- Email: [email protected]
Head injuries to bicyclists and the New Zealand bicycle helmet law
Accident Analysis and Prevention 32 (2000) 565 – 573 www.elsevier.com/locate/aap
Head injuries to bicyclists and the New Zealand bicycle helmet law Paul Scuffham a,b,* , Jonathan Alsop a, Colin Cryer c, John D. Langley a b
a Injury Pre6ention Research Unit, Uni6ersity of Otago, Dunedin, New Zealand Centre for Health Economics Research and E6aluation, Sydney Uni6ersity, Sydney, NSW, Australia c South East Institute of Public Health, Tunbridge Wells, Kent, UK
Received 1 February 1999; received in revised form 15 June 1999; accepted 21 June 1999
Abstract The purpose of this study was to examine the effect of helmet wearing and the New Zealand helmet wearing law on serious head injury for cyclists involved in on-road motor vehicle and non-motor vehicle crashes. The study population consisted of three age groups of cyclists (primary school children (ages 5–12 years), secondary school children (ages 13 – 18 years), and adults (19 + years)) admitted to public hospitals between 1988 and 1996. Data were disaggregated by diagnosis and analysed using negative binomial regression models. Results indicated that there was a positive effect of helmet wearing upon head injury and this effect was relatively consistent across age groups and head injury (diagnosis) types. We conclude that the helmet law has been an effective road safety intervention that has lead to a 19% (90% CI: 14, 23%) reduction in head injury to cyclists over its first 3 years. © 2000 Elsevier Science Ltd. All rights reserved. Keywords: Bicycle; Helmet; Head injury; Legislation
1. Introduction Head injuries due to cycle crashes have been identified as a significant cause of morbidity in New Zealand (Collins et al., 1993). Case – control studies show that the use of a correctly designed and fitted cycle helmet is an effective strategy for reducing the frequency of head and upper facial injury (Thompson et al., 1989, 1990, 1996a,b, Maimaris et al., 1994; Thomas et al., 1994 Acton et al., 1995). Cycle helmets are also thought to reduce the severity of head injury. Few studies examine the relationship between head injury severity and the protection afforded by helmet wearing. Thompson et al. (1996a) show that cycle helmets reduce the risk of a brain injury by 88% and the risk of a severe brain injury is reduced
* Corresponding author. Present address: School of Public Health, Queensland University of Technology, Victoria Park Road, Kelvin Grove, Qld 4059, Australia. Tel.: +61-2-38645724; fax: 61-238643369. E-mail address: [email protected] (P. Scuffham)
by 75%, (i.e. helmets afford less protection against more severe injuries). Shafi et al. (1998) showed that helmeted child cylists with a head injury admitted to hospital were more likely to have concussion and less likely to have a serious brain injury compared to head injured non-helmeted child cyclists (i.e. there is possibly a migration of injury severity from more to less severe). In an effort to reduce head injuries to cyclists, voluntary helmet use has been widely promoted in New Zealand. Sustained programmes promoting regular helmet use saw helmet wearing rates steadily increase from virtually zero in 1986 to 84%, 62%, and 39% for primary school children (aged 5–12 years), secondary school children (13–18 years), and adult commuters (over 18 years), respectively, in September 1992. This was one year prior to the introduction of the helmet law. The New Zealand cycle helmet regulation, effective from 1 January 1994, requires all cyclists to wear a standard approved cycle helmet for all on-road cycling. At the time when this legislation was introduced, helmet wearing rates increased to above 90% for the three cyclist age groups. For the first 2 years following the
0001-4575/00/$ - see front matter © 2000 Elsevier Science Ltd. All rights reserved. PII: S 0 0 0 1 - 4 5 7 5 ( 9 9 ) 0 0 0 8 1 - 0
566
P. Scuffham et al. / Accident Analysis and Pre6ention 32 (2000) 565–573
introduction of the law, wearing rates for three age groups of cyclists exceeded 92%, with a slight decline in the following year. An earlier New Zealand study of the effect of voluntary helmet wearing on serious head injury to cyclists (Scuffham and Langley, 1997) found little association between the percentage of cyclists wearing helmets and the percentage of head injuries to injured cyclists. This present study sought to examine the association between helmet wearing rates and the rates of admission to hospital for head injury to cyclists injured on public roads. Our principal hypothesis was that serious head injuries to cyclists decreased as cycle helmet wearing rates increased. We also sought to examine the effect of helmet wearing for different types of head injury (lacerations, intracranial injury including concussion, skull fractures). The type of injury is examined as it is more meaningful to discuss injury types rather than injury severity using, for example, AIS scores. On examination, we found that injury types tended to be described by one AIS score only, so that one is a proxy for the other. For example, 96% of all intracranial injury to cyclists received an AIS score of 2. In addition, we sought to estimate the change in head injuries to cyclists attributable to the helmet wearing law.
2. Method
2.1. Cycle helmet wearing data Cycle helmet wearing rates are measured by national surveys conducted by the Land Transport Safety Authority (LTSA). These surveys were performed in spring (September) and autumn (March) during morning and afternoon school and commuter rush hours between 1986 and 1994, and annually (September) since then. The surveys were conducted in 60 survey sites in towns and suburbs around New Zealand. On the day of the survey, LTSA staff observed cyclists for two 1-hour periods at each survey site during commuter rush hours. Whenever possible, the same persons in each locality conducted the survey. Approximately 10 000 cyclists were observed in each survey (mean 10 265, SD 1857). Cyclists were classified by age group since September 1989 only. Child cyclists were categorised into age groups by the surveyor based on the school uniform worn by the cyclist. All schools have their own uniform. Generally, school children in New Zealand attend Primary/Intermediate School up to age 12, then start High School at age 13. Identification of the uniform worn by the cyclist enabled the school to be identified and, consequently, the age group of the cyclist could be determined.
The reliability of the LTSA age group specific helmet wearing rates were tested by a comparison to the wearing rates obtained from an independent local authority helmet wearing survey for a large metropolitan centre. That centre uses the same survey design, frequency, and method to estimate helmet wearing rates. There were no significant differences between the LTSA and the metropolitan centre surveys for the three age groups (Scuffham and Langley, 1997). The LTSA helmet surveys were restricted to weekdays and may not reflect cyclist helmet wearing behaviour during recreational periods.
2.2. Cyclist injury data For the purposes of this study a serious injury was defined as that which resulted in inpatient treatment at a public hospital, excluding deaths. Fatalities averaged ten per 6-month period of the study, of which the majority had sustained severe injuries to multiple body regions, including the head. As such, the focus of this study is on hospitalisations. Hospitalisation data were obtained from the New Zealand Health Information Service (NZHIS) morbidity data files for 1988– 1996, inclusive. The NZHIS records data on all discharges from public and private hospitals in New Zealand. For convenience, these are referred to as hospitalisations rather than hospital discharges. The present study was restricted to discharges from public hospitals because discharge records from private hospitals seldom include information on the circumstances of injury. Furthermore, reference to injury diagnoses in private hospital records suggest that the majority of injury cases admitted to private hospitals are for follow-up treatment rather than acute treatment (Langley, 1995). As this study sought to determine the trends in the incidence of bicycling injuries, readmissions for the treatment of the same injury were excluded. Admissions for late effects of the initial injury were also excluded. Cases were identified from the hospital records using the International Classification of Diseases, Ninth Revision, Clinical Modification (ICD-9-CM) code. The ICD-9-CM system permits the classification of environmental events, circumstances, and conditions as the cause of injury, poisoning, and other adverse effects; these are known as E-codes and are recorded in a separate data field. Cases were selected if any of the diagnoses was an injury and the E-code was in the relevant range. Cases in the E-code range 810–819 (motor vehicle traffic crash) were included when the fourth digit in the E-code was ‘6’ (pedal cyclist). Other road vehicle crashes (E-code 826.1, pedal cycle accident) that occurred on public roads (identified via the location code in the NZHIS files) were also included if the injured person was a cyclist.
P. Scuffham et al. / Accident Analysis and Pre6ention 32 (2000) 565–573
2.3. Identification of head injuries The hospital morbidity file contains up to four diagnoses for each case and records are coded according to the ICD-9-CM. Reference was made to the principal diagnoses recorded to determine whether the victim had a head injury. The definition of ‘head injury’ as used in this study is the same as used by Scuffham and Langley (1997), constituting any area of the head that a helmet might be expected to protect. Included in this definition (and corresponding ICD-9 codes) are fractures to the skull (800-1, 803-4), intracranial injuries (850-4), and lacerations to the scalp including the ears (872, 873.0, 873.1, 873.8, 873.9). Injury data were grouped into quarterly intervals centred around the months of the helmet survey (March and September) and intervening months (June and December). For example, head injury data for February, March and April were grouped around the March helmet wearing survey data. These data were then grouped according to the age ranges as previously defined: primary school children (aged 5 – 12 years), secondary school children (13 – 18 years), and adults (older than 18 years of age).
2.4. The model We chose to model the data as event rates, where the numerator was the number of head injuries in cyclists. We also required a denominator, ideally this would be a person–time exposure, which in this case is the number of hours spent by cyclists on public roads. However, given the absence of cycling exposure data we chose to use the number of non-head injuries as a proxy for this. A critical assumption here is that cycle helmet wearing does not alter the number of non-head injuries in which case these would be proportional to the person – time exposure on public roads. To examine the effect of helmet wearing on different types of head injury we disaggregated head injuries into fractures, intracranial injuries and lacerations as defined by the ICD-9-CM codes above. The proportion of head injury admissions for non-cyclists has changed over time. We assumed that these patterns were a result of unaccounted changes in practices by hospital admitting doctors in New Zealand or other social and economic factors not accounted for. We hypothesised that these changes would affect cyclists also. Hence, we needed to control for this potential effect. We selected a comparison group to control for changes in the probability of being admitted to hospital with a head injury. The comparison group selected was all injured non-cyclists admitted to hospital. Examples of injured non-cyclists admitted to hospital include cases resulting from motor vehicle traffic crashes (other than cycle crashes), assaults, sporting
567
events, falls, poisonings, and so on. This was the same approach and comparison group used previously by Scuffham and Langley (1997). Given these conditions, models for the number of injured cyclists with a head injury were specified in the general linear form as: ln(CycHI)= b0 + ln(CycNonHI)/ + bp ln(AdmPol)+ bh(HWR)+ error
(1)
where CycHI was the number of cyclists admitted in each period with a head injury, b0 is the intercept term, CycNonHI was the number cyclists admitted in each period without a head injury, AdmPol was the proportion of head injuries present in the comparison group (injured non-cyclists) in each period and HWR was the percentage of cyclists wearing a helmet. Note that the parameter for the variable ln(CycNonHI) was omitted because, theoretically, this parameter must be restricted to unity to obtain the event rate. A restriction to unity on a parameter allows that parameter to be omitted from the estimating equation. This restriction is used to respecify the dependent variable as the log of the odds ratio, that is, ln(CycHI)− ln(CycNonHI)= ln(CycHI/ CycNonHI). In order to estimate the effect of the helmet law, forecasts of the number of head injuries expected to occur in the absence of the law were required. ‘Conditional forecasts’ based on the number of non-head injuries to cyclists and the proportion of head injuries to non-cyclists throughout the forecast period were estimated for each age group. To make conditional forecasts required that we ‘know’ the proportion of cyclists that would wear a helmet in the absence of the helmet law. We assumed, that in the absence of the helmet law, the helmet wearing proportion would remain at the proportion observed in the period 1 year prior to the introduction of the helmet law, that is, 83.5%, 62.4%, 39.1% and 64.9% for the primary school, secondary school, adult age groups and total, respectively, as observed in the September 1992 helmet survey. The difference between the actual number of head injuries for the 3 years from 1994 to 1996 and the total predicted number for the same period is the estimated number of head injuries averted by the helmet law. Injury data are the number of particular events — head injuries to cyclists — in a given period. Therefore, the dependent variable was in the form of non-negative integer values. The mean number of events per observation was small, and when head injuries were dissaggregated by age group and type of injury there were zero events in some observations. Therefore, methods of statistical analysis suitable for the discrete properties of the data were required.
P. Scuffham et al. / Accident Analysis and Pre6ention 32 (2000) 565–573
568
Fig. 1. Helmet wearing rates, 1988 – 1996, by age group.
2.5. Statistical analysis — the negati6e binomial regression model The specification of Eq. (1) is identical to the log of the odds ratio (i.e. a logit model). Because the outcome variable extends beyond the bounds of a binary variable (i.e. in the range from 0 to but with many observations close or equal to zero), a Poisson model was initially developed. However, preliminary analyses of the data suggested that the outcomes did not follow a Poisson distribution. Therefore, a negative binomial (NB) outcome was assumed. NB models have been used previously in epidemiological studies to examine road accidents (Miaou, 1994; Laberge-Nadeau et al., 1996) and hospital admissions (Glynn et al., 1993). The NB regression model has the following form:
G yi + p(Yi = yi )=
1 a
1 1 + ami
1/a
ami 1 +ami
yi
(2) 1 G(yi +1)G a z−1 i where G(z)= 0 t e dt and yi =0, 1, 2,… and where Yi is CycHI, xi,j are the ith observations on the jth explanatory variables in the model [i= 1, 2, 3,…, n; j= HWR, ln(AdmPol), ln(CycNonHI)], bj is the parameter estimate for xj and Var(Yi )= mi +am 2i and a ]0 (the over-dispersion parameter). A special case occurs when a =0, where the model reduces to the Poisson regression model. The NB model was estimated using the GENMOD
procedure in SAS (SAS Institute, 1993). In order to model the proportion of head injuries, the OFFSET feature in SAS was used to restrict the coefficient on ln(CycNonHI) to unity. To assess the ‘goodness-of-fit’ for a NB model estimated by maximum likelihood, the analysis of deviance with the associated degrees of freedom was used (Armitage and Berry 1987). The distribution of deviance follows an approximate x 2 distribution and can be used for testing the goodnessof-fit and over-dispersion of the model. The final specification for each of our models was based on these criteria. To compare the estimated parameters of one model with another, the models were estimated jointly by pooling the age group data and including age group interaction terms. When the interaction terms were individually and jointly insignificant, the interaction terms were collapsed to a single variable. The age group HWR and AdmPol variables were used as interaction terms in the pooled models. Individual parameter significance was given by the standard error of the parameter estimate, and joint significance was assessed by testing the change in deviance and change in degrees of freedom from a reduced model against a model fully specified with interaction terms, which as noted above has an approximate x 2 distribution. This technique used previously by Scuffham and Langley (1997), improved the efficiency of the models by increasing the degrees of freedom available and enabled multiple comparisons between parameter estimates.
P. Scuffham et al. / Accident Analysis and Pre6ention 32 (2000) 565–573
569
Fig. 2. Head injury and non-head injury in cyclists, 1988 – 1996 quarterly observations, all ages.
3. Results The changes in the helmet wearing rates for the three age groups are illustrated in Fig. 1. Fig. 2 depicts the number of head and non-head injuries to cyclists that resulted in admission to public hospitals over the sample period. There was a substantial seasonal pattern for cyclist head and non-head in1juries but overall there was a decrease in both types of injury to cyclists. The
decrease in head injuries to cyclists was greater than the decrease in non-head injuries to cyclists, as shown by the ratio of head:non-head injury presented in Fig. 3. However, the probability of a non-cyclist being admitted to hospital with a head injury also decreased (Fig. 3). Results for the pooled models are reported in Table 1. No seasonal effects in the residuals from the regression models were identified and there were no residual outliers.
Fig. 3. Head injury odds ratio in cyclists and non-cyclists, 1988 – 1996 quarterly observations, all ages.
P. Scuffham et al. / Accident Analysis and Pre6ention 32 (2000) 565–573
570
Table 1 Negative binomial estimation results for the pooled models, 1989–1996 Variable
Intercept HWRa AdmPolb Primary school age Secondary school age a (over-dispersion parameter) Deviance Degrees of freedom a b
All head injuries
Fractures
Intracranial injuries
Lacerations
Parameter estimate
p-value
Parameter estimate
p-value
Parameter estimate
p-value
Parameter estimate
p-value
0.5747 −0.0043 0.4375 −0.0586 −0.0292 0.0208
0.1208 0.0001 0.0201 0.7745 0.8970
−0.7607 −0.0099 0.3502 0.1211 −0.1510 0.0249
0.4952 0.0001 0.2104 0.5228 0.4073
0.3697 −0.0039 0.3806 −0.0045 0.0633 0.0233
0.3778 0.0006 0.0430 0.9839 0.7965
−3.6793 −0.0118 −0.0001 −0.2430 −0.1694 0.0720
0.1711 0.0188 0.9999 0.5328 0.6301
152.52 103
108.72 103
153.51 103
106.67 103
HWR =helmet wearing rate. AdmPol (admission policy) = proportion of head injuries present in the comparison group (i.e. injured non-cyclists).
Table 2 Negative binomial estimation results for the separate models, 1989–1996 All head injuries
Primary school age
Secondary school age
Adults
Variable
Parameter estimate
p-value
Parameter estimate
p-value
Parameter estimate
p-value
Parameter estimate
p-value
Intercept HWR AdmPol a (over-dispersion parameter) Deviance Degrees of freedom
0.4201 −0.0059 0.2477 0.0075
0.4366 0.0016 0.4746
0.6726 −0.0040 0.5901 0.0092
0.0022 0.0291 0.0150
0.2963 −0.0046 0.1742 0.0265
0.3638 0.0071 0.6279
0.1329 −0.0053 0.2090 0.0305
0.9002 0.0613 0.7026
270.57 32
216.88 32
The adult age group was selected as the reference age group. In all models estimated, there was no significant difference in the head injury rates between primary school age and adults or secondary school age and adults, adjusting for helmet wearing and other variables in the model. This result was found in all models where head injury was disaggregated into the three types of head injury (lacerations, intracranial injury and fractures). The helmet wearing rate parameter estimates were greatest for lacerations, then fractures, and smallest for intracranial injuries. These estimates are not strictly comparable because they were obtained from different models but they are indicative. In all models, the helmet wearing rate variables (HWR) and the admission practice variables (AdmPol) showed no significant interaction with each of the age group. These variables were then collapsed into single variables to increase the efficiency of estimation without loss of information. The HWR was negative and significant in all models. The estimated coefficient was not significantly different across age groups (as noted above). This indicates that increased helmet wearing rates were inversely associated with head injuries. A 1% increase in the helmet wearing
227.46 32
244.13 32
rate was associated with a reduction in overall head injuries by 0.43%, for all age groups. For two models, all head injury and intracranial injury, the (AdmPol) was significant (at the 5% level), positive and relatively large (Table 1). That is, changes in admission rates for head injury to cyclists are directly related to changes in admission rates for head injury to non-cyclists. There was a decreased propensity to admit patients with a head injury, including intracranial injury, irrespective of whether they are an injured cyclist or non-cyclist. To estimate the effect of the helmet law conditional forecasts were made. Separate models of all cyclist head injuries for each age group, and the three age groups combined, were estimated. These separate models included the HWR and AdmPol variables and results are reported in Table 2. Conditional forecasts for the 1994–1996 post-law period were made and the predicted number of head injuries to cyclists in the absence of the helmet law calculated. The difference between the predicted and the observed number of head injuries, that is head injuries averted by the law over the first 3 years of the law, are reported in Table 3. Upper and lower (95% CI) estimates were evaluated using a Monte
P. Scuffham et al. / Accident Analysis and Pre6ention 32 (2000) 565–573 Table 3 Head injuries to cyclists averted by the helmet law, 1994–1996 Age group
Number (%) of head injuries averted*
90% Confidence intervals Lower
5–12 13–18 19+ Total
12 31 85 139
(6.2%) (16.2%) (24.5%) (18.7%)
−5 20 66 95
(−2.8%) (11.1%) (20.1%) (13.6%)
Upper 30 42 105 184
(14.2%) (20.8%) (28.6%) (23.4%)
* The number of head injuries averted for the three age groups do not sum to the total because the estimates for the total were derived from a separate model.
Carlo approach, that is, a randomised resampling of the coefficients of the model (Manly, 1994). In each simulation, the HWR coefficients were ‘sampled’ from a normal distribution with the same mean and variance as the estimated coefficient from the original regression (Table 2)1. The upper and lower number of predicted head injuries was then obtained using the regression parameters. One thousand simulations were performed for each age group. The mean, upper and lower values were obtained using the 50th, 95th and 5th percentiles, respectively, of the simulated distributions (Table 3). In terms of reduction in head injuries, the age groups with the lowest pre-law helmet wearing rates benefited the most from the introduction of the helmet wearing law.
4. Discussion A general finding from our models was helmet wearing significantly reduces head injuries to cyclists, in all age groups. The helmet law was an effective strategy to increase helmet wearing and we estimated that the helmet law averted 139 head injuries over a 3-year period. These results are hampered by the lack of an accurate measure for exposure of cyclists to head injury. The proxy used, non-head injury, is a consistent measure of exposure and is a good capture of severe injury. For less severe injury this proxy may be less adequate. The statistical analysis in this present study employed negative binomial models. A negative binomial assumption may arise from ‘injury proneness’ (non-independent recurrent events) or mixed Poisson distributions (Maritz, 1950). We presumed that injury proneness
1 According to the central Limit Theorem the coefficients in a Generalised Linear model (such as the NB model) are asymptotically normally distributed. A large number of replications was chosen to ensure reasonably accurate estimates of percentiles.
571
(Glynn et al., 1996) did not occur here, but it is likely that some important factors were not taken into account in these regressions, such that injury events may be non-independent and recurrent. If each cyclist had injuries according to an individual Poisson injury rate and Poisson injury rates varied across cyclists, then the distribution of the total number of injuries follows a negative binomial distribution. The overdispersion parameters (Tables 1 and 2) were quite small suggesting that there is little variance in the Poisson injury rates across cyclists. A Poisson injury process may adequately model injuries to cyclists but similar parameter estimates can be obtained from the less restrictive negative binomial model and with standard errors potentially closer to the ‘correct’ standard errors. In this study primary diagnosis only was used for identifying head injuries from non-head injuries. Those admitted with both a head and non-head injury were classified according to their primary diagnosis. If victims with multiple (head and non-head) injuries are classified as head injury victims, and therefore excluded from the non-head injury count, preventing head injuries through helmet wearing would artificially raise the non-head injury count. Conversely, if victims with multiple injuries were included in both head and nonhead injury groups, preventing head injuries could reduce some victims’ overall injury profile below the threshold for hospital admission. This would artificially lower the non-head injury count. We examined the effect of using multiple diagnosis to identify head injuries, (30% of our sample had multiple injuries) and found that the use of primary diagnosis would have, at most, overestimated the effect of helmet wearing by 3.5% compared to using multiple diagnoses to identify head injury cases. We chose not to examine the effect of the law as a step function but rather as a continuous function of helmet wearing. This was because the increases in the helmet wearing proportion for the most part were gradual and were not uniformly absent or present. In addition, we chose not to include a linear variable to account for unobserved trends. During preliminary analyses, we noted that the addition of a time-trend component caused the helmet wearing proportion to become insignificant. That is, a time-trend variable ‘swamped’ the real effect. Scuffham and Langley (1997) used a similar approach for their earlier pre-law analyses. That is, they used the same definition of head injury as in this study, but reference was made to all diagnoses for a head injury (and not just the primary diagnosis as discussed). Further, their study was restricted to the period up to one year prior to the introduction of the helmet law. The Scuffham and Langley (1997) study showed that despite substantial increases in helmet wearing over
572
P. Scuffham et al. / Accident Analysis and Pre6ention 32 (2000) 565–573
time, only a small effect on head injuries was detected, in all of their models. (Some of their models included a time-trend variable that may have swamped the real effect, but their conclusions remained unchanged when no time-trend variables were included). The results from that earlier pre-helmet law evaluation study contrast with the results from this current study. The major difference in results can be attributed to the longer time frame used in this study, and possibly to the differences in case selection (as previously discussed). In this current study we have used an additional 4 years of observations on helmet wearing and hospitalisations. During these 4 years, the range in helmet wearing rates increased substantially and cyclist hospitalisations decreased. This present study has increased power to detect an association between aggregate helmet wearing rates and head injury rates given this increased variance. In addition, because the variances of the series were greater, Poisson models were over-dispersed. Thus, the negative binomial specification for the models used here, was appropriate. A further possible explanation is that pre-law cycle helmet users (i.e. ‘early adopters’) were more safety conscious with risk levels below average. Therefore, there would be little change in head injuries from increased voluntary helmet wearing. Our findings were consistent with those reported by Thompson et al. (1989, 1990, 1996a,b), Maimaris et al. (1994), Acton et al. (1995) and other ecological studies (e.g. Cameron et al., 1994; Pitt et al., 1994). That is, increased helmet wearing has reduced the incidence of head injuries to cyclists. However, our findings suggest that helmet wearing has had a greater effect on reducing lacerations as opposed to intracranial injury. Helmet laws have been shown to be an effective mechanism to increase helmet wearing (for example, Vulcan et al., 1992; Schieber et al., 1996; Ni et al., 1997). However, until now the evidence linking helmet wearing laws and reduced injury rates in a population was sparse. A recent study on cycle helmet effectiveness in New Zealand reported that the helmet law accounted for a 20% reduction in head injuries to cyclists involved in motor vehicle crashes (Povey et al., 1998). Using a different study design and statistical technique to Povey et al. (1998), we estimated that the reduction in head injuries for all cyclists (motor vehicle and non-motor vehicle crashes) attributable to the helmet law was 18.7%. Legislation can be used to require cyclists to purchase a cycle helmet (involuntarily) and increase helmet wearing rates. However, helmet wearing legislation may or may not be a cost-effective method for reducing head injuries (Hansen and Scuffham, 1995). The costeffectiveness of helmet legislation and the efficient allocation of public resources is a matter for further research, investigation and debate.
5. Conclusions Increases in helmet wearing in New Zealand have led to significant decreases in head injury to cyclists. The New Zealand helmet law was an effective strategy that substantially increased cycle helmet wearing rates and reduced head injuries in all age groups.
Acknowledgements The Injury Prevention Research Unit is funded jointly by the Accident Rehabilitation and Compensation Insurance Corporation (ACC) and the Health Research Council of New Zealand (HRC). The South East Institute of Public Health is supported by the NHS Executive South Thames Regional Office. This research was undertaken while the first author was funded by a Health Research Council of New Zealand Overseas Post-doctoral Fellowship. The use of the facilities by the first author at CHERE is greatly appreciated. The authors also thank the Land Transport Safety Authority for the provision of the helmet wearing data and for comments on a previous draft of this paper, and Paul Hansen (Economics Department, University of Otago) for his detailed comments. The authors are grateful for the pertinent comments received from the referees, especially the comment regarding case selection. The usual disclaimers apply.
References Acton, C., Thomas, S., Nixon, J.W., Clark, R., Pitt, W.R., Battistutta, D., 1995. Children and bicycles: what is really happening? Studies of fatal and non-fatal bicycle injury. Injury Prevention 1, 86 – 91. Armitage, P., Berry, G., 1987. Statistical methods in medical research, 2nd edition. Blackwell Scientific Publications, London. Cameron, M.H., Vulcan, A.P., Finch, C.F., Newstead, S.V., 1994. Mandatory bicycle helmet use following a decade of helmet promotion in Victoria Australia — an evaluation. Accident Analysis and Prevention 26, 325 – 377. Collins, B.A., Langley, J.D., Marshall, S.W., 1993. Injuries to pedal cyclists resulting in death and hospitalisation. New Zealand Medical Journal 106, 514 – 517. Glynn, R., Stukel, T., Sharp, S., Bubolz, T.A., Freeman, J.L., Fisher, E.S., 1993. Estimating the variance of standardized rates of recurrent events, with application to hospitalizations among the elderly in New England. American Journal of Epilepsy 137, 776 – 786. Glynn, R., Buring, J., 1996. Ways of measuing rates of recurrent events. British Medical Journal 312, 364 – 367. Hansen, P., Scuffham, P., 1995. The cost-effectiveness of compulsory bicycle helmets in New Zealand. Australian Journal of Public Health 19 (5), 450 – 454. Laberge-Nadeau, C., Dionne, G., Maag, U., Desjardins, D., Vanasse, C., Ekoe, J.M., 1996. Medical conditions and the severity of commercial motor vehicle drivers’ road accidents. Accident Analysis and Prevention 28, 43 – 51.
P. Scuffham et al. / Accident Analysis and Pre6ention 32 (2000) 565–573 Langley, J.D., 1995. Experiences using New Zealand’s hospital based surveillance system for injury prevention research. Methods of Informatics in Medicine 34, 340–344. Maimaris, C., Summers, C.L., Browning, C., Palmer, C.R., 1994. Injury patterns in cyclists attending an Accident and Emergency department: a comparison of helmet wearers and non-wearers. British Medical Journal 308, 1537–1540. Manly, B.F.J., 1994. Randomization and Monte Carlo methods in biology. Chapman and Hall, London. Miaou, S., 1994. The relationship between truck accidents and geometric design of road sections: Poisson versus negative binomial regressions. Accident Analysis and Prevention 26, 471–482. Maritz, J.S., 1950. On the validity of inferences drawn from the fitting Poisson and negative binomial distributions to observed accident data. Psychology Bulletin 47, 434–443. Ni, H., Sacks, J.J., Curtis, L., Cieslak, P.R., Hedberg, K., 1997. Evaluation of a statewide bicycle helmet law via multiple measures of helmet use. Archives of Pediatrc and Adolescent Medicine 151 (1), 59 –65. Pitt, W.R., Thomas, S., Nixon, J., Clark, R., Battistutta, D., Acton, C., 1994. Trends in head injuries among child bicyclists. British Medical Journal 308, 177. Povey, L.J., Frith, W.J., Graham, P.G., 1998. Cycle helmet effectiveness in New Zealand. Road Safety Conference: Research. Policy. Education., Wellington, 125–129. SAS Institute, 1993. SAS/ETS User’s Guide. Cary, SAS Institute Inc. Schieber, R., Kresnow, M., Sacks, J., Pledger, E., O’Neil, J., Toomey, K., 1996. Effect of a state law on reported bicycle helmet ownership and use. Archives of Pediatric and Adolescent Medicine 150 (7), 707 – 712.
.
573
Scuffham, P.A., Langley, J.D., 1997. Trends in cycle injury in New Zealand under voluntary helmet use. Accident Analysis and Prevention 29, 1 – 9. Shafi, S., Gilbert, J.C., Loghamanee, F., Allen, J.E., Caty, M.G., Glick, P.L., Carden, S., Azizkhan, R.G., 1998. Impact of bicycle helmet safety legislation on children admitted to a regional pediatric trauma center. Journal of Pediatric Surgery 33 (2), 317–321. Thomas, S., Acton, C., Nixon, J., Battistutta, D., Pitt, W.R., Clark, R., 1994. Effectiveness of bicycle helmets in preventing head injury in children: case – control study [see comments]. British Medical Journal 308, 173 – 176. Thompson, R.S., Rivara, F.P., Thompson, D.C., 1989. A case–control study of the effectiveness of bicycle safety helmets. New England Journal of Medicine 320, 389 – 397. Thompson, D.C., Rivara, F.P., Thompson, R.S., 1996a. Effectiveness of bicycle safety helmets in preventing head injuries: A case–control study. Journal of the American Medical Association 276 (24), 1968 – 1973. Thompson, D.C., Nunn, M.E., Thompson, R.S., Rivara, F.P., 1996b. Effectiveness of bicycle safety helmets in preventing serious facial injury [see comments]. Journal of the American Medical Association 276 (24), 1974 – 1975. Thompson, D.C., Thompson, R.S., Rivara, F.P., Wolf, M.E., 1990. A case – control study of the effectiveness of bicycle safety helmets in preventing facial injury. American Journal of Public Health 80, 1471 – 1474. Vulcan, A.P., Cameron, M.H., Watson, W.L., 1992. Mandatory bicycle helmet use: experience in Victoria Australia. World Journal of Surgery 16, 389 – 397.
Head injuries to bicyclists and the New Zealand bicycle helmet law Paul Scuffham a,b,* , Jonathan Alsop a, Colin Cryer c, John D. Langley a b
a Injury Pre6ention Research Unit, Uni6ersity of Otago, Dunedin, New Zealand Centre for Health Economics Research and E6aluation, Sydney Uni6ersity, Sydney, NSW, Australia c South East Institute of Public Health, Tunbridge Wells, Kent, UK
Received 1 February 1999; received in revised form 15 June 1999; accepted 21 June 1999
Abstract The purpose of this study was to examine the effect of helmet wearing and the New Zealand helmet wearing law on serious head injury for cyclists involved in on-road motor vehicle and non-motor vehicle crashes. The study population consisted of three age groups of cyclists (primary school children (ages 5–12 years), secondary school children (ages 13 – 18 years), and adults (19 + years)) admitted to public hospitals between 1988 and 1996. Data were disaggregated by diagnosis and analysed using negative binomial regression models. Results indicated that there was a positive effect of helmet wearing upon head injury and this effect was relatively consistent across age groups and head injury (diagnosis) types. We conclude that the helmet law has been an effective road safety intervention that has lead to a 19% (90% CI: 14, 23%) reduction in head injury to cyclists over its first 3 years. © 2000 Elsevier Science Ltd. All rights reserved. Keywords: Bicycle; Helmet; Head injury; Legislation
1. Introduction Head injuries due to cycle crashes have been identified as a significant cause of morbidity in New Zealand (Collins et al., 1993). Case – control studies show that the use of a correctly designed and fitted cycle helmet is an effective strategy for reducing the frequency of head and upper facial injury (Thompson et al., 1989, 1990, 1996a,b, Maimaris et al., 1994; Thomas et al., 1994 Acton et al., 1995). Cycle helmets are also thought to reduce the severity of head injury. Few studies examine the relationship between head injury severity and the protection afforded by helmet wearing. Thompson et al. (1996a) show that cycle helmets reduce the risk of a brain injury by 88% and the risk of a severe brain injury is reduced
* Corresponding author. Present address: School of Public Health, Queensland University of Technology, Victoria Park Road, Kelvin Grove, Qld 4059, Australia. Tel.: +61-2-38645724; fax: 61-238643369. E-mail address: [email protected] (P. Scuffham)
by 75%, (i.e. helmets afford less protection against more severe injuries). Shafi et al. (1998) showed that helmeted child cylists with a head injury admitted to hospital were more likely to have concussion and less likely to have a serious brain injury compared to head injured non-helmeted child cyclists (i.e. there is possibly a migration of injury severity from more to less severe). In an effort to reduce head injuries to cyclists, voluntary helmet use has been widely promoted in New Zealand. Sustained programmes promoting regular helmet use saw helmet wearing rates steadily increase from virtually zero in 1986 to 84%, 62%, and 39% for primary school children (aged 5–12 years), secondary school children (13–18 years), and adult commuters (over 18 years), respectively, in September 1992. This was one year prior to the introduction of the helmet law. The New Zealand cycle helmet regulation, effective from 1 January 1994, requires all cyclists to wear a standard approved cycle helmet for all on-road cycling. At the time when this legislation was introduced, helmet wearing rates increased to above 90% for the three cyclist age groups. For the first 2 years following the
0001-4575/00/$ - see front matter © 2000 Elsevier Science Ltd. All rights reserved. PII: S 0 0 0 1 - 4 5 7 5 ( 9 9 ) 0 0 0 8 1 - 0
566
P. Scuffham et al. / Accident Analysis and Pre6ention 32 (2000) 565–573
introduction of the law, wearing rates for three age groups of cyclists exceeded 92%, with a slight decline in the following year. An earlier New Zealand study of the effect of voluntary helmet wearing on serious head injury to cyclists (Scuffham and Langley, 1997) found little association between the percentage of cyclists wearing helmets and the percentage of head injuries to injured cyclists. This present study sought to examine the association between helmet wearing rates and the rates of admission to hospital for head injury to cyclists injured on public roads. Our principal hypothesis was that serious head injuries to cyclists decreased as cycle helmet wearing rates increased. We also sought to examine the effect of helmet wearing for different types of head injury (lacerations, intracranial injury including concussion, skull fractures). The type of injury is examined as it is more meaningful to discuss injury types rather than injury severity using, for example, AIS scores. On examination, we found that injury types tended to be described by one AIS score only, so that one is a proxy for the other. For example, 96% of all intracranial injury to cyclists received an AIS score of 2. In addition, we sought to estimate the change in head injuries to cyclists attributable to the helmet wearing law.
2. Method
2.1. Cycle helmet wearing data Cycle helmet wearing rates are measured by national surveys conducted by the Land Transport Safety Authority (LTSA). These surveys were performed in spring (September) and autumn (March) during morning and afternoon school and commuter rush hours between 1986 and 1994, and annually (September) since then. The surveys were conducted in 60 survey sites in towns and suburbs around New Zealand. On the day of the survey, LTSA staff observed cyclists for two 1-hour periods at each survey site during commuter rush hours. Whenever possible, the same persons in each locality conducted the survey. Approximately 10 000 cyclists were observed in each survey (mean 10 265, SD 1857). Cyclists were classified by age group since September 1989 only. Child cyclists were categorised into age groups by the surveyor based on the school uniform worn by the cyclist. All schools have their own uniform. Generally, school children in New Zealand attend Primary/Intermediate School up to age 12, then start High School at age 13. Identification of the uniform worn by the cyclist enabled the school to be identified and, consequently, the age group of the cyclist could be determined.
The reliability of the LTSA age group specific helmet wearing rates were tested by a comparison to the wearing rates obtained from an independent local authority helmet wearing survey for a large metropolitan centre. That centre uses the same survey design, frequency, and method to estimate helmet wearing rates. There were no significant differences between the LTSA and the metropolitan centre surveys for the three age groups (Scuffham and Langley, 1997). The LTSA helmet surveys were restricted to weekdays and may not reflect cyclist helmet wearing behaviour during recreational periods.
2.2. Cyclist injury data For the purposes of this study a serious injury was defined as that which resulted in inpatient treatment at a public hospital, excluding deaths. Fatalities averaged ten per 6-month period of the study, of which the majority had sustained severe injuries to multiple body regions, including the head. As such, the focus of this study is on hospitalisations. Hospitalisation data were obtained from the New Zealand Health Information Service (NZHIS) morbidity data files for 1988– 1996, inclusive. The NZHIS records data on all discharges from public and private hospitals in New Zealand. For convenience, these are referred to as hospitalisations rather than hospital discharges. The present study was restricted to discharges from public hospitals because discharge records from private hospitals seldom include information on the circumstances of injury. Furthermore, reference to injury diagnoses in private hospital records suggest that the majority of injury cases admitted to private hospitals are for follow-up treatment rather than acute treatment (Langley, 1995). As this study sought to determine the trends in the incidence of bicycling injuries, readmissions for the treatment of the same injury were excluded. Admissions for late effects of the initial injury were also excluded. Cases were identified from the hospital records using the International Classification of Diseases, Ninth Revision, Clinical Modification (ICD-9-CM) code. The ICD-9-CM system permits the classification of environmental events, circumstances, and conditions as the cause of injury, poisoning, and other adverse effects; these are known as E-codes and are recorded in a separate data field. Cases were selected if any of the diagnoses was an injury and the E-code was in the relevant range. Cases in the E-code range 810–819 (motor vehicle traffic crash) were included when the fourth digit in the E-code was ‘6’ (pedal cyclist). Other road vehicle crashes (E-code 826.1, pedal cycle accident) that occurred on public roads (identified via the location code in the NZHIS files) were also included if the injured person was a cyclist.
P. Scuffham et al. / Accident Analysis and Pre6ention 32 (2000) 565–573
2.3. Identification of head injuries The hospital morbidity file contains up to four diagnoses for each case and records are coded according to the ICD-9-CM. Reference was made to the principal diagnoses recorded to determine whether the victim had a head injury. The definition of ‘head injury’ as used in this study is the same as used by Scuffham and Langley (1997), constituting any area of the head that a helmet might be expected to protect. Included in this definition (and corresponding ICD-9 codes) are fractures to the skull (800-1, 803-4), intracranial injuries (850-4), and lacerations to the scalp including the ears (872, 873.0, 873.1, 873.8, 873.9). Injury data were grouped into quarterly intervals centred around the months of the helmet survey (March and September) and intervening months (June and December). For example, head injury data for February, March and April were grouped around the March helmet wearing survey data. These data were then grouped according to the age ranges as previously defined: primary school children (aged 5 – 12 years), secondary school children (13 – 18 years), and adults (older than 18 years of age).
2.4. The model We chose to model the data as event rates, where the numerator was the number of head injuries in cyclists. We also required a denominator, ideally this would be a person–time exposure, which in this case is the number of hours spent by cyclists on public roads. However, given the absence of cycling exposure data we chose to use the number of non-head injuries as a proxy for this. A critical assumption here is that cycle helmet wearing does not alter the number of non-head injuries in which case these would be proportional to the person – time exposure on public roads. To examine the effect of helmet wearing on different types of head injury we disaggregated head injuries into fractures, intracranial injuries and lacerations as defined by the ICD-9-CM codes above. The proportion of head injury admissions for non-cyclists has changed over time. We assumed that these patterns were a result of unaccounted changes in practices by hospital admitting doctors in New Zealand or other social and economic factors not accounted for. We hypothesised that these changes would affect cyclists also. Hence, we needed to control for this potential effect. We selected a comparison group to control for changes in the probability of being admitted to hospital with a head injury. The comparison group selected was all injured non-cyclists admitted to hospital. Examples of injured non-cyclists admitted to hospital include cases resulting from motor vehicle traffic crashes (other than cycle crashes), assaults, sporting
567
events, falls, poisonings, and so on. This was the same approach and comparison group used previously by Scuffham and Langley (1997). Given these conditions, models for the number of injured cyclists with a head injury were specified in the general linear form as: ln(CycHI)= b0 + ln(CycNonHI)/ + bp ln(AdmPol)+ bh(HWR)+ error
(1)
where CycHI was the number of cyclists admitted in each period with a head injury, b0 is the intercept term, CycNonHI was the number cyclists admitted in each period without a head injury, AdmPol was the proportion of head injuries present in the comparison group (injured non-cyclists) in each period and HWR was the percentage of cyclists wearing a helmet. Note that the parameter for the variable ln(CycNonHI) was omitted because, theoretically, this parameter must be restricted to unity to obtain the event rate. A restriction to unity on a parameter allows that parameter to be omitted from the estimating equation. This restriction is used to respecify the dependent variable as the log of the odds ratio, that is, ln(CycHI)− ln(CycNonHI)= ln(CycHI/ CycNonHI). In order to estimate the effect of the helmet law, forecasts of the number of head injuries expected to occur in the absence of the law were required. ‘Conditional forecasts’ based on the number of non-head injuries to cyclists and the proportion of head injuries to non-cyclists throughout the forecast period were estimated for each age group. To make conditional forecasts required that we ‘know’ the proportion of cyclists that would wear a helmet in the absence of the helmet law. We assumed, that in the absence of the helmet law, the helmet wearing proportion would remain at the proportion observed in the period 1 year prior to the introduction of the helmet law, that is, 83.5%, 62.4%, 39.1% and 64.9% for the primary school, secondary school, adult age groups and total, respectively, as observed in the September 1992 helmet survey. The difference between the actual number of head injuries for the 3 years from 1994 to 1996 and the total predicted number for the same period is the estimated number of head injuries averted by the helmet law. Injury data are the number of particular events — head injuries to cyclists — in a given period. Therefore, the dependent variable was in the form of non-negative integer values. The mean number of events per observation was small, and when head injuries were dissaggregated by age group and type of injury there were zero events in some observations. Therefore, methods of statistical analysis suitable for the discrete properties of the data were required.
P. Scuffham et al. / Accident Analysis and Pre6ention 32 (2000) 565–573
568
Fig. 1. Helmet wearing rates, 1988 – 1996, by age group.
2.5. Statistical analysis — the negati6e binomial regression model The specification of Eq. (1) is identical to the log of the odds ratio (i.e. a logit model). Because the outcome variable extends beyond the bounds of a binary variable (i.e. in the range from 0 to but with many observations close or equal to zero), a Poisson model was initially developed. However, preliminary analyses of the data suggested that the outcomes did not follow a Poisson distribution. Therefore, a negative binomial (NB) outcome was assumed. NB models have been used previously in epidemiological studies to examine road accidents (Miaou, 1994; Laberge-Nadeau et al., 1996) and hospital admissions (Glynn et al., 1993). The NB regression model has the following form:
G yi + p(Yi = yi )=
1 a
1 1 + ami
1/a
ami 1 +ami
yi
(2) 1 G(yi +1)G a z−1 i where G(z)= 0 t e dt and yi =0, 1, 2,… and where Yi is CycHI, xi,j are the ith observations on the jth explanatory variables in the model [i= 1, 2, 3,…, n; j= HWR, ln(AdmPol), ln(CycNonHI)], bj is the parameter estimate for xj and Var(Yi )= mi +am 2i and a ]0 (the over-dispersion parameter). A special case occurs when a =0, where the model reduces to the Poisson regression model. The NB model was estimated using the GENMOD
procedure in SAS (SAS Institute, 1993). In order to model the proportion of head injuries, the OFFSET feature in SAS was used to restrict the coefficient on ln(CycNonHI) to unity. To assess the ‘goodness-of-fit’ for a NB model estimated by maximum likelihood, the analysis of deviance with the associated degrees of freedom was used (Armitage and Berry 1987). The distribution of deviance follows an approximate x 2 distribution and can be used for testing the goodnessof-fit and over-dispersion of the model. The final specification for each of our models was based on these criteria. To compare the estimated parameters of one model with another, the models were estimated jointly by pooling the age group data and including age group interaction terms. When the interaction terms were individually and jointly insignificant, the interaction terms were collapsed to a single variable. The age group HWR and AdmPol variables were used as interaction terms in the pooled models. Individual parameter significance was given by the standard error of the parameter estimate, and joint significance was assessed by testing the change in deviance and change in degrees of freedom from a reduced model against a model fully specified with interaction terms, which as noted above has an approximate x 2 distribution. This technique used previously by Scuffham and Langley (1997), improved the efficiency of the models by increasing the degrees of freedom available and enabled multiple comparisons between parameter estimates.
P. Scuffham et al. / Accident Analysis and Pre6ention 32 (2000) 565–573
569
Fig. 2. Head injury and non-head injury in cyclists, 1988 – 1996 quarterly observations, all ages.
3. Results The changes in the helmet wearing rates for the three age groups are illustrated in Fig. 1. Fig. 2 depicts the number of head and non-head injuries to cyclists that resulted in admission to public hospitals over the sample period. There was a substantial seasonal pattern for cyclist head and non-head in1juries but overall there was a decrease in both types of injury to cyclists. The
decrease in head injuries to cyclists was greater than the decrease in non-head injuries to cyclists, as shown by the ratio of head:non-head injury presented in Fig. 3. However, the probability of a non-cyclist being admitted to hospital with a head injury also decreased (Fig. 3). Results for the pooled models are reported in Table 1. No seasonal effects in the residuals from the regression models were identified and there were no residual outliers.
Fig. 3. Head injury odds ratio in cyclists and non-cyclists, 1988 – 1996 quarterly observations, all ages.
P. Scuffham et al. / Accident Analysis and Pre6ention 32 (2000) 565–573
570
Table 1 Negative binomial estimation results for the pooled models, 1989–1996 Variable
Intercept HWRa AdmPolb Primary school age Secondary school age a (over-dispersion parameter) Deviance Degrees of freedom a b
All head injuries
Fractures
Intracranial injuries
Lacerations
Parameter estimate
p-value
Parameter estimate
p-value
Parameter estimate
p-value
Parameter estimate
p-value
0.5747 −0.0043 0.4375 −0.0586 −0.0292 0.0208
0.1208 0.0001 0.0201 0.7745 0.8970
−0.7607 −0.0099 0.3502 0.1211 −0.1510 0.0249
0.4952 0.0001 0.2104 0.5228 0.4073
0.3697 −0.0039 0.3806 −0.0045 0.0633 0.0233
0.3778 0.0006 0.0430 0.9839 0.7965
−3.6793 −0.0118 −0.0001 −0.2430 −0.1694 0.0720
0.1711 0.0188 0.9999 0.5328 0.6301
152.52 103
108.72 103
153.51 103
106.67 103
HWR =helmet wearing rate. AdmPol (admission policy) = proportion of head injuries present in the comparison group (i.e. injured non-cyclists).
Table 2 Negative binomial estimation results for the separate models, 1989–1996 All head injuries
Primary school age
Secondary school age
Adults
Variable
Parameter estimate
p-value
Parameter estimate
p-value
Parameter estimate
p-value
Parameter estimate
p-value
Intercept HWR AdmPol a (over-dispersion parameter) Deviance Degrees of freedom
0.4201 −0.0059 0.2477 0.0075
0.4366 0.0016 0.4746
0.6726 −0.0040 0.5901 0.0092
0.0022 0.0291 0.0150
0.2963 −0.0046 0.1742 0.0265
0.3638 0.0071 0.6279
0.1329 −0.0053 0.2090 0.0305
0.9002 0.0613 0.7026
270.57 32
216.88 32
The adult age group was selected as the reference age group. In all models estimated, there was no significant difference in the head injury rates between primary school age and adults or secondary school age and adults, adjusting for helmet wearing and other variables in the model. This result was found in all models where head injury was disaggregated into the three types of head injury (lacerations, intracranial injury and fractures). The helmet wearing rate parameter estimates were greatest for lacerations, then fractures, and smallest for intracranial injuries. These estimates are not strictly comparable because they were obtained from different models but they are indicative. In all models, the helmet wearing rate variables (HWR) and the admission practice variables (AdmPol) showed no significant interaction with each of the age group. These variables were then collapsed into single variables to increase the efficiency of estimation without loss of information. The HWR was negative and significant in all models. The estimated coefficient was not significantly different across age groups (as noted above). This indicates that increased helmet wearing rates were inversely associated with head injuries. A 1% increase in the helmet wearing
227.46 32
244.13 32
rate was associated with a reduction in overall head injuries by 0.43%, for all age groups. For two models, all head injury and intracranial injury, the (AdmPol) was significant (at the 5% level), positive and relatively large (Table 1). That is, changes in admission rates for head injury to cyclists are directly related to changes in admission rates for head injury to non-cyclists. There was a decreased propensity to admit patients with a head injury, including intracranial injury, irrespective of whether they are an injured cyclist or non-cyclist. To estimate the effect of the helmet law conditional forecasts were made. Separate models of all cyclist head injuries for each age group, and the three age groups combined, were estimated. These separate models included the HWR and AdmPol variables and results are reported in Table 2. Conditional forecasts for the 1994–1996 post-law period were made and the predicted number of head injuries to cyclists in the absence of the helmet law calculated. The difference between the predicted and the observed number of head injuries, that is head injuries averted by the law over the first 3 years of the law, are reported in Table 3. Upper and lower (95% CI) estimates were evaluated using a Monte
P. Scuffham et al. / Accident Analysis and Pre6ention 32 (2000) 565–573 Table 3 Head injuries to cyclists averted by the helmet law, 1994–1996 Age group
Number (%) of head injuries averted*
90% Confidence intervals Lower
5–12 13–18 19+ Total
12 31 85 139
(6.2%) (16.2%) (24.5%) (18.7%)
−5 20 66 95
(−2.8%) (11.1%) (20.1%) (13.6%)
Upper 30 42 105 184
(14.2%) (20.8%) (28.6%) (23.4%)
* The number of head injuries averted for the three age groups do not sum to the total because the estimates for the total were derived from a separate model.
Carlo approach, that is, a randomised resampling of the coefficients of the model (Manly, 1994). In each simulation, the HWR coefficients were ‘sampled’ from a normal distribution with the same mean and variance as the estimated coefficient from the original regression (Table 2)1. The upper and lower number of predicted head injuries was then obtained using the regression parameters. One thousand simulations were performed for each age group. The mean, upper and lower values were obtained using the 50th, 95th and 5th percentiles, respectively, of the simulated distributions (Table 3). In terms of reduction in head injuries, the age groups with the lowest pre-law helmet wearing rates benefited the most from the introduction of the helmet wearing law.
4. Discussion A general finding from our models was helmet wearing significantly reduces head injuries to cyclists, in all age groups. The helmet law was an effective strategy to increase helmet wearing and we estimated that the helmet law averted 139 head injuries over a 3-year period. These results are hampered by the lack of an accurate measure for exposure of cyclists to head injury. The proxy used, non-head injury, is a consistent measure of exposure and is a good capture of severe injury. For less severe injury this proxy may be less adequate. The statistical analysis in this present study employed negative binomial models. A negative binomial assumption may arise from ‘injury proneness’ (non-independent recurrent events) or mixed Poisson distributions (Maritz, 1950). We presumed that injury proneness
1 According to the central Limit Theorem the coefficients in a Generalised Linear model (such as the NB model) are asymptotically normally distributed. A large number of replications was chosen to ensure reasonably accurate estimates of percentiles.
571
(Glynn et al., 1996) did not occur here, but it is likely that some important factors were not taken into account in these regressions, such that injury events may be non-independent and recurrent. If each cyclist had injuries according to an individual Poisson injury rate and Poisson injury rates varied across cyclists, then the distribution of the total number of injuries follows a negative binomial distribution. The overdispersion parameters (Tables 1 and 2) were quite small suggesting that there is little variance in the Poisson injury rates across cyclists. A Poisson injury process may adequately model injuries to cyclists but similar parameter estimates can be obtained from the less restrictive negative binomial model and with standard errors potentially closer to the ‘correct’ standard errors. In this study primary diagnosis only was used for identifying head injuries from non-head injuries. Those admitted with both a head and non-head injury were classified according to their primary diagnosis. If victims with multiple (head and non-head) injuries are classified as head injury victims, and therefore excluded from the non-head injury count, preventing head injuries through helmet wearing would artificially raise the non-head injury count. Conversely, if victims with multiple injuries were included in both head and nonhead injury groups, preventing head injuries could reduce some victims’ overall injury profile below the threshold for hospital admission. This would artificially lower the non-head injury count. We examined the effect of using multiple diagnosis to identify head injuries, (30% of our sample had multiple injuries) and found that the use of primary diagnosis would have, at most, overestimated the effect of helmet wearing by 3.5% compared to using multiple diagnoses to identify head injury cases. We chose not to examine the effect of the law as a step function but rather as a continuous function of helmet wearing. This was because the increases in the helmet wearing proportion for the most part were gradual and were not uniformly absent or present. In addition, we chose not to include a linear variable to account for unobserved trends. During preliminary analyses, we noted that the addition of a time-trend component caused the helmet wearing proportion to become insignificant. That is, a time-trend variable ‘swamped’ the real effect. Scuffham and Langley (1997) used a similar approach for their earlier pre-law analyses. That is, they used the same definition of head injury as in this study, but reference was made to all diagnoses for a head injury (and not just the primary diagnosis as discussed). Further, their study was restricted to the period up to one year prior to the introduction of the helmet law. The Scuffham and Langley (1997) study showed that despite substantial increases in helmet wearing over
572
P. Scuffham et al. / Accident Analysis and Pre6ention 32 (2000) 565–573
time, only a small effect on head injuries was detected, in all of their models. (Some of their models included a time-trend variable that may have swamped the real effect, but their conclusions remained unchanged when no time-trend variables were included). The results from that earlier pre-helmet law evaluation study contrast with the results from this current study. The major difference in results can be attributed to the longer time frame used in this study, and possibly to the differences in case selection (as previously discussed). In this current study we have used an additional 4 years of observations on helmet wearing and hospitalisations. During these 4 years, the range in helmet wearing rates increased substantially and cyclist hospitalisations decreased. This present study has increased power to detect an association between aggregate helmet wearing rates and head injury rates given this increased variance. In addition, because the variances of the series were greater, Poisson models were over-dispersed. Thus, the negative binomial specification for the models used here, was appropriate. A further possible explanation is that pre-law cycle helmet users (i.e. ‘early adopters’) were more safety conscious with risk levels below average. Therefore, there would be little change in head injuries from increased voluntary helmet wearing. Our findings were consistent with those reported by Thompson et al. (1989, 1990, 1996a,b), Maimaris et al. (1994), Acton et al. (1995) and other ecological studies (e.g. Cameron et al., 1994; Pitt et al., 1994). That is, increased helmet wearing has reduced the incidence of head injuries to cyclists. However, our findings suggest that helmet wearing has had a greater effect on reducing lacerations as opposed to intracranial injury. Helmet laws have been shown to be an effective mechanism to increase helmet wearing (for example, Vulcan et al., 1992; Schieber et al., 1996; Ni et al., 1997). However, until now the evidence linking helmet wearing laws and reduced injury rates in a population was sparse. A recent study on cycle helmet effectiveness in New Zealand reported that the helmet law accounted for a 20% reduction in head injuries to cyclists involved in motor vehicle crashes (Povey et al., 1998). Using a different study design and statistical technique to Povey et al. (1998), we estimated that the reduction in head injuries for all cyclists (motor vehicle and non-motor vehicle crashes) attributable to the helmet law was 18.7%. Legislation can be used to require cyclists to purchase a cycle helmet (involuntarily) and increase helmet wearing rates. However, helmet wearing legislation may or may not be a cost-effective method for reducing head injuries (Hansen and Scuffham, 1995). The costeffectiveness of helmet legislation and the efficient allocation of public resources is a matter for further research, investigation and debate.
5. Conclusions Increases in helmet wearing in New Zealand have led to significant decreases in head injury to cyclists. The New Zealand helmet law was an effective strategy that substantially increased cycle helmet wearing rates and reduced head injuries in all age groups.
Acknowledgements The Injury Prevention Research Unit is funded jointly by the Accident Rehabilitation and Compensation Insurance Corporation (ACC) and the Health Research Council of New Zealand (HRC). The South East Institute of Public Health is supported by the NHS Executive South Thames Regional Office. This research was undertaken while the first author was funded by a Health Research Council of New Zealand Overseas Post-doctoral Fellowship. The use of the facilities by the first author at CHERE is greatly appreciated. The authors also thank the Land Transport Safety Authority for the provision of the helmet wearing data and for comments on a previous draft of this paper, and Paul Hansen (Economics Department, University of Otago) for his detailed comments. The authors are grateful for the pertinent comments received from the referees, especially the comment regarding case selection. The usual disclaimers apply.
References Acton, C., Thomas, S., Nixon, J.W., Clark, R., Pitt, W.R., Battistutta, D., 1995. Children and bicycles: what is really happening? Studies of fatal and non-fatal bicycle injury. Injury Prevention 1, 86 – 91. Armitage, P., Berry, G., 1987. Statistical methods in medical research, 2nd edition. Blackwell Scientific Publications, London. Cameron, M.H., Vulcan, A.P., Finch, C.F., Newstead, S.V., 1994. Mandatory bicycle helmet use following a decade of helmet promotion in Victoria Australia — an evaluation. Accident Analysis and Prevention 26, 325 – 377. Collins, B.A., Langley, J.D., Marshall, S.W., 1993. Injuries to pedal cyclists resulting in death and hospitalisation. New Zealand Medical Journal 106, 514 – 517. Glynn, R., Stukel, T., Sharp, S., Bubolz, T.A., Freeman, J.L., Fisher, E.S., 1993. Estimating the variance of standardized rates of recurrent events, with application to hospitalizations among the elderly in New England. American Journal of Epilepsy 137, 776 – 786. Glynn, R., Buring, J., 1996. Ways of measuing rates of recurrent events. British Medical Journal 312, 364 – 367. Hansen, P., Scuffham, P., 1995. The cost-effectiveness of compulsory bicycle helmets in New Zealand. Australian Journal of Public Health 19 (5), 450 – 454. Laberge-Nadeau, C., Dionne, G., Maag, U., Desjardins, D., Vanasse, C., Ekoe, J.M., 1996. Medical conditions and the severity of commercial motor vehicle drivers’ road accidents. Accident Analysis and Prevention 28, 43 – 51.
P. Scuffham et al. / Accident Analysis and Pre6ention 32 (2000) 565–573 Langley, J.D., 1995. Experiences using New Zealand’s hospital based surveillance system for injury prevention research. Methods of Informatics in Medicine 34, 340–344. Maimaris, C., Summers, C.L., Browning, C., Palmer, C.R., 1994. Injury patterns in cyclists attending an Accident and Emergency department: a comparison of helmet wearers and non-wearers. British Medical Journal 308, 1537–1540. Manly, B.F.J., 1994. Randomization and Monte Carlo methods in biology. Chapman and Hall, London. Miaou, S., 1994. The relationship between truck accidents and geometric design of road sections: Poisson versus negative binomial regressions. Accident Analysis and Prevention 26, 471–482. Maritz, J.S., 1950. On the validity of inferences drawn from the fitting Poisson and negative binomial distributions to observed accident data. Psychology Bulletin 47, 434–443. Ni, H., Sacks, J.J., Curtis, L., Cieslak, P.R., Hedberg, K., 1997. Evaluation of a statewide bicycle helmet law via multiple measures of helmet use. Archives of Pediatrc and Adolescent Medicine 151 (1), 59 –65. Pitt, W.R., Thomas, S., Nixon, J., Clark, R., Battistutta, D., Acton, C., 1994. Trends in head injuries among child bicyclists. British Medical Journal 308, 177. Povey, L.J., Frith, W.J., Graham, P.G., 1998. Cycle helmet effectiveness in New Zealand. Road Safety Conference: Research. Policy. Education., Wellington, 125–129. SAS Institute, 1993. SAS/ETS User’s Guide. Cary, SAS Institute Inc. Schieber, R., Kresnow, M., Sacks, J., Pledger, E., O’Neil, J., Toomey, K., 1996. Effect of a state law on reported bicycle helmet ownership and use. Archives of Pediatric and Adolescent Medicine 150 (7), 707 – 712.
.
573
Scuffham, P.A., Langley, J.D., 1997. Trends in cycle injury in New Zealand under voluntary helmet use. Accident Analysis and Prevention 29, 1 – 9. Shafi, S., Gilbert, J.C., Loghamanee, F., Allen, J.E., Caty, M.G., Glick, P.L., Carden, S., Azizkhan, R.G., 1998. Impact of bicycle helmet safety legislation on children admitted to a regional pediatric trauma center. Journal of Pediatric Surgery 33 (2), 317–321. Thomas, S., Acton, C., Nixon, J., Battistutta, D., Pitt, W.R., Clark, R., 1994. Effectiveness of bicycle helmets in preventing head injury in children: case – control study [see comments]. British Medical Journal 308, 173 – 176. Thompson, R.S., Rivara, F.P., Thompson, D.C., 1989. A case–control study of the effectiveness of bicycle safety helmets. New England Journal of Medicine 320, 389 – 397. Thompson, D.C., Rivara, F.P., Thompson, R.S., 1996a. Effectiveness of bicycle safety helmets in preventing head injuries: A case–control study. Journal of the American Medical Association 276 (24), 1968 – 1973. Thompson, D.C., Nunn, M.E., Thompson, R.S., Rivara, F.P., 1996b. Effectiveness of bicycle safety helmets in preventing serious facial injury [see comments]. Journal of the American Medical Association 276 (24), 1974 – 1975. Thompson, D.C., Thompson, R.S., Rivara, F.P., Wolf, M.E., 1990. A case – control study of the effectiveness of bicycle safety helmets in preventing facial injury. American Journal of Public Health 80, 1471 – 1474. Vulcan, A.P., Cameron, M.H., Watson, W.L., 1992. Mandatory bicycle helmet use: experience in Victoria Australia. World Journal of Surgery 16, 389 – 397.