Search methods for identification of studies

We reviewed all published systematic reviews since 2007 using the Montori systematic review filter⁷⁴ for MEDLINE and the Wilczynki filter⁷⁵ for EMBASE (both 75% sensitivity). Based on this overview, we opted to stratify results by statin use: populations taking statins and populations not taking statins. Consequently, two separate searches were carried out and two sets of results are presented, although the general methods remained the same for both groups. Dates for the searches were staggered, as the review was broken down into manageable tasks based on statin usage and existing reviews.

Searches were conducted in the Cochrane Central Register of Controlled Trials, MEDLINE, EMBASE, the Clinical Trials Register, the Current Controlled Trials register, and the Cumulative Index to Nursing and Allied Health Literature (CINAHL) to identify relevant studies.

All trials included in identified reviews were considered for inclusion, along with any trials identified by the work of the Cholesterol Treatment Trialists’ Collaboration.⁶⁶

Populations not taking statins A search was conducted (unrestricted for study design) from 1 September 2007 to 18 June 2013. This updated/extended the search carried out by the Emerging Risk Factors Collaboration (ERFC).⁷⁶ The search was restricted to English language only and did not include conference abstracts (see Appendix 3).

Populations taking statins Two strategies were used, based on study design type. For non-randomised controlled trials (non-RCTs) a search was completed for all publications up to 4 December 2012, with no restriction on language or publication type. For RCTs, all studies were included from previous reviews,^72,73 plus a search was conducted to extend these review’s searches up to 9 April 2013. In line with these reviews, the strategy was for English language-only publications (both strategies are presented in Appendix 4). An additional search was conducted within the international Clinical Trials Registry and ClinicalTrials.gov for any studies in progress from 2010 to April 2013.

The title and abstract of each paper identified in the searches was reviewed by two reviewers to identify potentially relevant references. The full text of potentially relevant studies was then obtained, and two reviewers independently selected studies to be included in the review using predetermined inclusion criteria. In all cases, disagreements about study inclusion were resolved by consensus and a third reviewer was consulted if disagreements persisted. Flow charts were constructed to show the selection process.

Study selection criteria

We included all prognostic cohort studies of any design (univariable or multivariable) that measured lipids in humans. Studies from any setting regardless of statin treatment were included. RCTs with adequate follow-up were treated as cohorts providing evidence to a specific study group [e.g. a statin vs. placebo RCT would provide evidence for both populations taking statins (statin arm) and not taking statins (placebo arm) separately].

Consistent with previous reviews in this area, we included studies where at least 1000 relevant participants were recruited per cohort/intervention arm and a minimum 2-year average follow-up duration.^66,72 There was no restriction on gender. We included studies with only two or more lipid measures assessed at baseline/relevant time point, as the focus of the review was to compare predictive performance across different lipid measures.

Target group in PECO (Population, Exposure, Comparator, Outcome) format:

• P Population with or without existing CVD.
• E Measurement of lipid measures at baseline/relevant time point. Ten lipids measures were studied including TC, HDL cholesterol, LDL cholesterol, non-HDL cholesterol, TGs, Apo B, Apo A-I, plus combinations of these measures as ratios TC/HDL cholesterol, LDL/HDL cholesterol and Apo B/Apo A-I.
• C Control or comparator may or may not be present.
• O Outcomes included cardiovascular events, cardiovascular mortality and all-cause mortality. Strokes were excluded if they were the sole outcome (included if treated as part of composite outcome with other CVD events).

In addition, we considered all studies from previous reviews for inclusion irrespective of whether or not they met the 1000-participant, minimum two lipid measures at baseline or 2-year follow-up thresholds to maximise comparability of our results with those published in previous reviews.

Included studies provided quantitative results for one or more of the following outcomes:

• cardiovascular events [CHD, non-fatal myocardial infarction (MI), stroke, angina attack, other CVD, coronary bypass]
• cardiovascular mortality (fatal MI, sudden cardiac death, stroke, congestive heart failure, other fatal CVD)
• all-cause mortality.

Outcome definition was based on that reported in the included studies.

Data extraction and management

For all included studies, two reviewers independently carried out data extraction. Details extracted included study name and year, cohort numbers, baseline age, gender, prevention group, statin usage and dose if applicable, follow-up period, blood sample details and lipid data. When several follow-up periods were reported then data for the longest follow-up were extracted. Disagreements were resolved by consensus with a third reviewer being consulted when required.

The methodological quality of included studies was assessed using the Quality in Prognostic Studies (QUIPS) assessment tool.⁷⁷ Quality of studies with populations taking statins was independently assessed by two reviewers and the level of agreement tested by the kappa statistic (κ). For studies of populations not taking statins, quality was assessed by one reviewer. Quality was assessed in six bias domains, with emphasis placed on prognostic factors and attrition:

• Study participants No quality bias was anticipated in this area, as all populations were included in the review. Studies were categorised by prevention group, which provided inception cohorts.
• Study attrition There is no clear guidance for the assessment of attrition rates in prognostic studies. Previous prognostic reviews have considered attrition rates of 10–20% as low risk of quality bias, but these have been based on small populations with relatively rare conditions and short follow-up periods.^78,79 As the new cohorts in this review have > 1000 participants with long follow-ups, we anticipated that dropout could be higher. Risk of study attrition bias was considered to be low – if it was documented – and < 25%, moderate if 25–40% or not reported, and high risk if > 40%.
• Prognostic factor Lipid measures were considered to be at low risk of quality bias if all tests were processed by a central accredited laboratory and procedures for different measures were described. Risk was assessed as moderate if a central laboratory was used but no further information given, and high risk of bias when no information was provided.
• Outcome measure This dimension was not considered to be significant, as it was a necessary requirement in the study entry criteria. However, outcome bias could have been introduced owing to clarity of outcome definitions; this was explicitly addressed in sensitivity analyses.
• Study confounding This too was not considered to be a significant concern, as study confounders were specially accounted for in the data analysis of each study: studies were divided into unadjusted and adjusted data.
• Statistical analysis As most of the included studies were not designed to report the outcome data required for this review, we did not assess the quality of the study in relation to how the study data were analysed; instead we addressed this issue separately by carrying out a sensitivity analysis considering the type of analysis used in the original study.

Data analysis

Data were sought from the included studies for predictive outcomes as follows:

1. Direct estimates for one standard deviation (SD) incremental change either as HRs, odds ratios (OR) or risk ratios (RRs) and their variances. These ratios were assumed to approximate the same measure of risk. This assumption was tested in a sample of studies with the highest event rates and the ratios showed little variation.
2. Data in an alternative format to incremental change by one SD were converted where possible.
3. Univariable and multivariable models (unadjusted and adjusted). Multivariable analyses included any studies with adjustment for one or more of the following covariates: age, gender, body mass index (BMI), smoking, etc. Where a study produced more than one multivariable model then the most advanced model was selected for data inclusion. However, we specifically excluded any model that adjusted for any other lipid measure.
4. IPD were not actively sought, but used when available.
5. When no direct estimates were reported, alternative indirect methods for estimation were used.^80,81 Additionally, two further alternative indirect methods for estimation were used based on (1) generalised least squares for trend^82,83 and (2) simulation methods. A brief description of these can be found in Appendix 5.

The primary analysis estimated the predictive strength (HR) for each lipid measure by each of the three outcomes. Outcomes are reported separately and stratified by unadjusted/adjusted estimates. Analyses were subgrouped by prevention group (primary or secondary prevention). Studies were considered to be only primary or secondary prevention if > 95% of the participants fell in that category; otherwise studies were considered as mixed. Numbers of studies, participants and study characteristics are presented in tables.

For each lipid measure, random effect meta-analyses were carried out to obtain summary HRs (by one SD) using Review Manager (RevMan) 5.2 (The Cochrane Collaboration, The Nordic Cochrane Centre, Copenhagen, Denmark) and results were displayed as forest plots. A random-effects model was selected because high heterogeneity was anticipated in both populations. Any risk of small study bias should be reduced by the majority of included studies having > 1000 participants. Overall summary plots presenting pooled estimates for all lipid measures (HDL and Apo A-I as inverse associations to allow clear comparisons with other measures) by outcome and by unadjusted/adjusted models were produced. Although pooling of data for lipid measures with only three or fewer studies could be potentially misleading, the figures were included in the summary plots as an indication of potential trends had there been more study data.

Heterogeneity was summarised using the I² and chi-squared statistics. Potential sources of any heterogeneity (clinical or statistical) were explored using subgroup and sensitivity analyses.

Subgroup analyses, where possible, were conducted by:

• pre-existing condition (e.g. diabetes)
• study type [prospective (including RCTs) or retrospective cohorts]
• age (threshold of median 40 years at inception).

Sensitivity analyses were conducted to determine the impact of study quality and data extraction on outcome. Focus was placed on consistency of overall results with those obtained in better-quality studies. When sufficient studies were available, sensitivity analyses were carried out by omitting:

• studies in which the predictive outcomes were defined by CHD, a surrogate measure of CVD
• studies assessed to have a higher/moderate risk of quality bias (based on prognostic factors and attrition quality domains)
• studies in which the majority of the lipid tests were not completed in a central laboratory (yes/no)
• studies in which summary data were extracted indirectly
• within the populations not taking statins, studies that confirmed the use of lipid modification therapy.

The search aimed to be as broad as possible to include as many studies as possible. However, it is acknowledged that there is considerable publication bias in prognostic studies that is difficult to characterise;⁸⁴ therefore we did not carry an analysis for potential publication bias.

A post hoc analysis was conducted as recommended by the external peer reviewer with the aim of exploring sources of potential heterogeneity. A meta-regression was completed to examine the associations between variables considered in the subgroup and sensitivity analyses. Meta-regression was completed, where possible, for those outcomes and variables with sufficient study data.

Studies based on populations not taking statins

Figure 3 summarises the selection process; 3776 publications were screened from the search, with a further 165 potential studies identified from existing reviews and hand searching. Of these, 90 studies met the inclusion criteria and provided results (see Appendix 8), with 16 studies providing two or more sets of data; this resulted in a total of 110 sets of data. Twelve studies divided the data by gender [TLGS,⁸⁸ Yao City,⁸⁹ AMORIS,⁹⁰ CB project,⁹¹ Copenhagen City,⁹² DUBBO,⁹³ Framingham Offspring,⁴⁰ IKNS (Yao),⁹⁴ MONICA,⁹⁵ Northwick Park I,⁹⁶ Reykjavik,⁹⁷ Rotterdam⁹⁸]. Two studies divided the data by gender and then further by ethnicity (Charleston⁹⁹) or presence of metabolic syndrome [European Prospective Investigation of Cancer (EPIC) – Norfolk study (EPIC-Norfolk)¹⁰⁰] so providing four data sets each. Puerto Rico¹⁰¹ divided the data by rural and urban populations, and MRFIT¹⁰² by black and white men.

FIGURE 3

Flow chart of study inclusion for populations not taking statins.

Full details of the characteristics of the 90 included studies are shown in Table 8. Of the included studies 73% were prospective cohorts and 17% were placebo arms of RCTs. The remaining nine studies were two retrospective cohorts (Swedish NDR;¹⁶⁴ Juntendo, Japan¹³⁴), three RCTs (of non-lipid modifying interventions) for which both arms were included in this review (SHEP,¹⁵⁹ TREAT,¹⁶⁷ Women’s Health¹⁷²) and four nested case–control studies (BUPA,¹¹³ GLOSTRUP Population,¹²⁴ Nurses’ Health,¹⁴⁷ Physicians’ Health¹⁵⁰). Studies provided data for between 1 and 10 lipid measures. Two studies provided data for all of the 10 lipid measures (FIELD,¹²¹ Women’s Health¹⁷²). Data details for all studies are shown in Appendix 8. Table 9 shows the number of studies and participants by the 10 lipid measures. Numbers are also shown for three additional measures (Apo B/HDL cholesterol, TG/HDL cholesterol/HDL cholesterol/TC) that were revealed by the review; however, data were available from only one study for one outcome for each measure and therefore no pooled estimates were obtained.

TABLE 8

Characteristics of included studies: populations not taking statins

TABLE 9

Number of included studies and participants by lipid measure: populations not taking statins

Outcome data were available for CVD events, CVD mortality and all-cause mortality. Data were further divided into unadjusted (CVD events, CVD mortality, all-cause mortality) and adjusted data [CVD events (adjusted), CVD mortality (adjusted), all-cause mortality (adjusted)] and then by lipid measure. Meta-analyses were not possible for all combinations of outcomes and lipid measures. Outcome data for all lipid measures were pooled for CVD events (except LDL/HDL cholesterol) and CVD events (adjusted). For CVD mortality, outcome data were available only for nine lipid measures (no studies examined the ratio Apo B/Apo A-I) and only HDL cholesterol, TGs and TC had sufficient (more than three studies) to pool data. Outcome data were available for all 10 lipid measures for CVD mortality (adjusted), but Apo A-I, LDL/HDL cholesterol, Apo B and Apo B/Apo A-I data could not be pooled due to an insufficient number of studies. Data were available for four lipid measures for all-cause mortality (LDL, HDL, TC, TGs), all of which had sufficient studies to pool data. Finally, for all lipid measures bar non-HDL cholesterol, outcome data for all-cause mortality (adjusted) were provided, although only HDL cholesterol, LDL cholesterol, TC and TGs could be pooled as a result of insufficient studies.

Heterogeneity

Heterogeneity was generally high across all prevention groups, outcome data and lipid measures with few exceptions. For CVD events, both adjusted and unadjusted, the overall heterogeneity was considerable (I² > 73%) for all lipid measures except Apo A-I in the unadjusted data (56%) and LDL/HDL cholesterol (55%) in the adjusted data where it was moderate. When further examined, the unadjusted CVD event data for primary prevention groups is considerably heterogeneous (77–97%) for all lipid measures except Apo A-I (5% based on five studies) and Apo B/Apo A-I (59% based on three studies). For adjusted CVD events outcome data in the primary prevention populations the heterogeneity is large (92–98%) for all lipid measures except LDL/HDL cholesterol were it is low (27%); though this is based on eight data sets, these were obtained from only three studies. There were insufficient data to examine heterogeneity in any of the lipid measures in the secondary prevention populations for CVD events adjusted outcome data. However, for these populations, in the unadjusted CVD events outcome data the heterogeneity was measurable in five out of the nine lipid measures and varied: LDL cholesterol was considerable (76%), TC, HDL cholesterol and TGs were moderate (34%, 52% and 63%) and for Apo B and Apo A-I there was no heterogeneity (0%). This was for the same two studies each (LIPID,¹³⁸ Thrombo¹⁶⁵). Heterogeneity for CVD mortality outcome data (unadjusted and adjusted) was substantial for all pooled lipid measures (67–91%). There was no heterogeneity statistic for secondary prevention populations owing to insufficient data. For the primary prevention groups in these data the heterogeneity ranged from 40% to 89%. Few lipid measures provided sufficient data for all-cause mortality and therefore assessment of heterogeneity was limited. For those that did, LDL cholesterol, TC, HDL cholesterol, TGs, heterogeneity for the summary estimate was moderate to considerable (52–95%), there were insufficient data to assess secondary prevention groups. All-cause mortality unadjusted data for primary prevention groups varied between two measures (LDL cholesterol, TC) showing considerable heterogeneity (> 95%) and two measures (HDL cholesterol and TGs) showing no heterogeneity (0%). Data from only two studies were available for each of these last two measures: of the two studies reporting HDL data, one study had 44,457 participants (AMORIS⁹⁰) and the other 1013 (InCHIANTI¹³¹); for TGs, the participant numbers were 1013 (InCHIANTI¹³¹) and 2086 (Lan 2007¹³⁶). Forest plots of all of these results appear in Appendix 9.

Figures 4–9 show the summaries of the overall results for all lipid measures for CVD events, CVD events (adjusted), CVD mortality, CVD mortality (adjusted), all-cause mortality and all-cause mortality (adjusted), respectively.

FIGURE 4

Summary results (pooled estimates) by all lipid measures as prognostic markers for cardiovascular events in populations not taking statins. All data sets include primary prevention data sets, secondary prevention data sets, and data sets with mixed or (more...)

FIGURE 9

Summary results (pooled estimates) by all lipid measures as prognostic markers for all-cause mortality (adjusted estimates) in populations not taking statins. All data sets include primary prevention data sets, secondary prevention data sets, and data (more...)

FIGURE 5

Summary results (pooled estimates) by all lipid measures as prognostic markers for cardiovascular events (after adjustment) in populations not taking statins. All data sets include primary prevention data sets, secondary prevention data sets, and data (more...)

FIGURE 6

Summary results (pooled estimates) by all lipid measures as prognostic markers for cardiovascular mortality in populations not taking statins. All data sets include primary prevention data sets, secondary prevention data sets, and data sets with mixed (more...)

FIGURE 7

Summary results (pooled estimates) by all lipid measures, as prognostic markers for cardiovascular mortality (adjusted estimates) in populations not taking statins. All data sets include primary prevention data sets, secondary prevention data sets, and (more...)

FIGURE 8

Summary results (pooled estimates) by all lipid measures as prognostic markers for all-cause mortality in populations not taking statins. All data sets include primary prevention data sets, secondary prevention data sets, and data sets with mixed or unreported (more...)

For unadjusted estimates – CVD events:

• When all of the data were combined for all of the lipid measures with more than three data sets, Apo B/Apo A-I (four studies: HR 1.62, 95% CI 1.07 to 2.45), followed by Apo B (10 studies: HR 1.54, 95% CI 1.31 to 1.81), non-HDL cholesterol (six studies: HR 1.49, 95% CI 1.25 to 1.77) and TC/HDL cholesterol (seven studies: HR 1.48, 95% CI 1.18 to 1.85) offer the highest predictive value.
• For primary prevention populations, of the single measures Apo B (five studies: HR 1.70, 95% CI 1.48 to 1.96) and non-HDL cholesterol (four studies: HR 1.57, 95% CI 1.33 to 1.85), offer the strongest predictive value, whereas for the ratios, TC/HDL cholesterol (five studies: HR 1.57, 95% CI 1.18 to 2.08) offered the strongest value.
• In the secondary prevention populations, only TC (five studies: HR 1.08, 95% CI 1.03 to 1.13) offers a significant association with CVD events.

For adjusted estimates – CVD events:

• All lipid measures have data for this outcome. When all data are combined, Apo B/Apo A-I offers the strongest predictive value based on 16 sets of data (HR 1.35, 95% CI 1.22 to 1.50). The other two lipid ratios (TC/HDL cholesterol and LDL/HDL cholesterol), TC, Apo B and non-HDL cholesterol all have similar predictive values based on similar or more studies (HR range 1.22–1.28).
• For primary prevention groups, the data follows the same pattern with Apo B/Apo A-I as the strongest predictor (HR 1.36, 95% CI 1.21 to 1.53). Of note in comparison with the unadjusted data, now Apo B/Apo A-I (14 studies) is the strongest predictor, followed by LDL/HDL cholesterol (eight studies: HR 1.29, 95% CI 1.27 to 1.32).
• For secondary prevention there were insufficient study data to provide results.

For unadjusted estimates – CVD mortality:

• The amount of evidence for CVD mortality is considerably smaller than for CVD events particularly in relation to lipid ratios.
• Only three lipid measures had sufficient data to meta-analyse (HDL, TGs, TC).
• HDL cholesterol offered the strongest predictive value for combined data (six studies: HR 1.28, 95% CI 1.11 to 1.48) with TC and TGs providing estimates of similar magnitude (combined: TC HR 1.14, 95% CI 1.14 to 1.42; five studies: TGs HR 1.24, 95% CI 1.08 to 1.42).
• In the primary prevention group, TGs offered the strongest predictive value (four studies: HR 1.30, 95% CI 1.14 to 1.49), but, as with the combined data, TC and HDL cholesterol have estimates of a similar magnitude (six studies: TC HR 1.27, 95% CI 1.01 to 1.60; four studies: HDL cholesterol HR 1.27, 95% CI 1.06 to 1.51).
• There were no data for secondary prevention populations.

For adjusted estimates – CVD mortality:

• Similar to the unadjusted data, the amount of evidence for CVD mortality (adjusted) is considerably smaller than for CVD events particularly in relation to lipid ratios.
• For all studies combined HDL (10 studies: HR 1.18, 95% CI 1.08 to 1.28), TC (19 studies: HR 1.17, 95% CI 1.10 to 1.24) and TGs (five studies: HR 1.17, 95% CI 1.04 to 1.31) offered the strongest predictive values.
• For the primary prevention group only HDL (five studies: HR 1.16, 95% CI 1.01 to 1.34) and TGs (four studies: HR 1.16, 95% CI 1.01 to 1.34) showed an association with adjusted CVD mortality data.
• Data were available for secondary prevention groups but with insufficient study numbers to obtain pooled results.

For unadjusted estimates – all-cause mortality:

• There were data for only four measures (HDL cholesterol, TC, LDL cholesterol, TGs).
• The associations followed a different pattern to that observed for CVD events and mortality (except for HDL cholesterol and TG).
• For all studies combined, only HDL cholesterol (nine studies: HR 1.25, 95% CI 1.10 to 1.43) and TGs (eight studies: HR 1.09, 95% CI 1.04 to 1.15) showed an association with all-cause mortality.
• For primary prevention populations there were insufficient data to provide results. There were no data for secondary prevention populations.

For adjusted estimates – all-cause mortality:

• Lipid measure data were available for all measures, except for non-HDL cholesterol, but in a small number of studies.
• TC (19 studies) was the only lipid measure reported in more than six studies.
• For all studies combined, only TGs (five studies: HR 1.09, 95% CI 1.06 to 1.13) and TC (19 studies, HR 1.05, 95% CI 1.02 to 1.09) showed an association with all-cause mortality (adjusted).
• For primary prevention groups there was no evidence of association or insufficient data.
• There were no data for secondary prevention populations.

Studies based on populations taking statins

The study selection process is summarised in Figure 10. In total, 2260 publications from the non-RCT search and 667 publications from the RCT search were screened, reducing to 25 included studies, with 28 sets of data meeting the inclusion criteria and providing results. Three studies (J-LIT,^174,175 Phase Z¹⁷⁶ and SEARCH⁵⁸) offered two sets of data; two studies had two intervention arms with different statin doses (Phase Z¹⁷⁶ and SEARCH⁵⁸). In addition, one study stratified results into separate primary and secondary prevention groups (J-LIT^174,175).

FIGURE 10

Flow chart of study inclusion for populations taking statins. CTT, Cholesterol Treatment Trialists’ Collaboration.

The characteristics of the included studies are shown in Table 10. Of the 25 studies, only three were non-RCTs (CLIP,¹⁷⁷ LIVES Ex,¹⁸² J-LIT^174,175), one was a nested case control from a RCT (TNT¹⁸⁵) and the rest were RCTs. Studies provided data for between one and seven lipid measures, and no single study examined all markers. No data were available for LDL/HDL cholesterol. No results were available for CVD mortality or all-cause mortality. Full details of the data can be seen in Appendix 11. Table 11 shows the number of studies and participants by lipid measure. Outcome data per lipid measure were separated into unadjusted CVD events (CVD events) and adjusted CVD events [CVD events (adjusted)]. Only half of the reported combinations of lipid measures and outcomes could be pooled using meta-analysis. Individual forest plots for these measures are presented in Appendix 12. The remaining data had fewer studies than our threshold for meta-analysis; however, a pooled estimate was obtained for all lipid measure summaries, but meta-analyses have not been presented.

TABLE 10

Characteristics of included studies: population taking statins

TABLE 11

Number of included studies and participants by lipid measure: populations taking statins

Heterogeneity

Across both prevention groups, heterogeneity was substantial for all but one lipid measure (non-HDL cholesterol). Only LDL cholesterol (CVD events) showed moderate heterogeneity for primary prevention populations and non-HDL cholesterol (CVD events) for secondary prevention populations.

Figure 11 shows a summary of the overall results for all lipid measures for CVD events, whereas Figure 12 shows the same, for adjusted estimates.

FIGURE 11

Summary results (pooled estimates) by all lipid measures as prognostic markers for cardiovascular events in populations taking statins. All data sets include primary prevention data sets, secondary prevention data sets, and data sets with mixed or unreported (more...)

FIGURE 12

Summary results (pooled estimates) by all lipid measures as prognostic markers for cardiovascular events (adjusted estimates) in populations taking statins. All data sets include primary prevention data sets, secondary prevention data sets, and data sets (more...)

For unadjusted estimates – CVD events:

• Overall HDL (19 studies: HR 1.12, CI 1.06 to 1.19) and TC (11 studies: HR 1.12, CI 1.05 to 1.20) offered the strongest association with CVD events.
• For primary prevention populations, LDL cholesterol (seven studies: HR 1.14, CI 1.06 to 1.23) offered the strongest predictive value. All other measures have a similar point estimate but are not significant.
• For secondary prevention groups LDL, Apo B, non-HDL cholesterol and HDL offered a similar predictive value (HR range 1.09–1.10).

For adjusted estimates – CVD events:

• Overall HDL (five studies: HR 1.61, CI 1.19 to 2.18) offered the strongest association with CVD events.
• For both primary and secondary prevention groups there were insufficient data to form conclusions.

Key finding for populations not taking statins

Results from the included studies in this population support Apo B/Apo A-I as the strongest predictor of CVD events (adjusted and unadjusted). Apo B, non-HDL cholesterol, LDL/HDL cholesterol and TC/HDL cholesterol offer similar predictive values and only slightly lower point estimates than Apo B/Apo A-I. However, all of the lipid measures offer positive associations with CVD events and there is considerable overlap in the CIs. This pattern is repeated when considering only primary prevention populations. There were insufficient data sets to pool data for Apo B/Apo A-I and LDL/HDL cholesterol for unadjusted data, but the limited data available do suggest these to be the strongest predictors, and this is reflected in the adjusted data. Although the number of studies were insufficient to pool data for Apo B/Apo A-I, the numbers of participants were comparable to those in the data for TC/HDL cholesterol and non-HDL cholesterol. Data for the secondary prevention group were very limited, but did appear to follow a similar trend, except for TGs, which appeared to have no association with CVD events.

The results from this review are in line with the latest research by the ERFC⁷⁰ and the 2011 review by Sniderman et al.⁷¹ The ERFC⁷⁰ examined the association among non-HDL cholesterol, Apo B, Apo A-I, TC/HDL cholesterol and Apo B/Apo A-I and CVD (adjusted outcome data) in 26 studies (primary and secondary prevention groups). The results in this review are almost identical for non-HDL cholesterol, Apo B, and Apo A-I, but the CIs are slightly narrower. Both reviews have found that the ratios TC/HDL cholesterol and Apo B/Apo A-I were more strongly associated with CVD events than other lipid measures. The ERFC’s final conclusion was that TC/HDL cholesterol offers a stronger association with CVD events than the individual measures Apo B and Apo A-I; our review agrees with these findings. However, the ERFC⁷⁰ did not include LDL/HDL cholesterol in their assessment and our review suggests that this is almost as predictive as Apo B/Apo A-I. Sniderman et al.⁷¹ only considered three lipid measures and their association with CHD. Although our outcome definition differs, results were similar, suggesting that Apo B is superior to non-HDL cholesterol and non-HDL cholesterol is superior to LDL in predicting events. Examination of diabetic populations not taking statins was limited by the lack of data. However, there is some indication that for CVD events (unadjusted) the lipid measures examined provide lower predictive values or no association for this population. Within the adjusted data, more lipid measures were represented, but trends were not clear: most, but not all, lipids were less predictive of CVD events in the diabetic populations than non-diabetic populations.

For CVD mortality (unadjusted and adjusted), HDL offered the strongest association with CVD mortality, but the CIs overlapped with those for TC and TGs. In the unadjusted primary prevention group the same three lipid measures showed an association, with overlapping CIs, but here TGs was the strongest, whereas in the adjusted primary prevention group TC no longer had an association, and HDL and TGs had exactly the same results. However, in the adjusted data, five other lipid measures (TC/HDL cholesterol, LDL/HDL cholesterol, LDL cholesterol, Apo A-I, Apo B) offered an indication that they could provide a similar, or stronger, predictive value in primary prevention groups and for all participants, if more study data were available.

The Asia Pacific Studies Collaboration (APSC)⁶⁷ assessed 17 studies, and examined the association between seven lipid measures (TC, TGs, HDL cholesterol, non-HDL cholesterol, TC/HDL cholesterol, TG/HDL cholesterol) and CVD mortality, and the association between LDL cholesterol and CHD. For the lipid measures for which we were able to pool data in our review (TC, TG, HDL), the HRs were only slightly greater than the APSC results,⁶⁷ and, similar to our review, the CIs in their data all overlapped. So, although the strength of our associations was slightly larger, there was similarity in the pattern of results. Results for TG, TC/HDL cholesterol and TG/HDL cholesterol were presented as log results, and therefore were not comparable, and we had no results for non-HDL cholesterol and CVD mortality (unadjusted data) in our review.

Post hoc meta-regression showed an association only between CVD events (adjusted) and using a surrogate measure for CVD events for four lipid measures (TC, non-HDL cholesterol, Apo B/Apo A-I, TC/HDL cholesterol). The analysis suggests that including the surrogate measure of CHD could potentially lead to an overestimation of the HR for CVD events (adjusted) by approximately 20%.

From the limited outcome data available for all-cause mortality (unadjusted and adjusted), only two lipid measures had an association. For unadjusted data, TGs had a stronger association, whereas for adjusted data, HDL cholesterol was more strongly associated. Although there were insufficient studies to analyse Apo B, Apo A-I and their ratio, these could potentially offer a stronger predictive value. Unexpectedly, higher levels of some lipid measures (LDL cholesterol, TC, LDL/HDL cholesterol) were associated with a lower risk of all-cause mortality. The level of comorbidities in these populations might account for this result. The IKNS study in Yao City⁹⁴ concluded that the decreased association with mortality was stronger in heavy drinkers and more significant in follow-up over 5 years. There is no existing review to compare results for all-cause mortality.

Key finding for populations taking statins

For all included studies, HDL and TC offered the strongest association with CVD events (unadjusted) and HDL cholesterol with adjusted data. For unadjusted data, only LDL cholesterol offered an association for primary preventions groups, whereas for secondary preventions groups Apo B, TC, LDL cholesterol, non-HDL cholesterol and HDL cholesterol all had a similar association. For adjusted data, although there were insufficient studies to produce combined results, there was some indication that HDL cholesterol for primary prevention groups and TC for secondary prevention groups could offer strong associations with CVD events.

Boekholdt et al.⁷² completed an IPD systematic review using adjusted data from eight studies to assess the association between LDL cholesterol, non-HDL cholesterol, and Apo B and CVD events. In comparison, although 25 studies contributed data to this review, the maximum numbers of studies contributing data to any single measure in our comparable analysis was five. Comparison with the work of Boekholdt et al.⁷² was only partially possible. This review’s LDL cholesterol point estimate was higher than the estimate of Boekholdt et al.⁷² and not significant; for both Apo B and HDL cholesterol our point estimates were higher, with large CIs and based on only one included study. Boekholdt et al.⁷² concluded that there was a stronger association for non-HDL cholesterol with CVD events than LDL and Apo B; this review suggests agreement with this finding but does not have sufficient evidence to confirm it. However, Boekholdt et al.⁷² did not assess HDL cholesterol in his research, and this review would suggest that HDL cholesterol has a higher predictive value than non-HDL cholesterol, LDL cholesterol or Apo B for CVD events in populations taking statins.

No outcome data for CVD mortality or all-cause mortality were available for this group.

The lipid measures with the higher associations with CVD events and CVD mortality are consistent with current NICE recommendations for monitoring.^2,62 Our results suggest that alternative measures such as Apo B or the ratios LDL/HDL cholesterol and Apo B/Apo A-I, could be given further consideration in the future based on their clinical validity; however, in 2014 these were unlikely to be affordable or practical.

Strengths and limitations

The main strength of this review is its extensive coverage of all study designs, populations taking and not taking statins, and the wide panel of 10 lipid measures. The review has subgrouped populations into primary and prevention groups to seek to understand whether or not the predictive ability of various lipids varies according to these populations. Previous reviews in this area often look at only a subset of lipid measures, which complicates comparisons and the development of clinical recommendations. There is a tendency for reviews to exclude studies if they do not have data for all of the lipids being assessed in the review. In this review we have included study data even if they contributed to only some of the lipid measures or outcomes under consideration. We have extracted data from publications using indirect methods, as well as direct and converted data.

To complete a review of the scale and complexity of this one, it is required to make decisions about the scope and focus that impact on study identification and selection. Although other lipid-lowering drug groups could have been included in this review (such as fibrates), together with the advisory board we decided to focus only on statins. It is possible that by searching from the dates of previous reviews and modifying the search strategy, studies may have been missed; however, as we have aimed to include all studies from previous reviews and larger trials of > 1000 participants, we believe it is unlikely that large relevant studies have been overlooked. From the existing reviews we were unable to trace 15 potential studies, accounting for 10% of our final sample. This could have been because studies were mentioned as joining collaborations, but the studies never completed or the findings were not published. Despite the large number of included studies in this review, the number of studies reporting data for each pooled estimate is much smaller. This is typical of prognostic reviews in the area; for example, the ERFC used only 68 out of a potential 112 studies.⁶⁹

Many included studies did not disclose details such as whether or not the participants had pre-existing CVD or whether or not the participants were taking statins. Assumptions have been made in the data analysis to accommodate these studies, and sensitivity analyses were conducted to account for these assumptions.

No outcome data for CVD mortality or for all-cause mortality were available in populations taking statins, and data for these outcomes were provided only in a smaller proportion of studies of populations not taking statins, compared with CVD events. As expected, heterogeneity was high for most analyses, and contributed to our decision to use a random-effects model for the pooled estimates. The populations included in the studies were hugely varied, with no specific inclusion criteria relating to baseline characteristics. Furthermore, it was expected that there would be significant variation in the way the lipid measures were tested. The strength of predictive value of the lipid measures was largely based on the point estimates, but it should be noted that the CIs were often overlapping and the degree of difference between lipid measures was on the whole small. In meta-regression we did not find any measures that considerably reduced heterogeneity.

All estimates were obtained independently for each lipid measure or ratio. Therefore, comparisons between these measures are all indirect comparisons. We did not explore the alternative approach based around network meta-analysis (NMA) of a subset of studies to make direct comparisons, because this approach would use fewer studies and the methods for NMA are better developed for treatment comparisons than for prognostic reviews.

Assessment of methodological quality of the individual studies was difficult because included studies were not designed to address the same objective as this review and, therefore, the information required for this review to assess quality might not have been a priority for the study authors. Statistical analysis was rarely in a form that was appropriate for our outcomes. Assessment of attrition quality bias had no precedent in previous reviews, and setting boundaries for levels of low, moderate and high attrition bias were proven, subsequently, to be too generous, as no studies had high attrition.

There were a number of changes from the original protocol. The search methods were modified to accommodate the scale of the review and recognise the publication of three reviews in 2012. Similarly, by introducing the minimum number of participants, a 2-year follow-up and a minimum two lipid measures at baseline it was anticipated this would reduce the inclusions of smaller, more-prone-to-bias studies. This also removed the need for sensitivity analysis for smaller studies. In the analysis, all outcome data for each lipid measure was pooled. Although acknowledging that conclusions drawn from three or fewer sets of data could be misleading, it was important to enable an overall comparison between lipid measures and their trends, and view potentially useful lipid measures for future research. The indirect methods for data extraction were changed once included studies were examined; the best methods were used to extract the maximum amount of data. Quality assessment of included studies was instead based on a more recent tool⁷⁷ that was specifically designed for prognostic reviews: QUIPS.

This review did not use IPD. It is well acknowledged that this is the ideal data source for these analyses and this means that a number of the limitations above would have been reduced or removed. Future research should look to utilise this approach if possible.

Prognostic systematic reviews are still in the early stages of methodological development. Search strategies, data extraction techniques, quality assessment tools and analysis software are not well established compared with other types of systematic reviews. This made each stage of the review challenging; as this area of research develops, it will provide a clearer path for undertaking this type of review.