Abstract
The Aspirin in Reducing Events in the Elderly (ASPREE) Trial recruited 19,114 participants across Australia and the United States during 2010–2014. Participants were randomized to receive either 100 mg of aspirin daily or matching placebo, with disability-free survival as the primary outcome. During a median 4.7 years of follow-up, 37% of participants in the aspirin group permanently ceased taking their study medication and 10% commenced open-label aspirin use. In the placebo group, 35% and 11% ceased using study medication and commenced open-label aspirin use, respectively. In order to estimate compliance-adjusted effects of aspirin, we applied rank-preserving structural failure time models. The results for disability-free survival and most secondary endpoints were similar in intention-to-treat and compliance-adjusted analyses. For major hemorrhage, cancer mortality, and all-cause mortality, compliance-adjusted effects of aspirin indicated greater risks than were seen in intention-to-treat analyses. These findings were robust in a range of sensitivity analyses. In accordance with the original trial analyses, compliance-adjusted results showed an absence of benefit with aspirin for primary prevention in older people, along with an elevated risk of clinically significant bleeding.
Keywords: aged, aspirin, causal inference, compliance, older persons, prevention, randomized trials, rank-preserving structural accelerated failure time models
Abbreviations
- AF
acceleration factor
- ASPREE
Aspirin in Reducing Events in the Elderly
- CVD
cardiovascular disease
- HR
hazard ratio
- ITT
intention-to-treat
- MACE
major adverse cardiovascular events
- OLA
open-label aspirin
- RPSFTM
rank-preserving structural failure time model
The use of aspirin for the secondary prevention of cardiovascular disease (CVD) is well-established. However, recent trials have indicated an absence of any benefit for primary prevention in high-risk populations (1–3). The Aspirin in Reducing Events in the Elderly (ASPREE) Trial was a double-blind, randomized, placebo-controlled trial conducted among 19,114 community-dwelling people in Australia and the United States, mostly aged ≥70 years, with no history of heart disease, dementia, or disability (1, 4–6), assessing whether 100 mg of aspirin daily could prolong disability-free life. In intention-to-treat (ITT) analyses, the group randomized to receipt of aspirin did not have improved disability-free survival as compared with the placebo group, but it had increased risk of clinically significant bleeding (1). There was weak evidence of reduced rates of CVD outcomes in the aspirin group but an increased mortality risk, driven predominantly by excess cancer mortality (4, 5).
During the follow-up period (median, 4.7 years), some participants permanently ceased using randomized study medication and some commenced open-label aspirin (OLA) use (1), resulting in treatment switching in both randomized groups. With switching, the ITT randomized treatment comparisons no longer estimate the effectiveness of taking aspirin, and in general are biased towards the null. To estimate the causal effect of daily aspirin use, which is arguably of greater interest for individuals considering aspirin use, we need to adjust for treatment switching. Given the confounding effect of prognostic factors on the relationship between compliance and outcomes, simply conditioning on compliance will not estimate a treatment effect with a valid causal interpretation (7). G-methods (7) may be used to account for this selection bias while adjusting for treatment switching with time-to-event outcomes. We applied one of these, the rank-preserving structural failure time model (RPSFTM) (8, 9), to the ASPREE Trial. This method allows estimation of the causal effects of treatment under the hypothetical scenario of perfect compliance, that is, if all participants adhered to the treatment protocol assigned to them, as recommended by the International Council for Harmonization (10). We refer to these causal effects as “compliance-adjusted treatment effects.” The estimand here is the causal effect of continuous usage of aspirin from trial enrollment to the occurrence of each event of interest, compared with not using aspirin at all, which has been termed the switching-adjusted estimand (11–13).
Here we summarize treatment compliance patterns in ASPREE and estimate compliance-adjusted treatment effects of aspirin on the primary endpoint of disability-free survival and a range of secondary outcomes.
METHODS
Study design and population
Recruitment into the ASPREE Trial took place in Australia (n = 16,703) and the United States (n = 2,411) between March 2010 and December 2014 (14, 15). Participants were aged ≥70 years (aged ≥65 years for US African-American and Latino participants) and were required to be in good health, free of preexisting major CVD, cognitively intact, and able to independently perform basic activities of daily living. Participants were required to have successfully completed a 4-week placebo run-in phase with pill-taking compliance of at least 80%. The study protocol and statistical analysis plan are described elsewhere (6, 16). A total of 19,114 participants were randomized to receive a daily tablet of either 100 mg of aspirin (n = 9,525) or placebo (n = 9,589). Participant demographic characteristics, comorbid conditions, clinical measures, and lifestyle factors were assessed at baseline and at annual visits, which were scheduled from the time of recruitment through to trial cessation in June 2017. Results of the ASPREE Trial are described elsewhere (1, 4–6).
Study outcomes
The primary study endpoint was disability-free survival, defined as survival free of incident dementia and persistent physical disability. Dementia was defined according to the Diagnostic and Statistical Manual of Mental Disorders, Fourth Edition, criteria (17). Persistent physical disability was defined as an inability to independently perform, or having severe difficulty performing, at least 1 of the 6 basic activities of daily living on 2 consecutive occasions 6 months apart (18).
Secondary endpoints included all-cause mortality, dementia, persistent physical disability, major hemorrhage, cardiovascular events, cancer incidence, cancer mortality, and an additional endpoint, major adverse cardiovascular events (MACE). CVD was a composite of fatal coronary heart disease, nonfatal myocardial infarction, fatal or nonfatal stroke, and hospitalization for heart failure. MACE was defined as a composite of death due to coronary heart disease (but not including death from heart failure), nonfatal myocardial infarction (19), and fatal and nonfatal ischemic stroke (20). Major hemorrhage was defined as hemorrhagic stroke, symptomatic intracranial bleeding, or extracranial clinically significant bleeding (defined as bleeding requiring transfusion, hospitalization or prolongation of hospitalization, or surgery, or bleeding causing death) (5, 21).
Event-adjudication committees reviewed primary and secondary endpoint events in a blinded manner according to definitions published elsewhere (1, 5, 21).
Compliance
Compliance with the study medication regimen was measured using an annual pill count from returned bottles (details are given in Web Appendix 1, available at https://doi.org/10.1093/aje/kwad168). At annual visits, participants indicated whether they had been taking OLA during that year. No details on dosage, dates, or length of time on OLA were collected throughout the trial.
Time to switching and time on aspirin
RPSFTMs require data on time on aspirin for each study participant. This was estimated assuming a maximum of 1 treatment switch per person. Time on study medication was estimated by adding the number of pills taken in a participant’s last year on medication to the number of days from randomization to the beginning of that year. Time on OLA was estimated for participants who took OLA over 2 consecutive calendar years. Annual visit dates for these calendar years were used as stopping and starting times to estimate time on OLA. OLA was assumed to be equivalent to study aspirin. If the data indicated that a participant took randomized aspirin and OLA on the same day, their total dosage that day was still assumed to be 100 mg.
For the aspirin group, the period of taking study medication plus any contiguous period of OLA use was their time on aspirin; the switch occurred upon cessation of study medication (or additional cessation of OLA if applicable). For the placebo group, participants switched when commencing OLA; time after that switch was time on aspirin. For each endpoint, if the switch occurred after the event of interest, it did not affect the analysis of that endpoint.
A second approach allowing for multiple treatment switches, using yearly pill count data in the aspirin group and estimated days on OLA for either group, was used as a sensitivity analysis. Details for both approaches are given in Web Appendix 2.
We used the estimated single switch times to determine how many participants switched before the trial ended. For both treatment groups, we summarized pill compliance in each year since randomization, length of time on study medication, time to commencement of OLA use, and estimated switching time.
Adjusting for compliance: the RPSFTM
An RPSFTM (8) is used to estimate the causal effect of treatment if everyone in the study had complied with the treatment protocol they were randomized to—that is, if no treatment switching had occurred. To estimate this treatment effect, we assume that OLA use is equivalent to study aspirin use and that taking placebo is equivalent to not taking any aspirin. Being fully compliant then translates to taking 100 mg of enteric-coated aspirin daily if randomized to the aspirin group and taking no aspirin if randomized to the placebo group.
The RPSFTM is an instrumental-variable estimation method (22, 23), used when it is difficult to measure all prognostic factors affecting both treatment and outcome, such as in treatment switching. The RPSFTM makes use of counterfactual survival times (i.e., potential outcomes) (8). These are the survival times we could potentially observe under each possible treatment history (including no treatment, full treatment, and switching), prior to any randomization. The RPSFTM treatment effect is estimated by assuming that counterfactual survival times under the “no treatment” scenario are independent of randomized treatment assignment. We will refer to this as the randomization assumption (24). If all participants did in fact comply perfectly, the RPSFTM estimate is equivalent to the ITT effect. The RPSFTM is the structural version of the accelerated failure time model (25):
 |
(1) |
where 
is the counterfactual survival time under no treatment for individual i, 
is the observed survival time, 
is the observed treatment at time t (0 for placebo, 1 for aspirin), and 
is the treatment effect to be estimated. The word “structural” applies, as we are modeling counterfactual survival times. With a binary treatment, equation 1 simplifies to
 |
(2) |
where 
and 
represent the time on treatment and time off treatment for individual i. The model assumes that if an individual receives no treatment, their untreated counterfactual survival time equals their observed survival time. The effect of treatment is the multiplying factor 
and is the compliance-adjusted treatment effect of interest. It is interpreted as an acceleration factor (AF), which either expands or contracts survival time relative to no treatment, if an individual is continually treated with the active treatment according to the protocol (13, 25). For example, AF = 2 implies that continuously taking 100 mg of aspirin daily will extend the untreated survival time by a factor of 2. AF < 1 implies that treatment causes harm. The AF is assumed to be constant, and it applies whenever an individual (from either randomization group) is taking aspirin, regardless of when that is, but ceases once treatment stops. This is sometimes called the common treatment effect assumption (24) and satisfies the “rank-preserving” part of the model (i.e., if we could rank the survival outcomes of all participants under their counterfactual outcome with no treatment, their rankings would remain the same if all were fully treated) (8).
The model parameter 
is estimated via g-estimation (12, 26) and requires the validity of the randomization assumption (24). In g-estimation, possible values of 
are entered into equation 2 to estimate 
for each participant. The value of 
that returns a log-rank test statistic of 0, and hence renders counterfactual survival times as independent of randomized treatment, is the estimate of 
. A hazard ratio (HR) can be estimated by using the estimate of 
to predict the counterfactual survival times that would have occurred if all of those randomized had fully complied with their assigned treatment (9). A Cox proportional hazards model is fitted to these predicted counterfactual survival times, and compliance-adjusted Kaplan-Meier curves (9, 27) are constructed. If AF > 1 and hence increases survival time in the treatment group, the corresponding HR will be less than 1—that is, the risk of the event in the treatment group is reduced, but there is no direct relationship between the AF and the HR values. As a randomization-based method, the RPSFTM retains the P value of the ITT analysis.
Implementing the RPSFTM
The RPSFTM was implemented in Stata/MP 15.1 (StataCorp LLC, College Station, Texas) (28) using the “strbee” (27) command. This command covers all aspects of implementing and assessing the model and obtaining compliance-adjusted treatment effect estimates (27). For each outcome, 500 bootstrap samples (stratified by randomized group) were used to estimate bias-corrected 95% confidence intervals for the compliance-adjusted AF and HR. The data were recensored using the administrative study end date before g-estimation was applied. Recensoring is necessary because observed survival times may be related to treatment; so, for example, if treatment were beneficial, participants might be more likely to be censored rather than experience the event, or vice versa if treatment were harmful. Therefore, censoring on the counterfactual scale of 
will be informative regarding treatment (because that censoring is estimated from observed survival times). To avoid this potential selection bias, the 
are recensored at the minimum possible censoring time based on all possible treatment histories (9, 27). No baseline variables were adjusted for in the RPSFTM analysis, to match the prespecified ITT analysis. Model performance was assessed by plotting the predicted counterfactual survival times under no treatment for each treatment group; comparable survival distributions imply that the model estimation procedure worked well (24), whereas differences may indicate invalid underlying model assumptions (13). The uniqueness of 
was checked by plotting estimates of 
against the associated log-rank test statistic.
Sensitivity analyses
Sensitivity analyses were performed to test various assumptions, as follows: 1) the model was fitted without recensoring, as information loss from recensoring may cause a bias that overestimates the treatment effect (29); 2) time on aspirin was estimated allowing for multiple treatment switches (see Web Appendix 2); 3) imputed pill counts were set to 0; and 4) switchers in the placebo group were assumed to have a 30% (or 50%) lower treatment effect than those randomized to aspirin use. Sensitivity analysis (4) assesses violation of the common treatment effect assumption. Participants may have switched from placebo to aspirin for clinical reasons, and the effect of aspirin on extending the time to other clinical events may be lessened in these participants. We also examined whether standard errors in the ITT analyses were sensitive to clustering of participants according to primary-care practice and whether the RPSFTM was sensitive to the competing risk of death.
RESULTS
Use of study medication and OLA
Rates of follow-up in ASPREE were high, with only 1.2% of participants (118 assigned to aspirin, 119 to placebo) withdrawing and 1.5% (139 assigned to aspirin, 157 to placebo) being uncontactable in the last 12 months of the trial and considered lost to follow-up (1). The numbers of annual visits completed (expressed as a percentage of the number of annual visits due) were 18,841 (99%) for year 1; 18,570 (99%) for year 2; 16,191 (98%) for year 3; 12,243 (98%) for year 4; 7,744 (97%) for year 5; and 2,738 (92%) for year 6 (1).
Figure 1 shows the number of participants who had permanently ceased using study medication before the end of the trial, or before death or withdrawal (37% in the aspirin group; 35% in the placebo group), and the number of participants who commenced using OLA (10% aspirin; 11% placebo). Reasons for ceasing study medication permanently are given in Table 1, for 83% of those who ceased use before the last year of the trial. Commencing use of aspirin (or another anticoagulant) was the most common reason given (38% aspirin; 44% placebo), with a contraindication for aspirin given for 8% and 5% in the aspirin and placebo groups, respectively.
Figure 1.
Use of study medication and open-label aspirin (OLA) in the ASPREE Trial, 2010–2017. Switchers were defined in the randomized placebo group as persons who commenced taking OLA and in the aspirin group as persons who stopped taking any study medication or OLA before the last year of the study, with the exception of deaths and withdrawals, who were only defined as switchers if they stopped taking aspirin more than 90 days before death/withdrawal. “Study end” = within 1 year of the study’s end. ASPREE, Aspirin in Reducing Events in the Elderly. a5 deaths/withdrawals overlapped; b7 deaths/withdrawals overlapped; c1 death/withdrawal overlapped.
Table 1.
Reasons for Ceasing Use of Study Medication Among ASPREE Trial Participants Who Stopped Using Medication Before the Last Year of the Trial, by Treatment Group, 2010–2017
| Treatment Group | ||||
|---|---|---|---|---|
| Aspirin (n = 9,525) | Placebo (n = 9,589) | |||
| Reason for Ceasing Study Medication a | No. | % | No. | % |
| Participant | ||||
| Bruising | 131 | 4.5 | 55 | 2.0 |
| Stomach upset | 87 | 3.0 | 91 | 3.3 |
| Nosebleeds | 58 | 2.0 | 32 | 1.2 |
| Personal reasonsb | 972 | 33.7 | 957 | 34.8 |
| GP/PCP | ||||
| Commenced use of aspirin/anticoagulantc | 1,108 | 38.4 | 1,218 | 44.2 |
| Aspirin now contraindicated (major bleed or anaphylaxis) | 220 | 7.6 | 133 | 4.8 |
| Otherd | 277 | 9.6 | 215 | 7.8 |
| Principal investigator | 31 | 1.1 | 52 | 1.9 |
| Total | 2,884 | 100.0 | 2,753 | 100.0 |
Abbreviations: ASPREE, Aspirin in Reducing Events in the Elderly; GP, general practitioner; PCP, primary-care physician.
a Reasons for ceasing study medication available for 83% of those who ceased before the last 12 months of the trial.
b For example, being unwilling to take the drug, or reasons not specified.
c Reasons specified were myocardial infarction, angina, transient ischemic attack, and stroke.
d For example, surgery or reasons not specified.
Observed data were consistent with a single treatment switch. In both randomized groups, when participants first stopped taking study medication, 98% ceased permanently. For those on assigned medication, average pill compliance was 83% or higher in all years after randomization (Table 2). In the placebo group, 860 (83%) of OLA users remained on OLA until <1 year from the end of the trial. OLA users in the placebo group were considered switchers (n = 1,039; 11%). Participants in the aspirin group who ceased study medication and any additional OLA use before the trial ended (or at least 90 days before death or withdrawal) were considered switchers (n = 2,780; 29%).
Table 2.
Mean Pill Compliancea in the ASPREE Trial Through Each Year Since Randomization, by Treatment Group, 2010–2017
| Year Since Randomization | Aspirin Group | Placebo Group | ||
|---|---|---|---|---|
| No. of Participants Taking Any Study Medication b | Pill Compliance Among Those on Study Medication, % c | No. of Participants Taking Any Study Medication b | Pill Compliance Among Those on Study Medication, % c | |
| First | 9,448 | 83 (21) | 9,505 | 83 (20) |
| Second | 8,003 | 87 (21) | 8,198 | 88 (20) |
| Third | 7,056 | 89 (18) | 7,321 | 89 (18) |
| Fourth | 5,676 | 89 (19) | 5,908 | 90 (18) |
| Fifth | 3,954 | 89 (19) | 4,146 | 90 (18) |
| Sixth | 2,230 | 90 (19) | 2,335 | 91 (18) |
| Seventh | 727 | 92 (17) | 732 | 92 (16) |
Abbreviation: ASPREE, Aspirin in Reducing Events in the Elderly.
a Pill compliance was defined as percentage of study pills taken in a given year.
b Only participants taking any study medication during the year are included.
c Values are expressed as mean (standard deviation).
Table 3 gives summary statistics for the time participants spent on study medication and OLA and for the time to switching by treatment group. Participants stayed on their study medication for 73% of their potential study time in the aspirin group and 75% of their potential study time in the placebo group. OLA usage was also similar. For switchers, on average, the aspirin group switched slightly earlier than the placebo group, possibly reflecting earlier adverse effects.
Table 3.
Estimated Amount of Time on Study Medication, Time to Starting Use of Open-Label Aspirin, Amount of Time Using Open-Label Aspirin, and Time to Switching Treatments, by Randomized Treatment Group, in the ASPREE Trial, 2010–2017
| Treatment Group | No. of Participants | Mean (SD) | Median (IQR) |
|---|---|---|---|
| Estimated Amount of Time on Study Medication, years | |||
| Aspirin | 9,525 | 3.37 (1.86) | 3.49 (1.90, 4.87) |
| Placebo | 9,589 | 3.47 (1.83) | 3.58 (2.25, 4.94) |
| Estimated Time on Study Medication as a Proportion of Potential Time in Study | |||
| Aspirin | 9,525 | 0.73 (0.35) | 0.97 (0.41, 1.00) |
| Placebo | 9,589 | 0.75 (0.34) | 0.98 (0.47, 1.00) |
| Estimated Time to Starting Use of OLA, years | |||
| Aspirin | 953 | 2.04 (1.36) | 1.98 (0.99, 2.99) |
| Placebo | 1,039 | 1.91 (1.35) | 1.94 (0.97, 2.97) |
| Estimated Time to Starting OLA as a Proportion of Potential Time in Study | |||
| Aspirin | 953 | 0.40 (0.24) | 0.38 (0.21, 0.61) |
| Placebo | 1,039 | 0.37 (0.24) | 0.36 (0.18, 0.56) |
| Estimated Amount of Time Taking OLA as a Proportion of Potential Time in Study | |||
| Aspirin | 953 | 0.45 (0.24) | 0.41 (0.24, 0.62) |
| Placebo | 1,039 | 0.48 (0.24) | 0.47 (0.27, 0.68) |
| Estimated Time to Switching Treatments, years | |||
| Aspirin | 2,780 | 1.66 (1.38) | 1.29 (0.50, 2.51) |
| Placebo | 1,039 | 1.91 (1.35) | 1.94 (0.97, 2.97) |
| Time to Switching as a Proportion of Potential Study Time | |||
| Aspirin | 2,780 | 0.33 (0.24) | 0.26 (0.10, 0.53) |
| Placebo | 1,039 | 0.37 (0.24) | 0.36 (0.18, 0.56) |
Abbreviations: ASPREE, Aspirin in Reducing Events in the Elderly; IQR, interquartile range; OLA, open-label aspirin; SD, standard deviation.
Compliance-adjusted results of RPSFTM
Table 4 displays the unadjusted ITT analyses and the compliance-adjusted treatment effects of aspirin for each outcome. The compliance-adjusted AFs and HRs were not materially different from the ITT results for the primary endpoint (disability-free survival), dementia, disability, MACE, CVD, or cancer. The compliance-adjusted AF and HR estimates for death, major hemorrhage, and cancer mortality showed an increased risk of these events’ occurring if participants were continuously taking aspirin daily as compared with taking no aspirin. The ITT AF estimate for time to major hemorrhage was 0.76 (95% confidence interval: 0.66, 0.86), implying a reduction of 24% in the expected time to a major hemorrhage associated with being assigned to aspirin compared with placebo; after adjustment for treatment switching, the AF was reduced to 0.65 (95% confidence interval: 0.54, 0.79), a 35% reduction in the expected time to major hemorrhage associated with continually taking aspirin as compared with the time expected if untreated.
Table 4.
Estimates of Treatment Effects on Incidence of Outcomes Due to Aspirin Use, Unadjusted and Adjusted for Compliance Using a Rank-Preserving Structural Failure Time Model, in the ASPREE Trial, 2010–2017
| Outcome | ITT Analysis Using the RPSFTM to Estimate the Acceleration Factor | Compliance-Adjusted Estimates of the Acceleration Factor Using the RPSFTM a | ITT Analysis Using the Cox Proportional Hazards Model | Compliance-Adjusted Estimates of the HR Using the RPSFTM a | ||||
|---|---|---|---|---|---|---|---|---|
| AF b | 95% CI c | Adjusted AF b | 95% CI c | HR | 95% CI | Adjusted HR | 95% CI c | |
| Disability-free survival (primary endpoint) | 0.99 | 0.95, 1.05 | 0.99 | 0.91, 1.11 | 1.01 | 0.92, 1.11 | 1.02 | 0.84, 1.20 |
| Death | 0.93 | 0.88, 0.99 | 0.89 | 0.78, 1.00 | 1.14 | 1.01, 1.29 | 1.25 | 1.00, 1.63 |
| Dementia | 1.01 | 0.93, 1.11 | 1.04 | 0.86, 1.17 | 0.98 | 0.83, 1.15 | 0.93 | 0.75, 1.35 |
| Persistent physical disability | 1.10 | 1.00, 1.22 | 1.11 | 1.00, 1.32 | 0.85 | 0.70, 1.03 | 0.82 | 0.64, 1.07 |
| Major hemorrhage | 0.76 | 0.66, 0.86 | 0.65 | 0.54, 0.79 | 1.38 | 1.18, 1.62 | 1.74 | 1.31, 2.19 |
| MACE | 1.10 | 0.99, 1.28 | 1.13 | 0.99, 1.45 | 0.89 | 0.77, 1.03 | 0.86 | 0.70, 1.08 |
| CVD | 1.04 | 0.95, 1.15 | 1.06 | 0.93, 1.25 | 0.95 | 0.83, 1.08 | 0.93 | 0.77, 1.11 |
| Cancer | 0.97 | 0.90, 1.06 | 0.95 | 0.86, 1.08 | 1.04 | 0.95, 1.13 | 1.06 | 0.95, 1.19 |
| Death from cancer | 0.85 | 0.76, 0.94 | 0.75 | 0.62, 0.91 | 1.31 | 1.10, 1.56 | 1.49 | 1.14, 1.91 |
Abbreviations: AF, acceleration factor; ASPREE, Aspirin in Reducing Events in the Elderly; CI, confidence interval; CVD, cardiovascular disease; HR, hazard ratio; ITT, intention-to-treat; MACE, major adverse cardiovascular events; RPSFTM, rank-preserving structural failure time model.
a Time on aspirin estimated from a single switching time with recensoring applied.
b Ratio of the expected survival time for the event if a person were continuously treated with daily aspirin to the expected survival time if the person were untreated.
c Bias-corrected bootstrap CIs (500 repetitions), stratified by treatment group.
Figures 2–5 show Kaplan-Meier observed and compliance-adjusted cumulative incidence curves for disability-free survival, death, major hemorrhage, and cancer mortality. The compliance-adjusted curves are the predicted counterfactual cumulative incidence curves assuming full compliance with the randomly assigned treatment. Adjusting for treatment switching made little difference in the curves for the primary endpoint. For death, major hemorrhage, and cancer mortality, adjusting for treatment switching increased cumulative incidence in the aspirin group. Corresponding graphs for dementia, disability, CVD, MACE, and cancer are given in Web Figures 1–5. For each of these outcomes, the compliance-adjusted curves overlaid the observed survival curves. The stepped feature of curves for dementia reflects the biennial screening of dementia events in the study (30).
Figure 2.
Observed and compliance-adjusted cumulative disability-free survival (primary endpoint), derived using a rank-preserving structural failure time model, in the ASPREE Trial, 2010–2017. ASPREE, Aspirin in Reducing Events in the Elderly.
Figure 5.
Observed and compliance-adjusted cumulative incidence of cancer mortality, derived using a rank-preserving structural failure time model, in the ASPREE Trial, 2010–2017. ASPREE, Aspirin in Reducing Events in the Elderly.
Figure 3.
Observed and compliance-adjusted cumulative incidence of death, derived using a rank-preserving structural failure time model, in the ASPREE Trial, 2010–2017. ASPREE, Aspirin in Reducing Events in the Elderly.
Figure 4.
Observed and compliance-adjusted cumulative incidence of major hemorrhage, derived using a rank-preserving structural failure time model, in the ASPREE Trial, 2010–2017. ASPREE, Aspirin in Reducing Events in the Elderly.
The estimated model parameters for each outcome were confirmed as unique solutions to the RPSFTM. Kaplan-Meier curves of the predicted counterfactual survival times under no treatment were compared between groups and indicated that the model estimation procedure worked well for all outcomes (Web Appendix 3, Web Figures 6–14).
Results of sensitivity analyses.
All sensitivity analyses resulted in conclusions similar to those of the primary analysis (Web Appendix 4, Web Tables 1–7). Importantly, assuming reduced treatment effects for placebo group switchers (sensitivity analysis 4) had no impact on the estimated treatment effects.
DISCUSSION
The ASPREE Trial was key to providing recommendations against the use of aspirin for primary prevention in the age group ≥70 years. In our analysis, we assessed the impact of compliance in the ASPREE Trial by estimating the causal effect of treatment if everyone in the study had complied perfectly with their assigned treatment, that is, if no treatment switching had occurred. Compliance-adjusted treatment effects were estimated for prespecified primary and secondary endpoints using the RPSFTM, assuming a single treatment switch—a simplification consistent with observed data. Assuming equivalence between study aspirin use and OLA use and between taking placebo and not taking aspirin, 29% of the aspirin group and 11% of the placebo group switched treatment. No major differences were seen between the compliance-adjusted and ITT effects of aspirin on disability-free survival or on the incidence of dementia, disability, MACE, CVD, or cancer. The compliance-adjusted treatment effects for death, major hemorrhage, and cancer mortality indicated greater risk with taking aspirin, as compared with placebo, than was reported in ITT analyses. The results were homogenous across a range of sensitivity analyses conducted to test modeling assumptions. Overall, compliance-adjusted effects are consistent with the published trial results, possibly because of the high levels of compliance but also because of relatively weak effects of daily aspirin use on the primary outcome and on dementia, disability, MACE, CVD, and cancer incidence. Hence, our results are consistent with those presented in the original ASPREE study and address recently raised concerns that nonadherence may have led to an absence of aspirin effects (31). Thus, our results strengthen the conclusion that there is no evidence of benefit for the use of aspirin for primary prevention among people aged ≥70 years who have no history of CVD.
Two examples of recent meta-analyses (32, 33) assessing the potential benefit-risk ratio of aspirin use for primary prevention of CVD are worth highlighting. These 2 meta-analyses included 13 aspirin primary prevention trials, including ASPREE, and found results comparable with those of the original ASPREE study for CVD and major hemorrhage. Gelbenegger et al. (32) concluded that the benefit-risk ratio was insufficient to recommend aspirin use for CVD primary prevention, and Zheng et al. (33) concluded that the absolute benefits and risks associated with aspirin use were comparable and that decisions should be made on an individual basis. In our analysis, the risk of CVD remained similar but the risk of major hemorrhage increased with aspirin.
Few aspirin trials have adjusted for treatment noncompliance. In the Women’s Health Study (34, 35), a large (n = 39,876) primary prevention trial comparing the occurrence of CVD events between low-dose aspirin users (100 mg every other day) and placebo users, researchers applied marginal structural models to adjust for treatment noncompliance and time-varying determinants of aspirin use. The compliance-adjusted effect of aspirin on CVD risk was unchanged from the primary analysis, but the risk of CVD mortality was decreased (35). Similar methods and results were found in the aspirin component of the Physicians’ Health Study, a double-blind placebo-controlled trial (n = 22,071) of 325 mg of aspirin every second day, when adjusting for the effect of treatment switching on CVD mortality (36).
In the ASPREE Trial, we applied the RPSFTM, one of two g-methods commonly used to adjust for treatment switching in randomized controlled trials, the other being the inverse-probability-of-censoring weights method. Analysis software code facilitating replication is provided in Web Appendix 5 (Web Table 8). Major strengths of the RPSFTM are that 1) treatment comparisons are based on the randomized groups and 2) the method does not require persons who switch to be comparable to those who do not switch (8, 9). The RPSFTM requires only the randomization, observed survival times, and treatment history for each participant. With inverse-probability-of-censoring weights, participants are censored at the time of the switch, and the observations are weighted by baseline and time-dependent prognostic factors of the probability of switching. This method relies on known switching times and assumes no unmeasured confounders.
Limitations
Our analysis had several limitations. Assuming treatment equivalence between placebo and no aspirin use and between study aspirin and OLA use may have been questionable. For example, participants who take a placebo may receive indirect psychological benefits through taking a daily pill. Similarly, Hawthorne effects are possible under OLA use relative to blinded randomized use. Nevertheless, with a pharmacological compound such as aspirin, these phenomena are likely to be muted in comparison with other forms of intervention, and in our opinion any bias that arises is likely to be relatively small.
Another limitation was the need to estimate aspirin use. While pill counts were only available on an annual basis, the high level of compliance is likely to have diminished this as a possible source of bias. Participants taking OLA may have taken a different dose of aspirin than the study drug, and the amount of time on OLA was estimated rather than observed. Assessing the impact of these assumption violations on the results is difficult, given that doses and dates of taking OLA were not consistently recorded. Assumptions made in imputing missing pill count data were addressed through sensitivity analyses, which gave similar estimates.
The common treatment effect assumption implies that the same treatment effect applies to all participants while taking aspirin, regardless of when that is and which randomization group they belong to, which may be implausible. For example, one may expect the effect of aspirin to differ in participants who switched due to a clinical event, although our sensitivity analysis assessing this yielded results consistent with the compliance-adjusted estimates reported. In a recent simulation study, Latimer et al. (37) found that the RPSFTM estimated unbiased treatment effects, as long as the switching proportion was less than 60%, and that this remained true even if the common treatment assumption was not satisfied, as long as the estimated AF was less than 2—conditions that were satisfied in our study. Although it is difficult to check the model assumptions, a comparison of the untreated counterfactual survival did not indicate any problems with the g-estimation procedure or the RPSFTM assumptions.
Other limitations of this study include the fact that the competing risk of death for nondeath outcomes (or noncancer deaths for the cancer death outcome) was not accounted for in the main analysis. The cumulative incidence plots from the “strbee” command will underestimate incidence for nondeath outcomes, since the overall 5.5% mortality rate will impart a competing-risk bias, although this bias will mostly arise later in follow-up, as the death rate is low in the early part of the follow-up period. The literature explaining how to account for competing risks within the causal framework is a developing area, and a robust methodological framework has not yet been established (38). In time-to-event analyses, a general approach to the dilemma of competing risks is to censor at the time of death and estimate cause-specific HRs, and this was the approach taken in the original ITT analyses; it has been argued (39) that this is reasonable when assessing disease etiology but may be less relevant for the purpose of prediction (which was not a focus here). A sensitivity analysis in which we reset the timing of competing-risk deaths showed results consistent with those of the main analyses. In addition, although adjusting for baseline covariates may increase statistical power, we chose to match the prespecified ITT analysis to enable direct comparison between the compliance-adjusted estimates and the ITT estimates. Further, models of the type we have fitted will be misspecified if they do not include all relevant time-dependent interactions (40). In ASPREE, there is potential time-dependent confounding of the relationship between aspirin use and incident CVD events such as ministroke or angina (that are not included in the definition of outcomes). Finally, loss to follow-up may cause selection bias for any analysis. However, in ASPREE, rates of withdrawal and loss to follow-up were very low at 1.5% and 1.2%, respectively, and are not expected to have affected the results.
Conclusion
These results reinforce the conclusions of the original published report on the ASPREE Trial (1, 5), namely that there is an absence of benefit with aspirin for primary prevention in older people and that there is the likelihood of elevated risk of significant bleeding. Although a set of assumptions was necessary for our analyses, the results were robust to those assumptions. These findings confirm the appropriateness of changes made to clinical guidelines that aspirin should not be recommended for primary prevention in older individuals (41, 42).
Supplementary Material
ACKNOWLEDGMENTS
Author affiliations: School of Public Health and Preventive Medicine, Monash University, Melbourne, Victoria, Australia (C. L. Smith, J. Kasza, R. L. Woods, J. E. Lockery, C. M. Reid, S. G. Orchard, A. M. Tonkin, J. J. McNeil, R. Wolfe); Berman Center for Outcomes and Clinical Research, Hennepin Healthcare Research Institute, Minneapolis, Minnesota, United States (B. Kirpach, A. M. Murray); Central Clinical School, Monash University, Melbourne, Victoria, Australia (E. Storey); Menzies Institute for Medical Research, University of Tasmania, Hobart, Tasmania, Australia (M. R. Nelson); Family Medicine and Rush Alzheimer’s Disease Center, Rush University Medical Center, Chicago, Illinois, United States (R. C. Shah); Department of Pharmacy Practice and Science, College of Pharmacy, University of Iowa, Iowa City, Iowa, United States (M. E. Ernst); and Department of Family Medicine, Carver College of Medicine, University of Iowa, Iowa City, Iowa, United States (M. E. Ernst).
ASPREE was supported by the National Institute on Aging and the National Cancer Institute, US National Institutes of Health (grant U01AG029824); the National Health and Medical Research Council of Australia (grants 334047 and 1127060); Monash University (Melbourne, Victoria, Australia); and the Victorian Cancer Agency (Melbourne, Victoria, Australia).
Data will be made available to investigators whose proposed use of the data, registered as a project through the ASPREE Access Management Site (https://ams.aspree.org/public/), has been approved by a review committee. Access will be allowed through a secure Web-based data portal (the ASPREE Safe Haven system) based at Monash University.
We acknowledge the ASPREE staff in Australia and the United States for their dedication and skill in the conduct of the trial. We are also most grateful to the ASPREE participants, who so willingly volunteered for this study, and the general practitioners and medical clinics who supported them.
ASPREE was approved by multiple institutional review boards in the United States and Australia, was registered with the International Standard Randomised Controlled Trial Number Registry (trial ISRCTN83772183) and clinicaltrials.gov (trial NCT01038583), and was undertaken in accordance with the principles of the Declaration of Helsinki.
Bayer AG (Leverkusen, Germany) provided aspirin and matching placebo for ASPREE but played no other role in the trial. M.R.N. received travel support from Bayer and personal fees related to advisory board activity from Sanofi S.A. (Paris, France) outside the scope of this work. A.M.T. received honoraria from Bayer for lectures and advisory board participation and honoraria from Pfizer, Inc. (New York, New York) and AMGen, Inc. (Thousand Oaks, California) for lectures.
REFERENCES
- 1. McNeil JJ, Woods RL, Nelson MR, et al. Effect of aspirin on disability-free survival in the healthy elderly. N Engl J Med. 2018;379(16):1499–1508. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 2. ASCEND Study Collaborative Group . Effects of aspirin for primary prevention in persons with diabetes mellitus. N Engl J Med. 2018;379(16):1529–1539. [DOI] [PubMed] [Google Scholar]
- 3. Gaziano JM, Brotons C, Coppolecchia R, et al. Use of Aspirin to Reduce Risk of Initial Vascular Events in Patients at Moderate Risk of Cardiovascular Disease (ARRIVE): a randomised, double-blind, placebo-controlled trial. Lancet. 2018;392(10152):1036–1046. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 4. McNeil JJ, Nelson MR, Woods RL, et al. Effect of aspirin on all-cause mortality in the healthy elderly. N Engl J Med. 2018;379(16):1519–1528. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 5. McNeil JJ, Wolfe R, Woods RL, et al. Effect of aspirin on cardiovascular events and bleeding in the healthy elderly. N Engl J Med. 2018;379(16):1509–1518. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 6. 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] [PMC free article] [PubMed] [Google Scholar]
- 7. Hernán MA, Robins JM. Per-protocol analyses of pragmatic trials. N Engl J Med. 2017;377(14):1391–1398. [DOI] [PubMed] [Google Scholar]
- 8. Robins JM, Tsiatis AA. Correcting for non-compliance in randomized trials using rank preserving structural failure time models. Commun Stat Theory Methods. 1991;20(8):2609–2631. [Google Scholar]
- 9. White IR, Babiker AG, Walker S, et al. Randomization-based methods for correcting for treatment changes: examples from the Concorde trial. Stat Med. 1999;18(19):2617–2634. [DOI] [PubMed] [Google Scholar]
- 10. Committee for Medicinal Products for Human Use, European Medicines Agency . ICH E9 (R1) addendum on estimands and sensitivity analysis in clinical trials to the guideline on statistical principles for clinical trials. Amsterdam, the Netherlands: European Medicines Agency; 2020. https://www.ema.europa.eu/en/documents/scientific-guideline/ich-e9-r1-addendum-estimands-sensitivity-analysis-clinical-trials-guideline-statistical-principles_en.pdf. Accessed June 2, 2023. [Google Scholar]
- 11. Hernan MA, Cole SR, Margolick J, et al. Structural accelerated failure time models for survival analysis in studies with time-varying treatments. Pharmacoepidemiol Drug Saf. 2005;14(7):477–491. [DOI] [PubMed] [Google Scholar]
- 12. Greenland S, Lanes S, Jara M. Estimating effects from randomized trials with discontinuations: the need for intent-to-treat design and G-estimation. Clin Trials. 2008;5(1):5–13. [DOI] [PubMed] [Google Scholar]
- 13. Sullivan TR, Latimer NR, Gray J, et al. Adjusting for treatment switching in oncology trials: a systematic review and recommendations for reporting. Value Health. 2020;23(3):388–396. [DOI] [PubMed] [Google Scholar]
- 14. McNeil JJ, Woods RL, Nelson MR, et al. Baseline characteristics of participants in the ASPREE (ASPirin in Reducing Events in the Elderly) study. J Gerontol A Biol Sci Med Sci. 2017;72(11):1586–1593. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 15. Lockery JE, Collyer TA, Abhayaratna WP, et al. Recruiting general practice patients for large clinical trials: lessons from the Aspirin in Reducing Events in the Elderly (ASPREE) study. Med J Aust. 2019;210(4):168–173. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 16. 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] [PMC free article] [PubMed] [Google Scholar]
- 17. American Psychiatric Association . Diagnostic and Statistical Manual of Mental Disorders (DSM-IV). 4th ed. Washington, DC: American Psychiatric Association; 2000. [Google Scholar]
- 18. Katz S. Assessing self-maintenance: activities of daily living, mobility, and instrumental activities of daily living. J Am Geriatr Soc. 1983;31(12):721–727. [DOI] [PubMed] [Google Scholar]
- 19. Alpert J, Thygesen K, Antman E, et al. Myocardial infarction redefined—a consensus document of the joint European Society of Cardiology/American College of Cardiology Committee for the Redefinition of Myocardial Infarction. J Am Coll Cardiol. 2000;36(3):959–969. [DOI] [PubMed] [Google Scholar]
- 20. World Health Organization . Stroke—1989. Recommendations on stroke prevention, diagnosis, and therapy. Report of the WHO Task Force on Stroke and Other Cerebrovascular Disorders. Stroke. 1989;20(10):1407–1431. [DOI] [PubMed] [Google Scholar]
- 21. Margolis KL, Mahady SE, Nelson MR, et al. Development of a standardized definition for clinically significant bleeding in the ASPirin in Reducing Events in the Elderly (ASPREE) trial. Contemp Clin Trials Commun. 2018;11:30–36. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 22. Angrist JD, Imbens GW, Rubin DB. Identification of causal effects using instrumental variables. J Am Stat Assoc. 1996;91(434):444–455. [Google Scholar]
- 23. Hernán M, Robins J. Instruments for causal inference: an epidemiologist’s dream? Epidemiology. 2006;17(4):360–372. [DOI] [PubMed] [Google Scholar]
- 24. Latimer NR, Bell H, Abrams KR, et al. Adjusting for treatment switching in the METRIC study shows further improved overall survival with trametinib compared with chemotherapy. Cancer Med. 2016;5(5):806–815. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 25. Cox DR. Analysis of Survival Data. 1st ed. Boca Raton, FL: Chapman & Hall/CRC Press; 1984. [Google Scholar]
- 26. Hernán MA, Robins J. Causal Inference: What If. Boca Raton, FL: Chapman & Hall/CRC Press; 2020. [Google Scholar]
- 27. White IR, Walker S, Babiker A. Strbee: randomization-based efficacy estimator. Stata J. 2002;2(2):140–150. [Google Scholar]
- 28. StataCorp LLC . Stata Statistical Software: Release 15. College Station, TX: StataCorp LLC; 2017. [Google Scholar]
- 29. Latimer NR, White IR, Abrams KR, et al. Causal inference for long-term survival in randomised trials with treatment switching: should re-censoring be applied when estimating counterfactual survival times? Stat Methods Med Res. 2019;28(8):2475–2493. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 30. Ryan J, Storey E, Murray AM, et al. Randomized placebo-controlled trial of the effects of aspirin on dementia and cognitive decline. Neurology. 2020;95(3):e320–e331. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 31. Kim K, Hennekens CH, Martinez L, et al. Primary care providers should prescribe aspirin to prevent cardiovascular disease based on benefit−risk ratio, not age. Fam Med Community Health. 2021;9(4):e001475. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 32. Gelbenegger G, Postula M, Pecen L, et al. Aspirin for primary prevention of cardiovascular disease: a meta-analysis with a particular focus on subgroups. BMC Med. 2019;17(1):198. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 33. Zheng SL, Roddick AJ. Association of aspirin use for primary prevention with cardiovascular events and bleeding events: a systematic review and meta-analysis. JAMA. 2019;321(3):277–287. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 34. Ridker PM, Cook NR, Lee I-M, et al. A randomized trial of low-dose aspirin in the primary prevention of cardiovascular disease in women. N Engl J Med. 2005;352(13):1293–1304. [DOI] [PubMed] [Google Scholar]
- 35. Cook NR, Cole SR, Buring JE. Aspirin in the primary prevention of cardiovascular disease in the Women’s Health Study: effect of noncompliance. Eur J Epidemiol. 2012;27(6):431–438. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 36. Cook NR, Cole SR, Hennekens CH. Use of a marginal structural model to determine the effect of aspirin on cardiovascular mortality in the Physicians’ Health Study. Am J Epidemiol. 2002;155(11):1045–1053. [DOI] [PubMed] [Google Scholar]
- 37. Latimer NR, Abrams KR, Lambert PC, et al. Assessing methods for dealing with treatment switching in clinical trials: a follow-up simulation study. Stat Methods Med Res. 2018;27(3):765–784. [DOI] [PubMed] [Google Scholar]
- 38. Young JG, Stensrud MJ, Tchetgen EJ, et al. A causal framework for classical statistical estimands in failure-time settings with competing events. Stat Med. 2020;39(8):1199–1236. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 39. Andersen PK, Geskus RB, de Witte T, et al. Competing risks in epidemiology: possibilities and pitfalls. Int J Epidemiol. 2012;41(3):861–870. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 40. Vansteelandt S, Joffe M. Structural nested models and G-estimation: the partially realized promise. Statist Sci. 2014;29(4):707–731. [Google Scholar]
- 41. Arnett D, Blumenthal R, Albert M, et al. 2019 ACC/AHA Guideline on the Primary Prevention of Cardiovascular Disease: a Report of the American College of Cardiology/American Heart Association Task Force on Clinical Practice Guidelines. J Am Coll Cardiol. 2019;74(10):e177–e232. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 42. Davidson KW, Barry MJ, Mangione CM, et al. Aspirin use to prevent cardiovascular disease: US Preventive Services Task Force recommendation statement. JAMA. 2022;327(16):1577–1584. [DOI] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.