The association of aspirin use with age-related macular degeneration progression in the Age-Related Eye Disease Studies: AREDS2 Report No. 20

Abstract

Purpose

To analyze the potential association between aspirin use and progression of age-related macular degeneration (AMD).

Design

Two prospective cohort studies within two controlled clinical trials of oral supplementation for age-related eye disease.

Participants

Age-Related Eye Disease Study (AREDS) participants, aged 55 to 80 years, and AREDS2 participants, aged 50 to 85 years.

Methods

Propensity scores for aspirin use were calculated for AREDS and AREDS2 participants, separately, by logistic regression. Of the participants without late AMD (geographic atrophy (GA) or neovascular AMD) in either eye at study baseline, aspirin users were matched 1:1 with non-users by propensity score (separately for AREDS and AREDS2). Proportional hazards regression was performed, adjusting for age, on the matched participants to evaluate associations between aspirin propensity score and progression to late AMD (and its subtypes).

Main outcome measures

Progression to late AMD on color fundus photographs, graded centrally.

Results

Of the 3734 eligible AREDS participants, 1049 (28.1%) were taking aspirin and 2685 (71.9%) were not. The equivalent figures in AREDS2 were 1198 (49.9%) and 1205 (50.1%), respectively. Following matching by propensity score, the characteristics of the users and non-users were similar, in both studies. Of the 1950 matched AREDS participants and 1694 matched AREDS2 participants, over median follow-up of 10.1 years (AREDS) and 5.0 years (AREDS2), the numbers who progressed to late AMD, GA, or neovascular AMD were 454 (23.3%), 345 (17.7%), and 278 (14.3%), respectively, in AREDS, and 643 (38.0%), 402 (24.6%), and 341 (20.1%), in AREDS2. The hazard ratios of progression in quintile 5 (highest propensity for aspirin use) versus 1 (reference) were 1.17 (p=0.35), 1.24 (p=0.25), and 0.95 (p=0.81), respectively, in AREDS, and 1.26 (p=0.09), 1.46 (p=0.03), and 1.12 (p=0.58), in AREDS2. No significant association with progression to late AMD was observed for any quintile 2–5, for any of the three outcomes, in either study.

Conclusions

Aspirin use was not significantly associated with progression to late AMD, or its subtypes, in either the AREDS or the AREDS2. Patients with AMD need not avoid aspirin for this reason, when its use is medically indicated.

Précis

Aspirin use was not associated with progression to late age-related macular degeneration (AMD), either geographic atrophy or neovascular AMD, in both AREDS and AREDS2 studies. Patients with AMD need not avoid aspirin when medically indicated.

Introduction

Age-related macular degeneration (AMD) is the leading cause of visual loss in developed countries (1, 2). Worldwide, the number of people with AMD is projected to be 196 million in 2020, rising to 288 million in 2040 (3). AMD arises from a complex interplay between genetic and environmental risk factors (4, 5), and is very strongly associated with aging; the prevalence of late AMD in Caucasian populations has been estimated at 1.4% at 70 years, rising to 5.6% at 80 years and 20.1% at 90 years (6). The disease is classified into early, intermediate, and late stages (7). Late AMD can occur in two forms: atrophic disease (characterized by enlarging lesions of geographic atrophy (GA)) and neovascular (or exudative) disease, though these two forms can coexist in the same eye.

Aspirin is a non-steroid anti-inflammatory drug (NSAID) with analgesic and anti-inflammatory properties. It is also an anti-thrombotic agent and has been used widely for many years in the treatment and prevention of cardiovascular and cerebrovascular diseases (8–11). Its mechanism of action includes both cyclooxygenase (COX)-dependent and COX-independent effects on platelets and other cells (12).

Conflicting reports have created ongoing controversy around the question of whether aspirin use may be associated with altered risk of AMD (13–16). This topic arose in 1988, when Kingham et al speculated that the availability of aspirin might ‘usher in a new wave of blindness’, based on their case series of 109 patients with macular hemorrhage and concomitant use of antiplatelet or anticoagulant drugs (17). Subsequent studies have examined whether aspirin use is associated with AMD prevalence or incidence; these have included case-control studies (18–21), cross-sectional studies (22, 23), cohort studies (24, 25), and randomized clinical trials (RCTs) (26, 27). However, the results of these studies have been inconsistent. Chong et al observed in their 2014 editorial that, in general, ‘the cross-sectional studies reported an association with early AMD, while this was not found in the cohort studies’, and that ‘the cohort studies found an association with only neovascular AMD … with no associations with early AMD or GA’ (14). Even meta-analyses conducted in this area have reported conflicting results (28–31). Two observed no significant association between aspirin use and AMD (28, 29), while one found a weak but statistically significant association (31). In the two that reported subgroup analysis of GA and neovascular AMD, both found a significant association with neovascular AMD but not with GA (29, 31).

This question is important since aspirin is used so widely in older individuals, particularly in developed countries, where cardiovascular disease is the leading cause of death. For example, in the US, 52% of adults aged 45 to 75 years reported taking daily aspirin (32). It has been suggested (33) that some patients with AMD have stopped taking aspirin without medical consultation following the publication of some previous studies (23, 24) and press releases (34, 35). If true, this might have caused excess mortality or morbidity in these individuals, particularly from cardiovascular events. On the other hand, it may also be important to consider whether aspirin might be associated with decreased AMD progression (perhaps for particular stages or subtypes), given renewed interest in the idea that AMD pathogenesis may have microvascular and hemodynamic components (36, 37). The question is also timely, since the role of daily low-dose aspirin in the medical field is currently the subject of major interest, particularly in primary prevention, following multiple reports from the ASPirin in Reducing Events in the Elderly (ASPREE) international RCT, as regards cardiovascular disease, dementia, and cancer (38–40).

The Age-Related Eye Disease Study (AREDS) was a multicenter prospective study of the clinical course of AMD and age-related cataract, as well as a phase III RCT designed to assess the effects of nutritional supplements on AMD and cataract progression (41). Similarly, the AREDS2 was a multicenter phase III RCT that analyzed the effects of different nutritional supplements on the course of AMD (42). The purpose of the current report was to examine potential associations between aspirin use and progression to late AMD in the AREDS and AREDS2.

Methods

Study population and procedures for AREDS

The study design for the AREDS has been described previously (41). In short, 4757 participants aged 55 to 80 years were recruited between 1992 and 1998 at 11 retinal specialty clinics in the United States. Based on fundus photographic gradings, best-corrected visual acuity, and ophthalmologic evaluations, participants were enrolled into one of several AMD categories. Of the 4757 participants, 3640 took part in the AMD trial. The participants were randomly assigned to placebo, antioxidants, zinc, or the combination of antioxidants and zinc.

The study design for the AREDS2 has also been described previously (42). In brief, 4203 participants aged 50 to 85 years were recruited between 2006 and 2008 at 82 retinal specialty clinics in the United States. Inclusion criteria at enrollment were the presence of either bilateral large drusen or late AMD (central GA or neovascular AMD) in one eye and large drusen in the fellow eye. The AREDS2 participants were randomly assigned to placebo, lutein/zeaxanthin, docosahexaenoic acid (DHA) plus eicosapentaenoic acid (EPA), or the combination of lutein/zeaxanthin and DHA plus EPA.

Institutional review board approval was obtained at each clinical site and written informed consent for the research was obtained from all study participants. The research was conducted under the Declaration of Helsinki and, for the AREDS2, complied with the Health Portability and Accessibility Act.

For both the AREDS and AREDS2, questionnaires administered at the baseline and subsequent study visits collected information that included medications, adverse events and treatment compliance. At baseline and annual study visits, comprehensive eye examinations were performed by certified study personnel using standardized protocols, and stereoscopic color fundus photographs were captured. Progression to late AMD was defined by the study protocol based on the grading of fundus photographs, as described previously (41, 42).

Participant cohorts and statistical methods: propensity score calculation, matching, and proportional hazards regression

The data for the AREDS and AREDS2 studies were analyzed separately but using the same methodology, which is illustrated here for the AREDS dataset. The unit of analysis was at the participant level. The number of participants who reported taking aspirin at least five times per week at the study baseline was calculated. A propensity score approach was used, since this controls for the presence of confounding in observational data (43), as described previously (44, 45). In this approach, a probability of aspirin use is computed for each participant, based on assessment of relevant covariates; aspirin users and non-users with similar scores can then be matched, to decrease confounding. Propensity scores for aspirin usage were calculated by logistic regression, using the following covariates: age, sex, race, education level, smoking status, history and control of hypertension (no history/positive history, no medication/positive history, on medication), history and control of diabetes mellitus (no history/positive history, no medication/positive history, on oral medication and/or insulin), history and control of hypercholesterolemia (for AREDS2 only; no history/positive history, no statin/positive history, on statin), and history of coronary artery disease (specifically angina, in AREDS, and angina/myocardial infarction/chronic heart disease/congestive heart failure, in AREDS2). For these analyses, all participants were used except those with missing covariate data.

The eligibility criteria for subsequent analyses were (i) participants without late AMD (defined as any GA or neovascular AMD) in either eye at study baseline, and (ii) presence of follow-up retinal imaging. Eligible participants taking aspirin were matched 1:1 without replacement to those not taking aspirin, using the greedy match algorithm (46). In order to assess balance, the characteristics of those taking and not taking aspirin were compared (in terms of the covariates listed above) by absolute standard differences, both before and after participant matching. In addition, in order to identify any significant differences present after participant matching, the characteristics of those taking and not taking aspirin were compared (in terms of the same covariates) by either the Wilcoxon rank-sum test (for age) or the chi-squared test (for categorical data).

Proportional hazards regression was performed on the dataset of matched participants, with progression to (i) late AMD, (ii) GA, and (iii) neovascular AMD as three separate outcomes, according to the aspirin propensity score; this was considered in quintiles, with quintile 1 (lowest propensity for aspirin use) as the reference quintile. Age was pre-specified to be included in all models. Assumptions of proportional hazards (for the propensity score quintiles) and of linearity (for age) were tested in all models. If these were not satisfied, additional models with a time-dependent aspirin variable were generated. The analyses were performed with and without adjustment for the competing risk of death by fitting a proportional subdistribution hazards model (47), using the macro developed by Kohl and Heinze (48). Significance was set by Bonferroni correction at p=0.008. All analyses were conducted using SAS version 9.4 (SAS Inc, Cary NC).

Results

Age-Related Eye Disease Study

Of the 4757 AREDS participants, 1337 (28.1%) reported taking aspirin at the study baseline and 3420 (71.9%) reported not taking aspirin. Propensity scores for aspirin use were generated for each participant using logistic regression. 3734 participants were eligible for subsequent analyses (i.e., had no GA or neovascular AMD at baseline), comprising 1049 (28.1%) taking aspirin and 2685 (71.9%) not. Prior to matching, the characteristics of those taking and not taking aspirin were compared (Table 1): the two groups were not balanced (i.e., absolute standard difference greater than 10%) with respect to age, sex, and history of angina. Of the 1049 participants taking aspirin, 975 (92.9%) were matched by propensity score to 975 (36.3%) of those not taking aspirin. Following matching, the groups taking and not taking aspirin were balanced (Table 1), with absolute standard differences well below 10% for all characteristics. In addition, no significant differences were present between the two groups (Table 2; p>0.3 for all characteristics).

Table 1.

Baseline characteristics of study participants according to aspirin use, before and after participant matching, for assessment of balance

	AREDS						AREDS2
	Original data			Matched data			Original data			Matched data
Characteristic	% in the aspirin group (n=1049)	% in the no aspirin group (n=2685)	Absolute standard difference *	% in the aspirin group (n=975)	% in the no aspirin group (n=975)	Absolute standard difference *	% in the aspirin group (n=1198)	% in the no aspirin group (n=1205)	Absolute standard difference *s	% in the aspirin group (n=847)	% in the no aspirin group (n=847)	Absolute standard difference *
Age (mean ± SD)	69.6±4.9	68.7±5.0	18.2	69.4±4.8	69.6±5.0	2.7	72.7±7.4	70.7±8.3	24.8	72.0±7.5	71.8±7.9	1.6
Male	52.5	39.6	26.2	50.4	50.7	0.6	47.4	35.4	24.7	40.3	40.4	0.2
White race	96.7	94.6	9.9	96.4	95.2	6.1	96.7	96.1	3.0	96.8	96.6	1.3
Education			1.8			0.8			1.7			0.6
HS or less	32.0	33.2		32.2	30.6		29.0	25.6		26.4	27.6
At least some college	27.9	31.1		28.1	30.1		47.4	50.7		49.2	49.8
Post-graduate	40.0	35.7		39.7	39.4		23.5	23.7		24.3	22.6
Smoking			4.5			0.5			2.4			0.7
Never	40.7	49.5		42.6	41.7		44.6	48.0		47.6	48.3
Former	53.3	44.1		51.4	52.4		50.7	45.6		46.6	45.2
Current	6.0	6.4		6.1	5.8		4.8	6.3		5.8	6.5
Hx angina/CAD**	19.4	5.2	44.1	13.2	12.8	1.2	19.8	6.8	38.9	8.1	9.6	5.0
Hx hypertension			5.8			0.6			10.3			0.9
No	54.1	66.7		56.0	55.4		36.1	56.1		44.4	45.7
Yes, no meds	6.4	6.2		6.6	8.3		2.9	4.2		3.5	4.0
Yes, on meds	39.5	27.0		37.4	36.3		60.9	39.7		52.1	50.3
Hx diabetes			0.5			0.2			1.7			0.1
No	90.3	92.7		90.5	90.8		85.3	92.4		89.4	89.6
Yes, no meds	2.8	2.3		2.8	3.3		12.6	6.7		9.3	9.1
Yes, on meds	7.0	5.0		6.8	5.9		2.1	0.9		1.3	1.3
Hx hypercholesterolemia									11.1			0.7
No							32.9	51.6		43.2	41.1
Yes, no meds							7.8	12.9		9.7	11.9
Yes, on meds							59.3	35.5		47.1	47.0

^*A value > 10% indicates imbalance

AREDS = Age-Related Eye Disease Study; AREDS2 = Age-Related Eye Disease Study 2; HS = high school; Hx = history of; SD = standard deviation

^**angina for AREDS, CAD=angina, CHF, CHD or MI for AREDS2

Table 2.

Baseline characteristics of study participants according to aspirin use, after participant matching, for assessment of significant differences

	AREDS			AREDS2 (for late AMD and neovascular AMD outcomes)			AREDS2 (for geographic atrophy outcome)
Characteristic	% in the aspirin group (n=975)	% in the no aspirin group (n=975)	P *	% in the aspirin group (n=847)	% in the no aspirin group (n=847)	P *	% in the aspirin group (n=817)	% in the no aspirin group (n=817)	P *
Age (mean ± SD)	69.4±4.8	69.6±5.0	0.61	72.0±7.5	71.8±7.9	0.93	71.9±7.5	71.8±7.9	0.83
Male	50.4	50.7	0.89	40.3	40.4	0.96	39.8	41.0	0.61
White race	96.4	95.2	0.18	96.8	96.6	0.79	96.7	96.6	0.89
Education			0.59			0.67			0.77
HS or less	32.2	30.6		26.4	27.6		26.2	27.1
At least some college	28.1	30.1		49.2	49.8		49.4	50.1
Post-graduate	39.7	39.4		24.3	22.6		24.4	22.9
Smoking			0.90			0.75			0.74
Never	42.6	41.7		47.6	48.3		47.6	48.3
Former	51.4	52.4		46.6	45.2		46.6	45.2
Current	6.1	5.8		5.8	6.5		5.8	6.5
Hx angina/CAD**	13.2	12.8	0.79	8.1	9.6	0.30	8.2	9.7	0.30
Hx hypertension			0.33			0.72			0.50
No	56.0	55.4		44.4	45.7		43.7	46.0
Yes, no meds	6.6	8.3		3.5	4.0		3.7	4.2
Yes, on meds	37.4	36.3		52.1	50.3		52.6	49.8
Hx diabetes			0.62			0.99			0.91
No	90.5	90.8		89.4	89.6		89.1	89.7
Yes, no meds	2.8	3.3		9.3	9.1		9.7	9.1
Yes, on meds	6.8	5.9		1.3	1.3		1.2	1.2
Hx hypercholesterolemia						0.30			0.28
No				43.2	41.1		43.7	41.5
Yes, no meds				9.7	11.9		9.5	11.9
Yes, on meds				47.1	47.0		46.8	46.6

^*comparison by chi-squared test, except for age (Wilcoxon rank-sum test)

AREDS = Age-Related Eye Disease Study; AREDS2 = Age-Related Eye Disease Study 2; AMD = age-related macular degeneration; HS = high school; Hx = history of; SD = standard deviation

^**angina for AREDS, CAD=angina, CHF, CHD or MI for AREDS2

Of the 1950 matched participants, the numbers who progressed to late AMD, GA, or neovascular AMD in either eye, over median follow-up of 10.1 years, were 454 (23.3%), 345 (17.7%), and 278 (14.3%), respectively. The results of proportional hazards regression, using the matched data, are shown in Table 3. The assumptions of proportional hazards (i.e., that the altered risk of progression to late AMD associated with aspirin propensity score quintiles is constant over time) and of linearity (i.e., that the relationship between age and risk of progression to late AMD is linear) were satisfied in all cases. No significant association with progression to late AMD was observed for any quintile 2–5 (with quintile 1 as the reference quintile), for any of the three outcomes, either at the Bonferroni-adjusted or the nominal level. The hazard ratios of progression to late AMD, GA, and neovascular AMD in quintile 5 (highest propensity for aspirin use) were 1.17 (p=0.35), 1.24 (p=0.25), and 0.95 (p=0.81), respectively. For all three outcomes, similar results were obtained when the analyses were adjusted for the competing risk of death (data not shown). In addition, sensitivity analyses were performed by repeating the proportional hazards regression using: (i) the full dataset, without matching, and (ii) a matched dataset, but using exact matching with several different combinations of characteristics. In all cases, the findings were similar to the original analyses and no significant results were observed (data not shown).

Table 3.

Progression to late age-related macular degeneration according to aspirin use: results of proportional hazards regression analyses according to quintiles of the aspirin propensity score

	AREDS		AREDS2
Outcome	HR (95% CL) *	P	HR (95% CL) *	P
Late AMD
Quintile 2 †	0.83 (0.57–1.19)	0.31	0.94 (0.74,1.20)	0.60
Quintile 3	1.01 (0.72–1.43)	0.94	1.05 (0.82,1.33)	0.72
Quintile 4	1.02 (0.72–1.42)	0.93	0.89 (0.69,1.14)	0.35
Quintile 5	1.17 (0.84–1.63)	0.35	1.26 (0.96,1.66)	0.09
Geographic atrophy
Quintile 2	0.81 (0.53–1.24)	0.33	0.94 (0.69–1.28)	0.70
Quintile 3	1.01 (0.68–1.50)	0.96	1.07 (0.78–1.45)	0.68
Quintile 4	0.95 (0.64–1.41)	0.81	0.83 (0.60–1.16)	0.27
Quintile 5	1.24 (0.86–1.81)	0.25	1.46 (1.04–2.04)	0.03
Neovascular AMD
Quintile 2	0.72 (0.44–1.15)	0.17	0.95 (0.68–1.33)	0.77 ‡
Quintile 3	0.96 (0.62–1.48)	0.84	1.19 (0.86–1.66)	0.30 ‡
Quintile 4	1.09 (0.71–1.65)	0.70	0.98 (0.69–1.40)	0.91 ‡
Quintile 5	0.95 (0.62–1.45)	0.81	1.12 (0.76–1.64)	0.58 ‡

^*adjusted for age

^†quin le 1 as reference, in all cases

^‡propor onal hazards assump on not sa sfied

AMD = age-related macular degeneration; AREDS = Age-Related Eye Disease Study; AREDS2 = Age-Related Eye Disease Study 2; CL = confidence limits; HR = hazard ratio

Age-Related Eye Disease Study 2

Of the 4203 AREDS2 participants, 2057 (48.9%) reported taking aspirin at the study baseline and 2146 (51.1%) reported not taking aspirin. Propensity scores for aspirin use were generated for each participant using logistic regression. 2403 participants were eligible for subsequent analyses of late AMD and neovascular AMD, comprising 1198 (49.9%) taking aspirin and 1205 (50.1%) not; 2327 participants were eligible for analyses of GA, comprising 1162 (49.9%) taking aspirin and 1165 (50.1%) not. Prior to matching, the characteristics of those taking and not taking aspirin were compared (Table 1): the two groups were not balanced with respect to age, sex, history/control of hypertension, history/control of hypercholesterolemia, and history of coronary artery disease.

For the analyses of late AMD and neovascular AMD, 847 (70.7%) of the 1198 participants taking aspirin were matched by propensity score to 847 (70.3%) of those not taking aspirin. Following matching, the groups taking and not taking aspirin were balanced (Table 1), with absolute standard differences well below 10% for all characteristics. In addition, no significant differences were present between the two groups (Table 2; p>0.3 for all characteristics). For the analyses of GA, 817 (70.3%) of the 1162 participants taking aspirin were matched by propensity score to 817 (70.1%) of those not taking aspirin. Again, following matching, no significant differences were present between the two groups (Table 2; p>0.25 for all characteristics).

Of the matched participants, the numbers who progressed to late AMD, GA, or neovascular AMD in either eye, over median follow-up of 5.0 years, were 643 (38.0%), 402 (24.6%), and 341 (20.1%), respectively. The results of proportional hazards regression, using the matched data, are shown in Table 3. The assumptions of proportional hazards and of linearity were satisfied for all three outcomes except neovascular AMD, for which the assumption of proportional hazards was not met. No significant association with progression to late AMD was observed for any quintile 2–5 (with quintile 1 as the reference quintile), for any of the three outcomes, at the Bonferroni-adjusted level. The hazard ratios of progression to late AMD, GA, and neovascular AMD in quintile 5 were 1.26 (p=0.09), 1.46 (p=0.03), and 1.12 (p=0.58), respectively. However, the latter result for neovascular AMD should be interpreted with caution, because of the assumption of proportional hazards. For this reason, additional models with a time-dependent aspirin variable were generated; in these models, again, the hazard ratios for progression to neovascular AMD were not significantly different to one for any quintile 2–5, at any time point, even at the nominal level (data not shown).

As regards progression to GA, the hazard ratio of 1.46 (1.04–2.04, p=0.03) for quintile 5 was borderline significant at the nominal but not the Bonferroni-adjusted level. However, no consistent dose-response association was observed across the quintiles. Similarly, with the propensity scores treated as a continuous variable rather than by quintiles, no significant association was observed at the nominal level (hazard ratio 1.89, 0.89–4.03; p=0.10). In addition, with inclusion of the competing risk of death, again, no significant association was present at the nominal level. Finally, analyses according to actual aspirin use (rather than propensity scores) in the matched cohort revealed no significant association at the nominal level.

For the other two outcomes (late AMD and neovascular AMD), similar results were obtained when the analyses were adjusted for the competing risk of death (data not shown). Sensitivity analyses were performed, as described above for the AREDS dataset; in all cases, the findings were similar to the original analyses and no significant results were observed (data not shown).

Discussion

In this study, longitudinal evaluation using propensity score matching and proportional hazards analysis demonstrated no evidence of elevated risk of progression to late AMD with aspirin use. In subtype analysis, this held true for both atrophic and neovascular forms of late disease. Importantly, the findings were also consistent between the AREDS and AREDS2. In addition, the results held true when analyses were adjusted for death as a risk in competition with progression to late disease. This factor is potentially important, given that participants taking aspirin might be more or less likely to die before capture of data on disease progression, through either increased prevalence of vascular disease or increased protection from disease, respectively.

As regards the finding of a hazard ratio that was borderline significant at the nominal level (for progression to GA only in the AREDS2 only), this may not be surprising in the context of multiple testing. The result was certainly not close to significance at the Bonferroni-adjusted level. In addition, the absence of similar findings in sensitivity analyses or in the AREDS adds confidence to the idea that this result is not clinically meaningful. Furthermore, previous concerns over increased risk of late AMD associated with aspirin usage have been surrounding neovascular AMD rather than GA (23–25).

Propensity score matching approach

In analyses of observational data, it is often essential to address confounding, such as by the propensity score approach used in this study. The importance is demonstrated by comparison of the baseline demographic and medical characteristics of the study participants, where significant and substantial differences are observed according to aspirin use. In addition to having significantly higher age, those taking aspirin had significantly higher prevalence of previous smoking, of cardiovascular disease, and of other risk factors for cardiovascular disease. Given that age and smoking are strong risk factors for AMD (6, 49), and that AMD (particularly neovascular AMD) is thought to share some risk factors with cardiovascular disease (49, 50), the potential difficulties of comparing AMD between these groups become apparent. In particular, there is strong potential for confounding by indication. This occurs when an exposure (e.g., aspirin use) appears to be associated with an outcome (e.g., late AMD), but the association is artefactual, originating from the true (causal) relationship between the indication for the exposure (e.g., combination of increased age, smoking, and cardiovascular disease/risk factors) and the outcome.

The advantage of propensity score matching is that it decreases or eliminates the effects of confounding in the analysis of observational data, by accounting for systematic differences in the baseline characteristics between subjects in two groups (51, 52). In this way, it allows observational (non-randomized) studies to mimic some aspects of an RCT, for improved assessment of causal relationships. Another advantage is that it permits simple analysis of covariate balance within the matched sample, as seen in Table 1.

Comparison with literature

As described above, four meta-analyses have been performed in this area (28–31), though only three of these examined risk of late AMD (28, 29, 31). Zhu et al found no significant association between aspirin use and risk of either early AMD or late AMD; the authors did not distinguish between GA and neovascular AMD (28). In the two meta-analyses that did distinguish between GA and neovascular AMD, both reported a significant association between aspirin use and increased risk of neovascular AMD (with relative risks of 1.59, 95% CI 1.09–2.31, and 1.95, 95% CI 1.40–2.72), but no significant association with GA (29, 31).

However, the meta-analyses are affected by the same potential limitations that apply to their constituent studies. Each of these two meta-analyses included five studies in their subgroup analysis of neovascular AMD, with most weight contributed by the cross-sectional study of de Jong et al (23), the cohort study of Klein et al (24), and the cohort study of Liew et al (25). However, observational studies such as these have inherent limitations, which have previously been discussed (13, 14). Amongst other limitations, the presence of residual confounding may substantially limit inferences regarding causality. A previous review article and editorial have demonstrated how, in statistical models of observational studies, residual confounding is likely to be present, despite attempts to adjust for cardiovascular risk factors (13, 14).

By contrast, RCTs have the major advantage of using randomization to eliminate confounding (both by known confounding variables such as age, smoking, and vascular risk factors, but also by unknown confounding variables), for more meaningful assessment of causal relationships. Two previous RCTs of aspirin found no increased risk of incident AMD (26, 27); the one of these that had some power to detect differences in the incidence of late disease found no increased risk (27). One ongoing RCT, the ASPREE-AMD study (53) is a sub-study of the ASPREE RCT mentioned above. It is examining the incidence and progression of AMD over 5 years in 5000 healthy participants aged 70 years and older. Until this RCT concludes, and in the absence of other available data from RCTs, analyses that incorporate propensity score matching or adjustment may be the best option available to address methodological problems from confounding.

We are aware of one other study that has used propensity score approaches to analyze the potential association between aspirin use and late AMD, though this report considered only neovascular AMD and not GA or late AMD overall (54). This large study used national health insurance data to follow a nationwide cohort in South Korea between 2010 and 2015. Using propensity score analysis (either for adjustment or for matching), the authors observed no elevated incidence of neovascular AMD in those taking long-term aspirin: hazard ratios were 0.97 (0.73–1.30) and 0.94 (0.70–1.28), respectively, i.e. consistent with our findings.

Finally, in the Comparison of AMD Treatments Trials (CATT), Ying et al analyzed the potential association between antiplatelet (and anticoagulant) drug use and retinal/subretinal hemorrhage in treatmentnaïve neovascular AMD (55). They observed that antiplatelet or anticoagulant use was not significantly associated with hemorrhage at baseline or at years 1 or 2. However, in the subgroup of participants with systemic hypertension at baseline, antiplatelet use was associated with increased prevalence but not increased size of hemorrhage (at baseline, but not in later years).

Strengths and limitations

The strengths of this study include its large size and long follow-up, meaning that the analyses were highly powered by the large number of progression events. In addition, the study benefitted from robust methods to capture data on disease progression; since these judgements were made through standardized protocols by reading center graders, without access to clinical information, the data were not subject to observer bias. The participants in this study were well characterized in terms of both eye disease and systemic comorbidities (including smoking status and cardiovascular disease, together with aspirin use); as this information was collected prospectively at baseline, it was not subject to recall bias. The use of propensity score matching was a strength in this study, as it addresses confounding by indication. Effective matching was verified by demonstration of covariate balance between groups.

Analyses were performed using both the AREDS and AREDS2 datasets for assessment of potential heterogeneity. This was considered important because the participants differed in several ways between the two studies (56). AREDS participants comprised individuals with a broad spectrum of AMD severity at baseline (including no disease), whereas AREDS2 participants were individuals with at least intermediate AMD. In addition, there was a substantial difference in aspirin usage between the two study populations (28% in AREDS and 49% in AREDS2). This difference mirrors national patterns observed in large surveys of aspirin usage in US adults: 28% in 1999 (57), 41% in 2004 (58), and 52% in 2012 (32). The probable reasons for this include expanded indications for aspirin usage (59, 60) and increased uptake of these recommendations over time. In addition, the higher age distribution (56) and higher prevalence of cardiovascular disease, hypertension, and diabetes in AREDS2 versus AREDS participants likely contributed to the difference. The negative results observed consistently in both studies may improve the confidence in and generalizability of the findings, since the AREDS2 results may be more generalizable to a population with a high burden of AMD, and the AREDS results to a population with a lower burden of AMD.

This analysis was limited by its retrospective (unplanned) nature. RCTs are the most effective way of making inferences regarding causal relationships, and this study did not involve allocation of aspirin treatment by randomization. However, propensity score matching is thought to mimic aspects of RCTs in the elimination of confounding by indication. Other limitations include the assumption that aspirin use (assessed at baseline) remained unchanged during follow-up and the fact the individuals studied, as clinical trial participants, may not be fully representative of the wider population. In addition, no data were available on the number of participants taking aspirin less frequently than five times per week. Hence, the control groups in our analyses likely contained some participants who were taking aspirin weekly or monthly. We have therefore not considered separately the anti-thrombotic properties of aspirin (of longer duration) versus its anti-inflammatory/analgesic properties (of shorter duration). However, a previous study showed that the proportion of older adults taking aspirin weekly (but not daily) is relatively low (23). Our definition of aspirin use was also in line with two previous studies that reported positive associations between aspirin use and AMD (23, 24).

Late AMD was defined principally by reading center grading of color fundus photographs, since spectral domain optical coherence tomography (SD-OCT) use was not widespread earlier on in the studies and no validated grading protocols existed. It is likely that some additional early cases of progression to GA and neovascular AMD may have been captured, had SD-OCT been performed on all participants. However, this limitation is partially mitigated by (i) long follow-up time, and (ii) use of anti-VEGF injection history, alongside color fundus photograph grading, to define progression to neovascular AMD. However, similar analyses based on data obtained from SD-OCT are an important future direction.

Conclusions

Aspirin use was not associated with increased risk of progression to late AMD, in either the AREDS or AREDS2. This held true for both the atrophic and neovascular forms of late disease. Propensity score matching was used to address the important issue of confounding by indication, which is strongly suspected in some previous observational studies. The benefits of aspirin in the secondary prevention of cardiovascular disease are well established. Based on the sum of existing data on AMD, we suggest that individuals with AMD need not avoid aspirin, when its use is medically indicated.