Using multivariable mendelian randomization to estimate the bmi-independent causal effect of bone mineral density on osteoarthritis

Purpose: Observational analyses have suggested that high Bone Mineral Density (BMD) is a risk factor for hip and knee osteoarthritis (OA), but it is unclear whether this represents a causal effect of BMD on OA or shared underlying biological pathways (genetic pleiotropy). It is also unclear whether these relationships are independent of the confounding effect of body mass index (BMI). Mendelian randomisation (MR) uses genetic variants robustly associated with an exposure to determine the causal effect of an exposure and an outcome, independent of confounding and reverse causality. We therefore performed two-sample (2S), one-sample (1S), and multivariable (MV)MR analyses to uncover the causal pathways between BMD, BMI and OA. Methods: Single nucleotide polymorphisms (SNPs) associated with BMD estimated from heel ultrasound (eBMD) in the latest genome wide association study (GWAS) in UK Biobank were used to instrument BMD. BMI was instrumented by SNPs from the largest Genetic Investigation of Anthropometric Traits (GIANT) consortium GWAS not including UK Biobank. Hip and knee OA were instrumented by SNPs identified from a meta-analysis of Genetics of Osteoarthritis (GO) consortium cohorts (excluding UK Biobank). Independent SNPs associated with the exposure at genome-wide significance were included as instruments. 362 SNPs were used to instrument eBMD, 69 for BMI, 10 for hip and 4 for knee OA. MR estimates were generated using inverse-variance weighting (IVW) fixed effects meta-analysis. MR-Egger regression was used to identify potential horizontal pleiotropy. 2SMR analyses were performed with the MR-Base R package. 1SMR analyses were performed using individual-level data from UK Biobank. Unweighted allele scores were generated for each exposure. A BMD allele score was generated using 43 independent femoral neck (FN) BMD SNPs identified by the Genetics of Osteoporosis (GEFOS) consortium. Two stage least squares regression was performed using the allele score as the instrument and adjusting for age, sex and 10 principal components (population stratification). 1S MVMR was performed, including both BMD and BMI and allele scores associated with them as instruments in the same model, to determine if there is a causal pathway from BMD to OA, independent of BMI. Results: In 2SMR analyses, there was strong evidence for a causal effect of eBMD on hip OA (OR per SD increase in eBMD=1.09 [1.03,1.16], p=0.002) but weaker evidence for a causal effect on knee OA (OR=1.04 [1.00,1.09], p=0.07). Directions of effect were consistent when using MR-Egger analysis. Effect sizes were stronger, for both outcomes, when restricting to 10 SNPs also associated with FN BMD (GEFOS) at genome-wide significance (p<5x10-8). Results suggested that the causal pathway between eBMD and knee OA was bidirectional, with evidence for a causal effect of knee OA on eBMD (β=0.13 [0.03,0.23], p=0.01, β represents the SD increase in BMD per doubling in OA risk). We did not find evidence for a causal effect of eBMD on BMI. BMI was strongly causally related to hip/knee OA and eBMD (OR=1.51 [1.28,1.78], OR=1.64 [1.44,1.86] and β=0.10 [0.06,0.15], respectively, all p<10-5), with consistent results when using the pleiotropy-robust MR-Egger method. Results were generally consistent using 1SMR, although there was additional evidence for causal effects of hip OA on eBMD and of hip and knee OA on BMI. MVMR analyses identified a BMI-independent causal pathway between eBMD and hip and knee OA (risk difference per SD increase in BMD=0.007 [0.001,0.013], p=0.03 and 0.015 [0.007,0.023], p=10-4, respectively). Conclusions: These Results demonstrate that previously observed relationships between higher BMD and OA partly reflect confounding by BMI. In addition, there is evidence of an independent causal effect of eBMD on OA, however this relationship appears to be bi-directional, suggesting that shared biological pathways contribute to both traits.