Estimating causal effects of C-reactive protein on disease and health outcomes using multivariable Mendelian randomization adjusting for heritable confounding.
C-reactive protein (CRP) is a marker of inflammation associated with autoimmune, cardiovascular, and neuropsychiatric disorders. However, it remains unclear whether CRP causally affects these traits or if observed associations result from reverse causation or confounding. Mendelian randomization (MR) uses genetic variants as instrumental variables to estimate causal effects and avoid the biases present in observational studies. Prior MR studies have suggested causal effects of CRP on several traits, including low-density lipoprotein (LDL) cholesterol, schizophrenia (SCZ), and knee osteoarthritis (OA). However, MR may produce biased results if traits that confound the exposure and outcome are heritable, resulting in horizontal pleiotropy. This is a major concern for studies of CRP, because CRP levels may increase in response to inflammation caused by a wide range of heritable conditions.
Multivariable Mendelian randomization (MVMR) can be used to eliminate bias from heritable confounding when genome-wide association study (GWAS) summary data are available for confounders. In this study, we used MVMR to estimate the causal effects of CRP on 16 outcomes with prior evidence of a causal or associational link to CRP. We used a novel computational pipeline to identify a broad set of potential heritable confounders between CRP and each outcome trait from studies in the Medical Research Council Integrative Epidemiology Unit (MRC-IEU) OpenGWAS database. We compared MVMR results with computationally selected confounders to univariable MR results and MVMR adjusting only for body mass index.
Univariable MR suggests evidence of potential causal effects of CRP on coronary artery disease, high-density lipoprotein (HDL) cholesterol, LDL cholesterol, triglycerides, type 2 diabetes, glycated hemoglobin (HbA1c), rheumatoid arthritis (RA), SCZ and OA at the nominal P < .05 significance level. However, after adjusting for computationally selected heritable confounders, only effects on HDL cholesterol (negative), HbA1c (positive), RA (risk increasing), and SCZ (risk decreasing) remain nominally significant. Using confounder-adjusted MVMR additionally reveals evidence of a protective effect of CRP on bipolar disorder not observed in the univariable analysis.
These results suggest that univariable MR analyses of CRP may be biased by high levels of heritable confounding, though CRP may indeed play a causal role in development of some diseases, potentially mediated by its role in innate immunity. These results also highlight the potential for automatic confounder selection to improve the robustness of MR analyses.
Multivariable Mendelian randomization (MVMR) can be used to eliminate bias from heritable confounding when genome-wide association study (GWAS) summary data are available for confounders. In this study, we used MVMR to estimate the causal effects of CRP on 16 outcomes with prior evidence of a causal or associational link to CRP. We used a novel computational pipeline to identify a broad set of potential heritable confounders between CRP and each outcome trait from studies in the Medical Research Council Integrative Epidemiology Unit (MRC-IEU) OpenGWAS database. We compared MVMR results with computationally selected confounders to univariable MR results and MVMR adjusting only for body mass index.
Univariable MR suggests evidence of potential causal effects of CRP on coronary artery disease, high-density lipoprotein (HDL) cholesterol, LDL cholesterol, triglycerides, type 2 diabetes, glycated hemoglobin (HbA1c), rheumatoid arthritis (RA), SCZ and OA at the nominal P < .05 significance level. However, after adjusting for computationally selected heritable confounders, only effects on HDL cholesterol (negative), HbA1c (positive), RA (risk increasing), and SCZ (risk decreasing) remain nominally significant. Using confounder-adjusted MVMR additionally reveals evidence of a protective effect of CRP on bipolar disorder not observed in the univariable analysis.
These results suggest that univariable MR analyses of CRP may be biased by high levels of heritable confounding, though CRP may indeed play a causal role in development of some diseases, potentially mediated by its role in innate immunity. These results also highlight the potential for automatic confounder selection to improve the robustness of MR analyses.