Key Points
Question
Among healthy older people, does the effect of low-dose aspirin (LDA) on cancer prevention vary based on individual characteristics?
Findings
In this comparative effectiveness study as a secondary analysis of a randomized clinical trial involving 9350 healthy adults 70 years or older, results calculated using an effect score model suggested heterogeneity in the individualized treatment effects (ITEs) of LDA (100 mg per day) on cancer prevention. Nonsmoking status and clonal hematopoiesis of indeterminate potential at a variant allele frequency of 10% or greater were associated with LDA benefit.
Meaning
The findings of this analysis suggest that the ITE of LDA on cancer prevention in older adults varies by participant characteristics.
This secondary analysis of the Aspirin in Reducing Events in the Elderly (ASPREE) randomized clinical trial explores whether individual characteristics are associated with estimated treatment effects of low-dose aspirin on 5-year cancer incidence in healthy older adults.
Abstract
Importance
The role of low-dose aspirin (LDA) in cancer prevention among older adults is unclear. Identifying individuals likely to experience benefit or harm is crucial for guiding personalized prevention strategies.
Objective
To identify subgroups of the older population who may benefit most from LDA for cancer prevention by developing and validating an effect score model, including individual characteristics such as clonal hematopoiesis of indeterminate potential (CHIP) and other factors, to predict the individualized treatment effects (ITEs) of LDA in cancer incidence.
Design, Setting, and Participants
This comparative effectiveness study is a nonprespecified secondary analysis of the Aspirin in Reducing Events in the Elderly (ASPREE) randomized clinical trial conducted between 2010 and 2014. Australian community-dwelling ASPREE participants (aged ≥70 years) provided biospecimen samples for CHIP testing. CHIP was measured in 18 genes using a targeted sequencing assay. Analyses were conducted between May 2023 and July 2025.
Exposures
LDA (100 mg per day) or placebo.
Main Outcomes and Measures
The primary outcome was overall cancer incidence at 5 years. Twelve candidate models for predicting ITEs were assessed; the final model was selected to maximize cancer risk reduction and internally validated using bootstrapping. Participants were grouped into treatment-favorable (ITE >0) and treatment-unfavorable (ITE <0) subgroups based on the final predictive model.
Results
A total of 9350 participants were included in the analysis (median age, 73.7 [IQR, 71.6-77.1] years; 5021 females [53.7%]) to LDA (4667 [49.9%]) or placebo (4683 [50.1%]). The median follow-up was 4.5 (IQR, 3.3-5.5) years. Factors associated with LDA benefit included older age, nonsmoking status, CHIP with a variant allele frequency of 10% or greater, family history of cancer, and lower body mass index. CHIP was the strongest predictor of benefit. Personalized treatment based on the model improved 5-year absolute cancer risk reduction by a median of 2.3% (IQR, 0.7%-3.7%) compared with a treat-all approach (LDA to all participants). LDA was associated with lower cancer risk (hazard ratio [HR], 0.85 [95% CI, 0.72-1.00]) in the treatment-favorable subgroup and higher cancer risk (HR, 1.14 [95% CI, 0.95-1.38]) in the treatment-unfavorable subgroup (P = .02 for heterogeneity).
Conclusions and Relevance
The findings of this study suggest significant heterogeneity in the effectiveness of LDA for cancer prevention among older adults, with CHIP identified as a key predictive factor. Further study is needed to fully understand the implications of these findings.
Trial Registration
ClinicalTrials.gov Identifier: NCT01038583
Introduction
Aspirin has emerged as a promising agent for the prevention of colorectal cancer,1,2 with meta-analyses reporting a 20% to 30% reduction in cancer incidence among regular users.3,4 However, the 2022 US Preventive Services Task Force concluded that evidence is insufficient to recommend low-dose aspirin (LDA) for cancer prevention, particularly in adults aged 60 years or older.5
The Aspirin in Reducing Events in the Elderly (ASPREE) study, a large randomized clinical trial (RCT) involving 19 114 relatively healthy adults aged 70 years or older (or ≥65 years for US Black and Latino individuals), found no overall difference in cancer incidence between participants randomized to LDA (100 mg per day) or placebo over a median follow-up of 4.6 years.6 This finding was consistent across various prespecified and post hoc subgroups. However, the limitations of one-by-one subgroup analyses, including low statistical power and increased risk of false-positive or false-negative results, restrict their interpretation.7 Effect score analysis has been proposed as a method to overcome these challenges.7,8 By considering multiple participant characteristics simultaneously, this approach estimates the individualized treatment effect (ITE), defined as the difference in potential outcomes under LDA and placebo for each participant. Hence, it offers a promising approach to identify patient subgroups who may benefit differentially from LDA for cancer prevention and thus have altered overall benefit-risk profiles to consider for treatment personalization.
Clonal hematopoiesis of indeterminate potential (CHIP) is an age-related phenomenon that has been shown to be associated with increased risk of both blood cancer and nonblood cancers9 as well as cardiovascular and other diseases. In a substudy of the ASPREE trial (ASPREE-CHIP), Z. McQuilten, MBBS, PhD, and colleagues (unpublished data, August 19, 2025) observed an association between the presence of CHIP at a variant allele frequency (VAF) of 10% or greater and increased risk of incident cancer.
In this study, we aimed to identify subgroups of the older population who may benefit most from LDA for cancer prevention by developing and validating an effect score model to predict the personalized treatment effect of LDA on cancer incidence. Given the emerging evidence of CHIP as a risk factor for incident cancer, including in the ASPREE cohort, we incorporated CHIP as a potential treatment modifier, along with other participant characteristics.
Methods
This nonprespecified secondary analysis of the ASPREE trial and the ASPREE-CHIP substudy was conducted under an approval from the Monash University Human Research Ethics Committee for secondary analysis of the ASPREE RCT. Participants provided informed consent to participate in the ASPREE RCT and, separately, to provide samples for genetic analysis. The trial protocol is available in Supplement 1. The participant flow diagram is presented in eFigure 1 of Supplement 2. This analysis followed the Predictive Approaches to Treatment Effect Heterogeneity (PATH) statement.
Study Population
ASPREE Trial
The ASPREE RCT investigated the effect of LDA (100 mg per day, enteric coated) on disability-free survival in otherwise healthy older people.10 The study recruited 19 114 participants between 2010 and 2014 from multiple centers across Australia and the US. Analyses were conducted between May 2023 and July 2025. The primary end point of the trial was a composite of death, persistent physical disability, and dementia. Participants were included in the trial if they were free of cardiovascular disease events, dementia, and independence-limiting physical disability and were expected to live for at least 5 years. Participants with a history of cancer diagnosis were not excluded from the study. Further details of the ASPREE trial, and the finding of no effect of aspirin on the primary end point, have been published elsewhere.10
ASPREE-CHIP Substudy of the ASPREE Trial
ASPREE-CHIP is a substudy of the ASPREE trial. In Australia, individuals who consented to participate in the ASPREE Biobank11 provided peripheral blood samples prior to randomization (approximately 65%) or within the first year after randomization. The ASPREE-CHIP substudy to measure CHIP in ASPREE participant blood samples was approved by the Alfred Health Human Research Ethics Committee. All participants gave separate written informed consent for genetic analysis.12
CHIP Testing
CHIP was measured in 18 CHIP-associated genes (eTable 1 in Supplement 2) using a custom targeted amplicon-based next-generation sequencing assay,13 designed to have a sensitivity of 1% variant allele fraction (VAF). Details of CHIP testing are provided in the eMethods in Supplement 2. Briefly, a variant was called as CHIP if it was a recognized canonical CHIP variant and met specific criteria (total read depth of ≥100, variant depth of ≥10 [or ≥20 for single nucleotide variants in DNMT3A, TET2, or TP53], VarDict quality score of ≥30, and VAF of 2%-45%). In this study, we defined CHIP as a VAF of 10% or greater.
Outcome
The primary end point was time to first incidence of cancer, defined as the first new cancer diagnosis reported and confirmed after randomization during the trial. In participants with a history of cancer prior to randomization, a new primary tumor diagnosis of the same or a different cancer type was counted as an event, whereas a local recurrence of the same type was not. Cancer incidence was one of the prespecified secondary end points of the ASPREE trial, and it was adjudicated by independent expert committees blinded to trial group assignments. Details of outcome adjudication and the main effects of LDA on cancer types are described elsewhere.6,10
Statistical Analysis
We selected candidate predictors for the LDA effect score model based on prespecified variables from the ASPREE trial,14 which included the following factors associated with cancer risk: age, sex, country, ethnicity, body mass index (BMI), smoking, alcohol drinking, anemia, polypharmacy, previous regular use of aspirin, personal history of cancer, family history of cancer, and previous cancer screening. For this study, we excluded country and ethnicity because all ASPREE-CHIP participants were Australian and predominantly White, and we omitted cancer screening due to substantial missing data. Anemia was replaced with baseline hemoglobin level to improve power. We also added CHIP (VAF ≥10%) as a binary predictor, based on prior evidence of treatment effect modification. All predictors were considered both prognostic factors (predictors of risk of incident cancer) and effect modifiers (predictors of the LDA effect).
We fitted multivariable regression models including all 2-way interactions between treatment and the selected predictors. The predicted ITE for each participant was the difference in the absolute risk of incidence cancer under 2 scenarios: (1) if everyone used placebo and (2) if everyone used LDA. This method of estimating ITE relates to the g-formula method15 and is referred to as the effect modeling approach in the PATH statement.16 A positive ITE value indicated a benefit from LDA, whereas a negative value indicated no benefit (and possible harm).
We explored several model specifications, differing in how continuous variables (age, BMI, and hemoglobin level) were handled, whether the competing risk of death was accounted for, and whether regularization was applied to mitigate overfitting. Specifically, continuous variables were modeled in 3 ways: (1) as continuous with a linear function, (2) as continuous with a nonlinear function (eg, restricted cubic spline function with 3 knots), or (3) as a categorical variable (only for BMI [calculated as weight in kilograms divided by height in meters squared], using the World Health Organization classification17: underweight, <18.5; normal, 18.5-24.9; overweight, 25-29.9; or obese, ≥30). We fitted Cox proportional hazards models both with and without accounting for the competing risk of death. In addition, we fitted models both with and without ridge penalization for regularizing regression coefficients to overcome potential overfitting.16 This led to 12 candidate models, which are specified in eTables 2, 3, and 4 in Supplement 2.
We selected the model with the highest decision accuracy from the 12 candidate models. Specifically, for each model, we defined a treatment decision rule to give LDA to individuals with a positive ITE (treatment favorable) and placebo otherwise (treatment unfavorable). Decision accuracy was defined as the improvement in incident cancer reduction if everyone followed this rule compared with a treat-all policy to give LDA to everyone. We used bootstrap validation with 1000 bootstrap samples to select the model that had the highest performance. Further details of the model selection process are presented in the eMethods in Supplement 2.
We assessed the final model’s discrimination by estimating absolute cancer risk using the Aalen-Johansen estimator, stratified by randomized treatment and the model-recommended treatment. Separation between the resulting risk curves indicated the model’s ability to distinguish those who benefited from LDA from those who did not.
To evaluate the calibration of the model, the absolute risk difference in cancer incidence at 5 years was compared with the average predicted ITEs within 5 groups defined based on quintiles of predicted ITEs.18 We chose 5 groups to warrant a reasonable number of events within each group for a valid estimation of the observed treatment effects.19
We constructed a variable importance plot to quantify the relative impact of each predictor variable on the predicted ITEs. Variable importance was determined by measuring how much the predicted ITEs for the entire cohort changed, on average, when we altered one specific variable at a time. For continuous variables, we replaced their values with the median values.20
We performed 2 sensitivity analyses: one excluding participants with a history of cancer, and another using models that did not include CHIP as a predictor. We also used the area under the targeting operator characteristic (AUTOC)21 as another decision accuracy metric to select the final model in the bootstrap validation. The AUTOC metric evaluates how well a treatment decision rule identifies individuals who benefit most from treatment.
Because the number of missing data values was low (<1%), all analyses were conducted on complete cases. Two-sided P < .05 was considered statistically significant. All analyses were performed using R, version 4.4.0 (R Project for Statistical Computing).
Results
Of the 9434 ASPREE-CHIP participants with baseline CHIP measurements available, 84 were excluded due to missing information. Therefore, this analysis included 9350 participants randomized to LDA (4667 [49.9%]) or placebo (4683 [50.1%]). The baseline characteristics of the cohort are presented in Table 1. The median age of all participants was 73.7 years (IQR, 71.6-77.1 years); 5021 participants (53.7%) were female and 4329 (46.3%) were male. A total of 905 participants (9.7%) had diabetes at baseline, and 529 (5.7%) had CHIP at a VAF of 10% or greater. The distribution of baseline characteristics was similar between randomized treatment groups. Participants were followed up for a median of 4.5 years (IQR, 3.3-5.5 years).
Table 1. Participant Baseline Characteristics, Overall and by Randomization Group.
| Characteristic | Participants, No. (%) | ||
|---|---|---|---|
| All participants (N = 9350) | Placebo group (n = 4683) | Aspirin group (n = 4667) | |
| Age, median (IQR), y | 73.7 (71.6-77.1) | 73.7 (71.6-77.1) | 73.7 (71.6-77.1) |
| Sex | |||
| Male | 4329 (46.3) | 2155 (46.0) | 2174 (46.6) |
| Female | 5021 (53.7) | 2528 (54.0) | 2493 (53.4) |
| BMI, median (IQR) | 27.4 (24.9-30.5) | 27.5 (25.0-30.4) | 27.4 (24.8-30.5) |
| Underweight (<18.5) | 161 (1.7) | 79 (1.7) | 82 (1.8) |
| Normal (18.5-24.9) | 2234 (23.9) | 1085 (23.2) | 1149 (24.6) |
| Overweight (25-29.9) | 4309 (46.1) | 2186 (46.7) | 2123 (45.5) |
| Obese (≥30) | 2646 (28.3) | 1333 (28.5) | 1313 (28.1) |
| Smoking status | |||
| Current | 296 (3.2) | 152 (3.2) | 144 (3.1) |
| Former | 3843 (41.1) | 1904 (40.7) | 1939 (41.5) |
| Never | 5211 (55.7) | 2627 (56.1) | 2584 (55.4) |
| Alcohol drinking | |||
| Current | 7468 (79.9) | 3722 (79.5) | 3746 (80.3) |
| Former | 439 (4.7) | 218 (4.7) | 221 (4.7) |
| Never | 1443 (15.4) | 743 (15.9) | 700 (15.0) |
| Diabetes | 905 (9.7) | 454 (9.7) | 451 (9.7) |
| Polypharmacy | 2382 (25.5) | 1170 (25.0) | 1212 (26.0) |
| Hemoglobin level, median (IQR), g/dL | 14.2 (13.4-15.1) | 14.2 (13.4-15.0) | 14.2 (13.4-15.1) |
| Previous regular use of aspirin | 695 (7.4) | 347 (7.4) | 348 (7.5) |
| Personal history of cancer | 1920 (20.5) | 960 (20.5) | 960 (20.6) |
| Family history of cancer | 5682 (60.8) | 2842 (60.7) | 2840 (60.9) |
| CHIP VAF ≥10% | 529 (5.7) | 270 (5.8) | 259 (5.5) |
| Outcome | |||
| Censored | 8146 (87.1) | 4070 (86.9) | 4076 (87.3) |
| Cancer | 1009 (10.8) | 513 (11.0) | 496 (10.6) |
| Death | 195 (2.1) | 100 (2.1) | 95 (2.0) |
Abbreviations: BMI, body mass index (calculated as weight in kilograms divided by height in meters squared); CHIP, clonal hematopoiesis of indeterminate potential; VAF, variant allele fraction.
SI conversion factor: To convert hemoglobin to g/L, multiply by 10.0.
During the follow-up, 1009 of 9350 participants (10.8%) developed incident cancer, and 195 (2.1%) died. For participants who provided their biospecimen sample during the first year after randomization, more did so after a first cancer diagnosis than before (208 vs 182). Estimated cumulative incidence risks of cancer and death are illustrated in eFigure 2 in Supplement 2. A summary of cancer types is presented in eTable 5 in Supplement 2. The most frequent cancer type was prostate cancer (200 [19.8%]), followed by colon or rectal cancer (144 [14.3%]), breast cancer (121 [12.0%]), melanoma (112 [11.1%]), and blood cancer (111 [11.0%]).
We evaluated the performance of the 12 candidate models in bootstrap internal validation to select the final model (eFigure 3 in Supplement 2). The Cox proportional hazards models with and without accounting for competing risks of death showed comparable performance, likely due to the low mortality rate in the study population (eFigure 2 in Supplement 2). In both cases, models incorporating restricted cubic spline functions for all continuous variables gave the best performance. For simplicity, we chose the Cox proportional hazards model without competing risk as the final model. Compared with the treat-all policy, following this model for treatment decision-making resulted in a median improvement of 2.3% (IQR, 0.7%-3.7%) in 5-year absolute cancer risk reduction. With AUTOC as the model performance metric, the same final model was selected (eFigure 4 in Supplement 2). The AUTOC of the final model was 0.014 (95% CI, 0.014-0.015), indicating that it significantly outperformed chance or random treatment allocation in stratifying individuals based on estimated treatment benefit.
Regression coefficients of the final model are presented in eTable 6 in Supplement 2. The corresponding distribution of the predicted ITE is shown in eFigure 5 in Supplement 2. Across the cohort, the predicted ITE of LDA in cancer risk reduction ranged from −23.0% to 21.5%. LDA was predicted to be beneficial for 59.1% of the study population (5529 of 9350), aligning with the null average treatment effect of LDA observed in the overall trial population (hazard ratio [HR], 0.96 [95% CI, 0.85-1.09]). However, the treatment effect varied significantly across benefit groups (P = .02 for heterogeneity) (Figure 1). Specifically, LDA was associated with a reduction in cancer risk in the treatment-favorable subgroup (HR, 0.85 [95% CI, 0.71-1.00]), whereas LDA was associated with an increase in cancer rates in the treatment-unfavorable subgroup (HR, 1.14 [95% CI, 0.95-1.38]).
Figure 1. Forest Plot of the Estimated Treatment Effect in Defined Subgroups.
The diamond indicates the overall estimate with the outer points indicating the 95% CI. HR indicates hazard ratio.
The discrimination performance of the final model is shown in Figure 2 and eFigure 6 in Supplement 2. Among individuals who were randomized to LDA, those with predicted benefit (predicted ITEs >0) exhibited a lower risk of cancer at 5 years (estimated risk, 0.10 [95% CI, 0.09-0.11]) compared with those with predicted harm (predicted ITEs <0) (estimated risk, 0.13 [95% CI, 0.12-0.15]). This difference in risk supports the model’s ability to discriminate between individuals with different treatment effects.
Figure 2. Aalen-Johansen Estimates of the Cumulative Incidence Function for Cancer Stratified by Placebo and Low-Dose Aspirin Groups.
Treatment favorable and treatment unfavorable indicate a predicted individualized treatment effect greater than 0 and less than 0, respectively. Shaded areas are 95% CIs.
A calibration plot of the final model is illustrated in eFigure 7 in Supplement 2. The predicted and observed treatment effects aligned well in the middle quantile groups. However, in the smallest and largest quantile groups, the predicted treatment effects slightly underestimated their observed counterparts.
A summary of participant characteristics at baseline by benefit group is presented in Table 2. Participants in the treatment-favorable subgroup were, on average, older, and they had higher median hemoglobin levels and higher prevalences of family history of cancer and CHIP. Participants in the treatment-unfavorable subgroup had a higher median BMI and higher prevalences of current smoking, diabetes, polypharmacy, and personal history of cancer. The impact of each variable on the average predicted ITEs is shown in Figure 3. Current smoking was the most important predictor for a detrimental effect of aspirin on cancer incidence, whereas CHIP was the most important predictor for a beneficial effect of aspirin, followed by age and a family history of cancer.
Table 2. Baseline Characteristics by Defined Subgroup.
| Characteristic | Subgroup participants, No. (%) | |
|---|---|---|
| Treatment-favorable subgroup (n = 5529)a | Treatment-unfavorable subgroup (n = 3821) | |
| Age, median (IQR), y | 74.1 (72.1-76.6) | 72.7 (71.2-78.6) |
| Sex | ||
| Male | 2534 (45.8) | 1795 (47.0) |
| Female | 2995 (54.2) | 2026 (53.0) |
| BMI, median (IQR) | 26.2 (23.9-29.7) | 28.7 (26.7-31.0) |
| Underweight (<18.5) | 157 (2.8) | 4 (0.1) |
| Normal (18.5-24.9) | 1885 (34.1) | 349 (9.1) |
| Overweight (25-29.9) | 2168 (39.2) | 2141 (56.0) |
| Obese (≥30) | 1319 (23.9) | 1327 (34.7) |
| Smoking status | ||
| Current | 51 (0.9) | 245 (6.4) |
| Former | 2366 (42.8) | 1477 (38.7) |
| Never | 3112 (56.3) | 2099 (54.9) |
| Alcohol drinking | ||
| Current | 4524 (81.8) | 2944 (77.0) |
| Former | 155 (2.8) | 284 (7.4) |
| Never | 850 (15.4) | 593 (15.5) |
| Diabetes | 372 (6.7) | 533 (13.9) |
| Polypharmacy | 826 (14.9) | 1556 (40.7) |
| Hemoglobin level, median (IQR) g/dL | 14.3 (13.6-15.1) | 13.9 (13.1-14.9) |
| Previous regular use of aspirin | 452 (8.2) | 243 (6.4) |
| Personal history of cancer | 956 (17.3) | 964 (25.2) |
| Family history of cancer | 4139 (74.9) | 1543 (40.4) |
| Treatment | ||
| Placebo | 2737 (49.5) | 1946 (50.9) |
| Aspirin | 2792 (50.5) | 1875 (49.1) |
| CHIP VAF ≥10% | 488 (8.8) | 41 (1.1) |
Abbreviations: BMI, body mass index (calculated as weight in kilograms divided by height in meters squared); CHIP, clonal hematopoiesis of indeterminate potential; VAF, variant allele fraction.
SI conversion factor: To convert hemoglobin to g/L, multiply by 10.0.
Treatment favorable and treatment unfavorable indicate a predicted individualized treatment effect greater than 0 and less than 0, respectively.
Figure 3. Variable Importance: Impact of Each Predictor on the Average Predicted Individualized Treatment Effects.
BMI indicates body mass index (calculated as weight in kilograms divided by height in meters squared); CHIP, clonal hematopoiesis of indeterminate potential; VAF, variant allele fraction.
SI conversion factor: To convert hemoglobin to g/L, multiply by 10.0.
In both sensitivity analyses, the same model form as in the main analysis was identified as the best model. The analysis excluding participants with a history of cancer resulted in a relative improvement in 5-year cancer risk reduction (median, 2.2%; IQR, 0.5%-3.8%) (eFigure 8 in Supplement 2) that was similar to the model on the full cohort. In contrast, the model without CHIP showed a lower overall population benefit (median, 1.8%; IQR, 0.4%-3.2%) (eFigure 9 in Supplement 2), highlighting the added value of CHIP for improving the model’s predictive performance.
Discussion
In this study, we developed a prediction model for the ITE of LDA on 5-year risk of incident cancer in an initially healthy older population. The findings of our analysis suggest the presence of heterogeneity in the treatment effect of LDA on cancer within this population. We identified a subgroup of participants more likely to benefit from aspirin in reducing cancer risk. These participants tended to be older or to have nonsmoking status, CHIP at a VAF of 10% or greater, family history of cancer, or lower BMI.
To our knowledge, this study is the first to identify CHIP as an effect modifier, particularly in combination with other factors. This observation suggests that CHIP may play a pivotal role in modulating aspirin’s efficacy, potentially through its involvement in inflammatory pathways. Specifically, evidence from a 2024 study linked CHIP to an increased risk of both hematologic and nonhematologic cancers.9 The mechanism for this association is unknown, but it is potentially due to the association between CHIP and elevated proinflammatory cytokines such as interleukin (IL)-6, IL-8, and tumor necrosis factor-α.22 Given aspirin’s well-established anti-inflammatory properties, which involve inhibiting cyclooxygenases (COX-1 and COX-2) and subsequently reducing prostaglandin-induced inflammation,23 our results suggest a plausible biological mechanism for the observed interaction between CHIP and aspirin. Although further research is needed to elucidate these complex pathways, our findings highlight CHIP as a promising biomarker for personalizing medicine in cancer prevention.
Consistent with these findings, aspirin was associated with greater absolute reduction of colorectal cancer risk among individuals with a greater number of lifestyle risk factors, including smoking, BMI, physical activity, diet quality, and alcohol drinking, in a previous study in 2 prospective cohorts.24 Our results also align with previous research demonstrating a potential association between family history of cancer and increased aspirin efficacy, as observed in the Women’s Health Initiative.25 However, other studies have suggested a more limited benefit of aspirin among those with a higher BMI26,27 and heavy smoking, 27,28 which may be due to differences in study populations, aspirin dosage, duration of use, or cancer types investigated.
It is crucial to emphasize that identifying subgroups more likely to benefit from LDA for cancer prevention should be considered as one component of a comprehensive benefit-risk assessment that also incorporates careful consideration of potential benefits and risks on other disease processes. In particular, the increased risk of bleeding associated with LDA use, as well as the potential benefit of LDA in cardiovascular disease prevention (eg, which might be seen in individuals with an elevated lipoprotein a genotype), should be considered.29 Although our model offers potentially valuable insights into cancer prevention over a 5-year period, it is essential to also consider longer-term outcomes.
Strengths and Limitations
A key strength of our study is the use of trial data with a relatively large sample size, which enhances the reliability of our ITE prediction in the absence of confounding. Furthermore, rigorous outcome adjudication strengthens the credibility of our analysis.
Our study has several limitations. The study population consisted of relatively healthy older White adults in Australia and investigated aspirin initiation after age 70 years, limiting the generalizability of our findings. Previous studies suggest that initiation of aspirin use at earlier ages may have differential benefits on long-term cancer risk compared with at older ages.30 Because biospecimens were collected before and during the first year after randomization, early cancer diagnosis could affect sample provision and introduce selection bias. However, this risk is likely minimal, given more participants provided samples after a diagnosis (208 vs 182) and the overall sample size was large (9350). The utility of our model in decision-making is yet to be tested on an independent population. To support its validation, we provide an equation in eTable 6 in Supplement 2 to calculate the predicted ITE for each individual. Further studies are also needed to demonstrate the usefulness of our model, such as an RCT comparing giving LDA with and without the model’s guidance.
Conclusions
This secondary analysis of an RCT provided evidence for the heterogeneity of the effect of LDA on cancer incidence in older adults. By incorporating CHIP and other clinical factors into a prediction model, we identified a subgroup of people who may derive greater benefit from aspirin therapy. These findings highlight the importance of considering heterogenous treatment effects in the design and analysis of RCTs, promote the application of personalized inference in clinical decision-making, and offer valuable insights for future research. Further investigation is warranted to fully understand the implications of our results.
Trial Protocol
CONSORT Flow Diagram
eMethods
eReferences
eTable 1. Frequency of the 18 Clonal Hematopoiesis of Indeterminate Potential (CHIP) Associated Genes
eTable 2. Parameterization of Candidate Effect Score Models
eTable 3. Statistical Models
eTable 4. Summary of the 12 Considered Models
eTable 5. Summary of Cancer Types
eTable 6. Regression Coefficients and Baseline Cumulative Hazard Function at 5 Years of the Final Model Predicting the Risk of Cancer Under 2 Scenarios: If Placebo Is Given and If Low-Dose Aspirin (LDA) Is Given
eFigure 2. Cumulative Incidence Function for Cancer
eFigure 3. Model Performance in Bootstrap Validation
eFigure 4. Model Performance Using AUTOC as the Performance Metric in Bootstrap Validation
eFigure 5. Histogram of the Predicted Individualized Treatment Effects
eFigure 6. Aalen-Johansen Estimates of the Cumulative Incidence Function for Cancer, Stratified by Defined Subgroups (Treatment-Favorable/Treatment-Unfavorable) and Randomization Groups (Aspirin/Placebo)
eFigure 7. Calibration Plot
eFigure 8. Sensitivity Analysis Excluding Participants With a History of Cancer: Model Performance in Bootstrap Validation
eFigure 9. Sensitivity Analysis Excluding Clonal Hematopoiesis of Indeterminate Potential (CHIP) as a Predictor: Model Performance in Bootstrap Validation
Data Sharing Statement
References
- 1.Drew DA, Chan AT. Aspirin in the prevention of colorectal neoplasia. Annu Rev Med. 2021;72:415-430. doi: 10.1146/annurev-med-060319-120913 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 2.Katona BW, Weiss JM. Chemoprevention of colorectal cancer. Gastroenterology. 2020;158(2):368-388. doi: 10.1053/j.gastro.2019.06.047 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 3.Rothwell PM, Price JF, Fowkes FGR, et al. Short-term effects of daily aspirin on cancer incidence, mortality, and non-vascular death: analysis of the time course of risks and benefits in 51 randomised controlled trials. Lancet. 2012;379(9826):1602-1612. doi: 10.1016/S0140-6736(11)61720-0 [DOI] [PubMed] [Google Scholar]
- 4.Bosetti C, Rosato V, Gallus S, Cuzick J, La Vecchia C. Aspirin and cancer risk: a quantitative review to 2011. Ann Oncol. 2012;23(6):1403-1415. doi: 10.1093/annonc/mds113 [DOI] [PubMed] [Google Scholar]
- 5.Dehmer SP, O’Keefe LR, Evans CV, Guirguis-Blake JM, Perdue LA, Maciosek MV. Aspirin use to prevent cardiovascular disease and colorectal cancer: updated modeling study for the US Preventive Services Task Force. JAMA. 2022;327(16):1598-1607. doi: 10.1001/jama.2022.3385 [DOI] [PubMed] [Google Scholar]
- 6.McNeil JJ, Gibbs P, Orchard SG, et al. ; ASPREE Investigator Group . Effect of aspirin on cancer incidence and mortality in older adults. J Natl Cancer Inst. 2021;113(3):258-265. doi: 10.1093/jnci/djaa114 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 7.Dahabreh IJ, Kazi DS. Toward personalizing care: assessing heterogeneity of treatment effects in randomized trials. JAMA. 2023;329(13):1063-1065. doi: 10.1001/jama.2023.3576 [DOI] [PubMed] [Google Scholar]
- 8.Wang G, Heagerty PJ, Dahabreh IJ. Using effect scores to characterize heterogeneity of treatment effects. JAMA. 2024;331(14):1225-1226. doi: 10.1001/jama.2024.3376 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 9.Singh J, Li N, Ashrafi E, et al. Clonal hematopoiesis of indeterminate potential as a prognostic factor: a systematic review and meta-analysis. Blood Adv. 2024;8(14):3771-3784. doi: 10.1182/bloodadvances.2024013228 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 10.McNeil JJ, Woods RL, Nelson MR, et al. ; ASPREE Investigator Group . Effect of aspirin on disability-free survival in the healthy elderly. N Engl J Med. 2018;379(16):1499-1508. doi: 10.1056/NEJMoa1800722 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 11.Parker EJ, Orchard SG, Gilbert TJ, et al. The ASPREE Healthy Ageing Biobank: methodology and participant characteristics. PLoS One. 2024;19(2):e0294743. doi: 10.1371/journal.pone.0294743 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 12.ASPREE Investigator Group . Study design of Aspirin in Reducing Events in the Elderly (ASPREE): a randomized, controlled trial. Contemp Clin Trials. 2013;36(2):555-564. doi: 10.1016/j.cct.2013.09.014 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 13.Kessler MD, Damask A, O’Keeffe S, et al. ; Regeneron Genetics Center; GHS-RGC DiscovEHR Collaboration . Common and rare variant associations with clonal haematopoiesis phenotypes. Nature. 2022;612(7939):301-309. doi: 10.1038/s41586-022-05448-9 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 14.Wolfe R, Murray AM, Woods RL, et al. The Aspirin in Reducing Events in the Elderly trial: statistical analysis plan. Int J Stroke. 2018;13(3):335-338. doi: 10.1177/1747493017741383 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 15.Nguyen TL, Collins GS, Landais P, Le Manach Y. Counterfactual clinical prediction models could help to infer individualized treatment effects in randomized controlled trials: an illustration with the International Stroke Trial. J Clin Epidemiol. 2020;125:47-56. doi: 10.1016/j.jclinepi.2020.05.022 [DOI] [PubMed] [Google Scholar]
- 16.Kent DM, Paulus JK, van Klaveren D, et al. The Predictive Approaches to Treatment effect Heterogeneity (PATH) Statement. Ann Intern Med. 2020;172(1):35-45. doi: 10.7326/M18-3667 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 17.Body mass index among adults. World Health Organization. Accessed August 18, 2025. https://www.who.int/data/gho/data/themes/topics/indicator-groups/indicator-group-details/GHO/bmi-among-adults
- 18.Efthimiou O, Hoogland J, Debray TPA, et al. Measuring the performance of prediction models to personalize treatment choice. Stat Med. 2023;42(8):1188-1206. doi: 10.1002/sim.9665 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 19.VanderWeele TJ, Luedtke AR, van der Laan MJ, Kessler RC. Selecting optimal subgroups for treatment using many covariates. Epidemiology. 2019;30(3):334-341. doi: 10.1097/EDE.0000000000000991 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 20.Buell KG, Spicer AB, Casey JD, et al. Individualized treatment effects of oxygen targets in mechanically ventilated critically ill adults. JAMA. 2024;331(14):1195-1204. doi: 10.1001/jama.2024.2933 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 21.Yadlowsky S, Fleming S, Shah N, Brunskill E, Wager S. Evaluating treatment prioritization rules via rank-weighted average treatment effects. J Am Stat Assoc. 2025;120(549):38-51. doi: 10.1080/01621459.2024.2393466 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 22.Nathan DI, Dougherty M, Bhatta M, Mascarenhas J, Marcellino BK. Clonal hematopoiesis and inflammation: a review of mechanisms and clinical implications. Crit Rev Oncol Hematol. 2023;192:104187. doi: 10.1016/j.critrevonc.2023.104187 [DOI] [PubMed] [Google Scholar]
- 23.Chung HY, Sung B, Jung KJ, Zou Y, Yu BP. The molecular inflammatory process in aging. Antioxid Redox Signal. 2006;8(3-4):572-581. doi: 10.1089/ars.2006.8.572 [DOI] [PubMed] [Google Scholar]
- 24.Sikavi DR, Wang K, Ma W, et al. Aspirin use and incidence of colorectal cancer according to lifestyle risk. JAMA Oncol. 2024;10(10):1354-1361. doi: 10.1001/jamaoncol.2024.2503 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 25.Seaton ME, Peters U, Johnson KC, Kooperberg C, Bafford A, Zubair N. Effects of colorectal cancer risk factors on the association between aspirin and colorectal cancer. Anticancer Res. 2019;39(9):4877-4884. doi: 10.21873/anticanres.13673 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 26.Rothwell PM, Cook NR, Gaziano JM, et al. Effects of aspirin on risks of vascular events and cancer according to bodyweight and dose: analysis of individual patient data from randomised trials. Lancet. 2018;392(10145):387-399. doi: 10.1016/S0140-6736(18)31133-4 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 27.Wang X, Chan AT, Slattery ML, et al. Influence of smoking, body mass index, and other factors on the preventive effect of nonsteroidal anti-inflammatory drugs on colorectal cancer risk. Cancer Res. 2018;78(16):4790-4799. doi: 10.1158/0008-5472.CAN-18-0326 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 28.Ishikawa H, Mutoh M, Suzuki S, et al. The preventive effects of low-dose enteric-coated aspirin tablets on the development of colorectal tumours in Asian patients: a randomised trial. Gut. 2014;63(11):1755-1759. doi: 10.1136/gutjnl-2013-305827 [DOI] [PubMed] [Google Scholar]
- 29.Lacaze P, Bakshi A, Riaz M, et al. Aspirin for primary prevention of cardiovascular events in relation to lipoprotein(a) genotypes. J Am Coll Cardiol. 2022;80(14):1287-1298. doi: 10.1016/j.jacc.2022.07.027 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 30.Guo CG, Ma W, Drew DA, et al. Aspirin use and risk of colorectal cancer among older adults. JAMA Oncol. 2021;7(3):428-435. doi: 10.1001/jamaoncol.2020.7338 [DOI] [PMC free article] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Supplementary Materials
Trial Protocol
CONSORT Flow Diagram
eMethods
eReferences
eTable 1. Frequency of the 18 Clonal Hematopoiesis of Indeterminate Potential (CHIP) Associated Genes
eTable 2. Parameterization of Candidate Effect Score Models
eTable 3. Statistical Models
eTable 4. Summary of the 12 Considered Models
eTable 5. Summary of Cancer Types
eTable 6. Regression Coefficients and Baseline Cumulative Hazard Function at 5 Years of the Final Model Predicting the Risk of Cancer Under 2 Scenarios: If Placebo Is Given and If Low-Dose Aspirin (LDA) Is Given
eFigure 2. Cumulative Incidence Function for Cancer
eFigure 3. Model Performance in Bootstrap Validation
eFigure 4. Model Performance Using AUTOC as the Performance Metric in Bootstrap Validation
eFigure 5. Histogram of the Predicted Individualized Treatment Effects
eFigure 6. Aalen-Johansen Estimates of the Cumulative Incidence Function for Cancer, Stratified by Defined Subgroups (Treatment-Favorable/Treatment-Unfavorable) and Randomization Groups (Aspirin/Placebo)
eFigure 7. Calibration Plot
eFigure 8. Sensitivity Analysis Excluding Participants With a History of Cancer: Model Performance in Bootstrap Validation
eFigure 9. Sensitivity Analysis Excluding Clonal Hematopoiesis of Indeterminate Potential (CHIP) as a Predictor: Model Performance in Bootstrap Validation
Data Sharing Statement
