Modelling of conspicuity-related motorcycle accidents in seremban and shah alam by Guia Buenas Practicas Motocicletas

Pergamon oool-4575(95)ooo71-2

MODELLING OF CONSPICUITY-RELATED MOTORCYCLE ACCIDENTS IN SEREMBAN AND SHAH ALAM, MALAYSIA RADIN

UMAR

R.S1,

MURRAY

MACKAY~

and

BRIAN

L. HILLS~

‘Accident Research Unit, Universiti Pertanian Malaysia, 43400 Serdang, Malaysia; ‘Accident Research Centre, University of Birmingham, Birmingham, England; 3Transport Research Laboratory, Crowthorne, Berks., England (Accepted 2 November 1995)

Abstract-Preliminary analysis of the short-term impact of a running headlights intervention revealed that there has been a significant drop in conspicuity-related motorcycle accidents in the pilot areas, Seremban and Shah Alam, Malaysia. This paper attempts to look in more detail at conspicuity-related accidents involving motorcycles. The aim of the analysis was to establish a statistical model to describe the relationship between the frequency of conspicuity-related motorcycle accidents and a range of explanatory variables so that new insights can be obtained into the effects of introducing a running headlight campaign and regulation. The exogenous variables in this analysis include the influence of time trends, changes in the recording and analysis system, the effect of fasting activities during Ramadhan and the “Balik Kampong” culture, a seasonal cultural-religious holiday activity unique to Malaysia. The model developed revealed that the running headlight intervention reduced the conspicuity-related motorcycle accidents by about 29%. It is concluded that the intervention has been successful in improving conspicuity-related motorcycle accidents in Malaysia. Copyright 0 1996 Elsevier Science Ltd Keywords-Motorcycle

accidents, Conspicuity,

Running headlights, Generalized linear modelling dents, were found to have dropped significantly (p<O.O05) after the campaign, based of accident figures 6 months before and 6 months after the intervention. In an attempt to evaluate the long-term impact of the RHL intervention, a more detailed analysis of the accident series has been carried out, using a generalized linear modelling technique (McCullagh and Nedler 1983). In recent years, this modelling technique has become a popular tool for the analysis of road accident data. The popularity of generalized linear models is primarily due to their statistical advantages. If before-and-after studies involve more than a single site or single period of time, for example, then the generalized linear modelling methodology provides a convenient way of analysing the data. Use of generalized linear model allows, in turn, the incorporation of dummy variables representing the influence of interventions to traffic accidents. The effectiveness of a scheme, a, can be estimated, together with the appropriate statistics for significance testing or for estimating confidence intervals (Maycock and Summersgill 1994). In addition to calculation of the effectiveness, a, in the context of a statistical model,

INTRODUCTION Nearly 60% of all vehicles in Malaysia are motorcycles. As a consequences, casualties among motorcyclists form the largest portion of the traffic injury problem. In 1992 alone, a total number of 39,272 motorcycles were reported to have been involved in traffic accidents resulting in 24,264 injuries (RMP 1992). Out of a total of 4557 fatalities and 31,705 serious and slight injuries throughout the country in 1992, approximately 51% (2307 cases) of all fatalities and 69% (21,957 cases) of all casualties were motorcycle riders and pillion passengers. In July 1992, a nationwide “daytime running headlight” intervention (RHL) was introduced. This intervention consists of a 3 month safety campaign, followed by the compulsory RHL regulation in September 1992. The aim of this programme was to improve motorcycle conspicuity and thus reduce accidents. Preliminary analysis of the short-term impact of the RHL intervention (Radin et al. 1995) revealed that there had been a sizeable drop in multipledaytime motorcycle accidents in two pilot study areas, Seremban and Shah Alam. Conspicuity-related acci325

RADM UMAR R.S. et al.

326

this technique allows other covariates to be included in the modelling process should this be required. This paper attempts to look in more detail at conspicuity-related accidents involving motorcycles. The aim of the analysis was to establish a model to describe the relationship between the frequency of conspicuity-related motorcycle accidents to a range of explanatory variables so that new insights can be obtained into the effects of RHL intervention. The explanatory variables in this analysis include the influence of time trends, changes in the recording system, the effect of fasting during the months of Ramadhan and the Balik Kampong culture, a seasonal cultural-religious holiday activity unique to the multiracial society of Malaysia. DATA AND CULTURAL

FACTORS

Since area-wide traffic flow data are not available, WEEK was used to reflect the growth of traffic and population over time. This approach is in line with the approach of Broughton (1985) who used the time trend term to model and evaluate the effectiveness of seat belt law in the U.K. Maycock and Summergill (1994) also suggested that if it is suspected that a time trend occurs in the data, and individual weekly or monthly accident data over a period of years are available, then a time trend term can be added to the model to calculate the effectiveness of the scheme taking the trend over time into account. This will allow both (a) the time trend step function due to the impact of intervention and (b) the long term trend due to natural growth of traffic and other exposures to be included in the analysis. The new police accident recording form (Radin 1992) known as POL27(Pin l/91), implemented in January 1992 (Fig. 1) was also expected to affect the quality and the quantity of accident data collected. Improvement in the recording system, RECSYS, would result in a higher number of recorded accidents in the database. This would influence the overall interpretation of the accident series over time. In Muslim fasting months, FAST (Fig. l), there is an increase in travelling activities generally in the late hours of the day. In these late hours, most Muslims and perhaps Malaysian society as a whole, take the opportunity to go out to buy specially made fresh foods at specially located stalls that are only available during these fasting months. For those who work in the city, these times are also the rush hour for them to buy those foods and join their family before the sunset. The influence of these activities in these particular months should be included to explain motorcycle accidents in the country. Another unique multi-cultural and religious activity in Malaysia is Balik Kampong, BLKG

(Fig. 1). This activity normally appears during national festivals and public holidays such as the Chinese New Year, Hari Raya, Deepavali, Christmas and school holidays. During these seasons, all Malaysians regardless of their ethnic groups or religion, take a one-week holiday, often to visit relatives and parents in the rural areas. Motorcyclists then tend to use public transport to travel with their large families and leave their motorcycles at home. Having defined the explanatory variables, a time series cross tabulation on motorcycle accidents was tabulated using an accident analysis package, MAAPS (Hills et al. 1994). A working file containing conspicuity-related motorcycle accidents was created and used in the selection of accident series. Conspicuity-related motorcycle accidents are defined as all accidents involving motorcycles travelling straight or turning onto a right of way and colliding with pedestrians and other vehicles (Radin et al. 1995). The above classification is necessary to differentiate between accident situations in which RHL could have potentially improved safety and those in which they probably would have been irrelevent. This classification was based on the detailed collision types available in the police sketch diagram and written description in POL27 (Pin l/91). A total of 4958 conspicuity-related accident series was available in the database which was later classified by week starting from January 1991 to December 1993. The reason for aggregating the series into the weekly blocks is that most of the above activities could be easily defined in weekly blocks. Two-level factors to represent the changes in the activities were used in the coding of the interventions according to the timings shown in Fig. 1. STATISTICAL

MODEL

To model the accident series, the generalized linear modelling incorporated GLIM software (Payne 1987) was used in the analysis. A log linear time series intervention model was adopted for the analysis. The dependent variable used was conspicuity-related motorcycle accidents per week, MSTOX, while the explanatory variables used to explain motorcycle accidents were WEEK, RECSYS, FAST, BLKG and RHL. The structure of the explanatory variables and their respective definitions as coded in GLIM are shown in Table 1. Accident frequencies have been related to the other variables and factors via a multiplicative or power model. Many studies, both theoretical and empirical, have been made on the relationship between accident frequencies and traffic characteristics, landuse, geometrical layouts and social characteristics. In modelling accident series, Scott (1983)

Modelhng

of conspicuity-related

motorcycle

accidents

in Seremban

(Weeks

and Shah Alam, Malaysia

327

MSTOX Accidents 20

Fig. 1. Timings

Table 1. Data Explanatory

variable

WEEK RECSYS FAST BLKG RHL

structure

and defination

in Ramadhan

Balik Kampong

culture

headlights

proposed a multiplicative model to relate accident series with traffic flows and interventions carried out. Broughton (1985) used his modelling approach to model casualty series in the evaluation of the effectiveness of seat belt legislation in U.K. In evaluating the impact of government road safety initiatives in New South Wales, Home1 (1994) employed a log linear analysis to model fatal crash series with different government interventions. To analyse this series of motorcycle accidents, a number of models were fitted which represent the trends and seasonal and cultural variations in the

101

111

121

131

141

151

variables.

of explanatory Two-level

Week of the year Recording system used

Running

for explanatory

Description

Fasting

NA 2 2 2 2

variables Coding

factors

(1) (2) (1) (2) (1) (2) (1) (2)

system

12,,,,,,>â&#x20AC;&#x2DC;, 3 4 5 6 7 8 9 ................ 156. Trial form + old form POL27(Pin l/91) Not fasting week Fasting week Not Balik Kampong season Bahk Kampong season Before intervention After intervention

series besides any possible effect of RHL interventions and recording system RECSYS. The model is similar to the one used by Broughton since the weekly series contain no information on traffic characteristics by time of day and week. The models are in the form of: Conspicuity - Related Accidents (MSTOX) = exp (a + /IWEEK + yRECSYS + GFASTING +cRHL + 0BLKG + e)

(1)

Where a, p, â&#x20AC;&#x2DC;ye6, E and 0 are coefficients to be estimated and (e) is the error terms representing the

RADINUYAR R.S. et al.

328

residual difference between the actual and predicted model. To express this model in terms that can be used in the GLIM analysis, the ERROR function is defined as Poisson and the LINK function as LOG and the linear equation becomes Log,MSTOX

= c1+ PWEEK

+ GFASTING

+ERHL

+?/RECSYS

+ BBLKG + e

(2)

The quality of fit for adding each term in the “null” model can be assessed by considering the difference made to the scaled deviance in the model (at a 5% level of significance), the ratio of estimates to standard errors and the ratio of residual deviance to degrees of freedom.

SIGNIFICANCE

TESTING

Where alternative forms are possible for the systematic component of a statistical model, it is necessary to be able to decide whether one model is significantly better than another. It is important also to be able to choose the variable or factor that contributes most to “explain” the variation of the data. When additional parameters are added or removed from the model, it is important to be able to decide whether any improvement in “fit ” is statistically significant. The statistic calculated by GLIM which forms the basis for testing the significance of adding terms to the model is called the scaled deviance. For Normal distribution, the deviance is just the residual sum of squares. For Poisson errors (Kimber and Kennedy 1988), the type of error which is commonly used to describe accident distribution, the deviance, is a maximum likelihood ratio statistic (D), defined by: Deviance(D)

c[ y ln( y/p) - (y - p)]

= 2 i

For a well-fitting model with appropriate link function, error distribution and functional form, the expected value of residual deviance should approximately be equal to the number of degrees of freedom, regardless of the value of p. Moreover, provided that the predicted mean value of the dependent variable is greater than 0.5, the scaled deviance with Poisson errors is asymptotically distributed as a chi-square variable with (n-p1) degrees of freedom (n is the number of data points and p is the number of independent variables fitted). In the present case the overall value of p is considerably greater that 1, so that the problems associated with low values of p are not of concern.

RESULTS Univariate analysis of the terms Table 2 summarizes the univariate analysis of conspicuity-related accidents involving motorcycles. It can be seen that all terms, except Balik Kampong, BLKG, are significant at the 5% level. Significance was assessed by (i) comparing the difference in scaled deviance with the 95% point of the chi-squared distribution, the appropriate degree of freedom being the difference in the degrees of freedom of the two models, and (ii) testing if the ratio of estimate to the standard error of each term is more than 1.96.

Multivariate analysis of the terms Table 3 shows the models and their respective terms included in the models. Model 1 shows the full model to explain the conspicuity-related accidents per week, MSTOX. As in the univariate analysis, all terms except BLKG are significant at the 5% level. The deviance changes from 309.0 to 211.8 with 5 df lost. The mean deviance, which is the ratio of residual deviance over degrees of freedom, ranged between 1.41 and 1.99. Since the residual deviance over degrees of freedom is not approximately equal to unity, it can be said that there has been some slight overdispersion in the assumed Poisson distribution. In Poisson distributions, the variance is equal to the mean. Over-dispersion can arise in three ways, viz. (i) the systematic component of the model may be incorrect, available variables may not be included or have not been included in the most appropriate form, (ii) significant variables have had to be omitted from the model and (iii) the assumed error structure is inappropriate. The first factor influencing over-dispersion could be dealt with as far as possible by attention to the range and the form of explanatory variables used, and by experimenting with alternative model specification. The most appropriate representation of the structure of residual variation, however, is the one that handles the combination of the second and the third factors. In order to overcome the over-dispersion problem, McCullagh and Nelder (1983) proposed a correction technique known as quasi-likelihood. This method assumes a common dispersion parameter which is independent of p, rather like the residual variance in a least square fit. Suppose that Yhas a described distribution with E(Y)=p, Var (Y)=o’ p and log ,u=/~x, then the quasi-likelihood estimate p is the same as for the Poisson distribution (Aitkin et al. 1989). The standard errors for p may be obtained by multiplying those

Modelling

of conspicuity-related

motorcycle

Table 2. Univariate Model No.

Explanatory variable

1 2

Constant Constant WEEK Constant RECSYS Constant FAST Constant BLKG

3 4 5

Estimates 2.3424 2.011 0.0038 1.992 0.4649 2.2950 0.3261 2.344 - 0.079

analysis

Explanatory variable

on conspicuity-related

and Shah Alam, Malaysia

motorcycle

Degrees of freedom

0.025 0.0544 0.00056 0.05116 0.05865 0.02641 0.08213 0.02888 0.05801

309.0

155

262.8

154

46.2

241.4

154

67.6

294.6

164

14.4

307.1

154

1.9

analysis

on conspicuity-related

329

accidents

Residual deviance

Dev. Diff

motorcycle

t-Values (uncorrected) 93.696 36.966 6.7857 38.937 7.9267 86.899 3.9705 81.163 - 1.362

p = 0.05

Mean deviance

Yes

1.99

Yes

1.71

Yes

1.57

Yes

1.91

1.99

accidents

Estimates

Standard errors

Residual deviance

Degrees of freedom

Dev. Diff

t-Values

p=o.o5

Constant WEEK RECSYS FAST BLKG RHL

1.850 0.00501 0.324 0.3331 - 0.0848 -0.3544

0.06622 0.00145 0.09398 0.08381 0.05859 0.09923

390.0 263.8 241.3 255.9 224.6 211.8

155 154 153 152 151 150

45.2 22.5 15.4 1.3 12.8

27.937 3.455 3.448 3.975 -1.447 - 3.572

Yes Yes Yes Yes No Yes

1.99 1.71 1.58 1.49 1.49 1.41

Constant WEEK RECSYS FAST RHL

1.835 0.00473 0.3372 0.3404 - 0.3405

0.0654 1 0.00143 0.09350 0.08370 0.09869

309.0 263.8 241.3 225.9 213.9

155 154 153 152 151

45.2 21.5 15.4 12.0

28.054 3.3054 3.6064 4.0669 - 3.450

Yes Yes Yes Yes Yes

1.99 1.71 1.58 1.49 1.42

Table 4. Corrected Model No.

in Seremban

Standard errors

Table 3. Multivariate Model No.

accidents

models

Mean deviance

for overdispersion

Estimates

Standard errors

Residual deviance

Degree of freedom

Dev. Diff

t-Values

Constant WEEK RECSYS FAST BLKG RHL

1.850 0.00501 0.324 0.3331 - 0.0848 -0.3544

0.1088 0.00171 0.11090 0.09892 0.06915 0.1171

225.5 191.8 173.2 162.2 161.2 152.0

155 154 153 152 151 150

33.7 18.6 11.0 1.0 9.2

20.267 2.9402 2.9216 3.3674 - 1.226 - 3.035

Yes Yes Yes Yes No Yes

1.45 1.25 1.14 1.07 1.07 1.01

Constant WEEK RECSYS FAST RHL

1.835 0.00473 0.3372 0.3404 -0.3405

0.07720 0.00169 0.11040 0.09879 0.1165

221.8 188.6 173.2 162.2 153.4

155 154 153 152 151

33.2 15.4 11.0 8.8

23.769 2.8033 3.0543 3.4457 - 2.923

Yes Yes Yes Yes Yes

1.43 1.22 1.13 1.07 1.02

Explanatory variable

from the Poisson model by an estimate of c*, a simple scale factor obtained by dividing the residual deviance or Pearson x2 by the residual degree of freedom. Table 4 shows the corrected model for the overdispersion using the quasi-likelihood technique. Note that the final scaled deviance is now almost equal to the degrees of freedom and the standard errors are redistributed in the models. By inspecting the deviance difference and the ratio of estimates over standard errors in Model lA, it can be seen that all terms, except BLKG, are significant at the 5% level. Thus BLKG term should be removed from the model. Model 2A shows the reduced model with the BLKG term removed from the full model. All terms

p=o.o5

Mean deviance

are now found to be highly significant at 5%, even after the correction for overdispersion. The scaled deviance reduces from 221.8 to 153.4 with 4 df lost. To further refine the model, additional analysis was carried out by adding the interaction terms. The point of fitting the interaction terms was to look more closely at variables in the model which contains the essential variables for interaction terms among variables (Hosmer and Lemeshow 1989). Table 5 shows the analysis of deviance ANODEV of conspicuity-related accidents with the assessment of the interaction terms. Based on the previous criteria of scaled deviance and the ratios of estimates to standard errors, all

RADINUMARR.S. et al.

330

Table 5. Anodev summary table on conspicuity-related

motorcycle accidents Significance of terms

Model fitted Mean WEEK + RECSYS + FAST + BLKG + RHL Assessment of terms -WEEK - RECSYS -FAST - BLKG - RHL WEEK+RECSYS+FAST+RHL Assessment of interaction terms + WEEK.RECSYS + WEEK.FAST + WEEK.BLKG + WEEK.RHL + RECSYS.FAST + RECSYSBLKG + RECSYS.RHL + FAST.BLKG + FAST.RHL +BLKG.RHL *Estimate/standard

Dev. Diff.

309.0 211.8

155 150

Not Applicable Not Applicable

223.8 223.6 226.3 213.9 224.6 213.9

151 151 151 151 151 151

12.0 12.1 14.5 2.1 13.6 Not Applicable

1 1 1 1 1

Yes Yes Yes No Yes

211.7 210.3 212.3 212.0 213.8 211.5 213.9 210.1 206.9 211.6

150 150 150 150 150 149 149 149 150 149

2.2 3.6 1.6 1.9 0.1 2.4 0.0 3.8 7.1 2.3

1 1 1 1 1 2 2 2 1 2

No No No No No No No No No* No

MSTOX Accidents

p = 5%

< 1.96.

terms are not significant at the 5% level. Although deviance for the interaction term FAST.RHL is significant, the t value obtained is less than 1.96. Model 2A can thus be considered as the best fit or parsimonious model to describe MSTOX accidents. Observed and modelled weekly motorcycle accidents are shown in Fig. 2. From this model, it can be said that the RHL intervention has resulted in about 29% less conspicuity-related accidents (exp (-0.3405)) to motorcyclists in the study areas. interaction

df lost, at

Deviance

Although extra care has to be taken when comparing results from different countries, as definitions, data classifications and traffic situations may be different, the above result is comparable with the findings reported by earlier researchers. In earlier RHL evaluations for motorcycles in the U.S.A., Robertson ( 1976) and Waller and Griffin (1977) suggested that about 20-25% of daytime multi-vehicle accidents with motorcycles can be prevented by headlight use laws. In Australia, Vaughan (1976) found a

Fig. 2. Observed and predicted MSTOX accidents in the pilot areas.

Modelling of conspicuity-related

motorcycle accidents in Seremban and Shah Alam, Malaysia

highly significant difference in headlamp usage between the 1104 motorcycles observed in a roadside survey and the 402 motorcycles involved in daytime motorcycle accidents (x2 = 58.6 df= 1, p<O.OOl). A motorcycle which does not have its light on appears to have 2.6 times the chance of being involved in a daylight multi-vehicle accident. These findings are also in agreement with a study conducted by Thomson (1980) in New Zealand who strongly argued that the introduction of the law is by far the best method of reducing motorcycle accidents compared with other countermeasures, and is very likely to have a benefit-cost ratio exceeding 1. In other studies, the effectiveness of laws has been found to have no effect or minor effects; Janoff and Cassel (1971) for example found a marginally significant reduction ( p < 0.1) in daytime accidents in “before” and “after” periods in four U.S. states. Muller (1984) also argued that motorcycle daytime headlight operation is either ineffective or of minor effectiveness in preventing multi-vehicle fatal collisions based on a study in several states in the U.S.A. It should be noted, however, that this present study was carried out in a country where: (i) before the intervention, no motorcyclists used headlights during the day; (ii) motorcycles constitute about 60% of the traffic population; and (iii) headlight usage increased significantly with about 82% using headlights during the daytime “after” the campaign (Radin et al. 1995). Thus RHL intervention should be able to reveal the full benefit of this road safety measure in terms of collision reduction in Malaysia. In addition, the accident series were specially classified into the conspicuity-related and non-conspicuity-related accidents, rather than the general multiple or single vehicle accidents. This classification has prevented any “misclassification” of conspicuity relevant accidents, as some multi-vehicle accidents such as rear-end should not be affected by daytime headlight use. The other main differences between this study and other studies, such as the one carried out by Muller (1984), are that (i) the dawn and dusk accidents and (ii) all categories of accidents, including injury and damage-only cases, were included in the analysis. The larger accident frequencies in each cell of the accident series make the analysis more meaningful statistically. CONCLUSION

AND SUMMARY

The best model to describe the conspicuityrelated accidents MSTOX was found to be model 2A shown in Table 4 and Fig. 2. The model gave the best reduction of scaled deviance per degree of freedom and all terms included were highly significant ( p<O.Ol). The estimates obtained in the model were

331

quite consistent and very much the same as in the univariate analysis. Detailed analysis of the interaction terms shows that all the two-way interaction terms were not significant at 5% level. A crucial question to answer in model building is how much deviance in the model can these interaction terms explain. The more variables included in a model, the greater the estimated standard errors become, and the more dependent the model becomes on the observed data. Thus a parsimonious model should aim to include only those terms that add meaning or substance to explain the data. Based on the above arguments, it can be concluded that the best model to explain conspicuityrelated accidents per week, MSTOX is MSTOX = 6.265 [e0.005WEEK]

[e0.337RECSYS]

~~0.340FAST1

fe-0.341RHL1 and the following conclusions can be made based on the model developed: 1. Week is an important

variable to reflect traffic growth and population. The average annual traffic growth is approximately 17% per year. The term is positively significant with an increase of about 0.5% MSTOX accidents per week. Aggregation of data according to weekly blocks proved to be convenient in describing the explanatory variables over time. The new recording system POL27(Pin l/91) has not only improved the quality of data but the quantity of the data collected. An increase of about 40% of MSTOX accidents was observed after the introduction of the new reporting system. The inclusion of this factor in the overall analysis of the accident series is important in interpreting the effectiveness of RHL intervention. Fasting in the month of Ramadhan is also an important factor to be considered in the overall analysis of RHL intervention. The number of accidents was found to have increased by about 41% in the fasting seasons. The reason for such an increase needs further analysis, though a change in the travelling and socialreligious activities and the presence of “rush” hours could be the explanation for the increase. Although Balik Kampong was thought to be an important factor to describe MSTOX accidents, detailed analysis proved that this term is not significant at the 5% level. This is shown in both univariate and multivariate analyses. The detailed analysis on the effect of RHL intervention revealed that the intervention

RADINUMARR.S. et al.

332

reduced the MSTOX accidents by about 29%. This finding is in line with the earlier analysis based on the local short-term data and with several overseas studies whose findings supported the hypothesis that improved motorcycle conspicuity reduces accidents. Since the population of motorcyclists is extremely high and their compliance to the RHL intervention has been maintained at about 82% over the whole period of observation, the possibility of the novelty effect is very unlikely. Thus it is concluded that the RHL intervention has been successful in improving conspicuity-related accidents in Malaysia. Acknowledgements-The authors are indepted to Geoff Maycock and Jeremy Broughton of TRL, Steve Lawson of the Automobile Association and Richard Hall of Southampton University for

assisting in modelling of the accident series. Data, grants and support for field data collections in the pilot project areas, Seremban and Shah Alam, Malaysia were provided by the Royal Malaysia Police, National Road Safety Council Malaysia, and National Science Council IRPA Malaysia

REFERENCES Aitkin, M.; Anderson, D.; Francis, B.; Hindle, J. Statistical modelling in GLIM. Oxford: Clarendon Press; 1989. Broughton, J. The effect of the 1983 changes to the law relating to drink/driving. Research Report 89. Crowthorne, U.K.: Transport and Road Research Laboratory; 1985. Hills, B.; Elliot, G.; Clark, D. MAAP 5 user guide v5. Crowthorne, U.K.: Overseas Centre, Transport Research Laboratory; 1994. Homel, R. Drink-driving law enforcement and the local blood limit in New South Wales. Accid. Anal. Prev. 26: 147-155; 1994. Hosmer, D.; Lemeshow, S.; Applied logistic regression. New York: Wiley-Interscience; 1989.

Janoff, M.S.; Cassel, A. Effect of daytime motorcycle headlight laws on motorcycle accidents. Highway Research Record 377: 53-63; 1971. Kimber, R.M.; Kennedy, J.V. Accident predictive relations and traffic safety. Proc. Int. Symposium on Traffic Safety Theory and Research Methods. The Netherlands: Inst. Road Safety Research, SWOV; 1988. Maycock, G.; Summersgill, I. Methods for investigating the relationship between accidents, road user behaviour and road design standards, Annex 4. International Research on Safety Effects of Road Design Standards; EC. Report Prepared by the Dutch; 1994. McCullagh, P.; Nelder, J.A. Generalised linear models. Monographs on statistics and applied probability. New York: Chapman and Hall; 1983. Muller, A. Daytime headlight operation and motorcycle fatalities. Accid. Anal. Prev. 16: 1-18; 1984. Payne, C.D. The GLIM system, release 3.77. Generalised linear interactive modelling, 2nd edition. Oxford: Numerical Algorithms Group; 1987. Radin, U.R.S. Detailed analysis on road accidents in pilot areas Seremban and Shah Alam. Research Report No. 4, JK3P. Kuala Lumpur: National Road Safety Council Malaysia; 1992. Radin U.R.S.; Mackay, G.M.; Hills, B. Preliminary analysis of motorcycle accidents: short-term impacts of the running headlights campaign and regulation. J. Traffic Med. U.K. 1995. Robertson, L.S. An instance of effective legal regulation: motorcycle helmet and daytime headlamp laws. Laws and Society Rev. 10: 467-477; 1976. Royal Malaysia Police (RMP). Statistical Report Road Accidents. Malaysia; 1992. Scott, P.P. Variations in two-vehicle accidents frequencies, 1970-1978. Laboratory Report 1086. Crowthorne, U.K.: Transport and Road Research Laboratory; 1983. Thomson, G.A. The role frontal motorcycle conspicuity has on road accidents. Accid. Anal. Prev. 12: 165-178; 1980. Vaughan, R.G. A study of motorcycle crashes, motorcycles and safety symposium. Australian Road Research Board; 1976. Waller, P.F.; Griffin, L.I. The impact of a motorcycle light-on law. Proc. of the Amer. Assoc. of Automotive Medecine, Vancouver; 1977: 14-15.