Systematic Review and Meta-analysis of Bicycle Helmet Efficacy to Mitigate Head, Face and Neck Injuries Prudence Creighton & Jake Olivier MATHEMATICS & THE UNIVERSITY OF NEW STATISTICS SOUTH WALES Creighton & Olivier (UNSW) Bicycle Helmets Efficacy November 2015 1 / 27
Motivation 1 Motivation 2 Methods 3 Results 4 Discussion Creighton & Olivier (UNSW) Bicycle Helmets Efficacy November 2015 2 / 27
Motivation Past Reviews of Helmet Efficacy There have been 2 systematic reviews and 3 meta-analyses that have assessed bicycle helmet efficacy Thompson, Rivara & Thompson, 1999 1 large protective effect very restrictive selection criteria Attewell, Glase & McFadden, 2001 2 protective effect did not exclude based on study design 1 Helmets for preventing head and facial injuries in bicyclists. Cochrane Database of Systematic Reviews 2 Bicycle helmet efficacy: a meta-analysis. Accident Analysis and Prevention Creighton & Olivier (UNSW) Bicycle Helmets Efficacy November 2015 3 / 27
Motivation Past Reviews of Helmet Efficacy Elvik, 2011 & 2013 3 Updated analysis of Attewell et al. No systematic review for additional included studies Did not follow PRISMA guidelines Three versions published due to numerical mistakes Concludes previous reviews are biased, and protection is reduced Based on analysis of additional studies, concluded that helmets do not have a statistically significant protective overall effect Other methodological errors 3 Corrigiendum to: Publication bias and time-trend bias in meta-analysis of bicycle helmet efficacy: A re-analysis of Attewell, Glase and McFadden, 2001 Accident Analysis and Prevention Creighton & Olivier (UNSW) Bicycle Helmets Efficacy November 2015 4 / 27
Motivation Current State of Literature No systematic search for studies since 1999 Assessment of publication bias requires identifying all relevant studies Lack of formal/standard methods for assessing publication or time-trend bias No reason to believe helmets affect different injuries in the same manner Past reviews lacked methodological rigour Creighton & Olivier (UNSW) Bicycle Helmets Efficacy November 2015 5 / 27
Methods 1 Motivation 2 Methods 3 Results 4 Discussion Creighton & Olivier (UNSW) Bicycle Helmets Efficacy November 2015 6 / 27
Methods Systematic Review Databases: MEDLINE, EMBASE, COMPENDEX & SCOPUS Broad search criteria: helmet* AND (cycl* OR bicycl*) Inclusion criteria Medical diagnosis of injuries, excluded self-reported injury data Helmet status and injuries known for individuals 2 2 table of injury by helmet status, or a reported odds ratio and 95% CI Any apparently relevant study with a full text, English language publication Authors contacted if relevant data not published but appear to meet other criteria Consulted with research librarian and adhered to PRISMA statement Search conducted 2 February 2015 Creighton & Olivier (UNSW) Bicycle Helmets Efficacy November 2015 7 / 27
Methods Injuries Included Head, face or neck injuries of any severity Serious head injuries AIS3+ Glasgow Coma Scale <8 Skull fractures and/or intracranial hemorrhage Reported as serious, severe or brain injury Fatalities where injuries are reported (any injury to the head was counted) Where possible No shell, foam or leather helmets were excluded Controls limited to those injured solely below neck Creighton & Olivier (UNSW) Bicycle Helmets Efficacy November 2015 8 / 27
Methods Meta-analysis Models For the observed log odds ratios y i with variance estimate v i, we fit the following models. Model 1: Baseline random effects model y i = θ i + e i, e i N(0, v i ) θ i = µ + u i, u i N(0, τ 2 ) Model 2: Meta-regression with injury type as a moderator µ = β 1 (head) + β 2 (serious head) + β 3 (face) + β 4 (neck) + β 5 (fatal) Model 3: Random effects for study j θ ij = + b 0j + b 1j (injury type) + u ij Final model chosen by likelihood ratio test and Akaike Information Criterion (AIC) Creighton & Olivier (UNSW) Bicycle Helmets Efficacy November 2015 9 / 27
Methods Assessment of Bias Publication Bias Funnel plot Egger s regression test for asymmetry (univariate models) Rank correlation test for asymmetry (multivariate models) Time Trend Bias Cumulative forest plot Include publication year as covariate in meta-regression (centred at 2014) Bias was assessed in overall model and individual models by injury type Creighton & Olivier (UNSW) Bicycle Helmets Efficacy November 2015 10 / 27
Results 1 Motivation 2 Methods 3 Results 4 Discussion Creighton & Olivier (UNSW) Bicycle Helmets Efficacy November 2015 11 / 27
Full-text articles assessed Full-text articles excluded, PRISMA Flowchart Screening Identification Records identified through database searching (n = 2312) Records after 1173 duplicates removed (n = 1146) Records screened (n = 1146) Additional records identified through other sources (n = 7) Records excluded (n = 1076)
Records screened (n = 1146) PRISMA Flowchart Records excluded (n = 1076) Full-text articles assessed for eligibility (n = 73) Full-text articles excluded, with reasons (n = 35) Studies included in qualitative synthesis (n = 38) Studies included in quantitative synthesis (meta-analysis) (n = 33)
Results Excluded Studies Injury by helmet use not reported per cyclist (n=11) Subset of included study (n=7) Case series (n=5) Self-reported data (n=5) Abstract only (n=3) Unreliable helmet data (n=2) Conflicting information in paper (n=2) There were too few helmeted cyclists or injuries to reliably estimate odds ratio for 5 studies and 2 injury types for an included study Creighton & Olivier (UNSW) Bicycle Helmets Efficacy November 2015 14 / 27
Results Excluded Studies Injury by helmet use not reported per cyclist (n=11) Subset of included study (n=7) Case series (n=5) Self-reported data (n=5) Abstract only (n=3) Unreliable helmet data (n=2) Conflicting information in paper (n=2) There were too few helmeted cyclists or injuries to reliably estimate odds ratio for 5 studies and 2 injury types for an included study 61 total effect sizes included (22 head, 20 serious head, 11 face, 6 neck and 2 fatal) Creighton & Olivier (UNSW) Bicycle Helmets Efficacy November 2015 14 / 27
Results Meta-Analysis Models Model AIC I 2 LRT df p-value Baseline 278.0 78.9% + injury type 166.8 66.1% 119.2 4 <0.0001 + random intercept 106.1 62.7 1 <0.0001 + random slope 103.9 4.2 1 0.041 Model chosen: y ij = θ ij + e ij θ ij = β 0 + β 1 (injury type) + b 0j + b 1j (injury type) + u ij Creighton & Olivier (UNSW) Bicycle Helmets Efficacy November 2015 15 / 27
Results Assessment of Bias Publication bias Rank test (overall model): p=0.22 Regression test (individual injuries): p >0.35 in each instance No strong visual evidence of funnel plot asymmetry Time trend bias Inclusion of year into overall model: p=0.09 Individual models: head (p=0.049), serious head (p=0.13), face (p=0.89) and neck (p=0.96) Year in overall model (AIC=103.2) vs year for head only (AIC=99.1) Final model includes year centred at 2014 for head injuries only (I 2 =61.4%) Creighton & Olivier (UNSW) Bicycle Helmets Efficacy November 2015 16 / 27
Funnel Plot for Final Model Standard Error 1.487 1.116 0.744 0.372 0.000 3.00 2.00 1.00 0.00 1.00 2.00 3.00 Residual Value
Head Injury
Serious Head Injury
Face Injury
Neck Injury
Summary Estimates
Discussion 1 Motivation 2 Methods 3 Results 4 Discussion Creighton & Olivier (UNSW) Bicycle Helmets Efficacy November 2015 23 / 27
Discussion Discussion Systematic review identified 73 relevant studies of which 33 were included in a meta-analysis Effectiveness of bicycle helmets differed by injury type Relative to head injury: face and fatal injuries were similar (p=0.28 and p=0.17), serious head and neck injuries differed substantially (p <0.0001 in each case) No evidence of publication bias Time trend bias only apparent for head injury Creighton & Olivier (UNSW) Bicycle Helmets Efficacy November 2015 24 / 27
Discussion Discussion Systematic review identified 73 relevant studies of which 33 were included in a meta-analysis Effectiveness of bicycle helmets differed by injury type Relative to head injury: face and fatal injuries were similar (p=0.28 and p=0.17), serious head and neck injuries differed substantially (p <0.0001 in each case) No evidence of publication bias Time trend bias only apparent for head injury Helmet use associated with odds reductions of 35% for head injury, 64% for serious head injury, 42% for face injury and 66% for fatal head injury. A non-significant 7% increase in neck injury was estimated Creighton & Olivier (UNSW) Bicycle Helmets Efficacy November 2015 24 / 27
Discussion Limitations Several relevant studies were not included due to lack of or unreliable published data (n=15) Authors were contacted with three supplying data Many studies published over 10 years ago Injury definitions varied substantially from study to study Medium amount of heterogeneity among the effect sizes (I 2 =61%) At least 4 relevant articles have been published since our search Creighton & Olivier (UNSW) Bicycle Helmets Efficacy November 2015 25 / 27
Discussion Acknowledgements I would like to thank Prudence Creighton, UNSW Australia Tim Churches, UNSW Australia Spiros Frangos, NYU School of Medicine, USA Nang Ngai Sze, University of Canterbury, NZ Justin Wagner, David Geffen School of Medicine at UCLA, USA Mike Bambach, University of Sydney Creighton & Olivier (UNSW) Bicycle Helmets Efficacy November 2015 26 / 27
Discussion Thank You! Questions? email: j.olivier@unsw.edu.au blog: injurystats.wordpress.com Creighton & Olivier (UNSW) Bicycle Helmets Efficacy November 2015 27 / 27