You are here

Article Text

Article menu
Clinical and epidemiological research
Extended report
Does smoking protect against osteoarthritis? Meta-analysis of observational studies
  1. Michelle Hui,
  2. Michael Doherty,
  3. Weiya Zhang
  1. Academic Rheumatology, University of Nottingham, Nottingham, UK
  1. Correspondence to Weiya Zhang, Academic Rheumatology, University of Nottingham, Clinical Sciences Building, City Hospital, Nottingham NG5 1PB, UK; weiya.zhang{at}nottingham.ac.uk

Abstract

Objectives To determine whether smoking is protective against the development of osteoarthritis (OA).

Methods Observational studies for the association between smoking and OA were systematically searched through Medline (1950–), Embase (1980–), Web of Science (1960–), PubMed, Google and relevant references. ORs and 95% CIs were directly retrieved or calculated. Current standards for reporting using MOOSE were followed. Quality-related aspects such as study design, setting, sample selection and confounding bias were recorded. Stratified and meta-regression analyses were undertaken to examine the covariates.

Results Of 48 studies (537 730 participants) identified from the systematic literature search, 8 were cohort, 21 cross-sectional and 19 case–control. There was an overall negative association between smoking and OA (OR=0.87; 95% CI 0.80 to 0.94) and subgroup analysis confirmed this in case–control studies (OR=0.82; 95% CI 0.70 to 0.95), but not in cohort (OR=0.92; 95% CI 0.81 to 1.06) or cross-sectional studies (OR=0.89; 95% CI 0.78 to 1.01). Within case–control studies a negative association occurred only in hospital settings (OR=0.65; 95% CI 0.52 to 0.81), not in community settings (OR=0.90; 95% CI 0.75 to 1.08). The association was also seen in knee OA, radiographic OA and smoking as a secondary exposure (covariate or confounding factor). Meta-regression analysis demonstrated that a hospital setting and smoking as a secondary exposure were the major source of the negative association.

Conclusions The protective effect of smoking in OA observed in some epidemiological studies is likely to be false. It may be caused by selection bias, often in a hospital setting where control subjects have smoking-related conditions and studies that are not primarily designed to investigate smoking. Critical appraisal of such studies is needed.

https://doi.org/10.1136/ard.2010.142323

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

    It is accepted that smoking is one of the major risk factors for many conditions such as cancer, diabetes and cardiovascular diseases.1 In contrast, smoking is reported to have a negative association with risk of developing certain conditions such as ulcerative colitis,2 Alzheimer's disease3 and Parkinson's disease.4 However, unlike the convincing evidence for the harm of smoking in many conditions, the evidence for the benefit of smoking is contentious. A recent study in Alzheimer's disease, for example, found that the negative association of this disease with smoking is biased by tobacco industry-funded/affiliated studies.3

    Smoking is associated with the elevated risk of back pain,5 6 chronic widespread pain7 and rheumatoid arthritis.8,–,10 Controversially, smoking has been reported to have a protective association with osteoarthritis (OA).11 12 However, not all studies consistently report this finding.13 14 It is unclear what mechanisms account for this phenomenon or, indeed, if any exist. Proposed reasons for a reduced risk of OA in smokers are the decreased physical sporting activity undertaken by smokers15 and that smokers tend to have lower body weight.16 However, the results are not supported after adjustment for physical activity17 and body weight.12 In contrast, some studies report that smoking may be associated with a greater risk both of cartilage loss and knee pain in OA.18 19

    The objective of this study was to assess whether there is a true negative association through a meta-analysis of observational studies. We hypothesised that a negative association between smoking and OA may result from selecting controls from hospital settings where patients from other departments (eg, cardiovascular) may have higher exposure to smoking than the general population.20 21

    Methods

    Literature search

    A systematic literature search was carried out in May 2009, and updated 15 May 2010, using Medline (1950–), Embase (1980–), Web of Science (1960–), PubMed and Google. The structured search strategy included (a) OA of the knee, hip, hand, spine; and (b) smoking; and (c) cohort, or (d) cross-sectional, or (e) case–control (online supplementary eTable 1). A further keyword search and review of bibliographies of retrieved studies was performed.

    Inclusion/exclusion criteria

    Only published epidemiological studies in OA including smoking as a primary exposure or secondary exposure (ie, covariate or confounding factor) were included. Review articles and editorials were excluded. Studies with more than one report were examined carefully and only one set of data was included. However, studies which reported the cross-sectional association in one report and, subsequently, the causal association (cohort) in another were included separately to examine differences between study designs. Any definition of OA, ranging from radiographic to self-reported physician-diagnosed, was included. Other recognised joint diseases, such as rheumatoid arthritis, were excluded. Studies allowing an assessment of the development or risk of OA were included. Those assessing progression were excluded as this would explore a different question relating both to the natural history of OA and treatment effects. There was no language restriction. Eligible studies in languages other than English were assessed by native-speaking medical colleagues (see ‘Acknowledgements’).

    Data extraction

    Data were fully extracted and assessed by one investigator (MH). Major outcome measures including study design and setting were double-extracted by WZ blinded to the results of MH. Disagreements were discussed and ratified.

    OR, or RR or HR, and their respective 95% CI were essential to allow further analysis to take place. These were either retrieved directly from the article or calculated from available data. Where available, OR adjusted for age, gender and/or body mass index (BMI) were used.

    Study characteristics were recorded (online supplementary eTable 2). Parameters captured included study design, setting from which the control population was drawn, whether smoking was the primary exposure, mean age of subjects, mean BMI, gender split, joint involved, OA definition and country.

    Smoking was defined as ‘ever’, ‘current’ or ‘past’. Results for ‘ever’ taken directly from studies usually represented current and past smokers. A semiquantitative measure for light (≤10 cigarettes/day) and heavy (≥20 cigarettes/day) smoking was also available from some studies for a dose–response analysis. Definitions of OA varied between studies. They were classified as radiographic, clinical, both radiographic and clinically defined, or self-reported.

    Funding resources was carefully reviewed for each paper, and a search for tobacco industry affiliation and affiliated authors was conducted at http://legacy.library.ucsf.edu.

    Quality assessment

    Study design, setting (community or hospital), sample size, case definition, exposure definition, confounding factors and adjustment were assessed. Quality scoring for studies was not performed as it is not possible or fair to assign equal weight to different quality aspects related to the study. However, current consensus standards of reporting meta-analysis of observational studies in epidemiology (MOOSE)22 were followed, and subgroup and sensitivity analysis were performed to examine the changes of the estimate according to different quality aspects.

    For case–control studies, setting was defined according to the source population of the control group. If control subjects were identified either from the community or general population (ie, subjects registered with general practices or termed ‘healthy’) the study was classified as community based. In contrast, if controls were selected from patients attending hospital, it was classified as hospital based.

    Data analysis

    ORs and 95% CIs were used to present the association between smoking and OA. The distribution of ORs and 95% CIs was represented using a forest plot. The Cochran Q test was used to estimate the p value for heterogeneity (pheter).23 The I2 statistic was calculated to demonstrate degree of heterogeneity—that is, the percentage of variation across studies that is not due to chance.24 When statistical pooling was required, the random-effects model was used for heterogeneous results, otherwise the fixed effects model was used.25 Publication bias was examined using the funnel plot and the Egger test.26 Stratified analysis was undertaken as appropriate—for instance, of specific joints, and the respective specific data used. The analysis was undertaken using StatsDirect V.2.6.1 (StatsDirect, Altringham, UK).

    A meta-regression was performed to adjust for covariates. Variables which showed different results in the stratified analysis (eg, negative association in one stratum but not in the other) were selected for the model. The log values of OR were used as dependent variables, and study level variables, including study design, setting, smoking as a secondary exposure, were used as independent variables for this analysis. To ensure a better power of this analysis, we predefined the maximum number of variables as 5 for the basic model according to the empirical evidence of 10 studies for each variable (we had 48 studies; therefore the maximum number of variables is 5).27 Additional variables (eg, adjustment for BMI) were included only to examine the effect of modification. The random-effects model was used to adjust for variances between studies. The analysis was undertaken using SPSS v14 (SPSS, Chicago, Illinois, USA).

    Results

    Selection and characteristics of studies

    In total, 831 citations were retrieved from the literature search. After reading abstracts, 716 were removed owing to irrelevance or duplication. Of 115 remaining articles, hard copies were obtained for further detailed assessment. Four were duplicate publications, 4 published for different purposes with the same smoking data, 5 were progression studies or assessed specifically cartilage volume rather than development of OA, 56 had no or inadequate quantitative data for smoking, and 1 was irretrievable. The author for two of the articles28 29 that reported only the OR without 95% CI was contacted but no response was received. The updated literature search found three more studies published between May 2009 and May 2010. Ultimately, 48 studies (online supplementary eTable 2) with available data for analysis met the inclusion criteria (figure 1).

    Figure 1
    Figure 1

    Flow diagram for study selection.

    Of 48 studies (537 730 subjects) included, 8 (32 298 subjects) were cohort, 21 (491 313 subjects) were cross-sectional and 19 (14 119 subjects) were case–control designs (table 1). Mean ages of participants ranged from 42 to 79 years. More women were found in case–control studies (63.8%), whereas there were more men in cross-sectional and cohort studies. Smoking was predominantly assessed as a secondary exposure or confounding factor; only six studies (12.5%) assessed smoking as a primary exposure.

    View this table:
    Table 1

    Characteristics of the studies

    Test for publication bias

    Although the funnel plot showed visually a symmetrical distribution of published studies for the association between smoking and OA, the Egger test suggested a statistically significant asymmetric distribution towards studies with negative association (Egger test: bias=−0.961; 95% CI −1.865 to −0.057; p=0.038 (figure 2).

    Figure 2
    Figure 2

    Funnel plot for publication bias.

    Association by study design and setting

    The overall summary result showed a negative association between smoking and OA (OR=0.87; 95% CI 0.80 to 0.94) (figure 3). However, there was considerable heterogeneity (I2=73.2%; pheter<0.001). Subgroup analysis according to study design demonstrated no association in cohort studies (OR=0.92; 95% CI 0.81 to 1.06; I2=58.9%; pheter=0.017), or cross-sectional (OR=0.89; 95% CI 0.78 to 1.01; I2=82.4%; pheter<0.001) but a negative association was found among the 19 case–control studies (OR=0.82; 95% CI 0.70 to 0.95; I2=47.9%; pheter=0.001). Further subgroup analysis within case–control studies revealed that only hospital-based case–control studies accounted for the negative association (OR=0.65; 95% CI 0.52 to 0.81; I2=0%; pheter=0.727), whereas community-based case–control studies showed no association (OR=0.90; 95% CI 0.75 to 1.08; I2=55.2%; pheter=0.008). Similarly, a negative association was seen in hospital-based cross-sectional studies (figure 3).

    Figure 3
    Figure 3

    Association between smoking and osteoarthritis.

    Association according to other study level variables

    Primary exposure

    The six studies that examined smoking as a primary exposure showed no association between smoking and OA (OR=1.05; 95% CI 0.83 to 1.32). Where smoking was assessed as a secondary exposure, the association was negative (OR=0.84; 95% CI 0.77 to 0.92) (table 2).

    View this table:
    Table 2

    Association between smoking and OA according to other study level variables

    Type and dose of smoking

    Only current smoking showed a negative association. However, there was no clear dose–response relationship between smoking and OA (table 2).

    OA definition

    A negative association was only seen in radiographic OA (OR=0.79; 95% CI 0.71 to 0.87), not clinical or symptomatic (symptoms plus radiographic changes) or self-reported OA (table 2).

    OA sites

    A negative association was only reported in knee OA (OR=0.86; 95% CI 0.77 to 0.96), not in hip, hand or spine OA (table 2). Further stratification of the 28 knee OA studies according to study design showed that the negative association was apparent only in case–control (OR=0.78; 95% CI 0.64 to 0.95), but not in the cross-sectional (OR=0.91; 95% CI 0.76 to 1.09) or cohort studies (OR=0.90; 95% CI 0.76 to 1.08). While one hospital-based case–control study of knee OA reported a negative association (OR=0.58; 95% CI 0.37 to 0.89), 10 community studies showed no significant association (OR=0.81; 95% CI 0.65 to 1.01).

    Confounding

    The negative association remained (OR=0.88, 95% CI 0.80 to 0.97) regardless of adjustment for any confounding factor (eg. age, gender or BMI), but disappeared when BMI was specifically adjusted (OR=0.93; 95% CI 0.82 to 1.06) (table 2).

    Funding from tobacco industry

    The Framingham study received funding from the Council for Tobacco Research in 1971.30 There was no other affiliation.

    Meta-regression

    As the negative association varied according to different variables, a meta-regression was undertaken to adjust for multiple covariates. Five significant variables selected from the stratified analysis were included in the basic model: case–control study design (yes/no), hospital-based study (yes/no), smoking as a secondary exposure (yes/no), radiographic OA (yes/no) and knee OA (yes/no). Given the adjustment for other variables, only hospital-based studies (OR=0.75; 95% CI 0.58 to 0.99) and studies using smoking as a secondary exposure (OR=0.77; 95% CI 0.62 to 0.95) showed significant negative association (table 3). However, after further adjustment for BMI, the significant negative association in hospital-based studies (OR=0.77; 95% CI 0.58 to 1.02) and studies using smoking as a secondary exposure (OR=0.79; 95% CI 0.63 to 1.00) became marginal, and the OR in the studies adjusted for BMI was 1.08 (95% CI 0.91 to 1.29).

    View this table:
    Table 3

    Meta-regression analysis

    Discussion

    Smoking has been reported to be a ‘protective’ factor against development of OA.11 12 31 However, our meta-analysis of 48 observational studies with over 500 000 participants confirms that this is true mainly in case–control studies, especially hospital-based ones. The association becomes neutral in cohort and cross-sectional studies especially in community-based settings, suggesting that the association is false negative. In addition, we also found a negative association in hospital-based cross-sectional studies; suggesting that the hospital setting may be the major source of bias regardless of study design. The results were confirmed by meta-regression analysis, where hospital setting and smoking measured as a secondary exposure continued to show a negative association, after adjustment for the potential confounding factors identified from the stratified analysis. Furthermore, we did not observe a dose–response relationship between smoking and OA. The negative association was only seen in current smoking but not in ‘ever’ and ‘past’ smoking. These further suggest that it is unlikely that smoking has a causal relationship with OA.

    It is speculated that the negative association between smoking and OA may be confounded by body weight—the major risk factor of OA. This is because smokers are normally thinner than non-smokers,16 and therefore less likely to develop OA. Our subgroup analysis showed that the adjustment for BMI indeed diluted the negative association both in the stratified analysis (table 2) and in the meta-regression analysis, suggesting that BMI may be a confounding factor for the negative association between smoking and OA that should be controlled for in future studies.

    The advantage of a meta-analysis approach to answer this question is that it can examine different study designs and settings by reviewing published literature. Such a question is rarely answered by primary research as each study is often of one single study design or setting. The limitation of hospital-based studies is well known.32 33 Hospital-referred cases are often unrepresentative of the general population. Studies recruiting hospital-referred patients as controls, who have diseases associated with the exposure in question, would lead to an overestimate or underestimate of the relative risk—Berkson's bias.34 35 This problem is particularly pertinent to smoking as smoking may be a risk factor for many conditions that lead people to attend general medical or surgical departments.20 As a result, a higher proportion of smokers are often found in the control group in the hospital setting, therefore generating a false (negative or positive) association with the disease of interest,36 in this case, the negative association with OA. Indeed, there was no mention that illnesses brought on by smoking-related conditions were excluded from the selection criteria of the 19 hospital-based case–control studies included in this analysis. The same problem may occur in a cross-sectional study undertaken with a hospital population. The overall larger negative association from case–control studies could be explained by risk estimate—the OR often overestimates the relative risk when a disease is common.37 Unlike case–control or cross-sectional studies, cohort studies initially separate the population according to exposure (eg, smokers vs non-smokers), before consequences of the exposure. Therefore, there is less likelihood of an effect of confounding by comorbidities that may be caused by the exposure. However, we did not find any hospital-based cohort study. Whether hospital-based cohort studies are prone to more bias remains to be determined.

    The report from the Framingham study was one of the first to generate the hypothesis that smoking may be protective for OA. The negative association was identified in both cross-sectional12 and cohort analyses.31 The reason for this association remains speculative. The Framingham prospective cohort was originally assembled in 1948 with cardiovascular events as the primary end point. A substudy was initiated 35 years later to assess OA as a secondary outcome. There is no evidence that funding for this study influenced the results, especially when the primary end point was not OA but cardiovascular diseases. Perhaps the large proportion of smokers at inception (65.9%)38 may have led to a large proportion of participants having smoking-related conditions and resultant deaths before the occurrence of OA—left censorship. That is, smokers are more likely to die earlier from smoking-related disease and associated lifestyle, whereas OA is a disease of older age so non-smoking ‘survivors’ appear to have a higher incidence of OA. However, this is rather hypothetical. A lifetime survival analysis to compare the incidence of OA, cardiovascular, cancer and other chronic conditions between smokers and non-smokers would be helpful in the future. It would be interesting to determine whether smoking is still found to be associated with OA in the Framingham cohort after adjusting for cardiovascular and other comorbidities.

    There are several limitations to our study. First, smoking is commonly a secondary measure and therefore not necessarily indexed in the databases, which may mean that not all relevant studies are captured. The funnel plot showed a marginal asymmetrical distribution of published studies (the Egger test p=0.038), suggesting that a potential publication bias in favour of the negative association cannot be excluded. Second, like other meta-analyses, results are often heterogeneous and overall statistical pooling is often misleading. We therefore undertook a number of subgroup analyses according to different quality aspects with regard to study design, setting (hospital vs community), primary or secondary exposure, disease classification and adjustment for confounding factors. Even so, we are still unable to explain some of the heterogeneity including that related to the same study design and setting. We pooled results only if heterogeneity could not be explained by potential covariates. Third, the dose–response analysis was based on different studies at different dose levels, therefore may carry significant bias. Fourth, meta-regression analysis was conducted in a small sample size (n=48). Although we restricted the maximum number of variables to 5, it is still under powered, so these results should be viewed with caution. Furthermore, some studies investigated the relationship between smoking and progression of OA.39 Our analysis only examined smoking as a risk factor for development of OA so the results cannot be extrapolated to OA progression. A meta-analysis of disease progression and smoking should be considered in the future to define further the true relationship between smoking and OA.

    In summary, the negative association between smoking and OA found from previous epidemiological studies is likely to be caused by selection bias in the control population. Care must be taken when selecting control subjects from the hospital setting as they may have diseases associated with the study exposure. Such bias may be minimised by post hoc adjustment for confounding factors such as comorbidities and BMI; but a study that is community-based, well-powered and primarily designed for the target question should be considered in the future to identify true association.

    Acknowledgments

    The authors would like to thank Helen Richardson for logistic support, Joanna Ramowski for data collection, Daniel McWilliams for producing the graphs, Tiraje Tuncer and Burkhard Leeb for their help with data extraction from non-English publications, the rheumatology department of Nottingham University Hospitals and the rheumatology department of Royal Derby Hospital for support to MH's research leave.

    References

    1. ↵
      1. Bartecchi CE,
      2. MacKenzie TD,
      3. Schrier RW
      . The human costs of tobacco use (1). N Engl J Med 1994;330:907–12.
      OpenUrlCrossRefPubMedWeb of Science
    2. ↵
      1. Mahid SS,
      2. Minor KS,
      3. Soto RE,
      4. et al
      . Smoking and inflammatory bowel disease: a meta-analysis. Mayo Clin Proc 2006;81:1462–71.
      OpenUrlCrossRefPubMedWeb of Science
    3. ↵
      1. Cataldo JK,
      2. Prochaska JJ,
      3. Glantz SA
      . Cigarette smoking is a risk factor for Alzheimer's Disease: an analysis controlling for tobacco industry affiliation. J Alzheimers Dis 2010;19:465–80.
      OpenUrlPubMedWeb of Science
    4. ↵
      1. Morens DM,
      2. Grandinetti A,
      3. Reed D,
      4. et al
      . Cigarette smoking and protection from Parkinson's disease: false association or etiologic clue? Neurology 1995;45:1041–51.
      OpenUrlAbstract/FREE Full Text
    5. ↵
      1. Deyo RA,
      2. Bass JE
      . Lifestyle and low-back pain. The influence of smoking and obesity. Spine 1989;14:501–6.
      OpenUrlCrossRefPubMedWeb of Science
    6. ↵
      1. Goldberg MS,
      2. Scott SC,
      3. Mayo NE
      . A review of the association between cigarette smoking and the development of nonspecific back pain and related outcomes. Spine 2000;25:995–1014.
      OpenUrlCrossRefPubMedWeb of Science
    7. ↵
      1. Andersson H,
      2. Ejlertsson G,
      3. Leden I
      . Widespread musculoskeletal chronic pain associated with smoking. An epidemiological study in a general rural population. Scand J Rehabil Med 1998;30:185–91.
      OpenUrlCrossRefPubMedWeb of Science
    8. ↵
      1. Sugiyama D,
      2. Nishimura K,
      3. Tamaki K,
      4. et al.
      Impact of smoking as a risk factor for developing rheumatoid arthritis: a meta-analysis of observational studies. Ann Rheum Dis 2010;69:70–81.
      OpenUrlAbstract/FREE Full Text
    9. ↵
      1. Heliövaara M,
      2. Aho K,
      3. Aromaa A,
      4. et al
      . Smoking and risk of rheumatoid arthritis. J Rheumatol 1993;20:1830–5.
      OpenUrlPubMedWeb of Science
    10. ↵
      1. Albano SA,
      2. Santana-Sahagun E,
      3. Weisman MH
      . Cigarette smoking and rheumatoid arthritis. Semin Arthritis Rheum 2001;31:146–59.
      OpenUrlCrossRefPubMedWeb of Science
    11. ↵
      1. Anderson JJ,
      2. Felson DT
      . Factors associated with osteoarthritis of the knee in the first national Health and Nutrition Examination Survey (HANES I). Evidence for an association with overweight, race, and physical demands of work. Am J Epidemiol 1988;128:179–89.
      OpenUrlAbstract/FREE Full Text
    12. ↵
      1. Felson DT,
      2. Anderson JJ,
      3. Naimark A,
      4. et al
      . Does smoking protect against osteoarthritis? Arthritis Rheum 1989;32:166–72.
      OpenUrlPubMedWeb of Science
    13. ↵
      1. Hart DJ,
      2. Spector TD
      . Cigarette smoking and risk of osteoarthritis in women in the general population: the Chingford study. Ann Rheum Dis 1993;52:93–6.
      OpenUrlAbstract/FREE Full Text
    14. ↵
      1. Wilder FV,
      2. Hall BJ,
      3. Barrett JP
      . Smoking and osteoarthritis: is there an association? The Clearwater Osteoarthritis Study. Osteoarthr Cartil 2003;11:29–35.
      OpenUrlCrossRefPubMedWeb of Science
    15. ↵
      1. Järvholm B,
      2. Lewold S,
      3. Malchau H,
      4. et al
      . Age, bodyweight, smoking habits and the risk of severe osteoarthritis in the hip and knee in men. Eur J Epidemiol 2005;20:537–42.
      OpenUrlCrossRefPubMedWeb of Science
    16. ↵
      1. Albanes D,
      2. Jones DY,
      3. Micozzi MS,
      4. et al
      . Associations between smoking and body weight in the US population: analysis of NHANES II. Am J Public Health 1987;77:439–44.
      OpenUrlCrossRefPubMedWeb of Science
    17. ↵
      1. Vrezas I,
      2. Elsner G,
      3. Bolm-Audorff U,
      4. et al
      . Case-control study of knee osteoarthritis and lifestyle factors considering their interaction with physical workload. Int Arch Occup Environ Health 2010;83:291–300.
      OpenUrlCrossRefPubMedWeb of Science
    18. ↵
      1. Amin S,
      2. Niu J,
      3. Guermazi A,
      4. et al
      . Cigarette smoking and the risk for cartilage loss and knee pain in men with knee osteoarthritis. Ann Rheum Dis 2007;66:18–22.
      OpenUrlAbstract/FREE Full Text
    19. ↵
      1. Ding C,
      2. Cicuttini F,
      3. Blizzard L,
      4. et al
      . Smoking interacts with family history with regard to change in knee cartilage volume and cartilage defect development. Arthritis Rheum 2007;56:1521–8.
      OpenUrlCrossRefPubMedWeb of Science
    20. ↵
      1. Morabia A,
      2. Stellman SD,
      3. Wynder EL
      . Smoking prevalence in neighborhood and hospital controls: implications for hospital-based case-control studies. J Clin Epidemiol 1996;49:885–9.
      OpenUrlCrossRefPubMedWeb of Science
    21. ↵
      1. Zhang W
      . Risk factors of knee osteoarthritis–excellent evidence but little has been done. Osteoarthr Cartil 2010;18:1–2.
      OpenUrlCrossRefPubMedWeb of Science
    22. ↵
      1. Stroup DF,
      2. Berlin JA,
      3. Morton SC,
      4. et al
      . Meta-analysis of observational studies in epidemiology: a proposal for reporting. Meta-analysis Of Observational Studies in Epidemiology (MOOSE) group. JAMA 2000;283:2008–12.
      OpenUrlCrossRefPubMedWeb of Science
    23. ↵
      1. Gavaghan DJ,
      2. Moore RA,
      3. McQuay HJ
      . An evaluation of homogeneity tests in meta-analyses in pain using simulations of individual patient data. Pain 2000;85:415–24.
      OpenUrlCrossRefPubMedWeb of Science
    24. ↵
      1. Higgins JP,
      2. Thompson SG,
      3. Deeks JJ,
      4. et al
      . Measuring inconsistency in meta-analyses. BMJ 2003;327:557–60.
      OpenUrlFREE Full Text
    25. ↵
      1. Whitehead A,
      2. Whitehead J
      . A general parametric approach to the meta-analysis of randomized clinical trials. Stat Med 1991;10:1665–77.
      OpenUrlCrossRefPubMedWeb of Science
    26. ↵
      1. Egger M,
      2. Davey Smith G,
      3. Schneider M,
      4. et al
      . Bias in meta-analysis detected by a simple, graphical test. BMJ 1997;315:629–34.
      OpenUrlAbstract/FREE Full Text
    27. ↵
      1. Borenstein M,
      2. Hedges LV,
      3. Higgins JPT,
      4. et al.
      Introduction to Meta-Analysis. Chichester, UK: John Wiley & Sons Ltd 2009.
    28. ↵
      1. Li NH,
      2. Xue QY,
      3. Wang KZ,
      4. et al.
      Non-conditional logistic regression analysis on risk factors of cervical osteoarthritis in the middle-aged and elderly from 6 cities of China. Chin J Clin Rehabil 2006;10:160–2.
      OpenUrl
    29. ↵
      1. Li NH,
      2. Zhang YN,
      3. Zhang Y,
      4. et al.
      Comparison of the risk factors for knee osteoarthritis in middle-aged and elderly populations from six cities of six administration regions of China. [Chinese]. J Clin Rehabil Tissue Eng Res 2007;11:7758–60.
      OpenUrl
    30. ↵
      Leonard Zahn and Associates Inc, Public relations counsel. New Tobacco-Health Grants Announced: One Helps Revive Framingham Heart Study. New York, USA. 9 November 1971.
    31. ↵
      1. Felson DT,
      2. Zhang Y,
      3. Hannan MT,
      4. et al
      . Risk factors for incident radiographic knee osteoarthritis in the elderly: the Framingham Study. Arthritis Rheum 1997;40:728–33.
      OpenUrlCrossRefPubMedWeb of Science
    32. ↵
      1. Kopec JA,
      2. Esdaile JM
      . Bias in case-control studies. A review. J Epidemiol Community Health 1990;44:179–86.
      OpenUrlFREE Full Text
    33. ↵
      1. Lubin JH,
      2. Hartge P
      . Excluding controls: misapplications in case-control studies. Am J Epidemiol 1984;120:791–3.
      OpenUrlFREE Full Text
    34. ↵
      1. Berkson J
      . Limitations of the application of fourfold table analysis to hospital data. Biometrics 1946;2:47–53.
      OpenUrlCrossRefPubMedWeb of Science
    35. ↵
      1. Feinstein AR,
      2. Walter SD,
      3. Horwitz RI
      . An analysis of Berkson's bias in case-control studies. J Chronic Dis 1986;39:495–504.
      OpenUrlCrossRefPubMedWeb of Science
    36. ↵
      1. Sadetzki S,
      2. Bensal D,
      3. Novikov I,
      4. et al
      . The limitations of using hospital controls in cancer etiology–one more example for Berkson's bias. Eur J Epidemiol 2003;18:1127–31.
      OpenUrlPubMedWeb of Science
    37. ↵
      1. Kleinbaum DG,
      2. Kuppler LL,
      3. Morgenstern H
      . Epidemiologic Research - Principles and Quantitative Methods. New York: John Wiley and Sons 1982.
    38. ↵
      1. Christakis NA,
      2. Fowler JH
      . The collective dynamics of smoking in a large social network. N Engl J Med 2008;358:2249–58.
      OpenUrlCrossRefPubMedWeb of Science
    39. ↵
      1. Schouten JS,
      2. van den Ouweland FA,
      3. Valkenburg HA
      . A 12 year follow up study in the general population on prognostic factors of cartilage loss in osteoarthritis of the knee. Ann Rheum Dis 1992;51:932–7.
      OpenUrlAbstract/FREE Full Text

    Supplementary materials

    Footnotes

    • Competing interests None.

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