Randomised by (your) god: robust inference from an observational study design

A data-adaptive method for investigating effect heterogeneity with high-dimensional covariates in Mendelian randomization

BACKGROUND

Mendelian randomization is a popular method for causal inference with observational data that uses genetic variants as instrumental variables. Similarly to a randomized trial, a standard Mendelian randomization analysis estimates the population-averaged effect of an exposure on an outcome. Dividing the population into subgroups can reveal effect heterogeneity to inform who would most benefit from intervention on the exposure. However, as covariates are measured post-"randomization", naive stratification typically induces collider bias in stratum-specific estimates.

METHOD

We extend a previously proposed stratification method (the "doubly-ranked method") to form strata based on a single covariate, and introduce a data-adaptive random forest method to calculate stratum-specific estimates that are robust to collider bias based on a high-dimensional covariate set. We also propose measures based on the Q statistic to assess heterogeneity between stratum-specific estimates (to understand whether estimates are more variable than expected due to chance alone) and variable importance (to identify the key drivers of effect heterogeneity).

RESULT

We show that the effect of body mass index (BMI) on lung function is heterogeneous, depending most strongly on hip circumference and weight. While for most individuals, the predicted effect of increasing BMI on lung function is negative, it is positive for some individuals and strongly negative for others.

CONCLUSION

Our data-adaptive approach allows for the exploration of effect heterogeneity in the relationship between an exposure and an outcome within a Mendelian randomization framework. This can yield valuable insights into disease aetiology and help identify specific groups of individuals who would derive the greatest benefit from targeted interventions on the exposure.

Genetically predicted vitamin K levels and risk of osteoarthritis: Mendelian randomization study

OBJECTIVE

Osteoarthritis (OA) is a progressive disease for which there is no disease modifying therapy. Vitamin K levels and vitamin K antagonism have been associated with risk and progression of OA which may have direct implications for clinical management, but these observational findings are susceptible to confounding. We aimed to estimate the causal association between vitamin K and OA risk using Mendelian randomisation (MR).

METHODS

We used data from the largest genome-wide association study (GWAS) of OA to date (up to 826,690 individuals) to estimate the effect of genetically predicted vitamin K level (instrumented using four variants derived from a GWAS of 2,138 individuals) on risk of all OA types, knee, hip, spine, hand OA, and total joint replacement. We employed the inverse-variance weighted method for the primary analysis and, in a series of sensitivity analyses, adjusted for sub-genome wide significant instruments and tested for potential bias from pleiotropy.

RESULTS

We showed that genetically predicted vitamin K levels were not causally associated with risk of OA overall (OR 0.98 per unit increase in log-transformed vitamin K1; 95%CI 0.96-1.01), knee (OR 0.98; 0.92-1.03), hip (OR 0.97; 0.88-1.07), spine (OR 0.97; 0.90-1.04), hand OA (OR 0.97; 0.91-1.04) or joint replacement (OR 0.96; 0.89-1.04). Results were similar across all sensitivity analyses.

CONCLUSION

We found little evidence of a causal association between genetically predicted vitamin K and OA risk. Larger genetic and interventional studies of vitamin K are required to confirm our findings.

Physical activity, sedentary behavior and risk of coronary artery disease, myocardial infarction and ischemic stroke: a two-sample Mendelian randomization study

AIMS

Observational evidence suggests that physical activity (PA) is inversely and sedentarism positively related with cardiovascular disease risk. We performed a two-sample Mendelian randomization (MR) analysis to examine whether genetically predicted PA and sedentary behavior are related to coronary artery disease, myocardial infarction, and ischemic stroke.

METHODS AND RESULTS

We used single nucleotide polymorphisms (SNPs) associated with self-reported moderate to vigorous PA (n = 17), accelerometer based PA (n = 7) and accelerometer fraction of accelerations > 425 milli-gravities (n = 7) as well as sedentary behavior (n = 6) in the UK Biobank as instrumental variables in a two sample MR approach to assess whether these exposures are related to coronary artery disease and myocardial infarction in the CARDIoGRAMplusC4D genome-wide association study (GWAS) or ischemic stroke in the MEGASTROKE GWAS. The study population included 42,096 cases of coronary artery disease (99,121 controls), 27,509 cases of myocardial infarction (99,121 controls), and 34,217 cases of ischemic stroke (404,630 controls). We found no associations between genetically predicted self-reported moderate to vigorous PA, accelerometer-based PA or accelerometer fraction of accelerations > 425 milli-gravities as well as sedentary behavior with coronary artery disease, myocardial infarction, and ischemic stroke.

CONCLUSIONS

These results do not support a causal relationship between PA and sedentary behavior with risk of coronary artery disease, myocardial infarction, and ischemic stroke. Hence, previous observational studies may have been biased.

Depression and interleukin-6 signaling: A Mendelian Randomization study

BACKGROUND

A large body of research has reported associations between depression and elevated interleukin-6 (IL-6), a cytokine with several roles including pro-inflammatory signaling. The nature and directionality of this relationship are not yet clear. In this study we use Mendelian Randomization to examine the possibility of a causal relationship between IL-6 and depressive symptoms, and to explore multiple signaling pathways that could serve as mechanisms for this relationship.

METHODS

This study uses a two-sample Mendelian Randomization design. Data come from the UK Biobank (n = 89,119) and published summary statistics from six existing GWAS analyses. The primary analysis focuses on the soluble interleukin-6 receptor (sIL-6R), which is involved in multiple signaling pathways. Exploratory analyses use C-reactive protein (CRP) and soluble glycoprotein 130 (sgp130) to further examine potential underlying mechanisms.

RESULTS

Results are consistent with a causal effect of sIL-6R on depression (PCA-IVW Odds Ratio: 1.023 (95% Confidence Interval: 1.006-1.039), p = 0.006). Exploratory analyses demonstrate that the relationship could be consistent with either decreased classical signaling or increased trans signaling as the underlying mechanism.

DISCUSSION

These results strengthen the body evidence implicating IL-6 signaling in depression. When compared with existing observational and animal findings, the direction of these results suggests involvement of IL-6 trans signaling. Further study is needed to examine whether IL6R genetic variants might influence IL-6 trans signaling in the brain, as well as to explore other potential pathways linking depression and inflammation.

The effects of age at menarche and first sexual intercourse on reproductive and behavioural outcomes: A Mendelian randomization study

There is substantial variation in the timing of significant reproductive life events such as menarche and first sexual intercourse. Life history theory explains this variation as an adaptive response to an individual's environment and it is important to examine how traits within life history strategies affect each other. Here we applied Mendelian randomization (MR) methods to investigate whether there is a causal effect of variation in age at menarche and age at first sexual intercourse (markers or results of exposure to early life adversity) on outcomes related to reproduction, education and risky behaviour in UK Biobank (N = 114 883-181 255). Our results suggest that earlier age at menarche affects some traits that characterize life history strategies including earlier age at first and last birth, decreased educational attainment, and decreased age at leaving education (for example, we found evidence for a 0.26 year decrease in age at first birth per year decrease in age at menarche, 95% confidence interval: -0.34 to -0.17; p < 0.001). We find no clear evidence of effects of age at menarche on other outcomes, such as risk taking behaviour. Age at first sexual intercourse was also related to many life history outcomes, although there was evidence of horizontal pleiotropy which violates an assumption of MR and we therefore cannot infer causality from this analysis. Taken together, these results highlight how MR can be applied to test predictions of life history theory and to better understand determinants of health and social behaviour.

Application of Mendelian Randomization to Investigate the Association of Body Mass Index with Health Care Costs

Causal effect estimates for the association of obesity with health care costs can be biased by reversed causation and omitted variables. In this study, we use genetic variants as instrumental variables to overcome these limitations, a method that is often called Mendelian randomization (MR). We describe the assumptions, available methods, and potential pitfalls of using genetic information and how to address them. We estimate the effect of body mass index (BMI) on total health care costs using data from a German observational study and from published large-scale data. In a meta-analysis of several MR approaches, we find that models using genetic instruments identify additional annual costs of €280 for a 1-unit increase in BMI. This is more than 3 times higher than estimates from linear regression without instrumental variables (€75). We found little evidence of a nonlinear relationship between BMI and health care costs. Our results suggest that the use of genetic instruments can be a powerful tool for estimating causal effects in health economic evaluation that might be superior to other types of instruments where there is a strong association with a modifiable risk factor.

Association of plasma C-reactive protein with ischaemic stroke: a Mendelian randomization study

BACKGROUND AND PURPOSE

Elevated C-reactive protein (CRP) is associated with an increased risk of ischaemic stroke (IS). However, the causality of this association is uncertain. The aim is to investigate whether genetically raised plasma CRP concentration levels are associated with IS on the basis of the Mendelian randomization method.

METHODS

Based on the National Center for Biotechnology Information single nucleotide polymorphism (SNP) database, the Chinese online genetic database as well as previously published studies, four CRP-associated SNP alleles (rs1130864, rs1205, rs876537 and rs3093059) with minor allele frequency ≥0.15 were selected and the concentration levels of CRP were measured in 378 first-ever IS patients and 613 healthy controls.

RESULTS

Three SNPs were chosen and used as instrumental variables. The adjusted odds ratios (ORs) [95% confidence interval (95% CI)] of IS per addition of the modelled allele were 1.07 (0.79-1.45) for rs876537, 0.99 (0.73-1.35) for rs1205 and 1.08 (0.71-1.65) for rs3093059. The OR (95% CI) of IS for plasma CRP ≥2.0 mg/l was 2.19 (1.06-4.53) compared with <2.0 mg/l. The adjusted OR (95% CI) of IS per genetically predicted 10% higher CRP concentration, based on the three SNPs as the instruments, was 1.02 (0.94-1.11). Furthermore, similar results were obtained with adjusted ORs (95% CI) of 1.00 (0.88-1.13) and 1.04 (0.93-1.16), respectively, for large-artery atherosclerosis and small-artery occlusion per genetically predicted 10% higher CRP concentration.

CONCLUSIONS

This Mendelian randomization study provides no clear support that elevated CRP concentration is causally associated with the risk of IS.

[Archibald Cochrane: evidence, effectiveness and decision-making in health]

Nowadays, Evidence-Based Medicine plays a fundamental role while making medical decisions, considering that through the methods of science, it attempts to justify the variety of alternatives that may be offered to patients. In order to understand the historical evolution of this way of practicing medicine, it is necessary to review the contribution of one of the main participants in this cultural movement: Archibald Leman Cochrane, who helped to define the theoretical framework that has allowed the integration of science into the practice of medicine. Since he insisted in the need of integrating scientific evidence into clinical experience, his role became a fundamental and decisive element in the development of a new discipline: Evidence-Based Medicine.

Designs combining instrumental variables with case-control: estimating principal strata causal effects

The instrumental variables framework is commonly used for the estimation of causal effects from cohort samples. However, the combination of instrumental variables with more efficient designs such as case-control sampling requires new methodological consideration. For example, as the use of Mendelian randomization studies is increasing and the cost of genotyping and gene expression data can be high, the analysis of data gathered from more cost-effective sampling designs is of prime interest. We show that the standard instrumental variables analysis does not appropriately estimate the causal effects of interest when the instrumental variables design is combined with the case-control design. We also propose a method that can estimate the causal effects in such combined designs. We illustrate the method with a study in oncology.

Missing data methods in Mendelian randomization studies with multiple instruments

Mendelian randomization studies typically have low power. Where there are several valid candidate genetic instruments, precision can be gained by using all the instruments available. However, sporadically missing genetic data can offset this gain. The authors describe 4 Bayesian methods for imputing the missing data based on a missing-at-random assumption: multiple imputations, single nucleotide polymorphism (SNP) imputation, latent variables, and haplotype imputation. These methods are demonstrated in a simulation study and then applied to estimate the causal relation between C-reactive protein and each of fibrinogen and coronary heart disease, based on 3 SNPs in British Women's Heart and Health Study participants assessed at baseline between May 1999 and June 2000. A complete-case analysis based on all 3 SNPs was found to be more precise than analyses using any 1 SNP alone. Precision is further improved by using any of the 4 proposed missing data methods; the improvement is equivalent to about a 25% increase in sample size. All methods gave similar results, which were apparently not overly sensitive to violation of the missing-at-random assumption. Programming code for the analyses presented is available online.

Realizing the promise of population biobanks: a new model for translation

The promise of science lies in expectations of its benefits to societies and is matched by expectations of the realisation of the significant public investment in that science. In this paper, we undertake a methodological analysis of the science of biobanking and a sociological analysis of translational research in relation to biobanking. Part of global and local endeavours to translate raw biomedical evidence into practice, biobanks aim to provide a platform for generating new scientific knowledge to inform development of new policies, systems and interventions to enhance the public's health. Effectively translating scientific knowledge into routine practice, however, involves more than good science. Although biobanks undoubtedly provide a fundamental resource for both clinical and public health practice, their potentiating ontology--that their outputs are perpetually a promise of scientific knowledge generation--renders translation rather less straightforward than drug discovery and treatment implementation. Biobanking science, therefore, provides a perfect counterpoint against which to test the bounds of translational research. We argue that translational research is a contextual and cumulative process: one that is necessarily dynamic and interactive and involves multiple actors. We propose a new multidimensional model of translational research which enables us to imagine a new paradigm: one that takes us from bench to bedside to backyard and beyond, that is, attentive to the social and political context of translational science, and is cognisant of all the players in that process be they researchers, health professionals, policy makers, industry representatives, members of the public or research participants, amongst others.

Use of genetic markers and gene-diet interactions for interrogating population-level causal influences of diet on health

Differences in diet appear to contribute substantially to the burden of disease in populations, and therefore changes in diet could lead to major improvements in public health. This is predicated on the reliable identification of causal effects of nutrition on health, and unfortunately nutritional epidemiology has deficiencies in terms of identifying these. This is reflected in the many cases where observational studies have suggested that a nutritional factor is protective against disease, and randomized controlled trials have failed to verify this. The use of genetic variants as proxy measures of nutritional exposure-an application of the Mendelian randomization principle-can contribute to strengthening causal inference in this field. Genetic variants are not subject to bias due to reverse causation (disease processes influencing exposure, rather than vice versa) or recall bias, and if obvious precautions are applied are not influenced by confounding or attenuation by errors. This is illustrated in the case of epidemiological studies of alcohol intake and various health outcomes, through the use of genetic variants related to alcohol metabolism (in ALDH2 and ADH1B). Examples from other areas of nutritional epidemiology and of the informative nature of gene-environment interactions interpreted within the Mendelian randomization framework are presented, and the potential limitations of the approach addressed.

Mendelian Randomization for Strengthening Causal Inference in Observational Studies: Application to Gene × Environment Interactions

Identification of environmentally modifiable factors causally influencing disease risk is fundamental to public-health improvement strategies. Unfortunately, observational epidemiological studies are limited in their ability to reliably identify such causal associations, reflected in the many cases in which conventional epidemiological studies have apparently identified associations that randomized controlled trials have failed to verify. The use of genetic variants as proxy measures of exposure -an application of the Mendelian randomization principle-can contribute to strengthening causal inference. Genetic variants are not subject to bias due to reverse causation (disease processes influencing exposure, rather than vice versa) or recall bias, and if simple precautions are applied, they are not influenced by confounding or attenuation by errors. The principles of Mendelian randomization are illustrated with specific reference to studies of the effects of alcohol intake on various health-related outcomes through the utilization of genetic variants related to alcohol metabolism (in ALDH2 and ADH1B). Ways of incorporating Gene × Environment interactions into the Mendelian randomization framework are developed, and the strengths and limitations of the approach discussed.

The association of C-reactive protein and CRP genotype with coronary heart disease: findings from five studies with 4,610 cases amongst 18,637 participants

Background

It is unclear whether C-reactive protein (CRP) is causally related to coronary heart disease (CHD). Genetic variants that are known to be associated with CRP levels can be used to provide causal inference of the effect of CRP on CHD. Our objective was to examine the association between CRP genetic variant +1444C>T (rs1130864) and CHD risk in the largest study to date of this association.

Methods and Results

We estimated the association of CRP genetic variant +1444C>T (rs1130864) with CRP levels and with CHD in five studies and then pooled these analyses (N = 18,637 participants amongst whom there were 4,610 cases). CRP was associated with potential confounding factors (socioeconomic position, physical activity, smoking and body mass) whereas genotype (rs1130864) was not associated with these confounders. The pooled odds ratio of CHD per doubling of circulating CRP level after adjustment for age and sex was 1.13 (95%CI: 1.06, 1.21), and after further adjustment for confounding factors it was 1.07 (95%CI: 1.02, 1.13). Genotype (rs1130864) was associated with circulating CRP; the pooled ratio of geometric means of CRP level among individuals with the TT genotype compared to those with the CT/CC genotype was 1.21 (95%CI: 1.15, 1.28) and the pooled ratio of geometric means of CRP level per additional T allele was 1.14 (95%CI: 1.11, 1.18), with no strong evidence in either analyses of between study heterogeneity (I² = 0%, p>0.9 for both analyses). There was no association of genotype (rs1130864) with CHD: pooled odds ratio 1.01 (95%CI: 0.88, 1.16) comparing individuals with TT genotype to those with CT/CC genotype and 0.96 (95%CI: 0.90, 1.03) per additional T allele (I²<7.5%, p>0.6 for both meta-analyses). An instrumental variables analysis (in which the proportion of CRP levels explained by rs1130864 was related to CHD) suggested that circulating CRP was not associated with CHD: the odds ratio for a doubling of CRP level was 1.04 (95%CI: 0.61, 1.80).

Conclusions

We found no association of a genetic variant, which is known to be related to CRP levels, (rs1130864) and having CHD. These findings do not support a causal association between circulating CRP and CHD risk, but very large, extended, genetic association studies would be required to rule this out.

Using genetic markers to orient the edges in quantitative trait networks: the NEO software

Background

Systems genetic studies have been used to identify genetic loci that affect transcript abundances and clinical traits such as body weight. The pairwise correlations between gene expression traits and/or clinical traits can be used to define undirected trait networks. Several authors have argued that genetic markers (e.g expression quantitative trait loci, eQTLs) can serve as causal anchors for orienting the edges of a trait network. The availability of hundreds of thousands of genetic markers poses new challenges: how to relate (anchor) traits to multiple genetic markers, how to score the genetic evidence in favor of an edge orientation, and how to weigh the information from multiple markers.

Results

We develop and implement Network Edge Orienting (NEO) methods and software that address the challenges of inferring unconfounded and directed gene networks from microarray-derived gene expression data by integrating mRNA levels with genetic marker data and Structural Equation Model (SEM) comparisons. The NEO software implements several manual and automatic methods for incorporating genetic information to anchor traits. The networks are oriented by considering each edge separately, thus reducing error propagation. To summarize the genetic evidence in favor of a given edge orientation, we propose Local SEM-based Edge Orienting (LEO) scores that compare the fit of several competing causal graphs. SEM fitting indices allow the user to assess local and overall model fit. The NEO software allows the user to carry out a robustness analysis with regard to genetic marker selection. We demonstrate the utility of NEO by recovering known causal relationships in the sterol homeostasis pathway using liver gene expression data from an F2 mouse cross. Further, we use NEO to study the relationship between a disease gene and a biologically important gene co-expression module in liver tissue.

Conclusion

The NEO software can be used to orient the edges of gene co-expression networks or quantitative trait networks if the edges can be anchored to genetic marker data. R software tutorials, data, and supplementary material can be downloaded from: .

Exploring the developmental overnutrition hypothesis using parental-offspring associations and FTO as an instrumental variable

BACKGROUND

The developmental overnutrition hypothesis suggests that greater maternal obesity during pregnancy results in increased offspring adiposity in later life. If true, this would result in the obesity epidemic progressing across generations irrespective of environmental or genetic changes. It is therefore important to robustly test this hypothesis.

METHODS AND FINDINGS

We explored this hypothesis by comparing the associations of maternal and paternal pre-pregnancy body mass index (BMI) with offspring dual energy X-ray absorptiometry (DXA)-determined fat mass measured at 9 to 11 y (4,091 parent-offspring trios) and by using maternal FTO genotype, controlling for offspring FTO genotype, as an instrument for maternal adiposity. Both maternal and paternal BMI were positively associated with offspring fat mass, but the maternal association effect size was larger than that in the paternal association in all models: mean difference in offspring sex- and age-standardised fat mass z-score per 1 standard deviation BMI 0.24 (95% confidence interval [CI]: 0.22 to 0.26) for maternal BMI versus 0.13 (95% CI: 0.11, 0.15) for paternal BMI; p-value for difference in effect < 0.001. The stronger maternal association was robust to sensitivity analyses assuming levels of non-paternity up to 20%. When maternal FTO, controlling for offspring FTO, was used as an instrument for the effect of maternal adiposity, the mean difference in offspring fat mass z-score per 1 standard deviation maternal BMI was -0.08 (95% CI: -0.56 to 0.41), with no strong statistical evidence that this differed from the observational ordinary least squares analyses (p = 0.17).

CONCLUSIONS

Neither our parental comparisons nor the use of FTO genotype as an instrumental variable, suggest that greater maternal BMI during offspring development has a marked effect on offspring fat mass at age 9-11 y. Developmental overnutrition related to greater maternal BMI is unlikely to have driven the recent obesity epidemic.

Assessing intrauterine influences on offspring health outcomes: can epidemiological studies yield robust findings?

The influence of factors acting during the intrauterine period on health outcomes of offspring is of considerable research and public health interest. There are, however, methodological challenges in establishing robust causal links, because exposures often act many decades before outcomes of interest, may act before it is evident that women are pregnant and would enter pregnancy birth cohorts, and may also be strongly related to other factors, generating considerable degrees of potential confounding. The degree of confounding can sometimes be estimated by comparing the association between exposures experienced by the mother during pregnancy and outcomes among the offspring with the association of exposures experienced by the father during the pregnancy period and offspring outcomes. If the effects are due to an intrauterine exposure, then maternal exposure during pregnancy should have a clearly greater influence than paternal exposure. A different approach is that of Mendelian randomization, which utilizes genetic variants of known functional effect that can proxy for modifiable exposures. If carried by the mother, these variants would influence the intrauterine environment experienced by her offspring. These genetic variants are stable over time and can be assessed after pregnancy is complete or even after outcomes in the offspring have been observed. The variants would also not generally be related to potential confounding factors. Other epidemiological strategies are briefly reviewed. It is concluded that the naïve acceptance of findings utilizing conventional epidemiological methods in this setting is misplaced.

Mendelian randomization: Using genes as instruments for making causal inferences in epidemiology

Observational epidemiological studies suffer from many potential biases, from confounding and from reverse causation, and this limits their ability to robustly identify causal associations. Several high-profile situations exist in which randomized controlled trials of precisely the same intervention that has been examined in observational studies have produced markedly different findings. In other observational sciences, the use of instrumental variable (IV) approaches has been one approach to strengthening causal inferences in non-experimental situations. The use of germline genetic variants that proxy for environmentally modifiable exposures as instruments for these exposures is one form of IV analysis that can be implemented within observational epidemiological studies. The method has been referred to as 'Mendelian randomization', and can be considered as analogous to randomized controlled trials. This paper outlines Mendelian randomization, draws parallels with IV methods, provides examples of implementation of the approach and discusses limitations of the approach and some methods for dealing with these.

Body mass index in middle life and future risk of hospital admission for psychoses or depression: findings from the Renfrew/Paisley study

BACKGROUND

There is evidence that greater body mass index (BMI) protects against depression, schizophrenia and suicide. However, there is a need for prospective studies.

METHOD

We examined the association of BMI with future hospital admissions for psychoses or depression/anxiety disorders in a large prospective study of 7036 men and 8327 women. Weight and height were measured at baseline (1972-76) when participants were aged 45-64. Follow-up was for a median of 29 years.

RESULTS

Greater BMI and obesity were associated with a reduced risk of hospital admission for psychoses and depression/anxiety in both genders, with the magnitude of these associations being the same for males and females. With adjustment for age, sex, smoking and social class, a 1 standard deviation (s.d.) greater BMI at baseline was associated with a rate ratio of 0.91 [95% confidence interval (CI) 0.82-1.01] for psychoses and 0.87 (95% CI 0.77-0.98) for depression/anxiety. Further adjustment for baseline psychological distress and total cholesterol did not alter these associations.

CONCLUSIONS

Our findings add to the growing body of evidence that suggests that greater BMI is associated with a reduced risk of major psychiatric outcomes. Long-term follow-up of participants in randomized controlled trials of interventions that effectively result in weight loss and the use of genetic variants that are functionally related to obesity as instrumental variables could help to elucidate whether these associations are causal.