Transportation Research Part F: Traffic Psychology and Behaviour
Personality versus traffic accidents; meta-analysis of real and method effects
Introduction
The present paper summarizes the literature on personality (in terms of the Big Five system) as a predictor of traffic accident involvement in a meta-analysis. Several methodological problems are considered, such as outliers, dissemination bias and conversion of data between different personality systems.
Personality as a phenomenon is multi-faceted, but can usually be defined as the stable behavioural tendencies of people over time, or the psychological traits which cause such behaviours. This has been conceptualised in many different ways through the years, but today it is agreed by most researchers that the most parsimonious description is by five dimensions; Openness, Agreeableness, Conscientiousness, Neuroticism and Extraversion. Most other systems map onto these dimensions, and results can therefore be converted between them.
Throughout the history of traffic safety, researchers have studied the influence of individual differences in personality on accident record (although at first the term ‘accident proneness’ was used; Greenwood & Woods, 1919; see also papers by Drake, 1940, Harris, 1950, Parker, 1953). Many researchers have proposed that certain personality features, in terms of recurrent behaviours, cause accidents. In terms of the Big Five model (and its facets), Clarke and Robertson (2005) summarised the theoretical basis for their traffic accident-causing properties thus; people high on Extraversion tend to be poor on vigilance and take more risks. Those high on Neuroticism have been suggested to be easily distracted, less likely to seek control of the environment and prone to react to stress. Conscientiousness features several inter-related concepts which are thought to make people safe, such as planning, self-control and decision-making, while lack of Agreeableness is associated with accidents by the mechanism of aggression in terms of emotion as well as behaviour. Finally, Openness has been suggested to be positively correlated with accidents, due to the routine character of driving, where traits such as experimentation and improvisation are not in accord with safe operation. However, most researchers who investigate the link between personality and accidents refer to previous significant associations reported, and describe the behaviours typical of a certain personality dimension (e.g. Arthur et al., 2001, Begg et al., 1999, Burns and Wilde, 1995, Clement and Jonah, 1984, Hartman and Rawson, 1992).
Many researchers also express a strong belief in the predictive capacity of tests of personality versus accidents (e.g. Arthur et al., 2001, Brandau et al., 2011, Hansen, 1988, Jonah, 1997). However, results, as always, have been mixed, and this belief may therefore be unfounded. For example, Shaw and Sichel, 1971, Shaw, 1965 reported correlations between .4 and .7 for their personality tests and accidents for bus drivers, while Carty, Stough, and Gillespie (1998) found a strong negative association (−.212) instead of the expected positive one for Extraversion, and many other such examples exist. Results are thus very heterogeneous, which make interpretation difficult. A meta-analytic approach is therefore needed, where the reasons for this apparent heterogeneity can be identified, and estimates of the true (population) effects calculated.
Two meta-analyses of personality versus accidents have already been published; Arthur et al., 1991, Clarke and Robertson, 2005. However, there are several reasons for why a new analysis of the personality-traffic accident association is needed. Apart from now being outdated, the Arthur et al. study used a personality taxonomy which excluded some available studies (e.g. Andersson et al., 1970, Jamison and McGlothlin, 1973, Quenault, 1967). Similarly, the Clarke and Robertson study excluded many available papers, while including some which used methodologies which were different from those of the majority. Furthermore, moderator effects and dissemination bias were not investigated in these studies.
We therefore wanted to undertake a new meta-analysis which used a very different approach to the problem of meta-analysing personality as a predictor of traffic accident involvement, taking into account not only the well-known problems of dissemination bias and methodological moderator effects, but also effects which are probably peculiar to accident prediction studies. The main aim of the study was therefore to estimate the population effect while keeping known or suspected moderators constant, as will now be described.
This section describes some of the methodological problems associated with meta-analysing data, under the general headings of trying to estimate a population effect, and the overall problem of heterogeneous data, i.e. very different results in different studies. Also, possible remedies are suggested.
In research on psychological mechanisms, it is usually the goal to infer from sampled data what all people are like in a defined population. For example, are high levels of empathy usually associated with low levels of aggression? In a meta-analytic context, it would specifically be asked what the effect size is, i.e. how strong is the link between the two concepts? When effect sizes from different studies are combined, however, it is important that the data included is actually drawn from the same population, meaning those who share this trait/mechanism. For example, the link between empathy and aggression might have different strength in different cultures. If studies from different cultures are then combined, the ensuing effect size will be slightly misleading, showing really the mean effect for two (or more) different populations. When effect sizes from different populations are mixed, it is said that the meta-dataset is heterogeneous, i.e. the numbers differ more between themselves than could be expected by random sampling (which can be ascertained by statistical testing).
Heterogeneity can also be caused by differences in methodology. For example, it can be expected that experiments and field studies will yield different effect sizes, although they are ostensibly studying the same problem, because part of the effect is actually created by the method used (e.g., a social science analogue to Heisenberg’s uncertainty principle).
If heterogeneity is detected in the data, moderator1 analysis should be applied to investigate the causes of the variance. For example, the pooled effects for experiments versus field studies can be compared, to see whether the estimated population effects differ significantly between these two conditions. If they do, it can be concluded that the methods used have had an influence on the results, a fact that needs to be considered when the true population effect size is identified.
When meta-data has been gathered, an important operation is therefore to detect whether effects differ more between themselves than could be expected from sampling error alone. However, before moderator analysis is undertaken, data should be checked for outlying values, i.e. values which differ very strongly from the majority, and could be suspected to be due to errors in the research process. If a few such values are found, these deviating numbers can be excluded (e.g. Bond and Smith, 1996, de Winter and Dodou, 2010, Eagly et al., 1992, Fournier et al., 2010, Groh et al., 2014), although most meta-analysts in social science do not proceed beyond concluding that there is heterogeneity between effects (while in ’hard’ sciences, such as physics, they do; Hedges, 1987).
In this section, the methodological problems of meta-analysis, with a special emphasis on the problems associated with accident prediction and personality, are further described. The solutions chosen for the present analysis are also described.
Variance in the accident variable has been shown to strongly affect effect sizes in accident prediction studies (af Wåhlberg, 2009, af Wåhlberg et al., 2015, af Wåhlberg and Dorn, 2009, Barraclough et al., 2016). In essence this is a problem of differences in restriction of variance between studies. This means that if a fair comparison of effects between sources of data or other moderators is to be undertaken in accident prediction studies, variance/restriction of range should be held constant. This can be undertaken by using a correction formula suggested by Hunter and Schmidt (1990). However, this formula uses the standard deviations of the samples, a statistic which is not always reported in accident prediction studies (61% in af Wåhlberg et al., 2015). Therefore, unless a large part of the available data is to be excluded, a different method is needed to correct for range restriction and make the results comparable between studies. In the present paper, the empirically derived association between the mean in the accident variable and the effect size for the predictor used (first calculated in af Wåhlberg, 2009) will be utilised.
The moderating effect of the variance in the accident variable also influences the statistics calculated, and the interpretation of the results. In standard meta-analysis, the goal is to calculate a mean effect over samples, and often to estimate a true population effect by correcting for measurement error (as advocated by Hunter & Schmidt, 1990). For individual differences in accident involvement, however, this is not meaningful, as any mean would only be ’true’ for a specific level of variance in the accident variable.2 Similarly, a correction for measurement error would involve the reliability of accidents, something which has not been established as a single value. Instead, a calculation of how the effect varies with the variation should be more relevant.
Yet another correction procedure which is sometimes used in meta-analysis is to adjust for the unreliability of the predictor, yielding an estimate of what the effect size would be if the predictor was perfectly measured. Given the state of the present data, it was chosen not to apply this method, as it would probably be impossible to retrieve reliability information about several of the instruments used. Instead, the main approach in this paper was to include as much data as possible, and leave correction procedures for the future, using a more restricted data set.
Errors in the research process can create very deviating values (outliers), which can unduly influence the population effect estimate (Hunter & Schmidt, 1990). However, as strong variance in effect sizes between studies due to differences in range restriction in the accident variable was expected in the present data (af Wåhlberg, 2009), standard univariate methods for outlier detection were not applicable. Instead, a new method for identifying suspect data points was applied in the present study. It uses the standard criterion of two standard deviations from the mean as a cut-off for outlying values, but applies this to bi-variate data points. This will be further described in the method section.
Dissemination bias (previously known as publication bias; Bax & Moons, 2011), such as when the results of a study influences its availability, is a problem in many areas of research (Ioannidis, Munafò, Fusar-Poli, Nosek, & David, 2014). Therefore, many different methods for detecting such bias have been invented, mostly based upon the assumptions of larger studies having more reliable results and studies with large effects being easier to publish (Møller & Jennions, 2001). This means, for example, that if the number of subjects and the effect sizes in published studies are negatively correlated, a number of small studies with small effects have probably not been published. However, these methods are not considered fully reliable (Pham et al., 2001, Song et al., 2010, Vevea and Woods, 2005), and tend to have low power (Macaskill et al., 2001, Sterne et al., 2000) and it is therefore preferable to try to actually locate unpublished data, a method which has previously yielded a significant amount of additional data (af Wåhlberg et al., 2015, Barraclough et al., 2016, Eyding et al., 2010, Judge et al., 2004).
There are good reasons to suspect that in studies using self-reported accident data as well as self-reported predictors, effects are artificially increased (af Wåhlberg, 2009, Hessing et al., 1988, Podsakoff et al., 2003, Schwartz, 1999). The problem of single-source data, especially self-reports, is that common method variance can influence the results, and sometimes substantially change the true effect size. This distorting effect is also possible for studies into personality, leading to the prediction that studies on personality as a predictor of accident involvement will have larger effect sizes if the criterion is self-reported than if it is objectively gathered data. Such an effect has previously been found in meta-analyses by Reijntjes, Kamphuis, Prinzie, and Telch (2010) for internalizing problems and peer victimisation, by af Wåhlberg et al. (2015) for a driver behaviour inventory and by Barraclough et al. (2016) for citations versus crashes, but see also Arthur et al., 1991, Morina et al., 2015 for less clear-cut results.
In summary, the present study sets out to meta-analyse the association between personality in terms of the Big Five dimensions, measured by standard personality scales, and traffic accident involvement (main aim). All other analyses (identifying suspicious values, and testing for various method effects and biases) were included to increase the precision of the population estimate. This included effects of variance in the accident variable and common method variance, differences been inventories, as well as the more commonly known problem of dissemination bias.
Section snippets
General
The preliminary work on this paper followed the standard guidelines for meta-analysis (e.g. Chung et al., 2006, Field and Gillett, 2010; and the discussion pieces of Aguinis et al., 2011, Huf et al., 2011, Noble, 2006, Orme-Johnson et al., 2014). Thereafter, it was mainly geared towards investigation of the effect of moderators, in similarity to, for example, Bond and Smith (1996) and as described in Steel and Kammeyer-Mueller (2002). One feature that was unusual was the great number of
General
Besides the overall aim of summarising the research information on personality as a traffic accident predictor, the present study also aimed to report all the information used in the analysis, so as to facilitate future analysis by other researchers. Also, as many decisions, for example about conversion of scales into Big Five, can be criticised, it was deemed important to report this in full detail, so that any errors can be corrected by other researchers.
As the amount of information for each
Results
This paper aimed to investigate the association between personality, as expressed in the Big Five dimensions, and traffic accident involvement. Although the objective was to estimate the population effect, most of the calculations were undertaken to check and control for possible moderators, including variance in the accident variable, source of the data and dissemination bias. After controlling for such effects, the population estimates were found to be in the range of .01–.07. How does this
Acknowledgements
A number of researchers have responded to e-mailed questions about their data, sent papers and/or further statistical results, and the authors would therefore like to thank the following colleagues; Mark Conner (University of Leeds, UK), Pauline Gulliver (New Zealand), Joan Harvey (Newcastle University, UK), Dragan Jovanovic (Department of Transport, Serbia), Timo Lajunen (Middle East Technical University, Turkey), Inese Muzikante (University of Latvia, Latvia), Laura Seibokaite (Vytautas
References (92)
Some methodological deficiencies in studies on traffic accident predictors
Accident Analysis and Prevention
(2003)The relation of non-culpable traffic incidents to bus drivers’ celeration behavior
Journal of Safety Research
(2008)A reporting guide for studies on individual differences in safety
Journal of Safety Research
(2010)- et al.
The driver behaviour questionnaire as accident predictor; a methodological re-meta-analysis
Journal of Safety Research
(2015) - et al.
Beyond publication bias
Journal of Clinical Epidemiology
(2011) - et al.
A longitudinal study of lifestyle factors as predictors of injuries and crashes among young adults
Accident Analysis and Prevention
(1999) - et al.
Personality subtypes of young moped drivers, their relationship to risk-taking behavior and involvement in road crashes in an Austrian sample
Accident Analysis and Prevention
(2011) - et al.
Risk taking in male taxi drivers: Relationships among personality, observational data and driver records
Personality and Individual Differences
(1995) - et al.
A practical guide to meta-analysis
Journal of Hand Surgery
(2006) - et al.
Field dependence, sensation seeking and driving behaviour
Personality and Individual Differences
(1984)