Mendelian Randomization Studies of Lifestyle-Related Risk Factors for Osteoarthritis: A PRISMA Review and Meta-Analysis

Risk factors for osteoarthritis (OA) often exert effects over protracted time-courses. Mendelian randomization (MR) studies therefore have an advantage over conventional observational studies when studying the causal effect of long-term lifestyle-related risk factors on OA. However, given the heterogeneous design of existing MR studies on OA, the reported causal estimates of these effects remain inconsistent, thus obscuring the true extent of the biological effects of OA lifestyle-risk factors. We conducted a PRISMA systematic review and specifically included MR studies that investigated the causal effect between lifestyle-related risk factors and OA, where causal estimates for various lifestyle factors were pooled for meta-analysis. Quality of studies was assessed according to STROBE-MR guidelines. A total of 1576 studies were evaluated and 23 were included. Overall, the studies included were of high quality and had a low risk of bias. Our meta-analysis demonstrates the positive causal effect of BMI (ORIVW-random effects 1.49 [1.23–1.80]) and negative causal effects of serum calcium (ORIVW-random effects 0.69 [0.57–0.83]) and LDL levels (ORIVW-random effects 0.93 [0.90–0.96]) on OA. Despite the heterogeneous designs and estimates of causal effects provided by various MR studies, our meta-analysis suggests that lifestyle-related risk factors in the form of BMI, serum calcium, and LDL have true biological effects on the development of OA.


Introduction
Osteoarthritis (OA) is a degenerative joint disease that ranks fifth amongst all causes of disability globally [1]. It is characterized by a loss of articular joint cartilage, thinning of the subchondral plate and the formation of osteophytes. The main problem for individuals who suffer from OA is chronic joint pain and impaired mobility that decreases quality of life-those morbid with OA are unable to perform major daily activities and 80% have movement limitations [2]. The World Health Organization predict that around 18% of all women and 10% of all men aged over 60 have osteoarthritis [3].
Whilst cartilage loss and injury are an inevitable result of ageing, not all individuals develop osteoarthritis, therefore research into modifiable lifestyle factors that predispose to the condition may provide ways of improving primary prevention and insight into putative biological pathways that may be targeted by new treatments [4]. The timeframe for risk factors impacting OA takes a long time to develop due to the etiology of the condition, therefore, observational studies in the form of case-control or prospective longitudinal studies may not always be feasible. Risk exposures such as smoking are known to result in other forms of harm if patients are subject to prolonged exposure and so are unethical.
Mendelian randomization (MR) studies are becoming increasingly prevalent and use genetic variants associated with a risk factor as instrumental variables to assess whether there is a causal effect, or if there are spurious associations due to retro-causality on a disease outcome in an observational setting [5,6]. While there exist various MR studies that study 1. Any studies that used MR to investigate the causal effect between any risk factor with osteoarthritis or related phenotypes; 2.
Any studies that used genetic variation as a proxy (Lawlor definition) for an exposure to make causal inferences (not necessarily quantifying causal effect) regarding the effect of the instrumental variable on the outcome of osteoarthritis [12]; 3.
Any studies with genome-wide or phenome-wide association studies (GWAS/PheWAS) that included MR as part of their analysis; 4.
Studies of all sex, age, cohorts, and ethnicities; 5.
The following were used as exclusion criteria: 6. Any studies that were unrelated to risk factors on osteoarthritis outcome (MRI, imaging studies); 7.
Any studies not conducted on humans; 8.
Any studies in which a full English manuscript was not accessible.

Data Extraction and Quality Assessment
Data from included studies were extracted (Table 1) A quality assessment was conducted based on adherence to the Strengthening the Reporting of Mendelian Randomization Studies (STROBE-MR) Guidelines [13]. The guidelines were adapted based on articles that reported quality assessment approaches utilized to document MR studies [14,15]. Upon conversion of the quality assessment score to a percentage, scores of <75%, 75-85% and >85% were considered to indicate high, medium and low risk of bias, respectively, in each of the 23 included studies (Table A1, Appendix B) . Causal effect between habitual coffee consumption and an increased risk of osteoarthrosis, arthropathy and obesity, but some reductions to postmenopausal bleeding.

Meta-Analysis
Data were pooled in a meta-analysis when at least 3 separate studies assessing whether there was a causal effect between the causal factor/genetic instrument with OA were identified. Meta-analyses were performed on data utilizing both the same risk factor and MR technique. Primary outcomes were defined as the association of genetic risk with osteoarthritis diagnosis or risk and were reported as OR with 95% confidence intervals. A two-sided p-value of <0.05 was considered statistically significant. Meta-analyses were performed using the 'meta' package from RStudio (Version 1.4.1717, RStudio, PBC, Boston, MA, USA). The inverse variance method with both fixed and random effects models was used alongside the DerSimonian-Laird estimator for tau 2 in random effects analysis and the Jackson method for confidence interval of tau 2 and tau. All ten studies were included for developing funnel plots to analyze publication bias.

Study Characteristics
The database searches returned a total of 1576 results, of which 560 duplicates were removed to yield 1016 unique records ( Figure 1).

Body Mass Index
We assessed the causal effect between BMI on all OA outcomes using values obtained only by the inverse variance-weighted method [18,23,24,27,28]; this suggested that there was a positive causal effect between BMI and all OA outcomes under both fixed effects

Body Mass Index
We assessed the causal effect between BMI on all OA outcomes using values obtained only by the inverse variance-weighted method [18,23,24,27,28]; this suggested that there was a positive causal effect between BMI and all OA outcomes under both fixed effects   We then meta-analyzed studies that used the MR-Egger methods to include all OA outcomes [17,24,27] to assess whether results would support or rebut those from IVW data. This again showed a positive causal effect between BMI and all OA outcomes under both fixed (OR: 1.03 [1.01-1.05]) and random effects models (OR: 1.80 [1.17-2.77]) ( Figure 3). Findings from the IVW and MR-Egger models were corroborated by MR studies that used various models to calculate OR and CI [17,21,24,27]. They corroborated the positive causal effect of BMI on OA when using fixed effects (OR: 1.03 [1.025-1.04]) or random effects models (OR: 1.36 [1.08-1.71]) ( Figure 4).
Samples were then stratified and analyzed by the ethnicity of the population investigated. In the meta-analysis of studies involving European individuals [21,23,24], there was a positive causal effect between BMI and all OA outcomes under both fixed (OR: 1.03 [1.027-1.04]) and random effects models (OR: 1.27 [1.20-1.34]) ( Figure 5). In the analysis of studies involving White-British subjects [18,27,28], the effect was more marked than in European populations for the fixed and (OR: 1. data. This again showed a positive causal effect between BMI and all OA outcomes under both fixed (OR: 1.03 [1.01-1.05]) and random effects models (OR: 1.80 [1.17-2.77]) ( Figure  3). Findings from the IVW and MR-Egger models were corroborated by MR studies that used various models to calculate OR and CI [17,21,24,27]. They corroborated the positive causal effect of BMI on OA when using fixed effects (OR: 1.03 [1.025-1.04]) or random effects models (OR: 1.36 [1.08-1.71]) ( Figure 4).   outcomes [17,24,27] to assess whether results would support or rebut those from IVW data. This again showed a positive causal effect between BMI and all OA outcomes under both fixed (OR: 1.03 [1.01-1.05]) and random effects models (OR: 1.80 [1.17-2.77]) ( Figure  3). Findings from the IVW and MR-Egger models were corroborated by MR studies that used various models to calculate OR and CI [17,21,24,27]. They corroborated the positive causal effect of BMI on OA when using fixed effects (OR: 1.03 [1.025-1.04]) or random effects models (OR: 1.36 [1.08-1.71]) ( Figure 4).

Serum Calcium
We assessed the causal effect between serum calcium on all OA outcomes using values obtained only by the inverse variance-weighted method [25,29,30].

Serum Calcium
We assessed the causal effect between serum calcium on all OA outcomes using values obtained only by the inverse variance-weighted method [25,29,30].  There are no standardized tools to ascertain the risk of bias in MR studies. Thus, to assess the quality of the included studies, we evaluated whether the three assumptions of MR were validated, and the method used for those validations. We constructed funnel plots to assess publication bias under both fixed effects ( Figure 9) and mixed effects ( Figure 10) meta-regression models when the standard error was used as the predictor. Under both conditions, there was no evidence of publication bias from the test for funnel plot asymmetry. The p-values for fixed and mixed effects (restricted maximum likelihood) were 0.159 and 0.822, respectively. We also constructed funnel plots when the sampling variance, inverse standard error and inverse sampling variance were used as predictor ( Figures A1 and A2, Appendix B).  There are no standardized tools to ascertain the risk of bias in MR studies. Thus, to assess the quality of the included studies, we evaluated whether the three assumptions of MR were validated, and the method used for those validations. We constructed funnel plots to assess publication bias under both fixed effects ( Figure 9) and mixed effects (Figure 10) meta-regression models when the standard error was used as the predictor. Under both conditions, there was no evidence of publication bias from the test for funnel plot asymmetry. The p-values for fixed and mixed effects (restricted maximum likelihood) were 0.159 and 0.822, respectively. We also constructed funnel plots when the sampling variance, inverse standard error and inverse sampling variance were used as predictor ( Figures A1 and A2, Appendix B).

Discussion
The main findings of this systematic review and meta-analysis were that BMI had a positive causal effect and that serum calcium and LDL levels had a negative causal effect on all outcomes of osteoarthritis. These findings were consistent with those reached in our

Discussion
The main findings of this systematic review and meta-analysis were that BMI had a positive causal effect and that serum calcium and LDL levels had a negative causal effect on all outcomes of osteoarthritis. These findings were consistent with those reached in our

Discussion
The main findings of this systematic review and meta-analysis were that BMI had a positive causal effect and that serum calcium and LDL levels had a negative causal effect on all outcomes of osteoarthritis. These findings were consistent with those reached in our qualitative analysis for BMI [17,18,21,23,24,27,28,35,38] and serum calcium [25,29,30]. Our findings for LDL matched those reached by Hindy et al. and Karhunen et al. [17,18], only differing from the lack of causality conclusion reported by Funck-Brentano et al. [23]. Overall, the studies included were of high quality and had a low risk of bias.

Body Mass Index
Our results supported those of the studies included in our systematic review that there is a direct causal effect of BMI on OA [17,18,21,23,24,27,28,35,38]. This is replicated in other genome-wide analyses [40]. OA could be in part due to systemic metabolic dysfunction secondary to high BMI, obesity and dyslipidemia [41,42]. The protective effect of education found by Karhunen et al. has been suggested to be mediated through BMI and smoking despite high uncertainty values of mediated proportion between 13% and 57% [18]. Education may have a downstream protective effect related to mechanisms such as greater engagement with healthcare practices and increased self-management, thus, reducing BMI. Funck-Brentano et al. showed that high BMI has a causal effect on increasing the risk of knee and hip OA but not hand OA [23]. Interestingly, after exclusion of genetic instruments that were also associated with BMI, other metabolic factors (HDL, triglyceride, CRP, type 2 diabetes) had no causal effect on developing OA in the same study.
BMI has its limitations as a measure of obesity as it does not account for body composition, age-related lean muscle/fat changes and sex differences so further investigation into how these factors alongside BMI affect OA risk is required [43]. Raud et al. demonstrated a dose-response relationship between BMI and knee OA [44], whilst Reyes et al. demonstrated an increased risk of hip and knee OA in overweight individuals [45]. Similar findings by He et al. from robust MR methods support the main implication that weight control would be a significant intervention in the management of OA [24]. Overall, the etiology of how BMI exerts a causal effect on OA risk has not been fully deduced. Therefore, it was postulated that whilst obesity could initially cause OA changes in the weight-bearing joints, there was a role of synergistic effects such as comorbidity, lifestyle factors and lack of mobility that resulted in further increasing the risk of severe OA symptoms.

Serum Calcium
For calcium, the results of our meta-analysis supported the conclusions reached in studies captured by our systematic review [25,29,30]. Previously, a cross-sectional study conducted on 2855 Chinese individuals indicated that serum calcium concentration was inversely associated with the risk of knee OA diagnosed through radiographic means [46]. Although the three studies included in our meta-analysis used different target populations, the causal effect found was similar [25,29,30]. However, Hunter et al. did not find a causal effect between serum calcium levels and OA in female Caucasian twins [47]. Similarly, Zoli et al. found no causal effect between serum calcium levels and hand OA in their case-control study [48]. Due to conflicting conclusions reached by multiple studies, further research into the mechanisms through which calcium has a causal role in OA at both the site and sex level is required. Furthermore, the modifiable nature of mineral status lends itself to major clinical implications in preventing and screening of OA if the correct causal effect is identified.

Low-Density Lipoproteins
The results of our meta-analysis were consistent with the conclusions reached in most studies included in our systematic review [17,18,23]. There was a weak negative causal effect between the reduction in LDL levels and the increase in risk of knee and hip OA found by Funck-Brentano et al. [23]. However, these results contrasted with previously conducted observational studies that suggested increased levels of serum cholesterol were a risk factor for OA [49,50]. It has been previously hypothesized that lipid accumulation in the cartilage of weight-bearing joints could contribute to the pathogenesis of OA [49] and similar results have been shown in experimental animal models [51]. Previous epidemiological studies have reached differing conclusions on the success of statins in preventing OA. Some found that statins lowered risk [52,53], whilst others have found that statins increased risk [54], or found no association [55]. Our findings from the meta-analysis support this alternative hypothesis, that increased LDL levels may serve as a protective effect in lowering the risk of OA diagnosis. There was only a modest OR of protective effect demonstrated for LDL on OA risk relative to the effects seen in education, BMI and smoking so further investigation is required to ascertain whether it is of clinical relevance [18].

Strengths and Limitations
To our knowledge, this systematic review and meta-analysis is the first to evaluate Mendelian randomization studies that assess how different causal risk factors affect the risk of OA, and the first to provide pooled estimates of the causal effect between BMI, serum calcium and LDL on the risk of OA. Prior to this review, causal estimates provided by various independently conducted MR studies, as demonstrated in Figures 2, 7 and 8 for BMI, serum calcium, and LDL, respectively, were heterogeneous and sometimes inconsistent. Moreover, previous observational studies detailed in Sections 4.1-4.3 provided conflicting associations and biological hypotheses. By providing pooled causal estimates, our review has disambiguated the true extent of the biological effect of lifestyle-related risk factors on OA.
Our quality assessment using STROBE-MR found that 20 out of 23 studies were at a low risk of bias in terms of study design, methodology, and quality of data (Table A1, Appendix B). In addition, no significant publication bias was found when examining the included studies in Figures 9 and 10. Overall, the low risk of methodological and publication bias suggests that our included MR studies are of high quality and is indicative of the robustness of our pooled causal estimates for BMI, serum calcium, and LDL on OA.
In studies utilizing the UK Biobank cohort [16,18,[20][21][22][23][24][25]27,29,30,[33][34][35][36][37][38], the causal effects represented only the selected European or White-British populations. These were selected predominantly to avoid population stratification. However, these conclusions may not be representative of more general or non-European populations. Similarly, a comparison of conclusions reached across different age demographics may not be feasible. Some studies identified in our meta-analysis utilized older GWAS datasets such as ArcOGEN [19,28,32] that are no longer the most comprehensive available to identify outcome SNPs. More updated datasets, such as those by Zengini et al., should be used [40]. The inclusion of different studies that use overlapping cohorts (particularly UK Biobank), while unavoidable, inevitably reduces the effective sample size and statistical power of our meta-analyses. BMI was the most investigated risk factor in our meta-analysis, but the SNPs used to instrument it varied across cohorts. In some cases, weak causal effects or the absence of any causal effects could not be confirmed since the genetic instruments and sample size used may have been underpowered to detect them [17]. Even when using the MR-Egger method that exploits the assumptions of instrumental variables to estimate pleiotropy, misclassification in these polymorphisms is still possible. Further analyses, utilizing multiple methodologies such as IVW, MR-Egger and median-based tests within a given study will allow the findings to be rigorously tested.
Sub-grouping by other individual-level baseline characteristics such as age, genealogical ancestry and sex may reveal further insight. The lack of sex-stratified GWAS data used in the studies of our systematic review may mean that the sex-specific phenotypic differences remain undetected.
Most of the meta-analyses displayed substantial degrees of heterogeneity and so the conclusions reached should be interpreted with caution despite the usage of random effects models [56]. This heterogeneity can have a greater impact on pooled estimates that report a weak causal effect. The range of meta-analyses that could be conducted was limited by the small number of previous studies that assessed the same causal risk factor. Pooling data on OA at both different sites and diagnosis methods may also have limited our findings as it is clear from our meta-analysis that determinants of OA vary by site. Studies that evaluated hospital-diagnosed OA may have been based on radiographic changes and are therefore likely to be reliable. Self-reported OA, on the other hand, can be unreliable due to its subjective nature and may be subject to recall bias or misclassification that could affect the estimates generated. Furthermore, self-reported cases could indicate other joint pathologies, especially in individuals who have not yet experienced any symptoms of pain but have other symptoms of OA, such as crepitus and swelling. This may exaggerate the reported causal effect.
The Mendelian randomization approach utilizes cumulative lifelong effects of genetic variants and therefore cannot be extrapolated to study the potential effects of clinical treatment [18]. This approximation of the average effect may not be useful in the case of blood mineral levels whose levels vary constantly throughout life [29]. If bidirectional associations are not assessed, reverse causation cannot be completely discounted even if the assumptions for instrumental variable assignment hold. MR approaches aim to remove confounding biases, but it is possible that null results could have been due to misclassification of OA diagnosis [23]. This remains a major challenge for observational and genetic studies of osteoarthritis.

Clinical Implications
Given the moderate causal effect of increasing BMI has on the development of OA, clinical guidelines ought to recommend weight control as a means of primary prevention. Indeed, institutions such as the European Alliance of Associations for Rheumatology and American College of Rheumatology have already recommended 'weight loss' as a non-pharmacological means for the self-management of OA [57,58]. We suggest that the next step would be to extend their recommendations to the general population to reduce the population-wide incidence of OA through more aggressive control of BMI in earlier life. Similarly, the moderate causal effect of lower serum calcium on OA would suggest that clinical recommendations pertaining to adequate dietary calcium intake and regular weight-bearing exercise to prevent calcium loss would be useful to the general population to reduce the incidence of OA as primary prevention.
Despite our finding of a weak causal effect of lower LDL on the development of OA, i.e., LDL is a weak protective factor of OA, we do not recommend increasing dietary LDL intake since it is significantly associated with an increased risk of numerous cardiovascular co-morbidities [59].

Future Research
Future research into more quantifiable measures for lifestyle-related risk factors with MR studies such as micronutrient intake of various minerals and vitamins would be useful to examine the impact of various diets and foodstuffs on OA or other musculoskeletal diseases. Furthermore, MR studies on OA from GWAS cohorts of other ethnicities such as East Asians, South Asians, South Americans, etc. would provide insight into how different genetic compositions and different environments/lifestyles stemming from inter-ethnic diversity may result in different magnitudes of causal effect for certain lifestyle-related risk factors on the development of OA.

Conclusions
Osteoarthritis (OA) is a significant public health issue affecting 303 million people. It is therefore critical that recommendations for management and primary prevention focused on modifiable risk factors are informed by robust research that establishes a causal effect with osteoarthritis, beyond the delimited information of non-causal associations reported by conventional observational studies [60].
Despite the heterogeneous designs and findings of independent MR studies studying the causal effect of various lifestyle-related risk factors on OA, our systematic review and meta-analyses of these MR studies have conclusively demonstrated the extent of the positive causal effect of BMI and that of the negative causal effects of serum calcium and LDL levels on osteoarthritis.
Given the established biological causality between lifestyle-related risk factors and the development of OA, future studies should aim to study how precisely modulating these risk factors with pharmacological and non-pharmacological strategies could modify the incidence and progression of osteoarthritis in various joints as per a dose-response relationship [61,62].