The Comparative Effect of Exposure to Various Risk Factors on the Risk of Hyperuricaemia

Diet Has a Weak Causal Effect

Ruth K. G. Topless; Tanya J. Major; Jose C. Florez; Joel N. Hirschhorn; Murray Cadzow; Nicola Dalbeth; Lisa K. Stamp; Philip L. Wilcox; Richard J. Reynolds; Joanne B. Cole; Tony R. Merriman


Arthritis Res Ther. 2021;23(75) 

In This Article

Participants and Methods

Participants and Data Collection

The attributable fraction analysis included four distinct cohorts of European ancestry (Table 1)—cohorts 1 to 3 are population-based and the fourth cohort is comprised entirely of people with gout.

Cohort 1 comprised 14,247 participants of European ancestry from the US population—7342 from the Atherosclerosis Risk in Communities (ARIC) Study, 1314 from the Coronary Artery Risk Development in Young Adults (CARDIA) Study, 2513 from the Cardiovascular Health Study (CHS) and 3078 from the Framingham Heart Study (FHS). These numbers exclude people without serum urate measurements or genome-wide genotypes, along with individuals aged under 18 years, people with kidney disease or gout and those taking urate-lowering therapy. People who answered less than 10% of the food frequency survey, those whose estimated average daily calorie intake was less than 600 kcal/day or greater than 4200 kcal/day and those whose questionnaire answers were deemed unreliable by the study interviewer at recruitment were also excluded. For ARIC, one person from each first degree-related pair was excluded.

Cohort 2, cohort 3 and the gout cohort were sourced from the UK Biobank resource. Subjects of European ancestry and who had urate measures and genotypes available were included in the analysis. Relatives with kinship coefficients > 0.177 were removed, and one person from each relationship was kept, with a preference for keeping gout-affected participants. Those who self-reported having kidney disease were also removed. The gout cohort comprised people who self-reported having gout at visit 0 or were being treated with urate-lowering therapy (n = 6781),[31] this case definition has been validated.[31,32] Cohort 2 consisted of UK Biobank participants who answered a 24-h dietary recall questionnaire during the assessment visit (n = 57,251) and cohort 3 consisted of the remaining UK Biobank subjects who answered a reduced food frequency questionnaire (n = 347,526). We excluded, from cohort 2, subjects who had energy intakes in excess of 18,000 kJ for females and 20,000 kJ for males based on their 24-h dietary recall data, those who had unreliable dietary data as flagged by the recruiter, or subjects not eating normally due to illness or fasting. No additional exclusions were applied for cohort 3 or the gout cohort. Participants used for the MR analysis were also from the UK Biobank, and the MR cohort has been described previously.[26]

Collection of dietary and serum urate data are described in the Supplemental material. Collection of genetic data for the ARIC, FHS, CHS and CARDIA cohorts is described in ref.[10] and for the UK Biobank in.[33]

Data Dichotomisation

To calculate a PAF, all exposure and outcome variables must be dichotomous. The outcome for this study was HU, defined as serum urate ≥0.42 mmol/L [≥7 mg/dL] for men and women.[34] For the genetic exposures, HU risk alleles were defined as urate-increasing under a dominant model.[35] Determination of dichotomised dietary exposures is described in the Supplemental material. Alcohol exposure was defined as > 1 drink per week, being overweight/obese as BMI ≥25 kg/m2, age was dichotomised as ≥50 versus < 50 years partly in order to capture menopause as a risk factor in women and diuretic use either self-reporting or not self-reporting diuretic intake—these variables were the same as those for which PAR estimates were calculated in the Third National Health and Nutrition Examination Survey in ref.[9] and for age. In the gout cohort, self-reported treatment with urate-lowering therapies, allopurinol (n = 4841), probenecid (n = 3) and sulphinpyrazone (n = 21) (the only three urate-lowering medications for which baseline medication data were available) was a dichotomised exposure variable.

Statistical Analysis

All analyses were performed using R v3.6.1 in RStudio 1.2.5019. For the various exposures, the PAF calculation was (frequency of exposure in cases) × (ORExposure – 1)/ORExposure).[36] Odds ratios for the risk of HU for these exposures were calculated in a logistic regression multivariable model including all other environmental and endogenous exposure variables and SLC2A9 rs12498742 genotype—this variant was chosen for individual focus because of its large effect on serum urate levels.[35] For the percent variance explained analysis (Table 3), effect sizes (β) on serum urate levels for the same exposures were calculated in a linear regression multivariable model including all other environmental and endogenous exposure variables and SLC2A9 rs12498742 genotype. Average attributable fractions (AAFs), adjusted for all other exposure variables and rs12498742 genotype (Table 2) were calculated using a multivariable model in the R function averageAF,[36] described in more detail in Supplemental material. Age was removed as it did not confer risk in the gout cohort. In general, AAFs are lower than PAFs calculated from the same data (as consistently observed here) and it has been proposed that they reflect the most plausible/reliable result across the many different methods of calculating attributable fractions.[37]

For genetic variants in Table S2, PAFs (and odds ratios) were calculated per allele under an additive model using the method of Rockhill et al.[36] ((frequency risk allele in cases) × ((ORRiskAllele – 1)/ORRiskAllele)). However, PAFs (and odds ratios) in Table 2, Table S1 and Figure 1 were calculated under a dichotomised (dominant) model with the exposed group defined as those with one or more risk allele and were adjusted for dichotomised age, sex, BMI, diet, alcohol and diuretic exposure risks. All AAF analyses used a dominant model for SNP analysis because risk was dichotomised.

Figure 1.

Population attributable fraction, average attributable fraction and variance explained for environmental and endogenous risk exposures for hyperuricaemia or serum urate levels. PAF – population attributable fraction; AAF – average attributable fraction; R 2 – partial R 2 value (R 2 B) converted to a percentage (R 2 * 100). Population attributable fraction and average attributable fraction values relate to hyperuricemia as a dichotomous variable where for the SLC2A9 rs12498742 A-allele PAF and AAF was calculated under a dominant model; R 2 relates to serum urate as a continuous variable. All risk factors are dichotomous. All analyses are adjusted in a multivariable model for the other risk factors. #Non-adherence to DASH guidelines; *Non-adherence to the Harvard Healthy Eating Pyramid guidelines. Missing bars in Figure 1d reflect age not being a risk factor in this group

Mendelian Randomisation

Two-sample MR using the MendelianRandomization R package[38] was tested for a causal role of dietary habits in determining urate levels. GWAS summary statistics for urate levels in Europeans were obtained from ref.[3] comprising 288,649 individuals; 101 loci with non-ambiguous lead SNPs where no strand assignation issues could have arisen by GWAS which can occur with A/T and G/C variants where the alternative allele is the same as the potential strand-flipped allele. This conservative approach was taken in order to harmonise effect alleles from a meta-analysis of multiple studies using multiple imputation panels. In order to test that this QC did not influence the results, we conducted a sensitivity analysis and repeated the primary Mendelian randomisation with the inclusion of all ambiguous SNPs (Figure S1)—the results were in high concordance. GWAS data for dietary habits were obtained from a study of 455,146 individuals of European ancestry from the UK Biobank.[26] The latter were derived from consumption patterns for 47 single foods asked about in the reduced food frequency questionnaire or 40 principal component-derived dietary patterns with ≥3 non-ambiguous genome-wide significant index SNPs reported in ref.[26] that were also present in the serum urate GWAS summary statistics.[3] Independence of loci was based on the distance-based pruning approach within the source GWAS[3,26]—loci had to be > 500 kb apart and if there were two in a 1 Mb interval independence was established by plotting and visualisation of inter-marker linkage disequilibrium. The MR was performed using three methods available in the MendelianRandomization R package—inverse-variance-weighted meta-analysis,[39] MR Egger (enables detection of pleiotropy[40]) and weighted median (robust to pleiotropy[41]). Significance thresholds were set at P < 0.05/87 (5.8 × 10− 4) for the inverse-variance-weighted analysis, P < 0.05 for the weighted median analysis and intercept P > 0.05 for the MR Egger analysis. Our strategy was to identify causal effects from the inverse-variance-weighted analysis corrected for multiple testing followed by sensitivity analysis by weighted median and MR Egger to test the robustness of the results. All three significance thresholds had to be met for a dietary habit to be considered causal. Multivariable MR was conducted to investigate the possible upstream impact of BMI as a common causal link between dietary patterns and serum urate levels using the likelihood-based method on summary data as described by Burgess and Thompson,[42] as implemented in the MendelianRandomization R package. This technique uses pleiotropic genetic variants to estimate the direct effect of multiple exposures on an outcome (e.g. BMI and diet on serum urate), with causal estimates representing the independent causal effect of each exposure on the outcome, not operating through the other exposure included in the analysis.