Table 2.
Approach | Brief description of approach | Key sources of bias and directiona | Duration of exposure assumption by Frerence et ala |
---|---|---|---|
Multivariable regression in prospective cohort studies54 | Prospective Cohorts Collaboration: an individual participant meta-analysis of 958 074 adults (61 studies) aged 40–69 with no previous history of CVD. Exposure = SBP (in both Ference paper and original paper); outcome = fatal CHD | Residual confounding by adiposity and height, both of which exaggerate any true causal effect. Repeat SBP measurements in a large subgroup were used to adjust the association for regression dilution bias; the estimate of duration of exposure is therefore unlikely to be biased by this | From baseline SBP assessment to death or end of follow-up (mean 13.2 years) |
IV of intermediate in RCTs32 | Systematic review and meta-analyses of 25 RCTs including 109 797 participants with no clinical evidence of cardiovascular disease before randomization. Authors of the original paper calculated ratios of difference in log odds CHD ÷ difference in SBP/DBP by randomized group for each antihypertensive, and meta-analysed these. Exposure = SBP (in Ference (triangulation) paper but is actually the combined SBP and DPB effect in the original paper); outcome = fatal or non-fatal CHD | Ference et al. assumed the results represented a risk reduction in CHD for 10-mmHg lower SBP, whereas they were the risk reduction in CHD for a 10-mmHg lower SBP or 5-mmHg lower DBP. Thus the (assumed) SBP effect is exaggerated in comparison with its true effect | From randomization to end of follow-up (mean 4.6 years) |
MR53 | Two-sample MR. Genetic variants from the International Consortium for Blood Pressure genome-wide association study (ICBP) that had reached genome-wide levels of statistical significance were used.77 The association of each of those variants with CHD was extracted from the Coronary ARtery Disease Genome-Wide Replication And Meta-Analysis (CARDIoGRAM) consortium database (22 223 fatal or non-fatal CHD cases and 64 762 controls).78 The authors calculated ratios of difference in log odds CHD ÷ difference in SBP for each genetic variant, and meta-analysed them | We undertook a number of sensitivity analyses to explore the possibility of bias due to: (i) the fact that the ICBP results were for SBP adjusted for BMI; and (ii) violation of the exclusion restriction criteria (Supplementary text and Figures S1 to S3). On the basis of these we concluded these results were unlikely to have major bias | Whole of the participant’s life, and so mean age at end of follow-up or becoming a CHD case (54.9 years) |
In Supplementary text (available at IJE online) we provide full details of how we assessed a range of potential key sources of bias and their likely direction; here we describe the ones that we concluded were the key sources.
For this example, the three approaches were compared within one paper by Ference and colleagues,53 who undertook the MR study themselves but used results from published meta-analyses of prospective cohort and RCT studies. We have reviewed both Ference et al’s paper53 and the two papers32,54 that they took results from in preparing this table and the detailed Supplementary text.