Oral contraceptives and the risk of subarachnoid hemorrhage

Nontraumatic subarachnoid hemorrhage (SAH) accounts for 5 to 10% of patients with stroke¹ and is fatal in 50% of patients.^2,3 The vast majority of patients with SAH are due to ruptured cerebral aneurysms.¹ The condition affects a relatively young population, producing more than 50% of mortality due to cerebrovascular disease in those younger than 35 years.⁴ In women aged 20 to 44 years, SAH is as common as ischemic stroke,⁵ and produces twice its mortality rate.⁶ Furthermore, SAH mortality and incidence are more than 50% greater in women than men in the United States.^6-8 Oral contraceptive (OC) use and hormonal factors have been suggested as possible causes of these age and gender differences.

The cost to society is substantial. An estimated $2.3 billion is lost each year in the United States when direct and indirect costs are tallied.⁹ In spite of its importance, little attention has been directed to the study of SAH. This may be due to its relative infrequency compared with ischemic stroke when all age groups are considered, or because SAH can be difficult to distinguish from other forms of intracerebral hemorrhage.¹⁰

On the contrary, many careful studies have examined the risks of OCs. These studies are observational in design because a randomized, placebo-controlled trial of OC use has not been possible. In early studies, an association between increased rates of SAH and OC use was found, and was one of the first major concerns about the safety of these agents.^11,12 Both case-control and cohort designs have evaluated the risk of SAH among OC users and have found relative risk (RR) as high as 6.5¹² and as low as 0.5.¹³ Recently, three large case-control studies suggested a small, nonsignificant elevated risk of SAH among OC users.^5,14,15

We performed a meta-analysis of published studies to determine whether OC use is a risk factor for SAH after accounting for the variability in study designs and results.

Methods. The medical literature, since the introduction of OCs in 1960, was searched in all languages using the MEDLINE database (January 1966 to December 1997) and Dissertation Abstracts On-line (January 1861 to December 1997 for North American universities) with the key words subarachnoid, cerebral, hemorrhage, aneurysm, risk factor, and oral contraceptive. The Cumulated Index Medicus (January 1960 to December 1965) was searched with the subject heading: contraceptives, oral. Additional references were retrieved from the bibliographies of reviews and other articles of interest.

All pertinent articles were reviewed independently by two investigators(S.C.J. and D.R.G.) and screened using the following predetermined inclusion criteria: 1) the study used a cohort or case-control design with control subjects gathered within 2 years of patients, 2) at least five patients with SAH were reported, 3) SAH was defined in the report and distinguished from intraparenchymal hemorrhage, 4) adequate information was given to determine the RR and CIs for SAH in OC users compared with nonusers, 5) age was controlled either in the design or the analysis, and 6) the data were not fully reported in a later study.

Data were extracted independently by two investigators (S.C.J. and J.M.C.). If a potential confounding risk factor was restricted or matched in the study design, or included in multivariate or stratified analysis, that factor was considered controlled in the study result. These methods attempt to eliminate confounding effects of other factors on the association of interest. Estrogen dose was determined by the prevailing use pattern at study onset, with OCs containing more than 50 µg of estrogen considered a high dose.^14,16 Study onset was chosen as a convenient reference point for estrogen dosage to accommodate potential effects of earlier exposures. Current users were those taking OCs at the time of presentation, and ever users included current users and those reporting past use, the minimum duration of which was variably defined. Definitions for smoking, hypertension, and alcohol also varied between studies and were not uniformly reported.

The most extensively controlled RR estimates were used as the primary outcome measures from individual studies. For the overall estimate, if a study provided an RR for SAH incidence rather than for mortality, the former category results were included to maximize the number of subjects. When data on current use were available, they were preferred over ever and past use because they were felt to be more reliable and possibly more relevant biologically. Variances were extracted directly, or calculated from CIs or two-by-two tables.¹⁷ Before the data were analyzed, a subgroup analysis of exposure, outcome, and cofactor measures was planned to examine potential sources of bias and heterogeneity between studies. This analysis included stratification by cohort or case-control design, prevailing estrogen dose usage, control for smoking and hypertension, exposure classification as current versus ever users, and outcome measure of incidence or mortality. Firm diagnoses of SAH patients were defined as those supported by cerebral angiography, CT, MRI, or autopsy. The source of SAH generally was not stated in the studies. Refusal rate was defined as the reported percentage of eligible subjects approached who refused to participate in the study.

Differences between investigators on study inclusion and abstraction were discussed, and full agreement on final coding was reached in all subjects.

Summary RR estimates were calculated as an average of individual study results weighted by the inverse of their variance using the random effects method of DerSimonian and Laird applied to odds ratio results.¹⁸ In this way, larger studies contribute more to the summary RR estimate. Further, this method assumes that the studies included in the meta-analysis are only a sample of all the studies to be performed, and heterogeneity between study outcomes "broadens" the CI of the summary estimate, providing a more conservative estimate of significance.¹⁷ Homogeneity was tested using the Mantel-Haenszel method,¹⁷ and study results were considered heterogenous (i.e., with differences unlikely to be due to chance alone) if the resultant p value was less than 0.10. Based on published simulation studies, we estimated more than 75% power to identify study heterogeneity in the overall estimate.¹⁹ Summary estimates from subgroups of studies were compared using a z statistic with unequal variances assumed.²⁰

Results. A total of 522 references were retrieved from MEDLINE(483), Index Medicus (2), Dissertation Abstracts On-line (1), and bibliographic review (36). Twenty-seven of these were case-control or cohort studies examining OC risk on cerebral hemorrhage. Ten references were excluded because the reported data were not adequate to determine the RR of SAH associated with OC use, most often because patients with SAH and intraparenchymal hemorrhage were not distinguished.^21-30 Due to redundancy in data reporting, the remaining 17 publications describe only 11 independent studies.^{5,11-15,31-41} One study had cohort and nested case-control components, and these are shown separately in tables 1 and 2, but only the later, more complete data were included in the overall meta-analysis. The 11 independent RR estimations used in the overall meta-analysis are italicized in table 2.

Nine of the 11 studies showed an increased risk of SAH in OC users, which reached statistical significance in subgroups of three studies (see table 2⁹). However, two studies found nonsignificant protective effects.^13,36 The RR estimates ranged from 0.5 to 6.5, and were statistically homogeneous by the Mantel-Haenszel test (p = 0.30), which evaluates the probability that differences in study outcomes arose by chance alone(table 3).

The summary estimate of effect for all studies was an RR of 1.42 (95% CI, 1.12 to 1.80; p = 0.004), indicating a statistically significant elevation in SAH risk in OC users with CIs excluding an RR of one. A plot of RR estimations against time of study onset showed a trend toward smaller RR through the years (data not shown). To examine sources of variation and to try to explain the decreasing trend of effect size, studies were stratified based on control of potential confounders, design features, and exposure and outcome measures (see table 3).

If only those study results controlling for smoking were included, evidence of homogeneity increased (from p = 0.30 to p = 0.69), and the summary RR remained highly significant (RR, 1.55; 95% CI, 1.26 to 1.91; p < 0.0001; see table 3). For the six studies controlling for hypertension and smoking, the summary RR was similar (RR, 1.49; 95% CI, 1.20 to 1.85;p = 0.0003).

A greater risk of SAH was reported in studies with higher estrogen doses(high-estrogen RR, 1.67; 95% CI, 1.10 to 2.53; low-estrogen RR, 1.23; 95% CI, 0.84 to 1.80), but the difference was not significant (p > 0.10). After excluding studies not controlling for smoking, the small, nonsignificant risk difference persisted (high-estrogen RR, 1.94; 95% CI, 1.06 to 3.56; low-estrogen RR, 1.51; 95% CI, 1.18 to 1.92; see table 3).

The cohort studies reported higher RR (1.92; 95% CI, 0.91 to 4.06) compared with the case-control studies taken as a whole (RR, 1.40; 95% CI, 1.10 to 1.78), but this difference was not significant (p > 0.10; see table 3). There was no difference in the RR summary estimation for studies using mortality (RR, 1.54; 95% CI, 1.12 to 2.14) rather than incidence (RR, 1.37; 95% CI, 0.89 to 2.09) as an outcome measure. There was a small, nonsignificant difference between RR estimations from studies examining outcome in current OC users (RR, 1.57; 95% CI, 1.25 to 1.99) compared with ever users (RR, 1.18; 95% CI, 0.79 to 1.77) and past users (RR, 1.16; 95% CI, 0.46 to 2.92), and the study results of current users were more homogenous (p = 0.63) than those of ever users (p = 0.05) and past users (p = 0.008).

Discussion. Taken together, the studies support a weak positive association between OC use and SAH risk with an RR of 1.42 (95% CI, 1.12 to 1.80; p = 0.004). The study results did not reveal evidence of heterogeneity by the Mantel-Haenszel test (p = 0.30), implying that the range of study outcomes may have arisen by chance alone. However, the estimations of risk varied from a protective effect with an RR of 0.5 in one population-based case-control study¹³ to a strong detrimental effect with an RR of 6.5 among current OC users in an early, nested case-control study¹² (see table 2), indicating large variation in study outcomes. Because we were concerned about potential bias in the summarized observational studies, the study results were stratified by features of design and analysis to examine the sources of this variation.

Cigarette smoking has been identified as an important risk factor for SAH, with an RR of 3.0 to 3.5.^42,43 In addition, OC use and cigarette smoking have been correlated with early studies showing higher rates of smoking among OC users,²⁸ and later studies showing lower rates as smoking became known as a relative contraindication to OC use.¹³ Therefore, studies in the late 1960s and early 1970s that failed to control adequately for smoking would tend to overestimate the SAH risk of OCs, whereas later studies would underestimate the effect because some of the effects of smoking would be attributed to OC use.

If only those studies controlling for smoking are included in the summary RR estimate, the study results are more homogeneous (p = 0.69) and the RR remains elevated and highly significant (RR, 1.55; 95% CI, 1.26 to 1.91; p < 0.0001). This suggests that failure to control for the confounding influence of smoking was responsible for some of the variability in study results and probably contributed to the declining trend in RR estimations through the years as an initial positive correlation between smoking and OC use was replaced with a negative correlation in the 1970s as contemporaneous use was discouraged. Effect modification with excess risk of SAH in OC users that smoke may also explain some of this downward trend.

Hypertension is also a risk factor for SAH, with an RR of 2.4 to 3.7 in one systematic review,⁴³ and has been considered a relative contraindication to OC use. Therefore, studies that fail to control for hypertension may underestimate the risk of OC use because hypertension will be overrepresented in the unexposed group. To determine whether this bias was important in the group of studies, results from the six studies controlling for hypertension and smoking were analyzed separately. The summary RR did not differ from that calculated by including all studies, suggesting that failure to control for hypertension did not affect the overall estimate of effect (see table 3).

Heavy alcohol consumption,⁴³ positive family history,⁴⁴ and lower socioeconomic status¹⁴ may also be risk factors for SAH. There were not enough studies controlling for these factors to determine whether they were confounding variables.

The current low-estrogen OCs, in widespread use since the late 1970s, are generally considered to be safer than their high-dose predecessors.¹⁴ To explore the relationship between estrogen dose and SAH risk, meta-analysis was stratified on dose of estrogen based on prevailing use pattern during the period of study.¹⁶ A greater risk was seen in studies with higher estrogen doses but the difference was not statistically significant, and there was heterogeneity among the low-estrogen studies (see table 3). After excluding studies not controlling for smoking, the results were homogeneous, and the risk difference persisted. The summary effect of the low-estrogen OCs used today in studies controlling for smoking remained significant but was small (RR, 1.51; 95% CI, 1.18 to 1.92; p = 0.0009). The transition from high- to low-estrogen OCs may have contributed to the decline in RR over time.

The RR estimations from studies examining outcome in current OC users was greater than that of ever users and past users (see table 3). This suggests that the effects of OCs on SAH risk may attenuate with time. The study results of current users were also more homogeneous than those of ever and past users, as might be expected given the greater timing and dose variability in those whose exposure is defined by ever using OCs. Further, two well-performed studies have even shown a nonsignificant protective effect of past OC use.^15,36

In this meta-analysis of observational studies, which is limited by the biases of the summarized studies, use of OCs appears to be a weak risk factor for SAH, particularly with the low-estrogen preparations used today. Although each of the studies examining the association between OCs and SAH risk has been flawed in some way, the vast majority has shown positive effects in the face of a shift toward fewer coincident risk factors in OC users. Only two studies^13,36 found a protective effect in current users, not statistically significant in either, and one of these was not controlled for hypertension and smoking at a time when their presence was likely negatively correlated with OC use.¹³ Perhaps the most carefully performed studies pertinent to practice today were the recent population-based case-control studies of Petitti et al.¹⁴ and Schwartz et al.¹⁵ It is reassuring that these authors found RRs of 1.5 and 1.3, respectively, very similar to the summary estimate of all studies.

Two other groups have produced systematic reviews of the literature on OC use and SAH.^16,43 The study of Teunissen et al.⁴³ did not find a significant effect of OC use on SAH risk, but the summary estimate only included five studies and failed to consider potential confounders. The review of Prentice and Thomas¹⁶ found an RR of 2.0 for current OC use in both cohort and case-control studies with a CI excluding 1.0. Again, only five studies, all published before 1984, were reviewed, the specific methodology used was not described, and potential confounders were not considered in the analysis.

The mechanism by which OCs could increase the risk of SAH is not clear. OCs produce mild elevations in blood pressure that could account for some of the increased risk.⁴⁵ Also, exogenous estrogen has been associated with a decrease in the carotid Doppler pulsatility index of postmenopausal women, suggesting that OC use may alter the elasticity of cerebral vessels,⁴⁶ perhaps predisposing young women to the formation of aneurysms. In contrast, postmenopausal hormonal replacement therapy has been associated with a reduced risk of SAH,³⁶ making a simple association between increased exogenous estrogen use and SAH implausible. The causal pathway between OC use and SAH is likely complex, involving the replacement of normal menses in young women and protection from intracranial atherosclerosis in older women. The form of progesterone in the OC may also be important to the risk of SAH, as suggested by one recent study reporting a significantly greater risk for levonorgestrel but not norethindrone-type progestins.¹⁵

If it is accepted that OCs produce an increased risk of SAH, how important is the effect? Using an RR of 1.5 taken from the low-estrogen estimate for studies controlling for smoking, the attributable risk is 33%, suggesting that one-third of SAH patients among OC users are due to OCs. Using an estimate of 12% OC usage in women age 15 to 44 years,¹⁴ the population-attributing risk percent is 5.7%, implying that 5.7% of SAH events in women of child-bearing age are caused by OC use. With an estimated background rate of 15/100,000 person-years in this group,^1,3,47 the added risk to an OC user of child-bearing age would be 5/100,000 person-years, and 20,000 women would need to be treated before seeing a single additional patient with SAH. This is certainly a small risk compared with potential health benefits of OCs.⁴⁸ In the United States, an additional 430 patients each year with SAH due to OC use would be predicted, with a total cost estimated at $43 million.⁹ With this small increase in disease burden, it is unlikely that SAH risk will have a large independent influence on the decision about OC use in most women.

Should OCs be avoided in women at relatively high risk of SAH? Perhaps the absolute risk increase produced in patients already at high risk is enough to discourage OC use in this group, but evidence of effect modification would be even more compelling. Very few data are available assessing effect modification of OC use on other SAH risk factors. Petitti and Wingird¹² found strong positive effect modification in cigarette-smoking OC users, with an RR of 21.9 (95% CI, 7.1 to 67.3) compared with an RR of 6.5 for OC users and an RR of 5.7 for smokers.¹² One of the studies excluded from the meta-analysis due to inadequate differentiation of SAH from intraparenchymal hemorrhage also found effect modification with highly elevated hemorrhagic stroke risk in OC users who were smokers but not in nonsmokers.^21,22 These data provide more support for contraindicating OC use in smokers. Studies looking at the SAH risk of OC use among those with hypertension, heavy alcohol use, known existing aneurysms, or a positive family history are not available.

This meta-analysis suggests that OC use is associated with a small increased risk of SAH. For most women, this risk is probably inconsequential in evaluating the decision about OCs compared with other forms of birth control. However, for women at high risk of SAH due to existing unruptured aneurysms, a strong positive family history, cigarette smoking, or hypertension, it may be advisable to consider alternative forms of contraception until more data are available.