Meta-analyses of prospective observational studies have demonstrated a positive association between circulating CRP and CHD risk.
However, it remains unclear whether this association is causal or explained by confounding factors or reverse causality, or even underestimated as a result of attenuation by errors.
The best method to establish causality for this association would be by randomization to an intervention that altered CRP levels but did not affect any other cardiovascular risk factors. To our knowledge no such trials have been conducted. The use of genetic variants as instrumental variables to determine the causal association between circulating CRP and CHD provides an alternative to such a randomized controlled trial, with the advantage that this approach can be performed in existing datasets and may be more generalisable than a randomized controlled trial.
It has also been suggested that CRP genetic profiling might have clinical utility in predicting CHD risk.
In this study of 4,610 cases (the study with the largest numbers to date to examine the association of variation in CRP
with CHD), we present evidence that higher circulating CRP level is unlikely to be an important causal risk factor for CHD. We found no association of a genetic variant (rs1130864) in CRP
with CHD events, despite this variant being consistently associated with circulating CRP. Using rs1130864 to explore the causal association of CRP levels with CHD risk is valid since individuals who are homozygous for the T allele (TT genotype) will have experienced on average higher levels of circulating CRP over their lifetime than other individuals (CT or CC genotype),
but potential confounding factors will be evenly distributed between these two groups of individuals (TT versus CT or CC), as demonstrated for all our cohorts in . Thus, the association of rs1130864 with CHD cannot be influenced by reverse causality, attenuation by errors or confounding. 
In this respect, our analysis of the association of rs1130864 with CHD can be compared to a randomised controlled trial of individuals who have been randomly allocated (or not) to a 21% higher CRP level on average across their lives, given that our pooled ratio of geometric means of CRP by genotype was 1.21.
The assumptions underlying the Mendelian randomization approach is that the genetic variant is associated with the modifiable risk factor (circulating CRP levels in this example) and that it is not related to the outcome of interest (CHD) other than through its association with the modifiable risk factor (i.e. there are no confounding factors relating genotype to CHD and genotype is not related to CHD through other pathways).
The rs1130864 SNP has been consistently shown to be associated with circulating CRP levels in males and females of European origin,
a finding that we have replicated here. Furthermore we have previously shown that the rs1130864 SNP is associated with a shift in the whole distribution of CRP levels and not an increase, for example, only at the higher end or a decrease only at the lower end of the CRP distribution.
This whole-population shift in CRP levels should clearly be related to a whole-population shift in CHD risk if CRP were causally related to CHD and therefore our null finding is unlikely to be explained by missing individuals with CRP levels above or below a given threshold in relation to genotype.
There are strong theoretical bases for believing that CRP
genotype will not be related to socioeconomic and behavioural confounding factors that tend to distort observational epidemiological studies of this association,
and we have empirically demonstrated here that whilst circulating CRP is related to these potential confounding factors, genotype is not ( and ).
The variant that we have used in this study is in close linkage disequilibrium (LD) with variation within a transcription factor binding site located 5′ of CRP
that is associated with circulating concentrations of CRP and thought to be functional.
It is unlikely that this functional variant, or one in close LD with it, will have pleiotropic effects. It is possible that another variant in another gene near CRP
is in LD with rs1130864 and provides another pathway to CHD. To explain our null result in the context of circulating CRP being truly causally related to CHD this alternative pathway would have to result in a decrease in CHD risk by a magnitude that exactly reversed the posited causal influence of circulating CRP on CHD risk. However, rs1130864 exhibits no major LD (correlation R2
>0.2) with SNPs in any other gene in its genomic region, except with a nearby gene (ENSG00000196401) identified in Ensembl and HapMap R2
0.8, but not shown in Entrez. For lower levels of LD (i.e. with R2
0.2), the effect of any confounding gene on CHD would have to be so large as to be implausible (no such gene has been found in any CHD genome wide studies to date 
). Variant ENSG00000196401 (XR 017178.1) shows sequence similarity to ribosomal protein L27 (LOC646446) mRNA, which is widely expressed. However, there is no strong evidence that ribosomal proteins in general, ribosomal protein L27 specifically, nor ENSG00000196401 gene are associated with CHD risk. Of note, none of the recent genome-wide association studies identified variants in ENSG00000196401 as being related to CHD risk.
It is therefore unlikely that the variant we have examined here (rs1130864) is linked via alternative molecular genetic pathways to CHD risk in such a way that these alternative pathways completely counter balance an important circulating CRP causal effect on CHD.
A number of SNPs related to CRP levels have been identified in CRP
and we would have had greater statistical power for our instrumental variables analyses had we constructed haplotypes using several SNPs as in one of our previous papers.
However, we were limited here to using a SNP that was typed in all studies included in the analyses. This should not have biased our results, which are consistent with other studies showing that genetic variation in several SNPs in CRP
are not associated with CHD in the way predicted by their association with CRP levels (see introduction
and further discussion below).
Within each of our cohort studies participants are described as being ‘white’, ‘Caucasian’, or of ‘European’ origin and the consistency of association between rs1130864 and both CRP levels and CHD events across our studies suggests that population stratification is unlikely to have importantly confounded our genetic association results. Developmental canalisation (the process by which target receptors or organs develop differently in response to varying levels of the exposure of interest during key developmental periods) might limit the Mendelian randomization process. The extent to which this occurs with modest effects such as the differences in CRP level by genotype is unclear.
For our main analyses we combined incident and prevalent cases. In conventional observational epidemiology incident cases are important for making causal inference and minimising any bias due to reverse causality. However, in genetic association studies reverse causality is not possible and survivor bias very unlikely.
Repeating our analyses with incident cases only did not substantively alter any of the point estimates for any of our analyses.
There was important heterogeneity between individual studies in the association of circulating CRP with CHD that was not explained by differences in the distribution of sex, age or the proportion of incident cases between studies. Adjustment for potential confounding factors reduced heterogeneity but some remained even in the confounder adjusted analyses, suggesting that residual confounding in some studies might explain this heterogeneity. Since our Mendelian randomization study relates the proportion of circulating CRP that is explained by rs1130864 to CHD risk, and there was no between study heterogeneity in the association of this genetic variant with CRP levels or with CHD, our Mendelian randomization results are unaffected by between study heterogeneity in the observational association of CRP with CHD.
Our instrumental variables analysis uses the proportion of the variation in CRP that is explained by rs1130864 to provide an estimate of causal effect that is not biased by confounding, reverse causality or attenuation by errors. However, the advantage of this approach of being less biased than a conventional multivariable regression analysis comes at the cost of reduced precision. Although the point estimate of the odds ratio per doubling of CRP from the instrumental variables analysis was virtually null (1.04), the 95% confidence interval (0.61, 1.80) includes the observational association, and the association found in the most recent meta-analysis of observational studies.
The confidence interval for the instrumental variables analysis is much wider than the confidence interval from the observational analyses presented here or in previous meta-analyses.
An alternative to the instrumental variable analyses in Mendelian randomization studies is to compare the observed genetic-disease association to that expected from the best observational studies, as done recently by Casas et al.
Whilst this approach appears to provide more precise estimates of the causal association of CRP with CHD and does not require that all studies included in the analyses have measurements of all three of CRP levels, CRP
genotype and CHD, the level of precision is in fact spurious since it does not fully account for the relatively small proportion of total variation in CPR accounted for by genotype and does not include uncertainty from all analyses.
The fact that rs1130864 explains less than 1% of the variation in CRP within each of our studies, whilst affecting statistical precision is unlikely to result in bias. As noted previously,
many medications that are used in randomised controlled trials to determine causality explain a similarly small proportion of variation in the potentially causal risk factor, but with adequate sample sizes (sometimes, as in our Mendelian randomization study presented here, obtained only through meta-analysis of data from a number of trials) provide precise and valid estimates of the causal effect on clinical endpoints.
For example, blood pressure lowering therapies explain ~2% of the variation in blood pressure, and in participants who are randomised to either active blood pressure lowering therapy or control in randomised controlled trials there will be many other environmental and genetic factors that influence variation in blood pressure. Nonetheless, an adequately powered randomised trial of the effect of blood pressure medication on stroke (or other cardiovascular outcomes) is, rightly, accepted as unbiased evidence of the causal effect of blood pressure on stroke risk.
With respect to other Mendelian randomization studies, single SNPs that have been shown to be robustly associated with low density lipoprotein cholesterol (LDLc), and that explain less than 1% of the variation in circulating LDLc, are robustly associated with CHD, with the magnitude of this association being somewhat larger than that predicted from the causal randomised controlled trial evidence relating statins (which reduce LDLc) to CHD.
It has been suggested that the somewhat stronger effects with these genetic variants relates to the fact that the randomised difference in LDLc occurring as a result of genetic variants is life-long, whereas that occurring as a result of statins is from mid-adult life only.
Similarly, we have recently shown that a single SNP in FTO
(which again explains less than 1% of variation in body mass index or total fat mass) is associated with a wide range of vascular and metabolic outcomes, including fasting glucose, insulin and lipids, with magnitudes of association that are the same as would be predicted from the association of FTO
with BMI and of BMI with these outcomes in observational studies and trials of weight reduction.
It would seem odd to us that the Mendelian randomization approach provides valid causal estimates in examples where causation is not controversial (i.e. the effect of LDLc on CHD and BMI on fasting glucose) but is selectively biased in more debatable areas such as that assessed here and in the companion paper
(i.e. the association of CRP with atherosclerosis and CHD risk).
Our Mendelian randomization findings are consistent with other studies that have used variants in CRP
to explore the causal association of CRP levels with continuously measured vascular and metabolic traits and which suggest that CRP is not causally related to blood pressure, metabolic syndrome components or carotid intima media thickness.
As discussed in the introduction a number of other studies, which have not directly employed Mendelian randomization approaches and have included between 210 to 1062 CHD cases, have also found genetic variants within the CRP
gene to be unrelated to prevalent and incident CHD events, despite these variants being associated with CRP levels,
or have found associations in the opposite direction to what one would anticipate if higher CRP levels were causally associated with increased CHD risk,
or associations only in very specific subgroups.
Whilst such subgroup analyses might represent true differences in the effect of CRP
or CHD, in general subgroup analyses fail to replicate and should be treated with caution until replicated in other independent studies. Furthermore, none of the large genome wide association studies of CHD have found variation in CRP
to be robustly associated with CHD.
By contrast variants associated with established risk factors for CHD (e.g. LDLc) are identified in these genome wide association studies CHD. Taking these findings together with our own results–the largest study to date (N
4160 cases; four times the largest previous published study) to relate variation in CRP
to CHD–there does not appear to be any strong evidence for an association of CRP
with CHD events, despite its association with CRP levels. These findings together would suggest that variation in circulating CRP level is not importantly causally related to CHD risk and that genetic profiling of CRP
is unlikely to be useful in the prediction of CHD. However, we acknowledge that a very large genetic (Mendelian randomization) study is required to definitively demonstrate whether there is no causal association of CRP with CHD. A newly established collaboration that will include at least 30,000 CHD events will over the coming years be able to provide this definitive answer.
In a companion paper we examined the association of 3 tag SNPs [+1444T>C (rs1130864); +2303G>A (rs1205) and +4899T>G (rs 3093077)] in the CRP
gene with serum CRP and carotid intima-media thickness (CIMT) in one of the studies included in this paper (the Whitehall II Study).
In that study there was no independent association of CRP concentration with CIMT once potential confounding factors had been taken into account and no evidence from analyses using haplotypes in CRP
as instrumental variables that CRP concentration is causally related to CIMT.
As CIMT is a marker of atherosclerosis, that paper,
together with other studies finding no association of genetic variation in CRP
fail to support CRP as a causal factor for atherosclerosis.
Whilst our findings, together with those of several other studies described above, suggest that circulating levels of CRP are not importantly causally related to the development of atherosclerosis or CHD, an effect of CRP on case-fatality in those with CHD, is possible. Such an effect is implicated by findings from rodent models of beneficial effects of post-myocardial infarction CRP lowering,
but to our knowledge has not be demonstrated in humans.
Our findings, together with those from a number of other studies examining the association of genetic variation in CRP
with CIMT and CHD,
and with findings from genome wide association studies of CHD,
suggest that circulating CRP does not have an important causal association with CHD. In our study the instrumental variables result (an odds ratio of 1.04 per doubling of CRP) is our best estimate of causal effect. However, very large genetic association studies are required to provide a precise estimate of this association and rule out possible modest causal effects.
Data Access: DAL had full access to all of the data from the BWHHS, Caerphilly and Speedwell studies used in this paper; MK and MGM had full access to all of the data from the Whitehall II study and NW and LJP had full access to all of the data from HIMS. These individuals take responsibility for the integrity of the data and the accuracy of the data analysis.