Bus driver accident record: the return of accident proneness
Summary (2 min read)
1.1 History and theory of accident proneness
- The beginning of safety research has often been traced to British studies about the distributions of accidents in various populations (mainly industrial workers) in the early 20th century (e.g. Greenwood & Woods, 1919; Newbold, 1927; Slocombe & Brakeman, 1930).
- The conclusion drawn from this result is that the liability to have accidents is not only differently distributed in a population, but it is also surprisingly stable over time within each individual (note that most drivers in a year will have no accidents, as will they in the next year, an amazing stability that is not reflected by a correlation).
- So, although most of the researchers who have used culpable accidents have found lower correlations between time periods, as compared to all accidents, and concluded that the effect of such restrictions are negative, they have not taken into account the effect of the lowered mean, or checked their culpability criterion for correctness.
- First, it was investigated whether the intercorrelations between accidents in different time periods are determined by the mean level of accidents in the sample.
2.1 Subjects and data
- A UK bus company made traffic accident data available for analyses from five different geographical regions for the time period 2001-2005.
- For the present samples, the number of routes driven per working week ranged from 2 to 15 and these tend to be 10 driven by the same bus drivers operating from the same rotas week in and week out.
- Given that there is a greater percentage of experienced drivers compared with novice drivers within a typical depot and swopping rotas tends to happen only occasionally, this eventuality is less common.
- The basic variable was thus the number of accidents within a year for each driver.
- For the present data, it has been shown that the criterion for culpability used by the company is too lenient (af Wåhlberg & Dorn, 2007), meaning that several percent of accidents that should have been classified as culpable have not been designated as such.
2.2 Methods of analysis
- There are different ways of calculating the stability of accident record over time.
- The simplest and most common is to take two time periods and correlate the numbers in each for the drivers.
- For the next step two-year periods would be used, and so on until only two periods were left.
- The only critical issue is whether there is enough difference in the accident means for an association to become evident.
2.3 Statistical methods
- The preferred statistical tool for the analysis was Pearson correlation.
- Results 13 First, accidents were correlated between single years for the two accident variables (All and culpable only) for all samples (see Table 2), and the means of these associations computed (first section of Table 3).
- Third, mean correlations for single years against single years (Table 2) were computed by their proximity in time to the other year, with results that are shown in Table 4.
- 14 Finally, experience was added as a factor in the analysis, by separating each sample into those drivers with less than a year’s experience in the beginning of 2001, and others.
- It can be noted that despite the low numbers of drivers in the inexperienced groups (N=31-77) and the resulting possibility for chance deviations, the correlations still largely conformed to a straight regression line, with half the variation in correlation size explained by the mean number of accidents in the period.
4.1 Results
- The results of the present analysis would seem to support the hypotheses stated in the introduction.
- In the present study, exposure was only handled by arranging data per year.
- This hypothesis is in principle testable in the present data, but the results would be hard to interpret, for several reasons.
- Third, as exposure data was not available, it was not possible to calculate how strong the stability of exposure effect could be expected to be.
- This would seem to indicate that some of the stability in accident record is indeed due to differential and stable exposure, but only for non-culpable crashes.
4.2 Methodology and consequences
- One important distinction between concepts needs to be made.
- After the many decades of criticism of the accident proneness theory, the authors tentatively suggest that the statistical arguments that caused it to fall into disrepute may well have been rather prematurely based.
- It might be noted here that there does not seem to be any effect of low means such that the correlation coefficients become smaller than expected at the low end, as could possibly be expected due to the increasing violation of the assumption of normality of the distribution.
- It is not the theory that is deficient, but the previous interpretations of the results.
- The authors are also grateful to Jenny Stannard (Cranfield University) for extracting the relevant data from the company database.
5. References
- Accident statistics and the concept of accident proneness.
- Driver behaviour and accident research methodology; unresolved problems. af Wåhlberg, A. E., & Dorn, L. (2007).
Did you find this useful? Give us your feedback
Citations
90 citations
58 citations
46 citations
Cites methods from "Bus driver accident record: the ret..."
...Also, the Pearson correlation is very robust, showing no tendency to being affected by distributions that are more skewed than those of the variables used in the present study (af Wåhlberg and Dorn 2009)....
[...]
39 citations
Cites background from "Bus driver accident record: the ret..."
...For example, actual, recorded, traffic collision involvement has a certain between-time periods correlation, which increases strongly with the extension of the periods (af Wåhlberg & Dorn, 2009; see also the meta-analysis in af Wåhlberg, 2009)....
[...]
38 citations
Cites background from "Bus driver accident record: the ret..."
...The effect size in accident prediction studies is mainly dependent upon the variation in the accident variable (af Wåhlberg, 2009; af Wåhlberg & Dorn, 2009; af Wåhlberg, Barraclough & Freeman, 2015; Barraclough, af Wåhlberg, Freeman, Watson & Watson, 2016)....
[...]
...Variance in the accident variable has been shown to strongly affect effect sizes in accident prediction studies (af Wåhlberg, 2009; af Wåhlberg & Dorn, 2009; af Wåhlberg, Barraclough & Freeman, 2015; Barraclough, af Wåhlberg, Freeman, Watson & Watson, 2016)....
[...]
References
48 citations
47 citations
"Bus driver accident record: the ret..." refers background in this paper
...…et al. 1986, Dobson et al. 1999), as well as the correlations between culpable and non-culpable in different studies (e.g. Goldstein and Mosel 1958, McBain 1970, Gully et al. 1995, Arthur and Graziano 1996) show that, in many cases, the criterion does not achieve what it set out to do with any…...
[...]
46 citations
44 citations
"Bus driver accident record: the ret..." refers background in this paper
...those of Häkkinen (1979) and Blasco et al. (2003), whose high-risk, longitudinal samples do not fit into the general picture (Figure 1), unless curvilinearity is assumed. It might be added that Peck (1993) presented calculations on the maximum attainable correlation between accidents and a ‘perfect predictor’, given the variance of the former....
[...]
...…lack of published data in the range of 15–20 accidents (i.e. samples with such means) lying between the results quoted here and those of Häkkinen (1979) and Blasco et al. (2003), whose high-risk, longitudinal samples do not fit into the general picture (Figure 1), unless curvilinearity is assumed....
[...]
...those of Häkkinen (1979) and Blasco et al. (2003), whose high-risk, longitudinal samples do not fit into the general picture (Figure 1), unless curvilinearity is assumed. It might be added that Peck (1993) presented calculations on the maximum attainable correlation between accidents and a ‘perfect predictor’, given the variance of the former. In essence, what Peck was saying was the same thing as what has been pointed out here: if variance is low, correlations will be small, no matter how good the predictor, i.e. the actual association between the variables. One of the conclusions drawn from some of the low-correlation-between-time-periods studies referenced in Section 1 was that it would be fairly useless to exclude the highaccident drivers from driving, because they were few and most of them did not repeat their high number of crashes in the next time period. However, the present results indicate that these results and thinking were faulty and simply artefacts of the short time periods and low-risk populations utilised. If longer periods had been used, the effect of removing drivers would have been larger. After the many decades of criticism of the accident proneness theory, it is tentatively suggested here that the statistical arguments that caused it to fall into disrepute may well have been rather prematurely based. The positive results of the policy to exclude high-accident-involved drivers from driving within a transportation company for an eight-year period was reported by Rawson (1944) but does otherwise not seem to have been evaluated or reported as being a policy currently advocated by transportation companies, perhaps because the consensus has been that such a policy would be futile given the early arguments surrounding the notion of accident proneness....
[...]
...…drivers have reported fairly strong (4 0.40) associations over time (e.g. Cresswell and Froggatt 1963, Bach et al. 1975, Milosevic and Vucinic 1975, Blasco et al. 2003); third, low correlations have almost all been found in car driver populations (e.g. Forbes 1939, Burg 1970, Stewart and Campbell…...
[...]
...those of Häkkinen (1979) and Blasco et al. (2003), whose high-risk, longitudinal samples do not fit into the general picture (Figure 1), unless curvilinearity is assumed....
[...]
43 citations