A systematic review and meta‐analysis of venous thrombosis risk among users of combined oral contraception

Abstract Background Combined oral contraceptives (COCs) containing various progestogens could be associated with differential risks for venous thromboembolism (VTE). Objective To evaluate the comparative risks of VTE associated with the use of low‐dose (less than 50 μg ethinyl estradiol) COCs containing different progestogens. Search strategy PubMed and the Cochrane Library were searched from database inception through September 15, 2016, by combining search terms for oral contraception and venous thrombosis. Selection criteria Studies reporting VTE risk estimates among healthy users of progestogen‐containing low‐dose COCs were included. Data collection and analysis A random‐effects model was used to generate pooled adjusted risk ratios and 95% confidence intervals; subgroup and sensitivity analyses assessed the impact of monophasic‐COC use and study‐level characteristics. Main results There were 22 articles included in the analysis. The use of COCs containing cyproterone acetate, desogestrel, drospirenone, or gestodene was associated with a significantly increased risk of VTE compared with the use of levonorgestrel‐containing COCs (pooled risk ratios 1.5–2.0). The analysis restricted to monophasic COC formulations with 30 μg of ethinyl estradiol yielded similar findings. After adjustment for study characteristics, the risk estimates were slightly attenuated. Conclusions Compared with the use of levonorgestrel‐containing COCs, the use of COCs containing other progestogens could be associated with a small increase in risk for VTE.

is much lower than the incidence of VTE during pregnancy and the postpartum period. 3,4 The effect of COCs on the risk of thrombosis was traditionally thought to be solely related to the effects of estrogen on hemostatic factors. However, studies have indicated that the risk of VTE varies among women using COCs containing different progestogens. Given the popularity and widespread use of COCs, any increase in the relative risk of VTE for particular COC formulations could translate to an excess absolute risk of important magnitude.
The present review was conducted for a consultation held by the WHO to examine the venous and arterial risks of COCs, as part of the process of updating the WHO Medical Eligibility Criteria for Contraceptive Use (WHO MEC) 5 ; the review and meta-analysis have been updated since the WHO consultation to include data published during the interim period. For women who wish to use COCs, the key clinical question is whether certain COC formulations might further increase the risk of VTE above that associated with other formulations. Although several other meta-analyses 6-9 on this question have been conducted, the present meta-analysis updates previous analyses and compares different formulations with a levonorgestrel user group rather than with a nonuser group. The objective of the present systematic review and meta-analysis was to estimate the risk for VTE among women using COCs containing different progestogens compared with COCs containing levonorgestrel.

| MATERIALS AND METHODS
In the present systematic review and meta-analysis, PubMed and the Cochrane Library databases were searched for all articles on the association between COC use and VTE in all languages published from database inception through September 15, 2016, using a combination of search terms for oral contraception and venous thrombosis (Table   S1). In addition, the reference lists from identified studies and key review articles were hand-searched for additional studies.
For the exposure, studies were included that reported results for users of COCs with low-dose ethinyl estradiol (dose <50 μg) coupled with one of the following progestogens: cyproterone acetate, desogestrel, dienogest, drospirenone, gestodene, norgestimate, or levonorgestrel. Studies were only included if the risk estimates were reported separately by COC formulation (including monophasic formulations that had the same dose of ethinyl estradiol in all active pills and multiphasic formulations that had varying doses of ethinyl estradiol throughout the cycle, provided they had the same progestogen). Studies were excluded if COCs containing 50 μg of ethinyl estradiol or more accounted for more than 10% of the total exposure. Also excluded were articles that only reported the risk of VTE among a mixed group of COC users with different progestogen-containing COCs (e.g., "third generation"), and articles where the reference group cited the use of COC with non-specified progestogens. [10][11][12] Five of the included studies did not clearly state the estrogen dose contained; one study 13 [18][19][20] The validation of VTE was factored into the study quality assessment, with VTE cases considered to be validated if they were identified in one of the following ways: (1) from anticoagulation clinics, VTE clinics, or physician report; (2) from discharge diagnosis codes of inpatient hospitalizations; or (3) from diagnosis codes of outpatient records plus additional validation through anticoagulation treatment, medical record review, imaging studies, or physician or patient confirmation. All inpatient VTE diagnoses were considered valid because the diagnostic codes are generally based on confirmed diagnoses. Codes found solely in outpatient data may include codes for both suspected and confirmed diagnoses; therefore, outpatient VTE codes were considered valid only if additional information was examined such as anticoagulation prescriptions, imaging studies, or physician or patient report, in order to exclude suspected VTE that was later ruled out. 21 Age, personal history of VTE, and recent pregnancy are important risk factors for VTE. Therefore, studies that did not adjust for age were excluded, 22 and studies were only included if pregnant or postpartum women and women with a history of VTE were excluded from analyses. Exceptions that were included despite these criteria were one study 3 in which the prevalence of prior VTE was less than 1% among the entire cohort, one study 23 in which all cases and controls were COC users (because COC use is contraindicated in the context of current or historical VTE 5 ), and one study 24 that excluded women with a recent hospitalization (because this also likely excluded women with a recent pregnancy). A sensitivity analysis was conducted excluding the two studies that did not account for prior VTE 23 or recent pregnancy 24 and noted little difference in estimates (data not shown).
When multiple studies were identified that reported results from the same study sample, older analyses 14,[25][26][27] were excluded and only the most recent analyses 23,28-31 were included. In addition, nine articles 13,16,23,29,[32][33][34][35][36] reported risk estimates for both cohort and nested case-control analyses for the same study population; in the present analysis, the risk estimates from the nested case-control studies were used because most of the cohort risk estimates were unadjusted.
The evidence was summarized and systematically reviewed using standardized abstraction forms. The studies were abstracted by two authors (MVD, NKT) and verified by another (KMC).
Potential sources of bias were assessed for individual studies and quality ratings (good, fair, or poor) were assigned using study-designspecific criteria developed by the United States Preventive Services Task Force. 37 When assessing selection bias in case-control studies, the potential for biased selection of cases and controls (for example, hospital controls vs community controls) and the response rate were considered. The assessment of selection bias in cohort studies involved consideration of whether the cohort represented the population it was taken from, whether the exposed and unexposed groups came from the same population, and whether the follow-up rate was adequate.
The assessment of information bias focused on the determination of contraceptive exposure (for example, self-report vs pharmacy codes vs medical records) and VTE outcome (for example, diagnostic codes only vs diagnoses objectively confirmed). Finally, it was assessed whether potential confounders were addressed through restriction, matching, or adjustment in analysis; studies that did not control for age, history of VTE, or recent pregnancy were excluded as described above, and other potential risk factors for VTE were considered as potential confounders.
The meta-analysis included relative risk estimates of VTE that reflected comparisons between pills containing specific progestogen and levonorgestrel formulations. The preferred risk estimates were those with a reference group of users of monophasic COCs containing 30 μg of ethinyl estradiol and levonorgestrel. In some cases, studies reported risk estimates compared with levonorgestrel-containing COCs as a group but noted that the monophasic preparation represented at least 50% of the total exposure; other studies presented risk estimates compared with any low-dose levonorgestrel COC. If studies reported multiple risk estimates for users of levonorgestrel-containing COCs, the risk estimates for the most specific formulations were chosen. In cases where nonusers were the reference group and risk estimates for levonorgestrel and other progestogen-containing COCs were available, risk estimates with levonorgestrel as the reference group were calculated for inclusion in the meta-analysis. When studies presented risk ratios for multiple COCs containing the same progestogen at a specific ethinyl estradiol dose, a combined risk ratio for that progestogen was calculated.
A random-effects model based on profile likelihoods was used to calculate pooled risk ratios. 38 The presence of statistical heterogeneity was assessed using the standard Cochran χ 2 test, and the magnitude of the heterogeneity was evaluated using the I 2 statistic. 39 The included studies reported different risk estimate measures (odds ratios, hazard ratios, relative risks, or rate ratios). Because VTE is very rare, all these measures provide similar estimates and were combined in a single meta-analysis.
The analyses were stratified by the specific progestogen formulations. For studies that reported multiple adjusted relative risk estimates, the maximally adjusted estimates were used in the primary analysis.
Sensitivity and subgroup analyses were conducted based on whether the study adjusted for body mass index, smoking, or duration of COC use (or assessed these variables as potential confounders and determined adjustment was not needed 3 ); the study design (case-control or cohort); the study quality; and the funding source (pharmaceutical industry or other). In addition, a subgroup analysis on users of monophasic COCs containing a standard dose of 30 μg of ethinyl estradiol was conducted to isolate any effect of the progestogen. For comparisons with at least 10 studies, funnel plots and the Egger linear regression method were used to test for small-study effects (a marker of potential publication bias). 40 All analyses were performed using Stata/IC version 13.1 (StataCorp, College Station, TX, USA). P<0.05 was considered statistically significant.
All but one of the studies were conducted in Europe or the USA. The remaining study, 41 sponsored by the WHO, included populations from Europe, Asia, Latin America, and the Caribbean. Most studies, of poor to good quality, retrospectively evaluated large administrative databases.
Two cohort studies 3,28 collected prospective population-based data with a specific study design; these studies were rated as being of "good" quality (Table S3). All case-control studies and one cohort study 31 offered evidence of "fair" quality, and the remaining two cohort studies were considered to be of "poor" quality (Tables S2 and S3). Flaws in the studies included the reliance on self-reported COC use, which could introduce recall bias, or on prescription information, which may not reflect actual use. Additional flaws included small numbers of outcomes and no adjustment for certain risk factors such as the body mass index or smoking.  Fig. S6). The heterogeneity was moderate (I 2 =30%-66%) for all pooled analyses except for norgestimate (I 2 =15%).
Because there was evidence of statistical heterogeneity among the individual studies for many of the progestogens, additional analyses were conducted to assess the possible sources of heterogeneity.
In analyses restricted to monophasic COCs with a standard dose of The findings were also generally consistent in other sensitivity and stratified analyses. The exclusion of poor-quality studies did not impact the pooled estimates of relative risks or reduce the heterogeneity (Table 1). In stratified analyses, the pooled estimates of risk were generally lower in studies that adjusted for the body mass index, smoking, or the duration of use than in studies that did not adjust for these factors ( Table 1). The risk estimates were similar when studies were stratified according to the use of a case-control or cohort design; however, with the exception of deosgestrel and norgestimate, there were some differences in heterogeneity in analyses stratified by study design. Pooled estimates based on studies sponsored by the pharmaceutical industry typically indicated lower risks for VTE but more heterogeneity compared with studies not sponsored by the pharmaceutical industry. None of the differences in the stratified analyses were statistically significant (

| DISCUSSION
The present meta-analysis indicated that the use of low-dose (less T A B L E 2 Pooled estimates (95% confidence intervals) of unadjusted risk ratios for venous thromboembolism among users of combined oral contraceptives by progestogen type compared with levonorgestrel in published meta-analyses. a Estimates are given as risk ratios.

Meta-analysis
The present meta-analysis used adjusted risk estimates to reduce potential confounding. Four other meta-analyses 6,7,9,49 have also estimated pooled relative risks for VTE associated with a specific COC formulation compared with a levonorgestrel-containing formulation, but have used unadjusted estimates (Table 2). Although each used different methods and varied in the individual studies included, the findings are generally similar. Two of the meta-analyses 7,9 found no increase in the risk of VTE with norgestimate-containing COCs compared with levonorgestrel-containing pills, which is consistent with the present findings. Similarly, the present findings of slight increases in relative risk for desogestrel, drospirenone, and gestodene are consistent with results from the previous metaanalyses, 6,7,9,49 which found small but significantly increased (range 1.3-1.9) relative risks associated with these progestogens. The present estimate for cyproterone acetate was slightly higher (risk ratio 2.0), but generally consistent with the estimates from two other analyses (risk ratio 1.6-1.7). 7,9 The present analysis had limitations. There are no data from ran- Despite the presence of heterogeneity, the findings were generally robust in the subgroup and sensitivity analyses. Most of the exposure and outcome information came from large administrative databases. Although these databases offer greater assurance for capturing specific formulations and duration of use compared with self-report, prescription data may not accurately represent actual COC use at the time of the VTE event. 50,51 In addition, the accuracy of administrative databases for the ascertainment of medical conditions such as VTE is variable; however, linking data from these databases to other sources (for example, physician report, evidence for anticoagulation treatment) to verify information reduces the likelihood for misclassification.
In conclusion, the present meta-analysis indicated that COCs con-

CONFLICTS OF INTEREST
The authors have no conflicts of interest.

SUPPORTING INFORMATION
Additional Supporting Information may be found online in the supporting information tab for this article.