|Home | About | Journals | Submit | Contact Us | Français|
Multiple randomized controlled trials (RCTs) have established the efficacy of statins for prevention of cardiovascular disease. The benefits observed are often framed in terms of percent reductions in low-density lipoprotein cholesterol (LDL-C) from baseline or percent reduction between control and treatment groups, even though epidemiological data suggest that absolute inter-group difference in LDL-C (ΔLDLControl-Rx) is the more informative measure. We conducted a systematic review of large-scale trials of statins versus placebo, usual care or active (lower-dose statin) control to calculate updated summary estimates of risk reduction in coronary heart disease (CHD) and all-cause mortality. Meta-regression analysis was used to ascertain the relations of different LDL-C metrics to outcome. In 20 eligible RCTs, there were significant overall reductions for CHD (OR=0.72, 95% CI=0.67–0.78) and mortality (OR=0.89, 95% CI=0.84–0.94), but with substantial variability in trial results. ΔLDLControl-Rx was the strongest determinant of CHD risk reduction, particularly after excluding active-comparator studies, and was independent of baseline LDL-C. By contrast, baseline LDL-C edged ΔLDLControl-Rx as the strongest determinant of mortality, but neither was significant after exclusion of active-comparator studies. Exclusion of 3 RCTs involving distinct populations, however, rendered ΔLDLControl-Rx the predominant determinant of mortality reduction. In conclusion, these findings underscore the primacy of absolute reductions in LDL-C to the design and interpretation of RCTs of lipid-lowering therapies, and to framing treatment recommendations based on the proven coronary benefits of these drugs.
A multitude of randomized controlled trials (RCTs) have documented the efficacy of statins in reducing coronary events in diverse populations.1 These benefits are well-established in RCTs comparing statins against placebo or usual care,1 with accumulating data suggesting that this is also true for high- versus low-intensity statin treatment.2–4 Demonstration of statin-related benefits versus control has not been uniform, however, even in such mega-trials.3–5 A potential explanation cited for null results is that, despite major percent reductions in LDL-C from baseline in the active group, the percent difference in LDL-C achieved between study arms has been more modest.5–6 Yet assessments of benefit of lipid-lowering agents, whether in RCTs5,7–10 or meta-analyses,11–14 have often been framed in terms of percent cholesterol reductions associated with active treatment. The construct of percent cholesterol lowering from baseline has similarly been used in national treatment guidelines.15 Focus on percent, rather than absolute, reductions is itself problematic, however, because available epidemiological evidence shows that the relation between cholesterol level and cardiovascular risk is curvilinear.15–16 Hence, for any absolute change in serum cholesterol irrespective of its starting clinical value, the corresponding change in the relative risk of CHD is constant. Consequently, percent cholesterol reductions, whether longitudinally in the intervention arm11 or between study arms,17 may not be the most apt metric for describing statin-related benefits. We sought to define in comparative fashion the relation between various measures of LDL-C lowering to outcomes using meta-regression in an updated systematic review of large-scale statin trials.
We conducted a systematic search of the English-language literature using PubMed, EMBASE, BIOSIS, Web of Science, Cochrane Systematic Reviews, DARE, Central Register of Controlled Trials and clinicaltrials.gov from January 1994 to December 2008 to identify pertinent studies. Nineteen relevant keywords were entered, and clinical trials, qualitative and quantitative reviews, and editorials thus identified were reviewed to select relevant studies. Reference sections of articles served to identify additional studies, as did knowledge of experts in the field (C.J.V., A.M.G., and R.C.P.).
We focused our systematic review on large-scale RCTs18 of statins versus placebo, usual care, or active comparator having clinical cardiovascular events or all-cause mortality as their primary endpoint. Eligible RCTs required a minimum enrollment of 1,000 participants at risk for, or with stable, CHD and follow-up duration >1 year. We excluded a priori clinical trials that evaluated statins in combination with other therapies, that were designed to assess impact on intermediate primary endpoints, or that focused on distinct clinical populations or acute settings (which could introduce additional heterogeneity), namely, patients with acute coronary syndromes or advanced renal disease, or patients undergoing cardiac catheterization.
After selection of RCTs meeting inclusion criteria, a single investigator (J.R.K.) extracted data from individual studies. The 2 endpoints of interest were the primary or secondary outcomes of major CHD events and all-cause mortality. Major CHD events comprised non-fatal myocardial infarction and fatal CHD, as defined in individual RCTs. When resuscitated cardiac arrest formed part of an RCT’s major CHD endpoint, and could not be clearly excluded from the composite CHD outcome based on the information presented, it was included in the analysis. If fatal CHD was not reported, fatal myocardial infarction was used instead. Mean or median LDL-C at baseline and following allocation to treatment or control was abstracted directly or computed from available information. Where the mean in-trial follow-up LDL-C in the placebo group was not provided, it was imputed from the baseline value.
Five measures of LDL-C were examined: (i) baseline LDL-C in the treatment group (baseline LDLRx); (ii) absolute change in LDL-C in the treatment group (ΔLDLBaseline-Final); (iii) percent change in LDL-C in the treatment group (%ΔLDLBaseline-Final); (iv) absolute difference between achieved in-trial LDL-C in the control versus treatment group (ΔLDLControl-Rx); (v) percent difference between achieved LDL-C in the control versus treatment group (%ΔLDLControl-Rx).
STATA, version 10.0 (College Station, Texas), was used in all analyses. Separate analysis of RCTs not having active comparator was planned a priori. For individual trials, the numbers of participants with and without events in the treatment and control arms at the conclusion of the study were recorded. Where ascertainment of follow-up LDL-C was performed prior to conclusion of the trial, events reported as of the time of such measurements were used in the analyses. In view of the presence of significant heterogeneity (Cochran’s Q test), a random-effects approach (DerSimonian and Laird) was used to calculate pooled odds ratios (Ors). The quantity I2, the component of heterogeneity not attributable to chance, was also computed.19 Because only mega-trials were included, an approach that avoids potential biases associated with smaller studies,18,20 assessment for publication bias was not undertaken.
To investigate the observed study heterogeneity, random effects meta-regression of RCT results was performed against different measures of LDL-C or trial duration.21–22 Sensitivity analysis of ORs, I2s and meta-regression coefficients entailed serial exclusion of studies or groups of studies from consideration.
The search yielded 31 candidate RCTs meeting sample size and duration criteria, of which 11 did not meet additional inclusion criteria (4 for acute coronary syndromes or enrollment following cardiac catheterization; 2 for assessing drug-combinations; 2 owing to a focus on end-stage renal disease or renal transplantation; 1 with an intermediate primary endpoint; 1 for failed randomization; and 1 because of incompleteness of endpoint ascertainment). The characteristics of the 20 large-scale trials2–5,7–9,23–35 selected for inclusion are presented in Table 1. There were a total of 155,255 participants, of whom 11,508 experienced major CHD events, and 13,687 had all-cause fatal events. With exclusion of the 3 active-comparator trials 2–4, there were 124,302 individuals, with 8,332 and 10,448 developing major CHD and fatal events, respectively.
Analysis of all 20 RCTs showed significant overall reductions in risk of CHD (OR=0.72, 95% CI=0.67–0.78) and mortality (OR=0.89, 95% CI=0.84–0.94), but this was associated with significant underlying heterogeneity for both outcomes (P≤0.005). The lack of consistency in study findings was not attributable to chance, with moderate-to-high (I2=69.7%) or moderate (I2=51.1%) variability in effect estimates across RCTs. Figure 1 shows the corresponding findings for the analysis focused on the 17 RCTs of statins versus placebo or usual care. Significant risk reductions were again achieved for CHD and mortality, with less, though still significant, heterogeneity for both outcomes.
The results of meta-regression modeling of risk reduction as a function of various measures of LDL-C or trial duration are shown in Table 2. Taking all trials into account, ΔLDLControl-Rx exhibited the strongest inverse association (highest negative standardized regression coefficient, β) with relative reduction in CHD events, while that for the percent difference (%ΔLDLControl-Rx) was minimally weaker. According to the random-effects model, every 39 mg/dL (1mmol/L) inter-group difference in achieved LDL-C was associated with a 25% relative reduction in CHD risk (OR=0.75, 95% CI=0.68–0.82). There were significant, but weaker associations for baseline LDLRx, and absolute – though not percent – reduction in LDL-C in the treatment arm (ΔLDLBaseline-Final), whereas trial duration bore no significant relation to CHD risk reduction.
When active-comparator trials were excluded, the absolute inter-group difference in LDL-C (ΔLDLControl-Rx) again showed the strongest association to CHD risk reduction, followed by %ΔLDLControl-Rx (Table 2). The associations between these inter-group differences, whether in absolute or percent terms, and CHD risk decreased with exclusion of active-comparator trials. Figure 2a shows that, based on the model, there was a 23% relative reduction in CHD risk for every 39 mg/dL absolute inter-group decrease in LDL-C (OR=0.77, 95% CI=0.66–0.89). As compared with ΔLDLControl-Rx, ΔLDLBaseline-Final was more modest and fell just short of significance. No other variables were significantly related to CHD risk reduction.
Regarding all-cause mortality, analysis of all 20 RCTs showed significant relations only for ΔLDLControl-Rx and baseline LDLRx, but it was baseline LDLRx that had the higher negative regression coefficient. Accordingly, for every 39 mg/dL higher value for baseline LDL-C there was an associated 10% reduction in the relative risk of death (OR=0.90, 95% CI=0.83–0.99). The corresponding reduction for every 39 mg/dL of ΔLDLControl-Rx was 12% (OR=0.88, 95% CI=0.78–0.99). Notably, analyses of the 17-RCT subset revealed no significant associations for any of the independent measures (Table 2 and Figure 2b).
The association between ΔLDLControl-Rx and relative risk of CHD was not meaningfully altered by adjustment for baseline LDLRx, whether or not active-comparator trials were taken into account (data not shown). With respect to mortality, inclusion of both ΔLDLControl-Rx and baseline LDLRx in the full meta-regression model rendered both variables non-significant (P≥0.116).
Since exclusion of active-comparator RCTs was found to reduce heterogeneity, sensitivity analyses focused on the 17 trials of statin versus placebo or usual care. Exclusion of each of these RCTs individually did not materially influence estimates of risk reduction for CHD or mortality (data not shown). Selected individual and multiple exclusions based in part on outliers in meta-regression analysis (Figure 2) are shown in Table 3.
For both outcomes, exclusion of ALLHAT-LLT 5 and the Scandinavian Simvastatin Survival Study (4S)23 individually led to a decrease in variability across studies. Omission of each also led to appreciable weakening in the relationship between ΔLDLControl-Rx and either outcome. So too did exclusion of the GREek Atorvastatin and Coronary-heart-disease Evaluation (GREACE),31 particularly for CHD, but this was accompanied by an even more marked decrease in effect estimate for ΔLDLBaseline-Final. The same effect was evident with exclusion of high-risk CHD trials, of which ALLHAT-LLT, 4S, and GREACE are part. By contrast, exclusion of the Management of Elevated cholesterol in the primary prevention Group of Adult Japanese (MEGA) study,34 the Prospective Study of Pravastatin in the Elderly at Risk (PROSPER)29, and the Controlled Rosuvastatin in Multinational Trial in Heart Failure (CORONA),8 both individually and collectively, resulted in strengthening of the associations between ΔLDLControl-Rx and both outcomes, as well as ΔLDLBaseline-Final with regard to CHD events. In fact, exclusion of all three RCTs uncovered a significant association between ΔLDLControl-Rx and mortality. Last, exclusion of low-risk CHD trials, led to a stronger relation for ΔLDLBaseline-Final with respect to CHD events, but this depended entirely on exclusion of the MEGA study (data not shown). It also resulted in a stronger and significant association of baseline LDLRx with mortality, driven by exclusion of the Justification for the Use of Statins in Prevention: an Intervention Trial Evaluating Rosuvastatin (JUPITER) study.7
The present study provides updated summary estimates regarding the impact of statins on major CHD and all-cause mortality, accounting for several important mega-trials published since completion of a prospective systematic review by the Cholesterol Treatment Trialists’ Collaboration (CTTC), based on patient-level data.1 Strikingly, after incorporating an additional 34,246 participants (65,199 including active-comparator trials), and despite the use of study-level data and somewhat divergent inclusion criteria, the resulting effect estimates were nearly identical to those reported previously by the CTTC,1 including in a recent update.4 These findings were robust to single trial exclusions, yet even though removal of active-comparator trials and of either ALLHAT-LLT or 4S dampened heterogeneity, moderate variability remained across studies.
Beyond confirming these results, the significance of the present work lies in its focus on an issue that has heretofore received insufficient attention, namely, understanding which measure of LDL-C concentration or lowering best relates to the benefits conferred by statins – and, by extension, other lipid-lowering therapies. To our knowledge, the current investigation is the first to undertake a direct comparison of such different LDL-C metrics in relation to the magnitude of risk reductions achieved by statin medications in RCTs.
Our meta-regression analyses show that different measures of LDL-C help to explain the heterogeneity observed in the effect estimates from different studies. In these analyses, ΔLDLControl-Rx emerged as the strongest determinant of CHD risk reduction in the full range of trials considered, as well as in the narrower set excluding trials with active-comparator arms. Importantly, these relationships were not influenced by baseline LDL-C level. The same was not true for total mortality, however, where ΔLDLControl-Rx was less strongly associated with outcome than baseline LDLRx in the broader meta-analysis, and none of these measures was significant in the more restrictive meta-analysis.
The observation regarding all-cause mortality runs contrary to expectation. Yet, to the extent that a higher baseline LDL-C is strongly linked to risk of coronary mortality, the component of all-cause mortality most directly influenced by statin-associated LDL-C lowering, this finding reflects a greater contribution of coronary death reduction to the overall mortality risk. That the relationship between baseline LDLRx and all-cause mortality was bolstered by exclusion of low-risk CHD trials, and particularly JUPITER, is in turn consistent with the removal of other influences on mortality risk, such as low-grade inflammation.7
Moreover, it is notable that exclusion of PROSPER, MEGA and, especially, CORONA heightened the relationship between ΔLDLControl-Rx and death (as well as CHD), with the association becoming both significant and stronger than the baseline LDL-C level when all 3 trials were omitted. These studies included special populations – older adults with8 or without 29 symptomatic heart failure, Japanese subjects34 – apt to have different relations of LDL-C to disease outcomes, as compared with participants in the majority of statin trials. In the case of PROSPER and CORONA, competing, non-coronary risks for mortality would alter the expected influence of LDL-C lowering on the outcomes considered here. Indeed, neither showed a significant reduction in all-cause mortality. By contrast, in MEGA, disproportionate reductions in CHD and total mortality in the face of modest LDL-C lowering would have the opposite distorting effect on the relation between ΔLDLControl-Rx and outcomes.
Another important message from these findings is that measures of absolute reduction generally have stronger relations with outcome than measures expressing these reductions as percents. This was the case for ΔLDLBaseline-Final in the 2 analyses of major CHD, as well as for ΔLDLControl-Rx (even if minimally) in all relevant instances but one.
Apart from the influence of the various LDL-C measures, key differences in the design features among the trials included are likely to account in substantial measure for the moderate degree of heterogeneity observed. These include the types and doses of statins evaluated; the approach to, and extent of, in-trial ascertainment of LDL-C (including in non-random subsets); duration of follow-up period; definitions and ascertainment of CHD outcomes; and conduct of the trials over the span of nearly 2 decades, with the associated influence of secular trends. In the context of such disparate trial-specific features, the pre-eminence of ΔLDLControl-Rx as a determinant of major CHD events and its robustness to sensitivity analysis are all the more notable, even if this measure proved to be of less consequence to all-cause mortality.
The finding that ΔLDLControl-Rx outperforms ΔLDLBaseline-Final, even in trials of statins versus non-active control, is expected because the former better reflects adherence to study assignment and extent of cross-over between study arms. In turn, the superiority of ΔLDLControl-Rx is consistent with observations from cohort studies and RCTs wherein the relation between cholesterol and risk of CHD is curvilinear.15–16 The implication of this log-linear relation is that the same absolute change at a different starting level of LDL-C will translate into an identical reduction in the relative risk of CHD. Using the percent change in lieu of the absolute change violates this relation.
The superior validity of ΔLDLControl-Rx is of special relevance because clinical trials,5,7–10 meta-analyses11 and even professional guidelines15 have often used the percent LDL-C reduction from baseline as the reference in discussing the magnitude of expected benefit. The present findings emphasize the importance of incorporating absolute changes in LDL-C and an assessment of baseline coronary risk in discussing the degree of risk reduction anticipated with statin use. They also underscore the need to consider post-treatment differences between comparison arms in designing future clinical studies of emerging therapies, whether for lipid lowering or otherwise.
A limitation of the present work is our lack of patient-level data, which would have permitted more thorough evaluation of lipid measures and other covariates, including sociodemographic and clinical subgroups. Individual-participant data also would have allowed assessment of the impact of estimated cardiovascular risk, because relating the observed risk in control groups to risk reduction is fraught with the potential for regression to the mean.36 More granular data could be used to probe further the questions addressed here.
Dr. Kizer was supported in part by Grant K23 HL070854 from the National Heart, Lung, and Blood Institute.
Publisher's Disclaimer: This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final citable form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.
Financial Disclosures: Dr. Kizer has received honoraria from Merck & Co., Inc., and research support from diaDexus, Inc. Dr. Gotto is a current consultant for KOWA Pharmaceuticals, Merck, and Roche Pharmaceuticals, and he is on advisory boards for DuPont and Novartis. Dr. Gotto serves on corporate boards for Aegerion Pharmaceuticals, Arisaph Pharmaceuticals, and Vatera Capital. Dr. Pasternak is an employee of Merck & Co., Inc. None of the other authors have potential conflicts of interest to report.