yekutieli (BY) adjusted p-value. Gray areas highlight drugs with a relatively large magnitude of association (OR⩾1.5 or ⩽0.7).

Drugs Positively Associated with Digestive Cancers

We conducted active comparator analyses for the 42 drugs positively associated with digestive cancer risk and associations for 20 drugs retained statistical significance, including 15 positive associations (Supplementary Table 3). After excluding tolazamide due to small sample size and applying additional PS adjustment (Supplementary Tables 4-5), 14 drug-cancer associations retained statistical significance. Of them, 10 drugs remained positively associated (vs. their comparators) and 4 drugs (tenofovir, metformin, pioglitazone, pentoxifylline) became inversely associated.

The 10 drugs that retained positive associations with digestive cancer were as follows: entecavir (OR: 1.44; 95% CI:1.10 to 1.87), glipizide (1.06; 1.02 to 1.09), nadolol (1.28; 1.20 to 1.37), furosemide (1.07; 1.05 to 1.09), spironolactone (1.10; 1.07 to 1.13), torsemide (1.10; 1.04 to 1.16), fosinopril (1.08; 1.02 to 1.14), cilostazol (1.14; 1.07 to 1.22), digoxin (1.05; 1.02 to 1.08), and azathioprine (1.46; 1.32 to 1.62) (Table 2). Cancer associations for metformin, pioglitazone, tenofovir, and pentoxifylline reversed direction when they were compared to their active comparators (Table 2). For instance, both metformin and pioglitazone use were inversely associated with digestive cancer compared to 2^nd generation sulfonylureas after PS adjustment.

Table 2.

Selected associations between potential cancer-risk drug vs. comparator and any digestive cancer risk, with and without propensity score (PS) adjustment^a

			Active Comparator Analysis		Active Comparator Analysis with PS Adjustmentb
Drug	Comparator	N (Drug vs. Comparator)	OR (95% CI)	BY-adj p	OR (95% CI)	BY-adj p
Neomycin	Rifaximin	2915 vs. 4338	0.84 (0.75, 0.95)	0.036	1.06 (0.93, 1.20)	>0.999
Entecavir	Tenofovirc	602 vs. 491	1.59 (1.23, 2.07)	0.006	1.44 (1.10, 1.87)	0.037
Tenofovirc	Entecavir	491 vs. 602	0.63 (0.48, 0.81)	0.006	0.69 (0.53, 0.90)	0.034
Glimepiride	Metformin	13449 vs. 236940	1.07 (1.02, 1.11)	0.027	1.04 (1.00, 1.08)	0.295
Glipizide	Metformin	25454 vs. 201881	1.09 (1.06, 1.13)	<0.001	1.06 (1.02, 1.09)	0.006
Glyburide	Metformin	11809 vs. 230036	1.07 (1.02, 1.12)	0.034	1.03 (0.98, 1.07)	>0.999
Metformin	2nd Generation Sulfonylureas	135961 vs. 45638	0.88 (0.86, 0.91)	<0.001	0.92 (0.90, 0.95)	<0.001
Pioglitazone	2nd Generation Sulfonylureas	24180 vs. 142026	0.90 (0.87, 0.93)	<0.001	0.92 (0.89, 0.95)	<0.001
Nadolol	Carvedilol	5656 vs. 112318	1.68 (1.59, 1.78)	<0.001	1.28 (1.20, 1.37)	<0.001
Propranolol	Carvedilol	25385 vs. 110822	1.06 (1.02, 1.10)	0.011	1.00 (0.96, 1.04)	>0.999
Furosemide	Hydrochlorothiazide	154493 vs. 416501	1.14 (1.12, 1.15)	<0.001	1.07 (1.05, 1.09)	<0.001
Spironolactone	Hydrochlorothiazide	30759 vs. 505038	1.29 (1.25, 1.32)	<0.001	1.10 (1.07, 1.13)	<0.001
Torsemide	Hydrochlorothiazide	7844 vs. 526824	1.16 (1.10, 1.23)	<0.001	1.10 (1.04, 1.16)	0.003
Fosinopril	Valsartan	9106 vs. 145535	1.10 (1.05, 1.16)	0.002	1.08 (1.02, 1.14)	0.022
Cilostazol	Pentoxifylline	15192 vs. 7953	1.15 (1.07, 1.22)	<0.001	1.14 (1.07, 1.22)	<0.001
Digoxin	Diltiazem	43720 vs. 70582	1.07 (1.04, 1.10)	<0.001	1.05 (1.02, 1.08)	0.021
Pentoxifylline	Cilostazol	7953 vs. 15192	0.87 (0.82, 0.93)	<0.001	0.88 (0.82, 0.94)	<0.001
Azathioprine	Methotrexate (oral)	2878 vs. 17345	1.53 (1.39, 1.69)	<0.0001	1.46 (1.32, 1.62)	<0.001
Ursodiold	No Ursodiold	1211 vs. 42407	1.23 (1.08, 1.40)	0.017	1.01 (0.89, 1.15)	>0.999

^a.Only drugs included in active comparator analysis with PS adjustment were selected. Full list of comparisons was shown in Supplementary Table 5. Sample sizes shown in this table were for active comparator analysis. Sample sizes for active comparator analysis with PS adjustment were shown in Supplementary Table 4.

^b.Only drugs with statistical significance in active comparator analysis were included, except for tolazamide, which was excluded due to sample size <300. Propensity scores were estimated by using Bayesian Additive Regression Trees with 200 trees and 1000 iterations. Associations between drug vs comparator and any digestive cancer were examined by using logistic regression with overlap weights of propensity scores.

^c.Only Tenofovir users with a hepatitis B diagnosis code.

^d.Only participants with at least two different healthcare visit records with a gallstone diagnosis code.

Long-term use analyses for drugs used for chronic conditions (N = 12) showed that nadolol (1.53; 95%CI:1.31 to 1.79), furosemide (1.11; 1.07 to 1.15), spironolactone (1.11; 1.04 to 1.17), torsemide (1.19; 1.06 to 1.33), and azathioprine (1.47; 1.20 to 1.80) remained statistically significantly associated with digestive cancers compared to their comparators (Supplementary Table 6).

Drugs Inversely Associated with Digestive Cancers

For the 85 drugs inversely associated with cancer from Stage II, active comparator analyses using statins as a comparator yielded 35 drugs that retained statistically significant associations (Supplementary Table 3). For these drugs (except bicalutamide and letrozole, which were excluded due to small sample size), additional PS adjustment yielded 24 statistically significant inverse associations (Supplementary Tables 4-5).

Following PS adjustment, drugs with inverse associations included antipsychotics (aripiprazole, haloperidol, olanzapine, quetiapine, and risperidone), antidepressants (escitalopram, mirtazapine, sertraline, and trazodone), anti-dementia drugs (donepezil, galantamine, memantine, and rivastigmine), anti-Parkinson drugs (carbidopa, levodopa, and rasagiline), other nervous system drugs (divalproex and sumatriptan), antimalarials (atovaquone, mefloquine, and proguanil), sex hormones (estradiol, esterified/conjugated estrogens), and an immunosuppressant (leflunomide) (Table 3).

Table 3.

Selected associations between potential cancer-protective drug vs. comparator and any digestive cancer risk, with and without propensity score (PS) adjustment^a

		Active Comparator Analysis		Active Comparator Analysis with PS Adjustmentb
Drug	N (Drug vs. Statins)	OR (95% CI)	BY-adj p	OR (95% CI)	BY-adj p
Sumatriptan	3205 vs. 856492	0.76 (0.69, 0.84)	<0.001	0.80 (0.72, 0.88)	<0.001
Divalproex	5832 vs. 852486	0.77 (0.72, 0.83)	<0.001	0.84 (0.78, 0.90)	<0.001
Carbidopa	9469 vs. 847094	0.85 (0.80, 0.89)	<0.001	0.81 (0.77, 0.86)	<0.001
Levodopa	9469 vs. 847094	0.85 (0.80, 0.89)	<0.001	0.82 (0.77, 0.86)	<0.001
Rasagiline	1182 vs. 860450	0.74 (0.63, 0.87)	0.006	0.74 (0.63, 0.87)	0.003
Aripiprazole	3297 vs. 855309	0.79 (0.72, 0.87)	<0.001	0.85 (0.77, 0.93)	0.004
Haloperidol	2780 vs. 857934	0.77 (0.69, 0.85)	<0.001	0.84 (0.76, 0.93)	0.007
Lorazepam	19685 vs. 820441	0.94 (0.91, 0.98)	0.024	0.97 (0.93, 1.00)	0.318
Olanzapine	5289 vs. 853524	0.84 (0.78, 0.90)	<0.001	0.89 (0.82, 0.95)	0.007
Quetiapine	11927 vs. 840132	0.76 (0.72, 0.80)	<0.001	0.83 (0.79, 0.88)	<0.001
Risperidone	8245 vs. 849093	0.78 (0.73, 0.82)	<0.001	0.84 (0.79, 0.89)	<0.001
Ziprasidone	849 vs. 860572	0.73 (0.60, 0.87)	0.011	0.79 (0.65, 0.95)	0.061
Donepezil	26756 vs. 807860	0.74 (0.72, 0.77)	<0.001	0.77 (0.74, 0.80)	<0.001
Galantamine	1677 vs. 858924	0.61 (0.53, 0.71)	<0.001	0.63 (0.54, 0.73)	<0.001
Memantine	14941 vs. 832586	0.69 (0.66, 0.72)	<0.001	0.72 (0.68, 0.75)	<0.001
Rivastigmine	1667 vs. 858703	0.67 (0.58, 0.76)	<0.001	0.70 (0.61, 0.80)	<0.001
Citalopram	26194 vs. 802313	0.94 (0.91, 0.97)	<0.001	0.96 (0.93, 1.00)	0.112
Escitalopram	21323 vs. 814700	0.93 (0.90, 0.97)	0.002	0.95 (0.91, 0.98)	0.018
Mirtazapine	15607 vs. 833007	0.89 (0.85, 0.93)	<0.001	0.93 (0.89, 0.97)	0.005
Sertraline	26733 vs. 798183	0.94 (0.91, 0.97)	0.001	0.94 (0.91, 0.97)	0.001
Trazodone	26834 vs. 802516	0.94 (0.91, 0.97)	0.002	0.95 (0.92, 0.98)	0.007
Venlafaxine	8799 vs. 840987	0.92 (0.87, 0.97)	0.035	0.93 (0.88, 0.98)	0.051
Atovaquone	2924 vs. 857451	0.78 (0.70, 0.86)	<0.001	0.85 (0.77, 0.94)	0.012
Mefloquine	1149 vs. 859892	0.63 (0.52, 0.74)	<0.001	0.65 (0.54, 0.77)	<0.001
Proguanil	2782 vs. 857688	0.76 (0.68, 0.84)	<0.001	0.83 (0.75, 0.93)	0.005
Leflunomide	1381 vs. 859687	0.80 (0.70, 0.93)	0.035	0.80 (0.69, 0.92)	0.013
Estradiolc	12119 vs. 423308	0.76 (0.72, 0.80)	<0.001	0.84 (0.80, 0.89)	<0.001
Estrogensc,d	12912 vs. 423395	0.82 (0.78, 0.86)	<0.001	0.88 (0.84, 0.92)	<0.001
Medroxyprogesteronec	4070 vs. 434757	0.86 (0.79, 0.94)	0.008	0.93 (0.86, 1.02)	0.450
Progesteronec	1461 vs. 438277	0.74 (0.64, 0.86)	0.002	0.84 (0.72, 0.97)	0.095
Levothyroxine	83723 vs. 696233	0.97 (0.95, 0.99)	0.024	0.98 (0.96, 1.00)	0.085
Thyroid	3508 vs. 858825	0.87 (0.80, 0.95)	0.035	0.91 (0.83, 0.99)	0.155
Hyoscyamine	4776 vs. 852997	0.88 (0.81, 0.94)	0.009	0.93 (0.86, 1.00)	0.198

^a.Only drugs included in active comparator analysis with PS adjustment were selected. Full list of comparisons was shown in Supplementary Table 5. Statins were used to control healthy user bias for all drugs in this table. Sample sizes shown in this table were for active comparator analysis. Sample sizes for active comparator analysis with PS adjustment were shown in Supplementary Table 4.

^b.Only drugs with statistical significance in active comparator analysis were included, except for bicalutamide and letrozole, which were excluded due to sample size <300. Propensity scores were estimated by using Bayesian Additive Regression Trees with 200 trees and 1000 iterations. Associations between drug vs comparator and any digestive cancer were examined by using logistic regression with overlap weights of propensity scores.

^c.For drugs mainly used by females (estradiol, estrogens, medroxyprogesterone, progesterone), analyses were restricted to female individuals.

^d.Estrogens referred to esterified or conjugated estrogens.

Long-term use analyses for drugs used for chronic conditions showed that 14 drugs remained statistically significantly associated with digestive cancers compared to statins, including quetiapine (0.80; 95%CI:0.71 to 0.89), mirtazapine (0.84; 0.77 to 0.92), donepezil (0.75; 0.70 to 0.80), carbidopa (0.84; 0.75 to 0.94), and divalproex (0.78; 0.66 to 0.91) (Supplementary Table 6).

Sensitivity analyses of the 4 anti-dementia and anti-Parkinson drugs with a 5-year washout period showed that associations for donepezil and memantine remained comparable to results of primary PS analysis (Supplementary Table 7).

Discussion

Principal Findings

In this hypothesis-generating study of 442 oral prescription drugs, after multi-stage targeted analyses using active comparators and PS adjustment, we identified 10 drugs that were positively associated with digestive cancers and 28 drugs that were inversely associated. By using a multi-stage process designed to reduce confounding and false positives, we detected robust associations that are more likely to be clinically meaningful and appropriate for confirmation in targeted prospective studies.

Our study confirmed some drugs previously shown in other studies (e.g., azathioprine, furosemide, spironolactone, donepezil, estradiol, and sertraline) as well as uncovered some potentially novel associations that merit additional follow-up due to a large number of older adults using those drugs (e.g., 44,583 individuals used mirtazapine and 44,338 used memantine in our study). Positive associations included drugs across antidiabetic, antihypertensive, diuretic, other cardiovascular, antiviral, and immunosuppressant classes, while inverse associations spanned antidiabetics, antipsychotics, antidepressants, cognitive enhancers, anti-Parkinson agents, antimalarials, sex hormones, anticonvulsants, antivirals, and vasoactive agents.

Comparison with previous drug-cancer screening studies

Several prior studies used hypothesis-generating approaches to identify hundreds of drug-cancer associations.(5,6,24,25) However, prior analyses had limited adjustment for confounding by drug indication, healthy user bias, or other factors.(35,42) Only two accounted for multiple testing.(24,25) As a result, those studies acknowledged that many signals may represent false positives.

To this end, a Kaiser Permanente study reported 101 associations from 105 drugs across 55 cancer types.(5) Two of these associations (diltiazem and metoprolol) were also detected in our initial scan but were not eligible for more targeted analyses due to small effect size. Similarly, a Danish registry study reported 510 drug-cancer signals, including 143 for digestive cancers.(6) Our scan replicated 21 signals in the initial scan, but only 6 (azathioprine, furosemide, spironolactone, donepezil, estradiol, and sertraline) persisted through targeted analyses, while metformin switched to an inverse association, aligned with prior epidemiologic studies.(43,44) Studies in Sweden²⁵ and Norway²⁴ examined drug classes rather than individual drugs, identifying 3 classes associated with colon cancer (osmotically acting laxatives, acetic acid derivatives, and opioids) that did not persist in our targeted analyses.

These comparisons show that many previously reported signals were attenuated or disappeared with more targeted adjustment for confounding and other sources of bias, highlighting the value of further refining initial signals to generate credible hypotheses for follow-up.

Potential links with cancer outcome

Digestive cancers are especially relevant to drug association studies due to well-established links between digestive cancers and metabolic and inflammatory pathways that can be affected by medications.(8–11)

Targeted analyses revealed inverse associations for metformin and pioglitazone vs. second-generation sulfonylureas. These findings support some preclinical evidence of anti-tumor activity, (45–48) although epidemiologic findings are inconsistent.(43,44,49–51) Active comparators allowed us to control for drug indication and unmeasured confounders, strengthening confidence in these signals and highlighting these drugs as candidates for future analyses and, eventually, chemoprevention studies.(49,52)

For antihypertensives, only nadolol, spironolactone, furosemide, torsemide, and fosinopril remained positively associated after targeted analyses. Preclinical and clinical studies suggest beta blockers may have anti-cancer activity via catecholamine blockade and improved immunosurveillance,(53) but nadolol has been rarely examined and observed associations may reflect differences relative to its better studied comparators carvedilol and propranolol.(54,55) Diuretics show limited evidence for digestive cancer risk,(56–59) although spironolactone has been linked to gastrointestinal (GI) bleeding and ulcers,(60–62) which some studies argue may increase digestive cancer risk.(63) However, bias may exist due to the possibility that GI bleeding can enhance surveillance and detection of GI cancers. ACE inhibitors like fosinopril have mostly been evaluated as a class and have shown inconsistent associations with digestive cancer.(64–66) Evidence for its comparator valsartan, an angiotensin II receptor blocker, has also been mixed.(67–71) These data reinforce the importance of examining individual drugs in addition to classes in future studies.

Cardiovascular drugs including digoxin (vs. diltiazem) and antivirals including entecavir (vs. tenofovir) were also positively associated. These findings are consistent with prior epidemiologic studies linking digoxin to colorectal, pancreatic, and esophageal cancers,(72–74) as well as entecavir (vs. tenofovir) to hepatocellular carcinoma in patients with hepatitis B, though residual confounding cannot be ruled out.(75,76)

Immunosuppressants such as azathioprine (vs. methotrexate) and leflunomide (vs. statins) showed positive and inverse associations, respectively. Prior studies link azathioprine to increased risk of lymphoma and other malignancies,(77–79) while preclinical studies suggest anti-tumor effects for drugs like leflunomide.(77,80,81)

Many drugs showed inverse associations, suggesting potential protective benefit. Antipsychotics (aripiprazole, haloperidol, olanzapine, quetiapine, risperidone) and antidepressants (escitalopram, mirtazapine, sertraline, trazodone) were inversely associated with digestive cancers compared with statins. Most of these observations are supported by preclinical and some epidemiologic studies.(82–92) Dopaminergic and cholinergic drugs also showed inverse associations, potentially reflecting anti-tumor activity observed in some preclinical studies,(93–95) although epidemiologic evidence is limited and healthy user bias cannot be ruled out.

Other inverse associations included antimalarials (atovaquone, mefloquine, proguanil), the anticonvulsant divalproex, and estrogens (estradiol, esterified/conjugated estrogens). Preclinical(96–98) and limited epidemiologic(99) data support the potential anti-cancer activity of antimalarial drugs in digestive cancers. There is also limited preclinical evidence of divalproex’s anti-tumor activity,(100–102) but only 2 epidemiologic studies have examined this association with mixed results.(103,104) In contrast, estrogens such as estradiol and conjugated estrogens plus progestin, have substantial preclinical and epidemiologic evidence for a potential protective effect on digestive cancers, particularly colorectal, among postmenopausal women.(105–110)

Overall, these findings highlight that many observed associations span digestive or broader cancer contexts. While some associations are supported by preclinical or epidemiologic evidence, others, such as cardiovascular and psychoactive drugs in particular, may reflect unmeasured confounding or healthy user effects, reinforcing the need for further validation and mechanistic investigation.

We report new epidemiological associations—where there was previously no evidence or only limited preclinical data—with 2 diuretics (furosemide, torsemide), 2 antidepressants (mirtazapine, trazodone), 1 analgesic (sumatriptan), 2 antimalarials (atovaquone, mefloquine), 2 cardiovascular drugs (cilostazol, pentoxifylline), 1 immunosuppressant (leflunomide), 3 anti-Parkinson drugs (carbidopa, levodopa, rasagiline), and 4 anti-dementia drugs (donepezil, galantamine, memantine, rivastigmine). Future studies may be needed to validate these associations.

Strengths and limitations

Studies examining the impact of these drugs on cancer risk are challenging because i) the prevalence of exposure requires large studies, and ii) events occur infrequently, and after long latency, and thus are difficult to capture in clinical trials. Large observational datasets like SEER-Medicare provide detailed prescription, healthcare, and longitudinal follow-up data, enabling the detection of potential long-term drug-cancer associations that complement preclinical and clinical research.

This study used SEER-Medicare data and a rigorous multi-stage approach to scan for novel drug associations with digestive cancers. The dataset provided detailed information on prescriptions and healthcare utilization, confirmed cancer outcomes, and supplied sufficient power to study hundreds of drugs. By integrating a hypothesis-generating framework established by GWAS with pharmacoepidemiologic tools, such as active comparators and propensity score adjustment, we refined associations to those likely to be more robust and meaningful.

There are several study limitations. Data on lifestyle factors (e.g., BMI, physical activity) was incomplete. Although we adjusted for obesity, alcohol use disorders, and tobacco use, these factors are often underreported in claims data, which may lead to residual confounding. Exposure misclassification is possible, as drug dispensing records may not reflect actual consumption, and we could not capture drug use before age 65. However, we expect this misclassification to be similar for cases and controls. To reduce the potential bias caused by differential duration of drug use, we conducted a long-term use analysis with a 12-month washout period for chronic medications to better capture initiation. However, we were unable to assess lifelong cumulative dose without prescription information before age 65. We were therefore unable to capture prescription drug use in earlier life periods prior to Medicare eligibility. These drugs may be especially relevant for digestive cancers, which often have a long latency period and may develop over decades. Thus, future research is critical to assess cumulative dose across the life span in relation to digestive cancer risk. The composite digestive cancer outcome, also used by other studies,(111,112) may dilute drug associations specific to an individual cancer type, although our broad approach maximized statistical power for hypothesis generation; future studies should explore specific cancers. Active comparators were implemented to control for confounding by indication and other sources of bias, including healthy user bias. However, we were unable to separate out how much of the association between the active comparator and digestive cancer is due to bias (e.g., confounding by indication or healthy user bias) versus a true biological effect. Potential underestimation of a drug-cancer association in active comparator analyses may exist if an active comparator has direct effects on cancer or the degree of confounding is different between a drug of interest and its comparator. Finally, we did not disaggregate results by sex (except sex hormones) to preserve power to uncover novel signals, but follow-up studies can explore sex-specific associations.

Conclusions

Using a novel, multi-stage, hypothesis-generating approach with rigorous refinement in SEER-Medicare data, we identified 38 oral prescription drugs associated with digestive system cancers: 10 positively and 28 inversely associated. Among them, 6 (azathioprine, furosemide, spironolactone, donepezil, estradiol, and sertraline) are consistent with prior screening studies, with the remaining representing novel associations that merit additional follow-up soon to clarify potential risk or identify potential chemopreventive strategies. These findings reinforce prior data and generate novel hypotheses, highlighting drugs that warrant further investigation using robust pharmacoepidemiologic designs that incorporate more complete data on lifestyle factors and earlier exposures. This study demonstrates the ability of a large-scale, multi-stage DWAS approach to efficiently prioritize drug-cancer signals for future research, which has important implications for chemoprevention and risk reduction and application across fields. Future research is needed to further validate our findings and evaluate specific digestive cancer types.