You are here

Article Text

Article menu

Download PDFPDF

Clinical and epidemiological research
Extended report
Smoking paradox in the development of psoriatic arthritis among patients with psoriasis: a population-based study
  1. Uyen-Sa D T Nguyen1,2,
  2. Yuqing Zhang2,3,
  3. Na Lu2,3,
  4. Qiong Louie-Gao2,
  5. Jingbo Niu2,4,
  6. Alexis Ogdie5,
  7. Joel M Gelfand5,
  8. Michael P LaValley2,
  9. Maureen Dubreuil2,
  10. Jeffrey A Sparks6,
  11. Elizabeth W Karlson6,
  12. Hyon K Choi2,3
  1. 1 Department of Orthopedics and Physical Rehabilitation, University of Massachusetts Medical School, Massachusetts, USA
  2. 2 Department of Medicine, Clinical Epidemiology Research and Training Unit, Boston University School of Medicine, Massachusetts, USA
  3. 3 Division of Rheumatology, Massachusetts General Hospital, Harvard Medical School, Massachusetts, USA
  4. 4 Department of Medicine, Baylor Medical School, Houston, Texas, USA
  5. 5 Departments of Medicine, Dermatology, Biostatistics and Epidemiology, Perelman School of Medicine-University of Pennsylvania, Philadelphia, Pennsylvania, USA
  6. 6 Department of Medicine, Division of Rheumatology, Immunology and Allergy, Brigham and Women’s Hospital, Harvard Medical School, Boston, Massachusetts, USA
  1. Correspondence to Dr Uyen-Sa D T Nguyen, Department of Orthopedics and Physical Rehabilitation, University of Massachusetts Medical School, 55 Lake Avenue North, Worcester, MA 01655, USA; uyensa.nguyen{at}umassmed.edu

Abstract

Objectives Smoking is associated with an increased risk of psoriatic arthritis (PsA) in the general population, but not among patients with psoriasis. We sought to clarify the possible methodological mechanisms behind this paradox.

Methods Using 1995–2015 data from The Health Improvement Network, we performed survival analysis to examine the association between smoking and incident PsA in the general population and among patients with psoriasis. We clarified the paradox using mediation analysis and conducted bias sensitivity analyses to evaluate the potential impact of index event bias and quantify its magnitude from uncontrolled/unmeasured confounders.

Results Of 6.65 million subjects without PsA at baseline, 225 213 participants had psoriasis and 7057 developed incident PsA. Smoking was associated with an increased risk of PsA in the general population (HR 1.27; 95% CI 1.19 to 1.36), but with a decreased risk among patients with psoriasis (HR 0.91; 95% CI 0.84 to 0.99). Mediation analysis showed that the effect of smoking on the risk of PsA was mediated almost entirely through its effect on psoriasis. Bias-sensitivity analyses indicated that even when the relation of uncontrolled confounders to either smoking or PsA was modest (both HRs=~1.5), it could reverse the biased effect of smoking among patients with psoriasis (HR=0.9).

Conclusions In this large cohort representative of the UK general population, smoking was positively associated with PsA risk in the general population, but negatively associated among patients with psoriasis. Conditioning on a causal intermediate variable (psoriasis) may even reverse the association between smoking and PsA, potentially explaining the smoking paradox for the risk of PsA among patients with psoriasis.

  • smoking
  • psoriatic arthritis
  • epidemiology

https://doi.org/10.1136/annrheumdis-2017-211625

Statistics from Altmetric.com

    Request Permissions

    If you wish to reuse any or all of this article please use the link below which will take you to the Copyright Clearance Center’s RightsLink service. You will be able to get a quick price and instant permission to reuse the content in many different ways.

    Introduction

    Psoriasis (PsO) is a chronic inflammatory and autoimmune skin disease that affects over 5 million Americans,1 and psoriatic arthritis (PsA) is a progressive and often destructive joint disease that has been linked to premature cardiovascular (CV) events and mortality.2–4 In most patients with PsA, symptoms do not appear until years after the onset of cutaneous PsO.5 The high risk of PsA among patients with PsO provides a unique opportunity to identify and prevent this serious arthropathy and its complications. However, few modifiable risk factors for PsA have been established among patients with PsO.

    The evidence for smoking as a risk factor for PsA among patients with PsO remains limited and contradictory. Several studies have shown that smoking is associated with an increased risk of PsO and PsA among the general population.6 7 Yet, smoking was found to be inversely8 or weakly associated with PsA in analyses restricted to patients with PsO6 (ie, smoking paradox). Given these conflicting data on this important risk factor on the risk of PsA, we sought to examine these associations in a general population context compared with that in patients with PsO to clarify the methodological reason behind this potential smoking paradox.

    The goals of our study were to examine the smoking and PsA paradox among patients with PsO in a general population context and clarify the possible methodological mechanisms behind the paradox. We first estimated the effect of smoking on risk of PsA in the general population and then among patients with PsO. We then clarified the underlying methodological mechanisms using a unified mediation analysis and bias analysis.

    Methods

    Study population

    Details of The Health Improvement Network (THIN) have been published previously.9 10 Briefly, THIN is a database of computerised medical records from approximately 648 general care practices in the UK. Most patients in the UK are registered with a general practitioner (GP) through the National Health Service. THIN is a population-based cohort of the general population representative of the UK population seen by GPs.9 The database included anonymised healthcare data from approximately 11 million patients, with information on patient demographics, diagnoses including details of the GP visits and specialists’ referrals and hospital admissions and additional health information such as height and weight and lifestyle factors such as smoking and alcohol intake. Moreover, Read codes were used to specify medical diagnoses. For this current analysis, 7 247 774 million people from THIN population met our eligibility criteria which included men and women aged 20–89 years, enrolled in THIN for at least 12 months between 1 January 1995 and 31 May 2015, free of PsA before study entry and included both prevalent and incident patients with PsO. Of these, 596 475 people (8.2%) had missing information on smoking; thus, we included a subset of 6 651 299 participants for the current study. We followed STrengthening the Reporting of OBservational studies in Epidemiology reporting guidelines for observational studies.

    Ascertainment of PsO and PsA

    We used THIN Read codes to define PsO and PsA. Validity of THIN Read codes for PsO and PsA has been previously reported.10 11 Specifically, Read codes for PsO and PsA in THIN showed a positive predictive value (PPV) of 90% and 85% of clinical confirmation, respectively. Also, we performed sensitivity analyses where a person is defined as having PsA if there is a Read diagnostic plus use of disease-modifying antirheumatic drugs (DMARD) within 1 year of the Read code. However, the main definition remains Read code only, as DMARD data definitions only slightly improved the specificity for the diagnosis, but at the cost of a dramatic reduction in sensitivity.12 Both PsO and PsA were classified as a dichotomous variable (Yes or No).

    Assessment of smoking and covariates

    Our exposure of interest was smoking recorded by physicians. Smoking status was divided into three categories: non-smoker, ex-smoker or current smoker. As a lifestyle exposure variable, smoking information in THIN has been collected prospectively and has been used successfully in previous analyses demonstrating anticipated relations of smoking to the risk of myocardial infarction13 and the risk of lung cancer.14 Occasionally, patients were coded in the database as a ‘non-smoker’ but also had a previous code for ‘smoker’ and thus would be recoded as an ‘ex-smoker’. Covariates included age, sex, body mass index (BMI), alcohol and history of trauma at baseline.12 13 15

    Statistical analysis

    Smoking and the risk of PsA in the general population versus among patients with PsO

    First, we calculated the person-time of follow-up and incidence rates of PsA by smoking status. The study entry date in the general population was the date the subject free of PsA met our inclusion criteria for age, study calendar time, 12-month enrolment criteria and the first recorded smoking status, whichever comes last. Follow-up ended when participants developed PsA, became 90 years of age, died, transferred out of the GP practice or the administrative end of follow-up (31 May 2015). Thus, for the purpose of studying incident PsA, patients did not have a Read code for PsA before the start of follow-up and had a Read code for PsA during follow-up.

    We used Cox proportional hazards regression models to estimate the HR for the effect of current or ex-smoking compared with never smoking on risk of incident PsA in the general population, adjusting for covariates. In addition to sex, covariates include age (continuous), BMI (continuous and defined as weight in kg divided by the square of height in metres), alcohol intake (current drinker, ex-drinker or non-drinker) and history of trauma (yes or no) from the most recent date before or on the follow-up start date. We took the same approach for our subgroup analysis to assess the association between smoking and risk of PsA restricted to patients with PsO. The patients with PsO were included in the general population analysis. We also repeated the analyses using five sets of multiple imputation for missing BMI and alcohol data, where covariates for the multiple imputation model included age, alcohol, BMI, sex, trauma, smoking, PsA and follow-up time.16

    Assessing for possible misclassification of smoking, PsO and PsA

    To verify the robustness of the main study findings, we conducted several sensitivity analyses to determine the impact of possible misclassification. First, we used a different smoking classification to define smoking status, that is, smoking was defined based on the last smoking status before diagnosis of PsA, as approximately 14% of individuals repeatedly quit then started smoking again over time. Second, we repeated the analysis by restricting study participants to those with incident PsO (instead of including both prevalent and incident patients with PsO). Third, we used both Read code and DMARD use within 1 year of Read code diagnosis to define PsA. Finally, we required the eligible population to be free of both PsA and PsO at start of follow-up (instead of being free of PsA only).

    Methodological clarification of smoking paradox using mediation analysis with marginal structural models and bias-sensitivity analysis

    Two major methodological reasons can explain this potential paradox: (1) the discrepancy between the intended research question and results obtained by the study design and analytical approach and (2) an index event bias (ie, selection or collider bias), which is introduced when conditioning on a causal intermediate factor. Further details are provided in the methods section in the online supplementary file 1. To examine the role of the first explanation, we conducted mediation analysis using marginal structural models (MSM)17 to partition the total effect of smoking on the risk of PsA into the indirect effect (ie, the effect of smoking on the risk of PsA via PsO) and the direct effect (ie, the effect of smoking on the risk of PsA independent of PsO) (see methods section and figure S1.a in the online supplementary file 1). Then, to address the potential impact of index event bias and quantify its magnitude, we conducted a ‘bias-sensitivity analysis’ as detailed in the methods section in the online supplementary file 1. Further, rheumatic disease examples have also been reviewed in detail, including the ‘obesity paradox’.18

    Supplementary file 1

    All analyses were performed using SAS V.9.3 and adjusted for sex, age, BMI, alcohol intake and history of trauma at baseline.

    Results

    We identified 225 213 participants with incident or prevalent PsO and 7057 with incident PsA over a mean total of 7.0 years (median of 5.5 years) and 46 524 609 person-years of follow-up. In the overall study population, the average age was 42 years, 53% were female, 13% were obese and 62% were current drinkers at baseline. Approximately, 56% participants were non-smokers, 16% were ex-smokers and 28% were current smokers. Among those with PsO, the average age was 45 years, 52% were female, 21% were obese and 64% were current drinkers. Also, about 46% participants were non-smokers, 19% were ex-smokers and 35% were current smokers (table 1).

    View this table:
    Table 1

    Baseline characteristics by smoking status, overall and restricted to participants with psoriasis

    The smoking paradox for PsA

    The associations between smoking and the risk of PsA in the general population and among patients with PsO are shown in table 2. The adjusted HR for the risk of incident PsA comparing current smoking with non-smoking in the general population was 1.27 (95% CI 1.19 to 1.36), but the corresponding HR of the association among patients with PsO was 0.91 (95% CI 0.84 to 0.99), indicating a paradoxical phenomenon. Similar findings were also observed for ex-smokers. Results from multiple imputation of missing data were very similar (table 2). Our results remained robust and inference did not change materially with the various sensitivity analyses, whether with using different definitions of smoking status, restricting the index group to people with incident PsO instead of including both incident and prevalent PsO, a more restrictive definition of PsA with DMARDS or defining the overall study population to be free of both PsA and PsO at the start of follow-up (see results section in the online supplementary file 1).

    View this table:
    Table 2

    Association between smoking and PsA in the general population and among patients with PsO

    Clarification of the smoking paradox for PsA using mediation analysis

    Results from the mediation analysis to clarify the smoking paradox are shown in table 3. The total effect or the net causal effect of current smoking compared with non-smokers on the risk of PsA in the general population was 1.27 (95% CI 1.19 to 1.36), the indirect effect mediated through PsO status was 1.31 (95% CI 1.26 to 1.37) and the direct effect independent of PsO status was 0.96 (95% CI 0.93 to 1.00). The corresponding effect estimates for ex-smokers were 1.32 (95% CI 1.22 to 1.43), 1.21 (95% CI 1.17 to 1.26) and 1.08 (95% CI 1.03 to 1.13), respectively.

    View this table:
    Table 3

    Partitioning the total effect into components of indirect and direct effects using mediation analysis

    Bias-sensitivity analysis to determine potential impact of index event bias

    Bias-sensitivity analyses indicated that even when the relation of uncontrolled confounders to either smoking or PsA was relatively modest (both HRs=~1.5), it could reverse a 10% protective effect of smoking among patients with PsO (ie, HR=0.9) (see figure S2 and results section in the online supplementary file 1).

    Discussion

    In this large cohort representative of the UK population, we found that current smoking was associated with a 27% increased risk of PsA in the general population. However, when we limited the study population to those with PsO, current smoking was associated with approximately a 10% lower (protective) risk of PsA, illustrating the smoking paradox.18 Further analysis revealed that the effect of smoking on the risk of PsA was mediated almost entirely through the effect of smoking on PsO. Moreover, uncontrolled confounding even at a modest level could account for the collider bias resulting in the inverse association between smoking and PsA when the study was restricted to an index event such as PsO. Together, our findings illustrate the smoking paradox associated with the risk of PsA among the patients with PsO in a general population context, and methodological limitations could potentially provide an enticing explanation for the seemingly paradoxical phenomenon.

    Findings of the association between smoking and risk of PsA among people with PsO are limited and inconsistent. For example, Tey et al 19 found no association between smoking and PsA when comparing patients with PsO with PsA (cases) and without PsA (controls). Results from Pattison et al 20 suggested that smoking protects against PsA among cases with PsA compared with PsO controls, that is, smokers had about a 50% reduced risk for PsA. Similarly, Eder et al suggested a protective effect of smoking on risk of PsA among patients with PsO.8 Li et al however, found that smoking had an increased risk of PsA among patients with PsO, but the magnitude of association was substantially smaller than that seen for the association between smoking and risk of PsA in the general population of the Nurses’ Health Study II.6

    Explanations for the lack of consistent findings vary. For example, the biological explanation for the protective effect of smoking was hypothesised that smoking decreased expression of interleukin (IL)-1b, IL-8 and altered response of Toll-like receptor pathways to noxious agents.8 It is unclear how smoking can have a protective effect among patients with PsO but not in the overall general population. Another explanation may be that the effect estimate of smoking on the risk of PsA among the general population may be different from that of the traditional studies restricted to patients with PsO, owing to differences in risk of PsA among non-smokers in the general versus PsO populations.

    Our results suggest that smoking increased risk of PsA in the general population but smoking appeared to be protective among patients with PsO. With the large sample size, both modest effects were statistically significant and we were able to clarify the paradox and showed that not only the measure of effect in those with PsO was that of the direct effect, but was also possibly biased by uncontrolled confounding. When studying the effect of smoking on the risk of PsA among patients with PsO, the goal is to assess the total effect of smoking on the development of PsA. To obtain such an effect estimate, investigators could enroll a group of patients with PsO and assess how changes in smoking status (ie, either smokers stopped smoking or non-smokers started smoking) after PsO diagnosis are associated with the risk of PsA. By doing so, the effect estimate of smoking change on the risk of PsA represents the total effect of smoking on the risk of PsA among patients with PsO. In contrast, traditional studies of the association between smoking and risk of PsA restricted to people with PsO often assess prevalent smoking status at baseline. If smoking is a risk factor for PsO and having PsO increases risk of PsA, then restricting a study to those with PsO would be conditioning on an intermediate in the causal pathway between smoking and PsA. Thus, the measure of effect represents the direct effect of smoking on risk of PsA (see figure S1.a in the online supplementary file 1) independent of PsO. Furthermore, conditioning on an intermediate may also result in collider bias (see figure S1.b in the online supplementary file 1).

    Several limitations warrant discussion. First, GPs in THIN did not regularly record smoking status and other lifestyle factors. It may lead to under-reporting of smoking status, especially among non-smokers or healthy people. Such under-reporting or misclassification of smoking status may result in the effect estimates being biased towards the null. We conducted additional analyses for possible misclassification of smoking and results did not change materially. Second, the diagnostic accuracy of PsA and PsO and disease severity are potential concerns in this study using medical record data. However, validation studies have shown high positive predictive value for PsO (90%) in THIN10 and PsA (>90%) in an electronic medical records database similar to THIN.10 21 Our sensitivity analyses using both Read codes and DMARDs for classification of PsA did not change the inference, as was using incident PsO as compared with both incident and prevalent PsO provided similar inference. While excluding prevalent cases of PsO could better ensure the temporal relation between smoking and the onset of PsO, it will exclude a substantial proportion of individuals with a causal intermediate (ie, prevalent PsO) for PsA endpoints. Furthermore, conditioning on causal intermediates (ie, excluding prevalent PsO individuals) could then lead to potential selection bias. As such, we pursued both analyses, which resulted in very similar findings. Despite these limitations, our study was conducted using a large population-based cohort of the UK population; thus, our findings may apply to a general population. Also, we were able to perform various sensitivity analyses, and the study inference remained the same.

    Conclusion

    Our study showed that traditional study design and analytical methods could result in a risk factor paradox in the context of smoking and risk of PsA among patients with PsO. Future work would need to determine appropriate study design and analysis to ascertain the total effect of smoking in the development of PsA among those with PsO, as this may have critical clinical implications.

    Acknowledgments

    We thank Christine Peloquin for her help with THIN data. Four preliminary findings from our research had been presented at the 2015 and 2016 American College of Rheumatology Annual Meetings.

    References

    1. ↵
      2. Kurd SK ,
      3. Gelfand JM
      . The prevalence of previously diagnosed and undiagnosed psoriasis in US adults: results from NHANES 2003-2004. J Am Acad Dermatol 2009;60:218–24.doi:10.1016/j.jaad.2008.09.022
      OpenUrlCrossRefPubMedWeb of Science
    2. ↵
      2. Gelfand JM ,
      3. Gladman DD ,
      4. Mease PJ , et al
      . Epidemiology of psoriatic arthritis in the population of the United States. J Am Acad Dermatol 2005;53:573.e1–573.e13.doi:10.1016/j.jaad.2005.03.046
      OpenUrlCrossRefPubMed
    3. ↵
      2. Tobin AM ,
      3. Veale DJ ,
      4. Fitzgerald O , et al
      . Cardiovascular disease and risk factors in patients with psoriasis and psoriatic arthritis. J Rheumatol 2010;37:1386–94.doi:10.3899/jrheum.090822
      OpenUrlAbstract/FREE Full Text
    4. ↵
      2. Friedewald VE ,
      3. Cather JC ,
      4. Gelfand JM , et al
      . AJC editor’s consensus: psoriasis and coronary artery disease. Am J Cardiol 2008;102:1631–43.doi:10.1016/j.amjcard.2008.10.004
      OpenUrlCrossRefPubMedWeb of Science
    5. ↵
      2. Ritchlin CT ,
      3. Colbert RA ,
      4. Gladman DD
      . Psoriatic arthritis. N Engl J Med 2017;376:957–70.doi:10.1056/NEJMra1505557
      OpenUrl
    6. ↵
      2. Li W ,
      3. Han J ,
      4. Qureshi AA
      . Smoking and risk of incident psoriatic arthritis in US women. Ann Rheum Dis 2012;71:804–8.doi:10.1136/annrheumdis-2011-200416
      OpenUrlAbstract/FREE Full Text
    7. ↵
      2. Setty AR ,
      3. Curhan G ,
      4. Choi HK
      . Smoking and the risk of psoriasis in women: Nurses' Health Study II. Am J Med 2007;120:953–9.doi:10.1016/j.amjmed.2007.06.020
      OpenUrlCrossRefPubMedWeb of Science
    8. ↵
      2. Eder L ,
      3. Shanmugarajah S ,
      4. Thavaneswaran A , et al
      . The association between smoking and the development of psoriatic arthritis among psoriasis patients. Ann Rheum Dis 2012;71:219–24.doi:10.1136/ard.2010.147793
      OpenUrlAbstract/FREE Full Text
    9. ↵
      2. Blak BT ,
      3. Thompson M ,
      4. Dattani H , et al
      . Generalisability of The Health Improvement Network (THIN) database: demographics, chronic disease prevalence and mortality rates. Inform Prim Care 2011;19:251–5.doi:10.14236/jhi.v19i4.820
      OpenUrlPubMed
    10. ↵
      2. Seminara NM ,
      3. Abuabara K ,
      4. Shin DB , et al
      . Validity of The Health Improvement Network (THIN) for the study of psoriasis. Br J Dermatol 2011;164:602–9.doi:10.1111/j.1365-2133.2010.10134.x
      OpenUrlPubMed
    11. ↵
      2. Ogdie A ,
      3. Alehashemi S ,
      4. Love TJ , et al
      . Validity of psoriatic arthritis and capture of disease modifying antirheumatic drugs in the health improvement network. Pharmacoepidemiol Drug Saf 2014;23:918–22.doi:10.1002/pds.3677
      OpenUrlCrossRefPubMed
    12. ↵
      2. Love TJ ,
      3. Zhu Y ,
      4. Zhang Y , et al
      . Obesity and the risk of psoriatic arthritis: a population-based study. Ann Rheum Dis 2012;71:1273–7.doi:10.1136/annrheumdis-2012-201299
      OpenUrlAbstract/FREE Full Text
    13. ↵
      2. Lewis JD ,
      3. Schinnar R ,
      4. Bilker WB , et al
      . Validation studies of the health improvement network (THIN) database for pharmacoepidemiology research. Pharmacoepidemiol Drug Saf 2007;16:393–401.doi:10.1002/pds.1335
      OpenUrlCrossRefPubMedWeb of Science
    14. ↵
      2. Toh S ,
      3. Rodríguez LA ,
      4. Hernández-Díaz S
      . Use of antidepressants and risk of lung cancer. Cancer Causes Control 2007;18:1055–64.doi:10.1007/s10552-007-9045-1
      OpenUrlCrossRefPubMedWeb of Science
    15. ↵
      2. Thorarensen SM ,
      3. Lu N ,
      4. Ogdie A , et al
      . Physical trauma recorded in primary care is associated with the onset of psoriatic arthritis among patients with psoriasis. Ann Rheum Dis 2017;76:521–5.doi:10.1136/annrheumdis-2016-209334
      OpenUrlAbstract/FREE Full Text
    16. ↵
      2. Raghunathan TE ,
      3. Lepkowski JM ,
      4. Van Hoewyk J , et al
      . A multivariate technique for multiply imputing missing values using a sequence of regression models. Survey Methodol 2001;27:85–95.
      OpenUrl
    17. ↵
      2. Lange T ,
      3. Vansteelandt S ,
      4. Bekaert M
      . A simple unified approach for estimating natural direct and indirect effects. Am J Epidemiol 2012;176:190–5.doi:10.1093/aje/kwr525
      OpenUrlCrossRefPubMedWeb of Science
    18. ↵
      2. Choi HK ,
      3. Nguyen US ,
      4. Niu J , et al
      . Selection bias in rheumatic disease research. Nat Rev Rheumatol 2014;10:403–12.doi:10.1038/nrrheum.2014.36
      OpenUrlCrossRefPubMed
    19. ↵
      2. Tey HL ,
      3. Ee HL ,
      4. Tan AS , et al
      . Risk factors associated with having psoriatic arthritis in patients with cutaneous psoriasis. J Dermatol 2010;37:426–30.doi:10.1111/j.1346-8138.2009.00745.x
      OpenUrlCrossRefPubMedWeb of Science
    20. ↵
      2. Pattison E ,
      3. Harrison BJ ,
      4. Griffiths CE , et al
      . Environmental risk factors for the development of psoriatic arthritis: results from a case-control study. Ann Rheum Dis 2008;67:672–6.doi:10.1136/ard.2007.073932
      OpenUrlAbstract/FREE Full Text
    21. ↵
      2. Singh JA ,
      3. Holmgren AR ,
      4. Krug H , et al
      . Accuracy of the diagnoses of spondylarthritides in veterans affairs medical center databases. Arthritis Rheum 2007;57:648–55.doi:10.1002/art.22682
      OpenUrlCrossRefPubMedWeb of Science

    Footnotes

    • Handling editor Tore K Kvien

    • Contributors Study conception/design: USDTN, YZ and HKC. Data coding/analysis: USDTN, NL, QLG and MD. Manuscript drafting: USDTN, YZ and HKC. All authors substantially contributed to the data interpretation, manuscript revising, critical review and final approval.

    • Funding This current study was partly funded by the American College of Rheumatology RRF Investigator Award, the NIH NIAMS K01AR064351, K23 AR069668, K24AR064310 and P60AR047785. The funders had no involvement in the study design; in the collection, analysis, or interpretation of the data; in the writing of the report; or in the decision to submit the article for publication.

    • Competing interests None declared.

    • Ethics approval Institutional Review Board from the University of Massachusetts and Boston University Medical Schools.

    • Provenance and peer review Not commissioned; externally peer reviewed.

    • Data sharing statement THIN is a licensed proprietary database from IMS Health Real World Evidence Solutions.