A Mendelian Randomization Study Provides Evidence That Adiposity and Dyslipidemia Lead to Lower Urinary Albumin-to-Creatinine Ratio, a Marker of Microvascular Function

The urinary albumin-to-creatinine ratio (ACR), a marker of diabetic nephropathy, is used as a proxy for damage to the systemic microcirculation (1) and predicts first myocardial infarction and mortality in those with diabetes, poststroke, and in the general population (2–4). There is evidence linking metabolic-related traits, including adiposity, dyslipidemia, and insulin resistance with elevated ACR levels and microvascular damage (5,6). It is well accepted that tight glucose control in patients with type 2 diabetes (T2D) reduces the risk of microvascular retinal complications (7,8), and there is evidence that adiposity per se is associated with increased ACR. For example, population studies suggest that microalbuminuria is associated with central adiposity (9), and results from the Framingham Heart Study show that visceral but not subcutaneous fat is associated with increased albuminuria (10). Not all evidence linking metabolic-related traits comes from randomized control trials, and in absence of these, the next best evidence of causality comes from genetic studies using a technique known as Mendelian randomization (MR) (Fig. 1).

In MR, genetic variants that are strongly associated with the risk factor of interest are used to test its causal effect on an outcome (11). The MR approach exploits the natural experiment of genetic variants being randomly assigned at conception, which means they are less likely to be associated with confounding factors and should not suffer from reverse causality (12). MR studies investigating the role of metabolic traits in increasing microvascular damage, including ACR, infer causal relationships for higher blood pressure (13) but not for lipids (14), but the latter study was small, limited in power, and focused only on people with diabetes.

Here, we used data from 743 participants in the SUrrogate markers for Micro- and Macro-vascular hard end points for Innovative diabetes Tools (SUMMIT) (15) study to first confirm that ACR is a suitable proxy for early systemic microvascular damage by testing its association with two validated measures of microvascular function: skin microvascular response to iontophoresis of vasodilators acetylcholine (ACH) (endothelial dependent) and sodium nitroprusside (endothelial independent). Second, we tested the observational associations between ACR and nine metabolic risk factors in 438,075 participants in the UK Biobank. Finally, we used MR to test the effects of 11 metabolic risk factors on microvascular function using ACR as a proxy in the UK Biobank and summary results for 54,451 participants in the CKDGen genome-wide association study (GWAS) results.

The UK Biobank study recruited >500,000 individuals aged between 37 and 73 years between 2006 and 2010. The study collected detailed information from all participants via questionnaires, interviews, and measurements (16). Here we used 438,075 individuals of white European ancestry (defined through principal component–based analyses (17) with ACR available. We also defined a subset of 368,754 unrelated individuals of European ancestry. Related individuals were defined using a KING (Kinship-based INference for Genome-wide association studies) kinship, and an optimal list of unrelated individuals was generated to allow maximum numbers of individuals to be included. Ancestral principal components were then generated within these identified individuals for use in subsequent analyses.

Data for observational association and functional measures of microvascular function were collected in 743 individuals from two centers (Exeter and Dundee) participating in the vascular imaging cohort of SUMMIT. SUMMIT is a multicenter study aiming at identifying markers that predict the risks of developing diabetes-related chronic micro- and macrovascular complications (15,18).

In SUMMIT, skin microvascular function in the forearm is measured using laser Doppler fluximetry. A laser Doppler imager (model LD12; Moor Instruments) was used to measure perfusion before and after iontophoresis of endothelium-dependent (ACH) and endothelium-independent (sodium nitroprusside) vasodilatory stimuli. The full protocol of the techniques used has previously been described (18).

ACR was measured in SUMMIT from random spot urine collection (Exeter Pathology Services, Royal Devon and Exeter NHS Foundation Trust, and Dundee Pathology Services, Ninewells NHS trust) in accordance with the U.K. National Quality Assessment Scheme. Analysis of albumin concentration was performed using an immunoturbidimetric method with a detection limit of 3.0 mg/L (cobas System; Roche), and creatinine was measured using the Jaffe methods. In order to maintain a consistent approach with UK Biobank analysis, we set values below the detection limit at 2.9 mg/L prior to the calculation of the ratio. The ACR variable was inverse normalized prior to analysis.

The relationship between the gold standard microvascular functional measures and ACR was explored using linear regression models, with age and sex included as covariates.

We selected 11 metabolic markers that have previously been associated with ACR and have strong genetic instruments available in the form of multiple variants (Supplementary Table 2) identified in large GWAS. More information on how the outcome and exposures were defined in the UK Biobank appears below.

A continuous measure of ACR was derived using urinary measures of albumin and creatinine. If albumin was <6.7 mg/L (the detection level of the assay in UK Biobank [https://biobank.ndph.ox.ac.uk/showcase/showcase/docs/urine_assay.pdf]) then the albumin was set at 6.7 mg/L prior to the calculation of the ratio—an approach consistent with that of previous studies (13,19). Albumin was measured using an immunoturbidimetric analysis method (Randox Biosciences, U.K.), while creatinine was measured using an enzymatic analysis method (Beckman Coulter, High Wycombe, U.K.). The ACR variable was inverse normalized prior to analysis.

Nine of the eleven metabolic markers were measured in the UK Biobank.

Serum concentrations of LDL cholesterol (LDL-C) (N = 417,386) were obtained using an enzymatic selective protection analysis method [cat. no. AU5800, Beckman Coulter], HDL cholesterol (HDL-C) (N = 382,598) using an enzyme immuno-inhibition analysis method [cat. no. AU5800; Beckman Coulter], and triglycerides (TG) (N = 417,825) using an enzymatic analysis method [cat. no. AU5800, Beckman Coulter]. More details on the acquisition of these biomarkers can be found here: https://biobank.ndph.ox.ac.uk/showcase/showcase/docs/serum_biochemistry.pdf.

We used three measures of body composition: BMI, waist-to-hip ratio (WHR) adjusted for BMI, and, based on genetics, a measure of higher body fat percentage but lower metabolic disease risk, termed “favorable adiposity.” BMI was calculated for all participants from measured weight and height (height in kilograms divided by the square of height in meters) and was available for 436,631 individuals with ACR and genetic data available. WHR was calculated from measured waist and hip circumference measures and adjusted for BMI; this was available in 436,530 individuals. Body fat percentage was calculated from bioelectrical impedance data collected using a Tanita BC-418MA Body Composition Analyzer and was available in 430,546 individuals.

Systolic blood pressure (N = 437,121) and diastolic blood pressure (N = 436,394) were measured. The blood pressure readings were obtained from averaging two readings obtained in a seated position 5 min apart using an automated blood pressure device (Omron 705 IT, Omron Healthcare Europe B.V., Hoofddorp, the Netherlands). For participants for whom only one valid blood pressure was available, this was used. Blood pressure medication use was accounted for by adding 10 and 15 mmHg to diastolic and systolic measures, respectively.

T2D cases were defined through self-report of diabetes using the baseline questionnaire. Case subjects were participants diagnosed at >35 years of age without reporting of insulin use within the first year of diagnosis (20). This resulted in 13,799 case and 415,908 control subjects (Table 1).

Two measures of glycemic control were not measured in the UK Biobank at the time of study: fasting glucose and fasting insulin.

For all continuous measurements in UK Biobank, values >4.56 SD away from the mean were excluded. These variables when then inverse normalized prior to analysis.

The observational associations between the measured exposures and ACR were tested in the UK Biobank using linear regression models, adjusted for age, sex, and assessment center.

For MR, independent genetic variants were selected from the UK Biobank imputation data set. Variants were excluded if imputation quality (INFO) was <0.3 or the minor allele frequency was <0.1%.

The genetic variants for the exposure traits were selected based on published GWAS. Genetic variants were selected and extracted for the 11 metabolic markers including lipid levels (TG, HDL-C, and LDL-C), BMI, favorable adiposity (genetic variants associated with higher body fat percentage but lower risk of metabolic disease [e.g., T2D, coronary heart disease]), WHR (adjusted for BMI), systolic and diastolic blood pressure, T2D, fasting glucose, and fasting insulin (Supplementary Table 2). Four variants were identified that were previously identified to associate with ACR at genome-wide significance: rs1047891 (HDL variant), rs4865796 (fasting insulin variant), rs109953111 (diastolic blood pressure variant), and rs2068888 (TG variant) (21).

The extracted genetic variants were used to create genetic risk scores (GRS) for each metabolic trait of interest. The variants were weighted by their effect size (β-coefficient) obtained from the primary GWAS, where possible using GWAS that did not include data from the UK Biobank (Eq. 1). The weighted score was then rescaled to reflect the number of trait-raising alleles (Eq. 2).

where SNP is single nucleotide polymorphism.

We used MR to test for causal relationships between our 11 metabolic risk factors as exposures and ACR as an outcome. MR relies on several assumptions as outlined in Fig. 1:

• The exposure GRS are robustly associated with the relevant measured exposure (Supplementary Table 1).

• The exposure GRS are not associated, independently of their effects on the exposure, with confounding factors that bias conventional epidemiological associations.

• The exposure GRS is only associated with the outcome via its effect on the modifiable exposure.

In this study, we used several methods of MR (one- and two-sample MR). The primary analyses used data from 438,075 UK Biobank participants with measured ACR. We extracted the genetic variants for the 11 known metabolic traits (Supplementary Table 2) from the BOLT-LMM (22) GWAS of ACR, which was adjusted for baseline age, sex, study center, and genotyping array (0 = BiLEVE, 1 = UK Biobank Axiom interim release, 2 = UK Biobank Axiom final release). We also extracted association statistics for the same SNPs from the largest GWAS of ACR (54,451 participants from CDKGen consortium meta-analysis [Teumer et al. (19)]) which did not include the UK Biobank.

Our primary MR approach was to use the inverse variance weighted (IVW) estimator. The IVW method involves a weighted regression of the effect sizes of variant-outcome associations against the effect sizes of the variant–risk factor associations constraining the intercept to zero. The β-coefficient from the weighted regression represents the SD change in the ACR per SD change in the outcome variable (with the exception of T2D, where we present our findings as an SD change in ACR per twofold higher genetic liability for T2D). Several sensitivity analyses were performed to test whether the MR IVW estimates are biased by genetic variants that affect the outcome independently of the exposure of interest (i.e., horizontal pleiotropy). These methods were MR-Egger regression (23) and the weighted median (WM) estimator (24). MR-Egger is similar to IVW, except that the intercept is unconstrained. The intercept in MR-Egger reflects the average pleiotropic effect across genetic variants. Hence, this method is less susceptible to potentially pleiotropic variants having a stronger effect on the outcome compared with their effect on the primary traits. The WM method is also more resistant to pleiotropy and gives consistent estimates even when 50% of the variants are invalid. Given these different assumptions, if all methods are broadly consistent it strengthens our causal inference. The R code for the various two-sample methods has previously been published (23,24).

We performed sensitivity analyses for the four traits where one variant was known to be associated with ACR at genome-wide significance. Here, the two-sample MR was repeated excluding that one variant.

The results from the two-sample MR in the UK Biobank and the GWAS were meta-analyzed using the metan command in Stata.

There is some overlap between the genetic variants for LDL-C, HDL-C, and TG. Therefore, as well as individually exploring the role of the LDL-C, HDL-C, and TG SNPs in the outcomes we also ran multivariate models adjusting for the other lipid associations (25). For example, when testing the causal role of LDL-C, we included the LDL-C–SNP–TG association and the LDL-C–SNP–HDL association as covariates in our model.

In an unrelated subset of the data, we also performed one-sample MR using the GRS and the ivreg2 command in Stata. In these models, age, sex, ancestral principal components, assessment center, and genotyping platform were included as covariates. In cases where the predictor was not measured in the UK Biobank, we explored the association of the GRS directly with the outcome. As with the two-sample MR, we performed multivariate analyses for the lipids by adjusting models for the other lipid GRS. For example, we performed MR to explore the causal role of LDL-C in ACR, adjusting our models for all the standard covariates and the HDL-C and TG GRS.

The UK Biobank resource can be used by any bona fide researcher, and access to all the genetic and phenotypic data used in this study is available upon application to the UK Biobank (https://www.ukbiobank.ac.uk/). The summary statistics from the CDKGen are available: https://ckdgen.imbi.uni-freiburg.de. SUMMIT data used in this study are available on request to the Diabetes and Vascular Research Centre, University of Exeter Medical School.

Characteristics for the 438,075 UK Biobank and 743 SUMMIT participants are presented in Table 1.

Results from SUMMIT support the use of ACR as a proxy for microvascular function with lower microvascular function associated with raised ACR levels. There was a negative association between ACR and skin microvascular function for both endothelium-dependent (ACH) and -independent (sodium nitroprusside) function. One SD lower response in endothelium-dependent microvascular function as measured by skin reactivity to iontophoresis of ACH was associated with a 0.155 SD higher ACR (95% CI 0.078, 0.230, P = 5.8E-05). One SD lower response in endothelium-independent microvascular function as measured by reactivity to sodium nitroprusside was associated with a 0.206 SD higher ACR (95% CI 0.131, 0.281, P = 1.1E-07). Taken together, these measures demonstrate that lower systemic microvascular response measured by skin reactivity to iontophoresis is associated with elevation in urinary ACR.

Data for observational analyses in UK Biobank were available for 9 of 11 exposure traits. Observational analyses provided evidence that higher HDL-C, higher systolic and diastolic blood pressure, higher WHR adjusted for BMI, and T2D were associated with elevated ACR (Table 2). Higher LDL-C, TG, BMI, and body fat percentage were associated with lower levels of ACR (Table 2). The inverse association between higher LDL-C, TG, BMI, and body fat with lower ACR was unexpected but maybe due to treatment effects, confounding, or survival bias, thus highlighting the importance of more robust approaches, like MR.

MR inferred a causal role of higher TG and LDL-C in elevating ACR, with the effect of TG more than twice that of LDL-C. A 1 SD higher TG (∼86 mg/dL) was associated with a 0.062 SD (95% CI 0.040, 0.083) higher ACR (∼9.3 mg/mmol) (Table 3 and Fig. 2), while a 1 SD higher LDL-C (∼37 mg/dL) was associated with a 0.026 (95% CI 0.008, 0.044) SD higher ACR. There was no evidence to infer that higher HDL-C altered ACR. The evidence for a causal role of higher TG in elevating ACR was strengthened using multivariate MR, which adjusted for the association of the TG SNPs with HDL-C and LDL-C. A 1 SD higher TG (adjusted for LDL-C and HDL-C) associated with a 0.094 (95% CI 0.073, 0.115) SD higher ACR (Fig. 2 and Supplementary Table 3). In contrast, multivariate analyses attenuated the association between LDL-C and ACR, with a 1 SD higher LDL-C (adjusted for TG and HDL-C) associated with a 0.018 SD (95% CI 0.001, 0.035) higher ACR (Fig. 2 and Supplementary Table 3). There was no evidence that higher HDL-C adjusted for LDL-C and TG altered ACR.

Results were generally consistent when the more pleiotropy robust methods were used (Table 3). The estimates from the two studies (UK Biobank and CKDGen) and the one-sample MR in UK Biobank were consistent, strengthening the causal inference between TG and ACR (Supplementary Table 4 and Supplementary Fig. 1). Findings for HDL and TG were the same when variants known to be associated with ACR were excluded.

We next tested three measurements of body size and composition: BMI, WHR (adjusted for BMI), and metabolically favorable adiposity.

The MR analyses suggested that higher WHR caused elevated ACR levels, independently of BMI. A 1 SD higher WHR adjusted for BMI was associated with a 0.040 SD higher ACR (95% CI 0.020, 0.059]) (Table 3 and Fig. 3).

MR using the favorable adiposity genetic variants (associated with higher body fat percentage but lower risk of metabolic diseases [26]) showed that metabolically favorable higher adiposity was associated with lower ACR (−0.157 [95% CI −0.256, −0.057], P = 0.002) (Fig. 3).

The MR results for higher BMI were not conclusive, although they were directionally consistent with the WHR results.

Results from alternative MR methods (Table 3) and the study-specific results from the UK Biobank, CKDGen, and the one-sample MR were generally consistent (Fig. 3 and Supplementary Table 4). However, there was weak evidence of heterogeneity for BMI (P = 0.013, I² = 83.9%) and favorable adiposity (P = 0.027, I² = 79.5%).

MR inferred that genetic liability to T2D caused elevated ACR levels, with a twofold higher genetic liability to T2D associated with 0.013 (95% CI 0.007, 0.018) SD higher ACR levels (Table 3 and Fig. 4). There was no evidence of a causal relationship between either fasting insulin or fasting glucose and ACR.

Results were consistent when alternative MR methods were used (Table 3 and Supplementary Table 4) and when the fasting insulin SNP that is also associated with ACR was excluded. The study-specific results from the UK Biobank and CKDGen are presented in Supplementary Table 4 and Fig. 4.

MR confirmed previous evidence (13) for the causal relationship between higher blood pressure and elevated ACR levels. A 1 mmHg higher systolic and diastolic blood pressure was causally associated with a 0.006 (95% CI 0.004, 0.008) and 0.009 (95% CI 0.006, 0.012) SD higher ACR, respectively (Table 3 and Fig. 5).

Results were consistent when alternative MR methods were used, although not all reached P < 0.05 (Table 3). Excluding the one diastolic blood pressure variant that was associated with ACR in an independent study did not alter our findings. Study-specific results from the UK Biobank and CKDGen and the one-sample MR methods in the UK Biobank were generally consistent (Fig. 5 and Supplementary Table 4), although there was evidence of heterogeneity for systolic blood pressure (P = 0.002, I² = 89.6%).

This study used genetic approaches to infer the causal role of 11 metabolic risk factors in ACR, which was considered as a proxy for microvascular dysfunction. First, we confirmed that ACR is a valid proxy for microvascular function, using two gold standard physiological measures of microvascular function in the SUMMIT study: skin endothelial-dependent and -independent microvascular function. We then used genetic variants as unconfounded proxies for the 11 metabolic risk factors to infer that 7 of the 11 metabolic risk factors cause elevated levels of ACR and thus cause microvascular dysfunction.

Skin microcirculation is an established model to investigate systemic microvascular function prior to the clinical manifestation of disease (27). Skin microvascular responses have been demonstrated to be reduced in people with T2D (18) and associated with coronary microvascular function (28). Results presented here support the use of ACR as a proxy for the systemic microcirculation and not just for renal microcirculation.

In keeping with the clinical data, we inferred a causal role of LDL-C and TG in raising ACR levels, with multivariate lipid analyses strengthening the TG association and attenuating the LDL association. Indeed, the effect of TG on ACR is twice as large as the effect of LDL. This contrasts with available evidence for coronary artery disease (CAD) where LDL levels have a larger effect on CAD risk than do TG levels.

While the effect sizes in our results can be seen as small, they represent clinically meaningful results. For example, previous studies have demonstrated that small changes in LDL-C (e.g., 0.2 magnitude lower LDL in mmol/L) results in a 5–10% reduction in the risk of CHD (29). The majority of our analyses look at SD changes in ACR per genetically instrumented SD change in the predictor. For LDL, this equates to ∼0.9 mmol/L higher LDL, which in previous studies would equate to a 15–40% higher risk of CHD.

These results are consistent with those from clinical trials of cholesterol-lowering medication. HMG-CoA reductase inhibitors (statins) predominantly lower LDL-C and have been demonstrated to reduce CAD risk. These drugs, however, only have a small effect on ACR (30) and a similarly small impact on other manifestations of microvascular dysfunction such as diabetic retinopathy (31). In contrast, PPARα antagonists such as fenofibrate, which act predominantly on TG levels, have been shown to have beneficial effect on diabetic nephropathy and retinopathy (32). Combined statin-fenofibrate therapies can provide more endothelial vascular benefits than can statin and fenofibrate alone (33), and, according to the recent results of Action to Control Cardiovascular Risk in Diabetes (ACCORD), fenofibrate appears to be safe with regard to the risk of myositis or rhabdomyolysis when used in combination with a statin (34). Our results suggest that combined therapies lowering TG as well as LDL levels could provide compound benefits by reducing the atherosclerotic burden, and thus CAD, while simultaneously reducing microvascular dysfunction, which has a greater impact on the quality of life on patients (35).

We used three complementary measures of body composition to test the role of adiposity and body fat distribution in the ACR. These three measures were BMI, WHR (adjusted for BMI) as a measure of central adiposity, and favorable adiposity as a measure of higher fat mass “uncoupled” from its adverse metabolic effects (26). Our MR analyses infer that higher WHR (adjusted for BMI) elevates ACR. In contrast, having more favorable adiposity alleles lowers ACR. The favorable adiposity variants are known to associate with higher subcutaneous fat but lower liver fat and lower visceral-to-subcutaneous adipose tissue ratio (26). This provides further evidence that body fat distribution may be important in albuminuria and microvascular problems. Previous studies have suggested a role for body fat distribution and visceral fat in albuminuria, although to date, these studies have had low numbers of participants and have only used observational data and so are subject to more biases than the genetic approach used in this study (10,36,37). A consistent trend was also noted for BMI, with higher BMI trending toward elevated ACR. These results suggest that adiposity and distribution of fat are important in elevating ACR and suggest a causal role for adiposity and fat distribution in microvascular dysfunction.

Our analyses strengthen previous work demonstrating that higher systolic and diastolic blood pressure cause albuminuria (13). Our results confirmed the direction and magnitude of the MR-inferred causal role of systolic and diastolic blood pressure in ACR recently reported (13) and support evidence from clinical trials showing that antihypertensive treatments acting on the renin-angiotensin system reduce ACR (38).

As expected, our MR results confirm that diabetes plays a major role in raising ACR levels. These results add genetic evidence to the large body of data from observational studies and clinical trials clearly showing the role of T2D in causing renal damage. There was no genetic evidence for fasting insulin or fasting glucose levels causing elevated ACR levels. This is in contrast with observational studies showing an association between fasting insulin or fasting glucose and ACR levels (39,40). This may indicate that these observational associations are driven by confounding factors.

The major strength of this study is the availability of data in the UK Biobank and a large independent GWAS sample for testing the causal relationships using two-sample MR approaches. Another strength is the use of multiple rigorous MR methods to establish causality in this analysis. MR provides the next best evidence of causality after randomized control trials and allows causal inferences on large-scale databases such as those used in study.

We acknowledge, however, some limitations. First, MR studies are not immune from some of the issues that affect observational studies. For example, it is possible that biases such as survival bias could have affected the MR as well as observational studies. If, for example, a high ACR and high LDL cholesterol level result in a high mortality rate due to microvascular disease (e.g., stroke), then genetic factors that raise LDL-C level could be depleted from the study and associations between LDL-C–raising alleles and ACR could be weakened. This type of bias has been pointed out before (41). Second, our analyses were restricted to individuals of Caucasian descent and the UK Biobank is restricted to participants born between 1938 and 1971; therefore, the generalizability of our findings may be limited. Third, although multivariate MR was used to explore the role of the three lipids on ACR, there remains the potential for some residual bias due to the pleiotropic associations of the lipid variants, although more pleiotropy-resistant methods generally provided consistent results. Finally, some of our instrumental variables explain only a small percentage of the variability of the outcome variable and therefore we might be underpowered to detect causal association in some of the analysis.

In conclusion, we have used a genetic approach to show the causal role of seven metabolic risk factors in ACR and provided evidence that dyslipidemia, adiposity, and distribution of adipose tissue cause elevations in ACR and thus cause microvascular dysfunction.

Acknowledgments. This research has been conducted using the UK Biobank resource under application no. 9072. The authors acknowledge the use of the University of Exeter High Performance Computing facility in carrying out this work.

Funding. F.C. and J.T. were supported by the Diabetes Research and Wellness Foundation. R.N.B. is funded by Wellcome Trust and Royal Society grant 104150/Z/14/Z. S.E.J. is funded by the Medical Research Council (grant MR/M005070/1). H.Y. is funded by a Diabetes UK RD Lawrence fellowship (17/0005594). A.R.W. and T.M.F. are supported by European Research Council grant SZ-245 50371-GLUCOSEGENES-FP7-IDEAS-ERC. F.C., H.Y., K.M.G., K.A., and A.C.S. are supported by the National Institute for Health Research (NIHR) Exeter Clinical Research Facility. SUMMIT was supported by the Innovative Medicines Initiative (the SUMMIT consortium [IMI-2008/115006]). SUMMIT presents independent research supported by the NIHR Exeter Clinical Research Facility.

The funders had no role in the study design, analysis, or interpretation. All authors confirm their independence from the funders. The views expressed in this publication are those of the authors and not necessarily those of the NIHR Exeter Clinical Research Facility, the National Health Service, the NIHR, or the Department of Health in England.

Duality of Interest. No potential conflicts of interest relevant to this article were reported.

Author Contributions. F.C., T.M.F., and J.T. designed the study. F.C., T.M.F., and J.T. wrote the manuscript. A.C.S., W.D.S., and A.T.H. edited the manuscript and helped interpret the data. F.C., J.T., A.R.W., S.E.J., R.N.B., H.Y., K.M.G., K.A., and F.K. performed data processing, statistical analyses, and interpretation. A.C.S., W.D.S., and K.M.G. obtained funding for and designed and supervised SUMMIT. J.T. is the guarantor of this work and, as such, had full access to all the data in the study and takes responsibility for the integrity of the data and the accuracy of the data analysis.