Abstract

Increasingly often, practitioners in neuropsychological rehabilitation centers are called upon to assess patients' fitness to drive after brain injury. There is, therefore, a need for valid and reliable psychometric test batteries that enable unsafe drivers to be identified. This article investigates the contribution of five driving-related personality traits to the prediction of fitness to drive in patients suffering from traumatic brain injuries (TBI) or strokes over and above cognitive ability traits that have already shown to be related to safe driving. A total of 178 patients suffering from either strokes or TBI participated in this study. All the participants completed a standardized psychometric test battery and subsequently took a standardized driving test. The contribution of the driving-related ability and personality traits to the prediction of fitness to drive was investigated by means of a logistic regression analysis and an artificial neural network. The results indicate that both cognitive ability and personality factors are important in predicting fitness to drive, although cognitive ability factors contribute slightly more to the prediction of patients' actual fitness to drive than personality factors. Furthermore, even though there are subtle differences in the predictive models obtained for the two subsamples (stroke and TBI patients), these differences are adequately accounted for by a more unitary model calculated by means of an artificial neural network that is capable of taking account of moderating effects between the predictor variables.

Introduction

Resuming driving after a traumatic brain injury (TBI) or stroke is an important step toward independent living and improvement in the quality of life for these patients. However, it is also an issue of personal and public safety because these patients exhibit perceptual and cognitive impairments that can be assumed to be related to safe driving (cf. Lezak, 1995; Stuss, Shallics, Alexander, & Picton, 1995). Under Austrian and German law, patients who have suffered a TBI or stroke are often instructed by the rehabilitation center to undergo a traffic psychological examination to determine their fitness to drive (cf. Hartje, 2004). Söderstrom, Perretson, and Leppert (2006) found that practitioners in neuropsychological rehabilitation centers and traffic psychological examination centers commonly resort to psychometric test batteries to evaluate patients' fitness to drive due to the high costs associated with the administration of standardized on-road driving tests. However, this procedure requires a sufficient correlation between measures of patients' actual fitness to drive and a psychometric test battery consisting of individual ability and personality tests that have proved to be significantly related to safe driving in healthy adults and patients suffering from TBI or strokes (Groeger, 1997).

Measures of Actual Driving Fitness

In the current research literature, various measures of fitness to drive have been proposed, ranging from accident rates and traffic violations to off- and on-road driving tests. Several studies have focused on accident rates as a measure of respondents' fitness to drive. However, accident rates have several shortcomings. One such shortcoming concerns the fact that accidents are rare events and multicausally determined (Klebelsberg, 1982). This results in unfavorable statistical characteristics such as the Poisson distribution. The second shortcoming concerns the measurement of accident rates. In previous studies, different measurement methods, such as data gathered from official sources or insurance company records and self-reports, have varied considerably in the level of agreement they exhibit (r = .11 to .67; e.g., Owsley, Ball, Sloane, Roenker, & Bruni, 1991; Szlyk, Seiple, & Viana, 1995). Furthermore, in comparing self-reports and tape-recorded driving diaries in a sample of healthy adults, Chapman and Underwood (2000) demonstrated that about 80% of less severe accidents and a major proportion of near-accidents had been forgotten within a period of 2 weeks. This latter problem might be even more pronounced in samples of patients suffering from TBI or strokes due to the fact that these patients often suffer from impairments in memory functioning (cf. Lezak, 1995).

In order to circumvent these problems, some researchers have proposed more behaviorally based measures of respondents' fitness to drive, such as standardized on-road driving tests (cf. Fox, Bowden, & Smith, 1998; Risser et al., 2008). From a theoretical point of view, the primary advantage of this measure of respondents' fitness to drive is the capability to take account of precursors of traffic accidents, such as near-accidents or risky driving habits (cf. Evans, 1991). Furthermore, given a sufficient degree of standardization and a broad and representative spectrum of driving tasks, standardized on-road tests can be assumed to provide a reliable and valid measure of patients' fitness to drive (Fox et al., 1998), which is also less prone to impression management than self-reported accident rates (Hjälmdahl & Várhelyi, 2004). In addition, patients' global scores in standardized driving tests turned out to be significantly correlated with the proficiency ratings of experienced driving instructors (for an overview, seeFox et al., 1998) and other measures of patients' actual fitness to drive, such as accident rates (e.g., Chaloupka & Risser, 1995; Risser & Brandstätter, 1985). Overall, these results argue for the validity of standardized driving tests.

Ability and Personality Determinants of Post-Injury Driving Fitness

Since human failure turned out to be a main cause in a significant proportion of traffic accidents (cf. Evans, 1991), researchers started to investigate the ability and personality determinants of fitness to drive in patients who had suffered a TBI or a stroke.

Predictive Validity of Driving-Related Personality Traits

Driving-related personality traits are assumed to exert their influence primarily on the strategic and tactical level of driving (Hatakka, Keskinen, Gregersen, Glad, & Hernetkoski, 2002; Michon, 1985). These two levels are concerned with general decisions about route and vehicle choice prior to driving and the adjustment of driving goals according to the current traffic situation. According to Lundquist, Gerdle, and Rönnberg (2000) and Sommer, Herle, Häusler, Risser, and colleagues (2008), decisions made at the tactical level should be influenced by driving-related personality traits such as emotional stability, social responsibility, self-control, and subjectively accepted level of risk because these personality traits determine the goals that need to be met. Evidence indicating that these personality traits may be important in distinguishing drivers judged to be more or less likely to be fit to drive stems mainly from behavioral observations of patients' performance in standardized on-road driving tests or driving simulators (cf. Lundquist, Alm, Gerdle, Levander, & Rönnberg, 1997; van Zomeren & Brouwer, 1994; van Zomeren, Brouwer, Rothengatter, & Snoek, 1988) and studies conducted with healthy adults and elderly drivers (for an overview, seeSommer & Häusler, 2005; Sommer, Herle, Häusler, Risser, et al., 2008). Unfortunately, there is currently no direct empirical evidence of the importance of these personality traits in predicting fitness to drive in TBI or stroke patients.

Predictive Validity of Driving-Related Ability Traits

Driving-related cognitive abilities, on the other hand, are assumed to affect primarily the tactical and operational level of driving because the adjustment of driving goals according to the current traffic situation and the implementation of driving maneuvers and instant reactions to the traffic environment crucially depend on the cognitive abilities of the driver. According to Lundquist and colleagues (2000) and Sommer, Herle, Häusler, Risser, and colleagues (2008), decisions made at the tactical level should be influenced by complex choice reaction, fluid intelligence, and divided attention while the implementation of driving maneuvers and instant reactions to the traffic environment that characterize the operational level of driving depend on selective attention, perceptual speed, and simple reaction speed.

Even though a lot of research has been conducted on the cognitive ability determinants of safe driving in stroke and TBI patients, the results regarding the predictive validity of these cognitive ability traits are not entirely unequivocal. The findings regarding the predictive validity of complex choice reaction and perceptual speed seem to be an exception from this norm. In the majority of the studies conducted thus far, these two driving-related cognitive abilities turned out to contribute significantly to the identification of unfit drivers in samples of stroke patients (complex choice reaction: Hartje, Pach, Willmes, Hannen, & Weber, 1991; Lundquist et al., 2000; perceptual speed: Hartje et al., 1991; Marshall et al., 2007; Mazer, Korner-Bitensky, & Sofer, 1998), TBI patients (complex choice reaction: Lundquist et al., 1997; perceptual speed: Brooke, Questad, Patterson, & Valois, 1992; Coleman et al., 2002; Galski, Bruno, & Ehle, 1992; Korteling & Kaptein, 1996; van Zomeren et al., 1988), and mixed samples consisting of stroke patients and TBI patients (complex choice reaction: Burgard, 2005; Hannen, Hartje, & Skeczek, 1998; perceptual speed: Bouillon, Mazer, & Gelinas, 2006; Burgard, 2005; Niemann & Döhner, 1999). Only a few studies failed to provide evidence of the predictive validity of these two cognitive abilities (cf. Hannen et al., 1998; Mazer et al., 1998; Söderstrom et al., 2006).

The situation is less clear in the case of the driving-related cognitive ability traits fluid intelligence, divided attention, selective attention, reaction speed, and visual/perceptual abilities. Although some studies conducted with stroke patients (Nouri & Lincoln, 1992; Nouri & Tinson, 1988) and TBI patients (Coleman et al., 2002; Galski et al., 1992) argue for the validity of fluid intelligence in predicting fitness to drive, others (Hartje et al., 1991; Radford, Lincoln, & Murray-Leslie, 2004) failed to find evidence for the predictive validity of this driving-related ability trait. Similar results have been obtained with regard to divided attention, selective attention, and reaction speed. Although some studies conducted with stroke patients (divided attention: Akinwuntan et al., 2002; Lundquist et al., 2000; selective attention: Burgard, 2005; Lundquist et al., 2000; McKenna, Jefferies, Dobson, & Frude, 2004; Nouri & Lincoln, 1992; Nouri & Tinson, 1988; reaction speed: Akinwuntan et al., 2002; Bouillon et al., 2006; Hannen et al., 1998; Korner-Bitensky et al., 2000; Mazer et al., 1998) or TBI patients (divided attention: Lundquist et al., 1997; Radford et al., 2004; selective attention: Coleman et al., 2002; Galski et al., 1992; Lundquist et al., 1997; Radford et al., 2004; van Zomeren et al., 1988; reaction speed: Bouillon et al., 2006; Galski et al., 1992; Hannen et al., 1998; van Zomeren et al., 1988) argue for the predictive validity of these measures, others (cf. Korteling & Kaptein, 1996; Niemann & Döhner, 1999) fail to do so even though these measures turned out to be correlated with measures of fitness to drive.

Formulation of the Problem

Although current theoretical models of driving (Groeger, 2000; Michon, 1985) emphasize the importance of taking driving-related ability and personality traits into account in predicting fitness to drive, none of the currently available studies included personality tests or questionnaires in their test batteries. The present study thus attempts to fill this gap by investigating the contribution of driving-related personality traits to the prediction of patients' actual fitness to drive. The currently available models of driving behavior have not been elaborated with a level of detail that enables the type of structural relationship between all relevant driving-related ability and personality traits to be deduced precisely. The contribution of driving-related ability and personality traits will therefore be investigated by means of a logistic regression analysis and an artificial neural network. The advantage of the artificial neural network is its ability to model complex nonlinear relations without the need to specify the type of relations in advance (Rakes, 1991; Somers, 2001). Following prior research with healthy adult drivers (Risser et al., 2008; Sommer, Herle, Häusler, Risser, et al., 2008), the results obtained with both methods will be compared in order to determine the most parsimonious and valid predictive model.

Furthermore, since previous research (cf. McKenna et al., 2004; Söderstrom et al., 2006) found differences between stroke and TBI patients in the contribution of cognitive ability traits to the prediction of patients' actual fitness to drive, the calculations outlined above will be carried out for both subsamples separately. The results obtained in both subsamples will be compared with each other with regard to the contribution of the cognitive and personality traits to the prediction of fitness to drive and the generalizability of the results across both subsamples. Furthermore, we will investigate whether the structural relation between these two kinds of predictor variables and the criterion measure can be adequately modeled by a unitary model calculated on the basis of the combined sample.

Materials and Methods

The data in this study were obtained in two phases in the course of a multicenter study investigating the predictive validity of a test battery taken from the Expert System Traffic (XPSV; Schuhfried, 2005). The data were gathered between October 2005 and June 2008 through the Generation Research Program (GRP) in Bad Tölz (Germany) and the “Praxis für Allgemeine und Neuropsychologische Psychotherapie” (Office for General and Neuropsychological Psychotherapy) in Aachen (Germany); seven different rehabilitation centers and clinics and two specialist consultants were involved. The administration of the standardized on-road driving test and the psychometric test battery took place in two different locations (Bad Tölz and Aachen). Participants were recruited and informed about the study during their rehabilitation stay or thereafter. In order to be included in the sample participants had to fulfill the following inclusion criteria: (1) Participants are neurological patients who have suffered either a stroke or a TBI, (2) the question of participants' driving fitness has arisen in the course of their neuropsychological rehabilitation, (3) participants have held a European Union Class B driving license for at least 5 years prior to their brain injury, and (4) participants are aged 21 to 70 years. These inclusion criteria were used in order to enhance the ecological validity of the results obtained in this study and their practical relevance to professionals working in the field of driving rehabilitation in Germany and Austria.

In addition to these inclusion criteria, the following exclusion criteria were used in line with German government regulations on fitness to drive after TBI or strokes (FeV Anlage 5): (1) Visual acuity with the assistance of spectacles ≥.5 diopter in one eye and ≥.2 diopter in the other eye, (2) visual impairments such as hemianopsia or visual neglect, (3) coronary heart disease (e.g., cardiac arrhythmia with unforeseeable blackouts, angina pectoris during phases of relaxation or physiological excitation), (4) neuronal diseases such as epileptic seizures or sudden blackouts, (5) mental disorders or disabilities (severe mental handicaps, severe and pathological premature aging, severe personality-related lack of judgment), (6) kidney disease (renal failure), (7) substance abuse or dependency (e.g., alcohol, drugs), and (8) prescribed medication which impairs one's fitness to drive.

All patients took part in a structured interview to assess relevant aspects of their driving history before their brain injury. Information on the patients' medical status was also obtained from their medical records by means of a standardized questionnaire handed to the clinicians working in the rehabilitation centers and clinics participating in the study. In addition, all participants completed a standardized psychometric test battery and subsequently took a standardized driving test. The collection of test data and the observation of driving behavior were carried out by independent test administrators.

Measures

Description of the Structured Interview and Questionnaire Data

The structured interview was divided into questions about the patient's medical condition and questions about his or her driving experience. The sections dealing with driving experience elicited the number of traffic accidents in which the patient had been at fault, the number of times the patient's driving license had been withdrawn prior to the TBI or stroke as well as the average annual mileage driven before the injury. Patients were also asked whether they currently drove a car and which traffic environment they normally had to deal with, e.g., interstate, country roads, city traffic.

The medical section of the interview dealt with issues relating to the interval between the injury and the present study and also elicited information on patients' current medication and on any neurological conditions that by law preclude people from driving in Germany and Austria. This section also contained questions put to the treating physicians about the type and severity of the brain injury. The treating physicians had to consult existing patient records in order to answer the questions in this section of the interview. The etiological subtypes of strokes were classified using the ICD-10 scheme. Ischemic strokes were further subdivided by means of the Trial of Org 10172 in Acute Stroke Treatment (TOAST) system (Adams et al., 1993). The interview also contained a section enabling TBI to be classified according to their severity using the Glasgow Coma Scale. This classification of severity describes the severity of the injury at the time the patients were referred to the respective rehabilitation center and thus does not necessarily reflect the current severity of the injury. Besides classification of the severity of the injury, the interview also obtained data which enabled us to distinguish between open and closed TBI and between diffuse and focal TBI. In addition, the Barthel Index (Mahoney & Barthel, 1965) was used to provide a measure of the extent of the patient's disability at the time of the study.

Description of the Test Battery

To assess the ability and personality determinants of patients' fitness to drive, various tests were selected from the XPSV (Schuhfried, 2005), which is a computerized test battery designed to identify unsafe drivers. All measures used in this study proved to contribute significantly to the prediction of fitness to drive in studies conducted with healthy adults (e.g., Häusler & Sommer, 2005; Risser et al., 2008; Sommer, Herle, Häusler, Risser, et al., 2008). In addition, the driving-related ability tests used in this study also contributed significantly to the prediction of fitness to drive in samples of TBI or stroke patients (e.g., Burgard, 2005; Hannen et al., 1998; Hartje et al., 1991). The following tests were used.

Adaptive Matrices Test (S1)

This test was used to assess inductive reasoning, which is a marker of fluid intelligence (Gf). The items resemble classical matrices. The respondent's task is to identify the rules which govern the given matrices items and fill out the empty space in each item by selecting from eight answer alternatives the one that completes the matrix. There is no time limit on solving the problems. The test was administered as a computerized adaptive test (CAT). The 1PL Rasch model person parameter estimate represents the main variable of this test, which is calculated taking into account the difficulty of the items administered and whether each item has been solved by the respondent or not. In contrast to classical matrices tests, the Adaptive Matrices Test is based on an explicit construction rationale which specifies the cognitive processes and solution strategies that respondents use to solve these items. The empirically estimated item difficulty parameter estimates based on the 1PL Rasch model correlate at .70 with the item difficulty parameters predicted on the basis of the theoretical model. Furthermore, confirmatory factor analyses confirm that this test indeed measures inductive reasoning and is a marker of the higher-order factor Gf (Hornke, Etzel, & Rettig, 2003). In sum, the available evidence indicates that construct validity can be assumed for this test. The target reliability was set at α = .70.

Determination Test (S1)

This test is commonly used in traffic psychological examinations in Austria and Germany to measure the “complex choice reaction” of healthy and brain-injured respondents (cf. Golz, Huchler, Jörg, & Küst, 2004; Schuhfried, 1998). The respondent's task is to react to nine different stimuli by pressing the corresponding button on the response panel of the Vienna Test System. Five visual stimuli (white, yellow, blue, green, and red light), two acoustic stimuli (high and low frequency tone), and two stimuli requiring a response with the left or right foot pedal are presented in a predetermined sequence. The test is administered as a CAT, so that the stimulus presentation time adjusts to the respondent's reaction speed. However, unlike classic CATs, this test form presents the stimuli a little faster than would be optimal for the respondent's reaction speed, thus resulting in a condition of sensory stress. The main variable “correct reactions” was used on the basis of prior research. Prior research had already indicated that this test contributes significantly to the prediction of fitness to drive as measured in a standardized driving test in patients suffering from TBI or strokes (cf. Burgard, 2005; Hannen et al., 1998; Hartje et al., 1991). In the present study, the reliability (Cronbach's α) of the main variable “correct reactions” was α = .98.

Reaction Test (S3)

This test measures simple choice reaction time (Golz et al., 2004; Schuhfried & Prieler, 1997). Auditory (high and low frequency tones) and visual stimuli (yellow and red light) are presented to the respondent in any combination of auditory and visual stimuli. During presentation, the respondent places his finger on a rest button. Whenever a yellow circle appears at the same time as a high frequency tone, the patient must remove his finger from the rest button and press a defined button immediately above it on the response panel of the Vienna Test System. The mean reaction time from the onset of the relevant stimulus configuration to the lifting of the finger from the rest button serves as a measure of “decision speed.” “Physical motor speed” is measured by the latency from the start of the finger-lifting movement to the moment when the response key is pressed. In prior studies investigating the predictive validity of this test in identifying unsafe drivers suffering from TBI or strokes, both main variables were found to be significantly correlated with the patients' performance in a standardized driving test (cf. Burgard, 2005; Hannen et al., 1998; Hartje et al., 1991). The reliability coefficients (Cronbach's α) obtained in the present study were α = .94 (decision speed) and .98 (physical motor speed).

Tachistoscopic Traffic Perception Test (S1)

In this test, pictures of traffic scenes are presented to the patient for 1 s each. The test was administered as a linear test containing a total of k = 20 items. After presentation of each of the k = 20 traffic scenes, the patient has to select from a list of five different categories all the object classes that were visible in the picture (e.g., vehicles, bikes, pedestrians). The main variable “overview” represents the number of items for which all visible object classes were correctly identified by the respondent and no object class that was not visible in the traffic scene was marked. This variable serves as a measure of “perceptual speed” (cf. Biehl, 1996; Golz et al., 2004). The reliability coefficient (Cronbach's α) of the main variable “overview” obtained in the present study was α = .82. The construction of the items is based on an explicit construction rationale which specifies the cognitive processes respondents use to solve these items. The empirically estimated item difficulty parameters based on the 1PL Rasch model correlate at .89 with the item difficulty parameters predicted on the basis of the construction rationale (Sommer, Herle, Häusler, & Arendasy, 2008). Furthermore, the test performance of patients suffering from TBI or strokes was found to correlate significantly with their performance in a standardized driving test and to contribute incrementally to the identification of patients judged to be unfit to drive based on their performance in the standardized driving test (cf. Burgard, 2005; Hannen et al., 1998; Hartje et al., 1991).

Cognitrone (S11)

This test was used to measure “selective attention.” The respondent's task is to compare an abstract reference figure with four comparison figures and to determine whether it matches one of the four comparison figures. If it does, the respondent must press the green button on the response panel of the Vienna Test System. Otherwise, the red button must be pressed. The items are administered without any time limit. The mean time of correct rejections serves as a measure of “selective attention” (Wagner & Karner, 2001). The reliability coefficient (Cronbach's α) of the main variable “mean time of correct rejections” obtained in the present study was α = .95. Prior research conducted by Burgard (2005) indicated that the test performance of patients suffering from TBI or strokes was significantly correlated with their performance in a standardized driving test.

Peripheral Perception Test

This test is used to assess “field of vision” and “divided attention” (Schuhfried, Prieler, & Bauer, 2002). Light stimuli move at a fixed speed along a panel that is positioned on the periphery of the respondent's field of vision. Whenever such a light stimulus appears, the respondent must react by pressing the foot pedal. At the same time, the respondent is required to perform a central tracking task. The main variables “field of vision” and “tracking deviation” are used as predictor variables. Tracking deviation is used as a measure of “divided attention” (Golz et al., 2004). Burgard (2005) was able to demonstrate that the performance in this test of patients suffering from TBI or strokes correlated significantly with their performance in a standardized driving test. The reliability coefficients (Cronbach's α) of the two main variables obtained in the present study were .96 (field of vision) and .98 (tracking deviation).

Vienna Risk-Taking Test Traffic

The Vienna Risk-taking Test Traffic (Hergovich, Bognar, Arendasy, & Sommer, 2005) is an objective personality test that deduces respondents' subjectively accepted level of risk (Wilde, 1994) from their responses to videos of various traffic situations. The traffic situations can be categorized into (1) speed choice and overtaking situations and (2) decisions at intersections. The situations also vary with regard to weather conditions. The respondents first receive verbal descriptions of traffic situations outlining a particular driving maneuver to be carried out. The respondent is then shown a video of the traffic scene described previously. Following this initial presentation, the same video is presented for a second time and the respondent has to indicate at what point the intended driving maneuver would be too risky to perform. The mean latency of response to the videos presented is used as a measure of respondents' subjectively accepted level of risk. The internal consistency of this measure is given due to a fit of the latency model (Scheiblechner, 1985). The reliability coefficient (Cronbach's α) of the main variable “subjectively accepted level of risk” obtained in the present study was α = .92. Thus far, this personality test has not been used to identify unsafe drivers suffering from TBI or strokes. Nevertheless, research conducted with healthy adults (cf. Sommer & Häusler, 2005; Sommer, Herle, Häusler, Risser, et al., 2008) indicates that this measure contributes significantly to the prediction of fitness to drive, thereby supporting the criterion validity of this test.

Inventory of Driving-Related Personality Traits

This standardized questionnaire consists of 39 items measuring four driving-related personality traits: Sensation-seeking, social responsibility, self-control, and emotional stability (Herle, Sommer, Wenzl, & Litzenberger, 2004). Respondents assess the extent to which 39 different statements apply to them. The answer is entered on an answer bar with a sliding marker. This response is recalculated dichotomously by the evaluation program. The reliability coefficients (Cronbach's α) obtained in the present study range from .69 (self-control) to .76 (social responsibility and sensation-seeking). Although this questionnaire has not been used to predict fitness to drive in samples of patients suffering from TBI or strokes, research conducted with healthy adults argues for the relevance of these driving-related personality traits in samples of healthy adults (cf. Sommer & Häusler, 2005; Sommer, Herle, Häusler, Risser, et al., 2008).

Description of the Standardized Driving Test

In the present study, we used a version of the Vienna Driving Test adapted to local conditions in Bad Tölz and Aachen that will be referred to as the Bad Tölz Driving Test (cf. Bukasa, Wenninger, & Brandstätter, 1990; Burgard, 2005; Chaloupka & Risser, 1995). The standardized driving test took about 45 min and was carried out in driving-school cars in the presence of driving instructors and two independent raters. Both a car with manual gear change and an automatic were available. Patients had the opportunity to familiarize themselves with the vehicle before the standardized test was conducted. The observation of driving behavior took place along a predefined route which involved a wide range of driving tasks representative of those encountered in Germany (e.g., lane keeping, lane choice before obstacles and intersections, speed, distance from the cars in front and from persons and objects on the roadside, overtaking maneuvers, interaction with vulnerable road users, etc.). The predefined route was divided into 52 sections each containing several driving tasks (for further details, seeBurgard, 2005). Care was taken to ensure that the driving tasks themselves and their sequence within the standardized on-road driving test conducted in Bad Tölz and Aachen are identical to each other. This aim was accomplished by means of expert ratings in selecting and designing the route of the standardized on-road driving test conducted in these two locations. Patients' performance in these driving tasks was evaluated by two independent observers using a standardized observation sheet. The observers had previously received intensive training in conducting behavioral observations using the Bad Tölz Driving Test. The observers were unaware of the respondents' test results. The standardized observations sheet contained the following nine evaluation criteria for each section: “Lateral distance and distance from the vehicle in front,” “lane keeping,” “communication with other traffic participants,” “speed choice and overtaking behavior,” “safety behavior,” “behavior at intersections,” “orientation behavior,” “anticipatory driving behavior,” and “operation of the vehicle.” Some of these criteria can be evaluated with certainty in a given section, whereas others can possibly be evaluated, depending on the current traffic situation. Each of the evaluation criteria was scored as either “not observable,” “very well handled,” “handled,” or “error” by both observers for each driving task. For each of the nine evaluation criteria, a sum score was calculated for both observers by totaling the scores on each evaluation criterion and dividing the results by the number of sections in which the evaluation criterion was observable. Given that the inter-rater reliabilities obtained in the present study for these nine evaluation criteria range from .85 to .91, this procedure seems to be justified from a methodological point of view. In addition, prior confirmatory factor analysis using samples of neurological patients and healthy adults indicates that these nine evaluation criteria obtained from two independent raters load onto a single factor (Sommer, Herle, & Häusler, 2008). From these empirical results, we then calculated a global score by means of a multiple regression using the factor loadings as regression weights. In addition, we also obtained a global evaluation of the patient's driving performance at the end of the standardized on-road test from both raters using a standardized 6-point scale which is based on the Austrian and German practical driving test (for details, seeRisser et al., 2008). A global evaluation of 4 or 5 reflects driving behavior that would cause the respondent to fail the national driving test. A global evaluation of 6 indicates that the driving test had to be terminated prematurely because the driver was putting himself or others at risk. Global evaluations of 1, 2, and 3 correspond to very good, good, and satisfactory driving behavior, respectively. The inter-rater reliability of this global evaluation obtained in the present study amounts to .91. The global score and the global evaluation were found to be highly correlated (r = .96). Following Lundquist and colleagues (2000), the cut-off value for classifying patients as “more likely to be fit to drive” or “less likely to be fit to drive” was determined using the more subjective global evaluation and determining the corresponding cut-off value for the more objective global score. In line with previous studies (cf. Risser et al., 2008; Sommer, Herle, Häusler, Risser, et al., 2008), the cut-off value for the global evaluation was set to 3.33. To obtain this score or a worse one, at least one of the assessors must has awarded a global assessment of 4 or 5.

Sample

The sample consists of 109 (61.2%) stroke patients and 69 (38.8%) patients suffering from a TBI. All the patients were drivers before their injury. The median annual distance driven by the patients before their injury was 20,000 km. Table 1 shows some descriptive statistics as well as differences between the two subsamples.

Table 1.

Sample characteristics of the total sample and both subsamples separated for those who passed or failed the standardized driving test

Variable Total sample
 
Stroke patients
 
TBI patients
 
Differences between stroke and TBI
 
 Total (n = 178) Passed (n = 124) Failed (n = 54) Total (n = 109) Passed (n = 85) Failed (n = 24) Total (n = 69) Passed (n = 39) Failed (n = 30) Method Statistics
 
           χ2/Z p-value 
Gender 
 Men (n [%]) 145 (81.5) 104 (83.9) 41 (75.9) 88 (80.7) 70 (82.4) 18 (75.0) 57 (82.6) 34 (87.2) 23 (76.7) Cross-tabs 0.098 .754 
 Women (n [%]) 33 (18.5) 20 (16.1) 13 (24.1) 21 (19.3) 15 (17.6) 6 (35.0) 12 (17.4) 5 (12.8) 7 (23.3) 
Educational level 
 Educational level 2 (n [%]) 20 (11.2) 14 (11.3) 6 (11.1) 11 (10.1) 10 (11.8) 1 (4.2) 9 (13.0) 4 (10.3) 5 (16.7) Cross-tabs 6.24 .100 
 Educational level 3 (n [%]) 98 (55.1) 71 (57.3) 27 (50.0) 62 (56.9) 46 (54.1) 16 (66.7) 36 (52.2) 25 (64.1) 11 (36.7) 
 Educational level 4 (n [%]) 26 (14.6) 13 (10.5) 13 (24.1) 11 (10.1) 7 (8.2) 4 (16.7) 15 (21.7) 6 (15.4) 9 (30.0) 
 Educational level 5 (n [%]) 33 (18.5) 25 (20.2) 8 (14.8) 24 (20.0) 21 (24.7) 3 (12.5) 9 (13.0) 4 (10.3) 5 (16.7) 
 No information provided (n [%]) 1 (0.6) 1 (0.8) – 1 (0.9) 1 (1.2) – – – – 
Age 
 Mean 45.37 46.70 42.30 51.39 50.98 52.83 35.86 37.38 33.87 Mann–Whitney U-test −8.44 <.001 
 Standard deviation 11.92 10.94 13.53 8.92 8.73 9.63 9.68 9.44 9.77 
 Min 21 21 21 24 24 36 21 21 21 
 Max 68 65 68 68 65 68 59 59 57 
Time since injury 
 Mean 31.35 22.17 52.43 12.39 9.80 21.58 61.29 49.13 77.10 Mann–Whitney U-test −7.68 <.001 
 Standard deviation 50.84 40.44 64.70 20.24 19.72 19.74 67.76 57.85 76.97 
 Min 
 Max 300 228 300 120 120 75 300 228 300 
Withdrawal of driving license 
 ≥1 withdrawal prior to the injury (n [%]) 34 (19.1) 21 (16.9) 13 (24.1) 20 (18.3) 17 (20.0) 3 (12.5) 14 (20.3) 4 (10.3) 10 (33.3) Cross-tabs 1.00 .605 
 no withdrawal prior to the injury (n [%]) 144 (80.9) 103 (83.1) 41 (75.9) 89 (81.7) 68 (80.0) 21 (87.5) 55 (79.7) 35 (89.7) 20 (66.7) 
Resumption of driving 
 Yes (n [%]) 75 (42.1) 61 (49.2) 40 (74.1) 40 (36.7) 36 (42.4) 4 (16.7) 35 (50.7) 25 (64.1) 10 (33.3) Cross-tabs 3.23 .072 
 No (n [%]) 103 (57.9) 63 (50.8) 14 (25.9) 69 (63.3) 49 (57.6) 20 (83.3) 34 (49.3) 14 (35.9) 20 (66.7) 
Barthel Index 
 Mean 97.57 98.91 88.30 97.47 98.85 87.56 99.00 100.00 98.00 Mann–Whitney U-test −0.49 .792 
 Standard deviation 9.04 6.35 17.16 9.32 6.54 18.03 2.24 0.00 0.00 
 Min 45 50 45 45 50 45 95 100 98 
 Max 100 100 100 100 100 100 100 100 98 
Variable Total sample
 
Stroke patients
 
TBI patients
 
Differences between stroke and TBI
 
 Total (n = 178) Passed (n = 124) Failed (n = 54) Total (n = 109) Passed (n = 85) Failed (n = 24) Total (n = 69) Passed (n = 39) Failed (n = 30) Method Statistics
 
           χ2/Z p-value 
Gender 
 Men (n [%]) 145 (81.5) 104 (83.9) 41 (75.9) 88 (80.7) 70 (82.4) 18 (75.0) 57 (82.6) 34 (87.2) 23 (76.7) Cross-tabs 0.098 .754 
 Women (n [%]) 33 (18.5) 20 (16.1) 13 (24.1) 21 (19.3) 15 (17.6) 6 (35.0) 12 (17.4) 5 (12.8) 7 (23.3) 
Educational level 
 Educational level 2 (n [%]) 20 (11.2) 14 (11.3) 6 (11.1) 11 (10.1) 10 (11.8) 1 (4.2) 9 (13.0) 4 (10.3) 5 (16.7) Cross-tabs 6.24 .100 
 Educational level 3 (n [%]) 98 (55.1) 71 (57.3) 27 (50.0) 62 (56.9) 46 (54.1) 16 (66.7) 36 (52.2) 25 (64.1) 11 (36.7) 
 Educational level 4 (n [%]) 26 (14.6) 13 (10.5) 13 (24.1) 11 (10.1) 7 (8.2) 4 (16.7) 15 (21.7) 6 (15.4) 9 (30.0) 
 Educational level 5 (n [%]) 33 (18.5) 25 (20.2) 8 (14.8) 24 (20.0) 21 (24.7) 3 (12.5) 9 (13.0) 4 (10.3) 5 (16.7) 
 No information provided (n [%]) 1 (0.6) 1 (0.8) – 1 (0.9) 1 (1.2) – – – – 
Age 
 Mean 45.37 46.70 42.30 51.39 50.98 52.83 35.86 37.38 33.87 Mann–Whitney U-test −8.44 <.001 
 Standard deviation 11.92 10.94 13.53 8.92 8.73 9.63 9.68 9.44 9.77 
 Min 21 21 21 24 24 36 21 21 21 
 Max 68 65 68 68 65 68 59 59 57 
Time since injury 
 Mean 31.35 22.17 52.43 12.39 9.80 21.58 61.29 49.13 77.10 Mann–Whitney U-test −7.68 <.001 
 Standard deviation 50.84 40.44 64.70 20.24 19.72 19.74 67.76 57.85 76.97 
 Min 
 Max 300 228 300 120 120 75 300 228 300 
Withdrawal of driving license 
 ≥1 withdrawal prior to the injury (n [%]) 34 (19.1) 21 (16.9) 13 (24.1) 20 (18.3) 17 (20.0) 3 (12.5) 14 (20.3) 4 (10.3) 10 (33.3) Cross-tabs 1.00 .605 
 no withdrawal prior to the injury (n [%]) 144 (80.9) 103 (83.1) 41 (75.9) 89 (81.7) 68 (80.0) 21 (87.5) 55 (79.7) 35 (89.7) 20 (66.7) 
Resumption of driving 
 Yes (n [%]) 75 (42.1) 61 (49.2) 40 (74.1) 40 (36.7) 36 (42.4) 4 (16.7) 35 (50.7) 25 (64.1) 10 (33.3) Cross-tabs 3.23 .072 
 No (n [%]) 103 (57.9) 63 (50.8) 14 (25.9) 69 (63.3) 49 (57.6) 20 (83.3) 34 (49.3) 14 (35.9) 20 (66.7) 
Barthel Index 
 Mean 97.57 98.91 88.30 97.47 98.85 87.56 99.00 100.00 98.00 Mann–Whitney U-test −0.49 .792 
 Standard deviation 9.04 6.35 17.16 9.32 6.54 18.03 2.24 0.00 0.00 
 Min 45 50 45 45 50 45 95 100 98 
 Max 100 100 100 100 100 100 100 100 98 

Note: TBI = traumatic brain injury.

As can be seen in Table 1, the stroke patients differ significantly from the TBI patients with regard to age and time between the injury and the participation in this study. Furthermore, a significantly higher proportion (χ2[2] = 9.207, p = .002) of stroke patients was judged to pass the standardized driving test.

Three (2.7%) stroke patients had had an ischemic attack, 14 (12.7%) stroke patients had had a cerebral hemorrhage, 9 (8.2%) stroke patients had had a subarachnoidal hemorrhage, and 86 (78.2%) stroke patients had had cerebral infarction of various causes. Using the TOAST classification system for cerebral infarction (Adams et al., 1993), 12.2% had had an atherothrombotic cerebral infarction, 22.0% had had a cardioembolic cerebral infarction, 24.4% had had a microangiopathic cerebral infarction, 11.1% had suffered from cerebral infarctions of other known causes, 12.2% had had a cerebral infarction of unknown cause, and in the case of 18.1%, the medical reports did not contain any information on the cause of the cerebral infarction.

A total of 18 (25.7%) TBI patients had had an open brain injury, whereas 52 (74.3%) TBI patients had suffered from a closed brain injury. On the basis of a CT-scan classification, 36 (51.4%) TBI patients had had a focal brain injury whereas 25 (35.7%) TBI patients had had a diffuse brain injury. In nine (12.9%) cases, the medical reports did not contain enough detailed information to classify the TBI. The severity of the TBI was evaluated by means of the Glasgow Coma Scale. A total of 9 (12.9%) TBI patients were classified as having a mild injury (GCS: 13–15), 9 (12.9%) TBI patients had a moderate injury (GCS: 9–12), and 52 (74.3%) TBI patients had a severe injury (GCS: 3–8).

As can be seen from these descriptive statistics and an inspection of the exclusion criteria, the majority of the TBI patients had suffered a severe brain injury from which they had recovered reasonably well at the time of their participation in this study. At the time of their participation in this study, the majority of the respondents were classified by the Barthel Index as having only minor disabilities or none at all. The sample used in this study can therefore not be regarded as representative of TBI and stroke patients in the more acute phases of their injury. This selection bias is due to the legal regulations in Austria and Germany.

Methods of Analyzing the Predictive Validity of Psychometric Test Batteries

Given a sufficient sample size, newer studies have tended to use multivariate statistical methods to evaluate the incremental validity of psychometric tests. Within this approach, one may distinguish between (1) classical linear multivariate methods and (2) non-linear multivariate methods. Two of the most commonly used classical linear multivariate methods are discriminant analysis and logistic regression. However, the application of these classical multivariate methods is only appropriate in cases where the data fulfill a range of preconditions such as homogeneity of the variance–covariance matrices and multivariate normal distribution of the predictor variables in the subsamples (Bortz, 1999; Brown & Wicker, 2000; Venter & Scott, 2000). Furthermore, discriminant analysis and logistic regression assume linear or logit correlations between the predictor variables and the criterion which implies linear-additive relationships between the criterion and the predictor variables (cf. Rakes, 1991; Somers, 2001). Although complex nonlinear effects, interactions, and compensatory effects can be modeled within the framework of classical multivariate methods, these effects need to be specified in advance based on current theoretical models and thus often remain unconsidered. This problem is particularly prevalent in traffic psychological assessment due to the lack of a unified and coherent theoretical model of driving behavior which takes ability and personality variables into account and specifies the structural relations between these predictor variables and the criterion measure (cf. Ranney, 1994; Risser, 1997). In order to overcome this problem, Risser and colleagues (2008) proposed comparing the results of classical multivariate methods with those obtained from artificial neural networks. In general, artificial neural networks can be viewed as robust multivariate classification algorithms that assign respondents to predefined categories on the basis of a set of predictor variables (cf. Anderson & Rosenfeld, 1988; Bishop, 1995; Kinnebrock, 1992; Mielke, 2001; Rojas, 2000). Artificial neural networks consist of several elements or units which are organized into layers according to their function. The input layer represents the predictor variables, whereas the output layer represents the criterion measure. Between the input layer and the output layer, one hidden layer is usually interposed. The units of these three layers are linked to each other by weighted feed-forward full connections. In other words, all units in a layer transmit information to the units of the following layer. The data transmitted are weighted, summed, and transformed using a transformation function before being forwarded to the elements of the next highest layer. In the case of a dichotomous criterion variable, Softmax is commonly used as a transformation function (Bridle, 1990). This general structure of an artificial neural network illustrated in Fig. 1 is commonly referred to as a multilayer perceptrone (Kinnebrock, 1992).

Fig. 1.

Artificial neural networks incorporate hidden layers of functional neurons (N1 and N2) between the predictors (P1, P2, and P3) and the criterion (C). The number functional neurons in the hidden layer determines how many weights need to be optimized and hence the economy of the model.

Fig. 1.

Artificial neural networks incorporate hidden layers of functional neurons (N1 and N2) between the predictors (P1, P2, and P3) and the criterion (C). The number functional neurons in the hidden layer determines how many weights need to be optimized and hence the economy of the model.

The number of hidden layer units can be selected at will and determines the complexity of the structural relationships between the predictor variables and the criterion that can be modeled by an artificial neural network. However, it must also be borne in mind that a higher number of hidden layer units also increase the risk of an overfit, making it impossible to generalize the results obtained beyond the existing sample (Mielke, 2001). The choice of an appropriate number of hidden layer units is thus crucial in setting up an artificial neural network. In order to determine the number of hidden layer units, Häusler and Sommer (2005) and Risser and colleagues (2008) proposed a brute force algorithm that enables the most parsimonious artificial neural network to be selected by determining an optimal number of predictor variables and hidden layer units using the Bayesian Information Criterion (BIC: Schwarz, 1978) and the adjusted R2; by this means, several convergent artificial neural networks can be compared with regard to their predictive validity and parsimony. The BIC is based on the log-likelihood of the data under the hypothesis that perfect prediction of the criterion by the predictors is possible; it takes the complexity of the predictive model into account. The larger the number of weights which need to be optimized, the more “penalty points” are awarded to take account of the fact that complex models can more readily fit the data. These two statistical indices have proved to be effective in striking an optimal balance between validity and parsimony of artificial neural networks in simulation studies and various empirical applications (cf. Bishop, 1995; Häusler & Sommer, 2005).

Once the number of predictor variables and the number of hidden layer units have been determined, the weights of the feed-forward full connections need to be optimized by means of the so-called learning algorithms. During this process, the weights are optimized iteratively so that the difference between the artificial neural network's estimate of the criterion score and the actual criterion score is minimized. This involves strengthening useful paths and weakening the contribution of less significant ones. This process is sometimes described as “supervised learning.” Examples of learning algorithms commonly recommended for application in statistical judgment formation are QuickProp (Fahlman, 1988) and scaled conjugate gradient (Masters, 1995). The latter algorithm is particularly useful for a higher number of weights have to be optimized using comparably smaller sample sizes.

In a next step, the stability of the results obtained with particular artificial neural network architecture needs to be investigated. This can be done by means of a jackknife validation and a bootstrap validation. If both methods support the stability of the results, the researcher may calculate the incremental validity and the relative relevance of the predictor variables. The incremental validity and relative relevance enable the relevance of the predictor variables to be assessed by taking simultaneous account of the direct and indirect effects of the predictor variable on the criterion measure. Even though these incremental validities thus do not correspond to the beta-weights in classical multivariate methods, there are some similarities in the sense that both the incremental validates and the beta-weights cannot be interpreted in absolute terms but need to be interpreted in relation to all the other predictor variables used in the model (c.f. Bortz, 1999; Venter & Scott, 2000).

Results

Descriptive Results

In all analyses reported in this article, the test scores were corrected for age using the formula provided by the test authors. This formula were calculated on the basis of the German and Austrian norm data (for a summary, seeSchuhfried, 2005) to ensure that the effect of the brain injury on patients' test performance is not confounded with age effects. This is particularly important since stroke and TBI patients have been shown to differ significantly with regard to age in our sample.

In the first phase, we calculated descriptive statistics for all the main predictor variables used in the present study and analyzed differences between stroke and TBI patients with regard to their performance in these main predictor variables. This was done to provide the first overview of the present data. The results are summarized in Table 2.

Table 2.

Descriptive statistics of the predictor variables corrected for age

 Total sample (n = 178)
 
Stroke patients (n = 109)
 
TBI patients (n = 69)
 
Differences between stroke and TBI
 
 Mean SD Min Max Mean SD Min Max Mean SD Min Max Z-value p-value 
AMT: Fluid intelligence −1.45 1.02 −3.72 1.64 −1.57 1.05 −3.72 1.63 −1.25 0.94 −3.20 0.87 −2.276 .023 
DT: Complex choice reaction 194.52 47.52 60 325 194.78 47.45 86 326 194.10 47.96 60 272 −.078 .938 
RT: Simple choice reaction 471.03 99.76 292 847 459.96 94.88 313 726 488.51 105.37 292 847 −1.791 .073 
PP: Field of view 155.82 21.46 65 187 155.69 19.63 86 187 156.04 24.21 65 183 −0.92 .358 
PP: Divided attention 13.50 4.16 34 13.51 4.21 28 13.48 4.11 34 −0.378 .706 
TAVT: Perceptual speed 10.01 3.05 16 9.86 2.94 16 10.23 3.23 16 −0.721 .471 
COG: Selective attention 3.70 1.75 1.68 19.46 3.61 1.33 1.68 9.08 3.84 2.27 1.86 19.46 −0.142 .887 
WRBTV: Subj. accepted level of risk 6.44 1.61 2.80 11.21 6.28 1.54 2.80 11.21 6.70 1.68 3.13 10.80 −1.814 .070 
IVPE: Emotional stability 2.49 22.24 11 2.25 1.82 2.88 2.74 11 −0.963 .336 
IVPE: Social responsibility 6.61 2.86 10 6.88 2.91 10 6.19 2.73 10 −2.095 .036 
IVPE: Self-control 4.85 1.64 4.99 1.53 4.64 1.80 −1.163 .245 
IVPE: Sensation seeking 3.97 2.39 10 3.56 2.25 4.62 2.48 10 −2.784 .005 
 Total sample (n = 178)
 
Stroke patients (n = 109)
 
TBI patients (n = 69)
 
Differences between stroke and TBI
 
 Mean SD Min Max Mean SD Min Max Mean SD Min Max Z-value p-value 
AMT: Fluid intelligence −1.45 1.02 −3.72 1.64 −1.57 1.05 −3.72 1.63 −1.25 0.94 −3.20 0.87 −2.276 .023 
DT: Complex choice reaction 194.52 47.52 60 325 194.78 47.45 86 326 194.10 47.96 60 272 −.078 .938 
RT: Simple choice reaction 471.03 99.76 292 847 459.96 94.88 313 726 488.51 105.37 292 847 −1.791 .073 
PP: Field of view 155.82 21.46 65 187 155.69 19.63 86 187 156.04 24.21 65 183 −0.92 .358 
PP: Divided attention 13.50 4.16 34 13.51 4.21 28 13.48 4.11 34 −0.378 .706 
TAVT: Perceptual speed 10.01 3.05 16 9.86 2.94 16 10.23 3.23 16 −0.721 .471 
COG: Selective attention 3.70 1.75 1.68 19.46 3.61 1.33 1.68 9.08 3.84 2.27 1.86 19.46 −0.142 .887 
WRBTV: Subj. accepted level of risk 6.44 1.61 2.80 11.21 6.28 1.54 2.80 11.21 6.70 1.68 3.13 10.80 −1.814 .070 
IVPE: Emotional stability 2.49 22.24 11 2.25 1.82 2.88 2.74 11 −0.963 .336 
IVPE: Social responsibility 6.61 2.86 10 6.88 2.91 10 6.19 2.73 10 −2.095 .036 
IVPE: Self-control 4.85 1.64 4.99 1.53 4.64 1.80 −1.163 .245 
IVPE: Sensation seeking 3.97 2.39 10 3.56 2.25 4.62 2.48 10 −2.784 .005 

Notes: TBI = traumatic brain injury; AMT = Adaptive Matrices Test; DT = Determination Test; RT = Reaction Test; PP = Peripheral Perception; TAVT = Tachistoscopic Traffic Perception Test; COG = Cognitrone; WRBTV = Vienna Risk-Taking Test Traffic.

As can be seen in Table 2, the two subsamples differ with regard to “fluid intelligence,” as well as their scores in the subscales “social responsibility” and “sensation seeking” in the Inventory of Driving-Related Personality Traits (IVPE; Herle et al., 2004). TBI patients performed better in our measure of fluid intelligence—even after correcting for age—but were more adventurous and in need of excitement. Furthermore, stroke patients exhibited a higher degree of social responsibility compared with TBI patients.

We also calculated the correlations between the psychometric tests, the variable “time between the injury and participation in this study” and the global evaluation of patients' performance in the standardized driving test for the total sample and the two subsamples. Time since injury was included as predictor variable on the basis of previous studies and the finding that stroke and TBI patients differed significantly in this measure in the present study. The results are shown in Table 3.

Table 3.

Intercorrelations between the predictor variables in the total sample (above) and Spearman correlation coefficients between the psychometric tests and the global score (below)

  AMT: Fluid intelligence DT: Complex choice reaction RT: Simple choice reaction PP: Field of view PP: Divided attention TAVT: Perceptual speed COG: Selective attention WRBTV: Subj. accepted level of risk IVPE-PS: Emotional stability IVPE-VB: Social responsibility IVPE-SK: Self-control IVPE-TA: Sensation seeking 
Correlation coefficients between the predictor variables 
 AMT: Fluid intelligence –            
 DT: Complex choice reaction .39** –           
 RT: Simple choice reaction −.20* −.36** –          
 PP: Field of view .26* .33** −.30** –         
 PP: Divided attention −.32** −.387** .22* −.395** –        
 TAVT: Perceptual speed .41** .49** −.30* .39** −.35** –       
 COG: Selective attention −.34** −.44** .39** −.35** .30** .49* –      
 WRBTV: Subj. acc. level of risk .10 −.06 .05 −.07 −.02 −.03 −.12 –     
 IVPE: Emotional stability −.06 −.14 .06 −.11 .01 −.11 −.02 −.22* –    
 IVPE: Social responsibility −.03 .03 .03 .03 −.02 .12 .03 −.21* .04 –   
 IVPE: Self-control .04 .21* −.07 −.01 .03 .08 −.07 −.31* −.06 .53* –  
 IVPE: Sensation seeking .07 .10 −.14 .10 −.19* .08 −.18* .30* −.04 −.18* −.29* – 
Correlation between the predictor variables and patients' performance in the standardized driving test 
Sample Time since injury AMT: Fluid intelligence DT: Complex choice reaction RT: Simple choice reaction PP: Field of view PP: Divided attention TAVT: Perceptual speed COG: Selective attention WRBTV: Subj. accepted level of risk IVPE: Emotional stability IVPE: Social responsibility IVPE: Self-control IVPE: Sensation seeking 

 
 Total .292** −.105n.s. −.422** .214** −.277** .281** −.281** .250 **. .181* −.015 n.s −.207** −.139n.s. .081 n.s 
 Stroke .251** −.177n.s. −.527** .271** −.309** −.340** −.338** .301** .251* −.097n.s. −.207* −.055n.s. .085n.s. 
 TBI .235** −.155n.s. −.307** .096n.s. −.168n.s. .098n.s. −.265* .195n.s. .182n.s. .006n.s. −.145n.s. −.101n.s. .049n.s. 
  AMT: Fluid intelligence DT: Complex choice reaction RT: Simple choice reaction PP: Field of view PP: Divided attention TAVT: Perceptual speed COG: Selective attention WRBTV: Subj. accepted level of risk IVPE-PS: Emotional stability IVPE-VB: Social responsibility IVPE-SK: Self-control IVPE-TA: Sensation seeking 
Correlation coefficients between the predictor variables 
 AMT: Fluid intelligence –            
 DT: Complex choice reaction .39** –           
 RT: Simple choice reaction −.20* −.36** –          
 PP: Field of view .26* .33** −.30** –         
 PP: Divided attention −.32** −.387** .22* −.395** –        
 TAVT: Perceptual speed .41** .49** −.30* .39** −.35** –       
 COG: Selective attention −.34** −.44** .39** −.35** .30** .49* –      
 WRBTV: Subj. acc. level of risk .10 −.06 .05 −.07 −.02 −.03 −.12 –     
 IVPE: Emotional stability −.06 −.14 .06 −.11 .01 −.11 −.02 −.22* –    
 IVPE: Social responsibility −.03 .03 .03 .03 −.02 .12 .03 −.21* .04 –   
 IVPE: Self-control .04 .21* −.07 −.01 .03 .08 −.07 −.31* −.06 .53* –  
 IVPE: Sensation seeking .07 .10 −.14 .10 −.19* .08 −.18* .30* −.04 −.18* −.29* – 
Correlation between the predictor variables and patients' performance in the standardized driving test 
Sample Time since injury AMT: Fluid intelligence DT: Complex choice reaction RT: Simple choice reaction PP: Field of view PP: Divided attention TAVT: Perceptual speed COG: Selective attention WRBTV: Subj. accepted level of risk IVPE: Emotional stability IVPE: Social responsibility IVPE: Self-control IVPE: Sensation seeking 

 
 Total .292** −.105n.s. −.422** .214** −.277** .281** −.281** .250 **. .181* −.015 n.s −.207** −.139n.s. .081 n.s 
 Stroke .251** −.177n.s. −.527** .271** −.309** −.340** −.338** .301** .251* −.097n.s. −.207* −.055n.s. .085n.s. 
 TBI .235** −.155n.s. −.307** .096n.s. −.168n.s. .098n.s. −.265* .195n.s. .182n.s. .006n.s. −.145n.s. −.101n.s. .049n.s. 

Notes: TBI = traumatic brain injury; AMT = Adaptive Matrices Test; DT = Determination Test; RT = Reaction Test; PP = Peripheral Perception; TAVT = Tachistoscopic Traffic Perception Test; COG = Cognitrone; WRBTV = Vienna Risk-Taking Test Traffic.

**p < .01.

*p < .05.

n.s.p > .10.

In general, both the magnitude of the individual correlation coefficients and the pattern of the inter-correlations of the main variables of the psychometric test battery are similar to the results obtained in previous studies conducted with healthy adults (cf. Risser et al., 2008; Sommer, Herle, Häusler, Risser, et al., 2008). The cognitive ability traits were correlated with each other in an expected magnitude, whereas the personality traits turned out to be virtually unrelated to our cognitive ability measures.

Furthermore, with the exception of fluid intelligence, all the cognitive ability measures were significantly correlated with patients' performance in the standardized driving test in the total sample and the sample of stroke patients. In the TBI sample, only complex choice reaction and perceptual speed turned out to be significantly related to patients' performance in the standardized driving test. However, the patterns of the correlations obtained in the two samples seem to resemble each other, although the magnitudes of the correlation coefficients were higher in the subsample of stroke patients.

Regarding the criterion validity of our driving-related personality trait measures, the results indicate that the variable “social responsibility” from the Inventory of Driving-Related Personality Traits (IVPE) and the variable “subjectively accepted level of risk” from the Vienna Risk-taking Test Traffic (WRBTV) correlated significantly with the global evaluation of patients' performance in the standardized driving test in the total sample as well as the sample of stroke patients. In contrast, none of the driving-related personality trait measures turned out to be significantly correlated with patients' performance in the standardized driving test in the TBI sample. However, a comparison of the magnitude of the correlation coefficients between these three samples reveals that the correlation coefficients obtained for the variables “social responsibility” and “subjectively accepted level of risk” in the sample of TBI patients are still within the range of the total sample.

Additionally, the magnitude of the correlation coefficients between our cognitive tests and patients' performance in the standardized on-road driving test was higher than the one obtained between our criterion variable and patients' scores in the driving-related personality measures in all three samples. This suggests that cognitive abilities are generally more important in predicting patients' actual fitness to drive (cf. Anstey, Wood, Lord, & Walker, 2005; Sommer, Herle, Häusler, Risser, et al., 2008). Nevertheless, the finding that driving-related personality traits and patients' performance in the standardized driving test are significantly correlated combined with the finding that these two kinds of predictor variables are virtually unrelated suggests that personality traits might contribute significantly to the prediction of fitness to drive in patients suffering from TBI or strokes.

The present results also show that the variable “time between the injury and participation in this study” is significantly correlated with our criterion measure in all three samples. Thus, the more time that elapsed between the injury and the participation in this study, the worse patients' performance in the standardized driving test was rated.

Multivariate Results

The multivariate results were calculated by means of a logistic regression and an artificial neural network. In both cases, the main variables of the psychometric tests—corrected for age—and “time between the injury and participation in this study” served as independent variables whereas the dichotomized global score of the standardized driving test was used as the dependent variable. The analyses were performed for both subsamples and the total sample separately. In the first step, the resulting predictive models are compared with each other with regard to their stability and psychometric quality (here: Validity coefficient, classification rate, sensitivity, specificity) to identify the most stable and parsimonious model that accurately captures the structural relations between the predictor variables and our criterion measure. Afterwards, the relative importance of the predictor variables according to each of the predictive models is outlined. This provides first insights into the cognitive and personality determinants of fitness to drive according to each of the six predictive models. Finally, we investigate the generalizability of each of the six predictive models across the two subsamples investigated in this study to investigate the degree to which cognitive and personality determinants of fitness to drive are similar in samples of stroke and TBI patients.

Results Obtained with Classic Multivariate Statistics

The logistic regression analyses were calculated by means of SPSS 15.00. All the predictor variables were entered into the analyses simultaneously. For the total sample (−2 log-likelihood = 151.975; χ2[13] = 66.500, p < .001; Nagelkerkes R2 = .441) and the subsample of stroke patients (2 log-likelihood = 42.995; χ2[13] = 71.922, p < .001; Nagelkerkes R2 = .741) significant predictive models were obtained. In contrast, the predictive model calculated for the subsample of TBI patients (−2 log-likelihood = 78.558; χ2[13] = 15.919, p = .254; Nagelkerkes R2 = .276) failed to reach the level of significance. The stability of these results was assessed by means of a jackknife validation as well as an internal bootstrap validation using 1,000 bootstrap samples. The validity coefficients, classification rates, sensitivity, and specificity for the simple prediction as well as the jackknife and bootstrap validation are shown in Table 4.

Table 4.

Validity coefficient, classification rate, sensitivity, specificity, and stability of the simple prediction and according to the jackknife validation and bootstrap validation for the logistic regression

Sample Simple prediction
 
Jackknife validation
 
Bootstrap validation
 
 Validity coefficient Classification rate (%) Sensitivity (%) Specificity (%) Validity coefficient Classification rate (%) Sensitivity (%) Specificity (%) Stability Validity coefficient Classification rate (%) 
Total 0.53 79.8 53.7 91.1 0.53 79.2 53.7 91.1 0.99 [0.40; 0.67] [71.2; 83.5] 
Stroke 0.82 94.5 83.3 97.6 0.81 93.6 79.2 97.6 0.99 [0.79; 0.81] [89.8; 96.3] 
TBI 0.45 65.2 53.3 74.4 0.47 66.7 56.7 74.4 0.87 [0.30; 0.50] [60.3; 69.8] 
Sample Simple prediction
 
Jackknife validation
 
Bootstrap validation
 
 Validity coefficient Classification rate (%) Sensitivity (%) Specificity (%) Validity coefficient Classification rate (%) Sensitivity (%) Specificity (%) Stability Validity coefficient Classification rate (%) 
Total 0.53 79.8 53.7 91.1 0.53 79.2 53.7 91.1 0.99 [0.40; 0.67] [71.2; 83.5] 
Stroke 0.82 94.5 83.3 97.6 0.81 93.6 79.2 97.6 0.99 [0.79; 0.81] [89.8; 96.3] 
TBI 0.45 65.2 53.3 74.4 0.47 66.7 56.7 74.4 0.87 [0.30; 0.50] [60.3; 69.8] 

Notes: TBI = traumatic brain injury. The validity was calculated as the correlation between actual evaluation of performance in the standardized driving test and the classification probabilities based on the respondents' test performance.

As can be seen in Table 4, the results obtained for the total sample and the sample of stroke patients can be regarded as stable, which was not the case for the sample of TBI patients. However, there is an imbalance between sensitivity and specificity in the total sample, indicating that the predictive model seems to have problems in identifying those who are less likely to be judged fit to drive on the basis of their performance in the standardized driving test. In order to analyze whether the predictive model obtained for the total sample covers both subsamples equally well, we analyzed whether stroke and TBI patients are equally likely to be misclassified. The results indicate that TBI patients are indeed more often misclassified than stroke patients (χ2[1] = 4.305, p = .034).

In the next step, we calculated the contribution of the predictor variables to the prediction of patients' actual fitness to drive for all three predictive models. The results are summarized in Table 5.

Table 5.

Regression coefficient (B), Wald statistic, degrees of freedom (df), and significance level (p) for the main variables corrected for age according to the logistic regression for the total sample (above) and both subsamples (below)

Predictors B SE Wald df p-value 
Total sample (n = 178) 
Time since injury 0.013 0.006 4.445 .035 
AMT: Fluid intelligence 0.218 0.256 0.729 .393 
DT: Complex choice reaction −0.022 0.007 9.200 .002 
RT: Simple choice reaction 0.001 0.003 0.171 .679 
TAVT: Perceptual speed −0.026 0.097 0.070 .975 
PP: Field of vision −0.004 0.015 0.075 .996 
PP: Divided attention 0.027 0.065 0.176 .675 
COG: Selective attention 0.092 0.242 0.143 .705 
WRBTV: Subj. accepted level of risk 0.377 0.157 5.778 .016 
IVPE: Emotional stability −0.073 0.117 0.390 .930 
IVPE: Social responsibility −0.155 0.113 1.901 .168 
IVPE: Self-control −0.111 0.186 0.357 .550 
IVPE: Sensation seeking 0.008 0.114 0.005 .942 
Stroke sample (n = 109)/TBI sample (n = 69) 
Time since injury 0.029/0.006 0.022/0.005 1.776/1.720 1/1 .183/.190 
AMT: Fluid intelligence 0.460/−0.137 0.431/0.356 1.141/0.149 1/1 .286/.700 
DT: Complex choice reaction −0.056/−0.289 0.018/0.156 9.310/3.422 1/1 .002/.064 
RT: Simple choice reaction 0.001/−0.003 0.005/0.004 0.038/0.494 1/1 .846/.482 
TAVT: Perceptual speed −0.556/−0.158 0.266/0.112 4.373/2.031 1/1 .037/.154 
PP: Field of vision −0.015/−0.007 0.031/0.016 0.239/0.207 1/1 .625/.649 
PP: Divided attention 0.065/0.060 0.125/0.087 0.270/0.484 1/1 .603/.486 
COG: Selective attention 0.174/0.092 0.387/0.258 0.203/0.127 1/1 .652/.721 
WRBTV: Subj. accepted level of risk 1.371/−.011 0.423/0.008 10.491/1.868 1/1 .001/.172 
IVPE: Emotional stability −0.308/0.040 0.278/0.123 1.230/0.105 1/1 .735/.746 
IVPE: Social responsibility −0.399/−0.086 0.186/0.163 4.373/0.276 1/1 .037/.599 
IVPE: Self-control −0.334/−0.187 0.312/0.234 1.147/0.637 1/1 .284/.425 
IVPE: Sensation seeking 0.055 / −.025 0.189/0.202 0.084/0.016 1/1 .773/.901 
Predictors B SE Wald df p-value 
Total sample (n = 178) 
Time since injury 0.013 0.006 4.445 .035 
AMT: Fluid intelligence 0.218 0.256 0.729 .393 
DT: Complex choice reaction −0.022 0.007 9.200 .002 
RT: Simple choice reaction 0.001 0.003 0.171 .679 
TAVT: Perceptual speed −0.026 0.097 0.070 .975 
PP: Field of vision −0.004 0.015 0.075 .996 
PP: Divided attention 0.027 0.065 0.176 .675 
COG: Selective attention 0.092 0.242 0.143 .705 
WRBTV: Subj. accepted level of risk 0.377 0.157 5.778 .016 
IVPE: Emotional stability −0.073 0.117 0.390 .930 
IVPE: Social responsibility −0.155 0.113 1.901 .168 
IVPE: Self-control −0.111 0.186 0.357 .550 
IVPE: Sensation seeking 0.008 0.114 0.005 .942 
Stroke sample (n = 109)/TBI sample (n = 69) 
Time since injury 0.029/0.006 0.022/0.005 1.776/1.720 1/1 .183/.190 
AMT: Fluid intelligence 0.460/−0.137 0.431/0.356 1.141/0.149 1/1 .286/.700 
DT: Complex choice reaction −0.056/−0.289 0.018/0.156 9.310/3.422 1/1 .002/.064 
RT: Simple choice reaction 0.001/−0.003 0.005/0.004 0.038/0.494 1/1 .846/.482 
TAVT: Perceptual speed −0.556/−0.158 0.266/0.112 4.373/2.031 1/1 .037/.154 
PP: Field of vision −0.015/−0.007 0.031/0.016 0.239/0.207 1/1 .625/.649 
PP: Divided attention 0.065/0.060 0.125/0.087 0.270/0.484 1/1 .603/.486 
COG: Selective attention 0.174/0.092 0.387/0.258 0.203/0.127 1/1 .652/.721 
WRBTV: Subj. accepted level of risk 1.371/−.011 0.423/0.008 10.491/1.868 1/1 .001/.172 
IVPE: Emotional stability −0.308/0.040 0.278/0.123 1.230/0.105 1/1 .735/.746 
IVPE: Social responsibility −0.399/−0.086 0.186/0.163 4.373/0.276 1/1 .037/.599 
IVPE: Self-control −0.334/−0.187 0.312/0.234 1.147/0.637 1/1 .284/.425 
IVPE: Sensation seeking 0.055 / −.025 0.189/0.202 0.084/0.016 1/1 .773/.901 

Notes: TBI = traumatic brain injury; AMT = Adaptive Matrices Test; DT = Determination Test; RT = Reaction Test; PP = Peripheral Perception; TAVT = Tachistoscopic Traffic Perception Test; COG = Cognitrone; WRBTV = Vienna Risk-Taking Test Traffic.

As can be seen in Table 5, “time between the injury and participation in this study”, “complex choice raction (DT)”, and “subjectively accepted level of risk (WBRTV)” contribute significantly to the prediction of patients' actual fitness to drive in the total sample. In contrast to this, “time between the injury and participation in this study” did not contribute to the prediction of patients' actual fitness to drive in either of the two subsamples. However, “complex choice reaction (DT)” and “subjectively accepted level of risk (WBRTV)” also contribute to the prediction of fitness to drive in the sample of stroke patients. In addition, “perceptual speed (TAVTMB)” and “social responsibility (IVPE-VB)” also turned out to be relevant to the prediction of patients' actual fitness to drive in this specific subsample. As would have been expected due to the lack of statistical significance of the predictive model calculated for the TBI sample, none of the predictor variables contributed significantly to the prediction of actual fitness to drive in the sample of TBI patients.

Results Obtained with Artificial Neural Networks

We also calculated an artificial neural network for both subsamples and the total sample separately. Scaled conjugate gradient (Masters, 1995) was used as the learning algorithm in all three cases. The number of hidden layer elements was determined by comparing artificial neural networks with varying numbers of hidden layer elements using the BIC. In the total sample, this resulted in an optimal number of six predictor variables and three hidden layer elements. In contrast to this, the relation between the predictor variables and the criterion measure is best modeled by five predictor variables and two hidden layer elements in the subsample of stroke patients and six predictor variables and two hidden layer elements in the subsample of TBI patients. The stability of these results was assessed by means of a jackknife validation as well as an internal bootstrap validation. The results of these analyses are summarized in Table 6.

Table 6.

Validity coefficient, classification rate, sensitivity, specificity, and stability of the simple prediction and according to the jackknife validation and bootstrap validation for artificial neural network

Sample Simple prediction
 
Jackknife validation
 
Bootstrap validation
 
 Validity coefficient Classification rate (%) Sensitivity (%) Specificity (%) Validity coefficient Classification rate (%) Sensitivity (%) Specificity (%) Stability Validity coefficient Classification rate (%) 
Total .81 89.9 77.8 95.2 .81 89.9 77.8 95.2 .99 [.71; .91] [83.8; 93.9] 
Stroke .95 98.2 95.8 98.8 .95 98.2 95.8 98.8 .99 [.91; .99] [95.4; 99.1] 
TBI .82 86.9 80.6 89.7 .78 84.1 74.2 89.7 .87 [.62; .98] [75.8; 96.3] 
Sample Simple prediction
 
Jackknife validation
 
Bootstrap validation
 
 Validity coefficient Classification rate (%) Sensitivity (%) Specificity (%) Validity coefficient Classification rate (%) Sensitivity (%) Specificity (%) Stability Validity coefficient Classification rate (%) 
Total .81 89.9 77.8 95.2 .81 89.9 77.8 95.2 .99 [.71; .91] [83.8; 93.9] 
Stroke .95 98.2 95.8 98.8 .95 98.2 95.8 98.8 .99 [.91; .99] [95.4; 99.1] 
TBI .82 86.9 80.6 89.7 .78 84.1 74.2 89.7 .87 [.62; .98] [75.8; 96.3] 

Notes: TBI = traumatic brain injury. The validity was calculated as the correlation between actual evaluation of performance in the standardized driving test and the classification probabilities based on the respondents' test performance.

As can be seen in Table 6, the results obtained for the predictive models obtained on the basis of either the total sample or the stroke patient sample can be regarded as stable, whereas those obtained for the TBI sample alone seem to be less stable. Again, in the total sample, there is a slight imbalance between sensitivity and specificity. However, the imbalance is much less pronounced then was the case in the classical multivariate analyses. We also analyzed whether the proportions of stroke and TBI patients misclassified by the predictive model obtained for the total sample differ from each other. In contrast to the results obtained for the classical multivariate predictive model, stroke and TBI patients seem to be equally likely to be misclassified (χ2[1] = 3.045, p = .081) when using the more complex multivariate predictive model calculated by means of an artificial neural network. In a next step, we compared the predictive models calculated by means of a logistic regression analysis and the artificial neural network for the three samples. As can be seen in Tables 4 and 6, the confidence intervals calculated for the validity coefficient and the classification rate do not overlap in the case of the predictive model calculated for the total sample and the sample of TBI patients. This indicates that the predictive models calculated by means of an artificial neural network outperform the predictive models calculated by means of a logistic regression analysis. This is not the case for the stroke patient sample since both the confidence intervals for the validity coefficient and the classification rate calculated by a logistic regression and an artificial neural network overlap in this subsample. This indicates that the relation between the predictor variables and the criterion is modeled equally well by the two multivariate statistical approaches.

The incremental validities of the predictor variables for the total sample and both subsamples are presented in Table 7.

Table 7.

Incremental validity and relative relevance of the main variables corrected for age according to the artificial neural network for the total sample (above) and both subsamples (below)

Predictors Incremental validity Relative relevance (%) 
Total sample (n = 178) 
Time since injury .102 25.4 
AMT: Fluid intelligence – – 
DT: Complex choice reaction .071 25.2 
RT: Simple choice reaction .019 6.9 
TAVT: Perceptual speed .009 12.8 
PP: Field of vision – – 
PP: Divided attention – – 
COG: Selective attention – – 
WRBTV: Subj. accepted level of risk .072 25.5 
IVPE: Emotional stability – – 
IVPE: Social responsibility .012 4.3 
IVPE: Self-control – – 
IVPE: Sensation seeking – – 
Stroke sample (n = 109)/TBI sample (n = 69) 
Time since injury –/– –/– 
AMT: Fluid intelligence –/– –/– 
DT: Complex choice reaction .254/.178 27.7/33.4 
RT: Simple choice reaction .127/.258 14.9/18.8 
TAVT: Perceptual speed .215/.282 24.0/20.3 
PP: Field of vision –/– –/– 
PP: Divided attention –/– – 
COG: Selective attention –/.124 –/9.8 
WRBTV: Subj. accepted level of risk .183/.328 20.8/23.0 
IVPE: Emotional stability –/– –/– 
IVPE: Social responsibility .107/.190 12.7/14.4 
IVPE: Self-control –/– –/– 
IVPE: Sensation seeking –/– –/– 
Predictors Incremental validity Relative relevance (%) 
Total sample (n = 178) 
Time since injury .102 25.4 
AMT: Fluid intelligence – – 
DT: Complex choice reaction .071 25.2 
RT: Simple choice reaction .019 6.9 
TAVT: Perceptual speed .009 12.8 
PP: Field of vision – – 
PP: Divided attention – – 
COG: Selective attention – – 
WRBTV: Subj. accepted level of risk .072 25.5 
IVPE: Emotional stability – – 
IVPE: Social responsibility .012 4.3 
IVPE: Self-control – – 
IVPE: Sensation seeking – – 
Stroke sample (n = 109)/TBI sample (n = 69) 
Time since injury –/– –/– 
AMT: Fluid intelligence –/– –/– 
DT: Complex choice reaction .254/.178 27.7/33.4 
RT: Simple choice reaction .127/.258 14.9/18.8 
TAVT: Perceptual speed .215/.282 24.0/20.3 
PP: Field of vision –/– –/– 
PP: Divided attention –/– – 
COG: Selective attention –/.124 –/9.8 
WRBTV: Subj. accepted level of risk .183/.328 20.8/23.0 
IVPE: Emotional stability –/– –/– 
IVPE: Social responsibility .107/.190 12.7/14.4 
IVPE: Self-control –/– –/– 
IVPE: Sensation seeking –/– –/– 

Notes: TBI = traumatic brain injury; AMT = Adaptive Matrices Test; DT = Determination Test; RT = Reaction Test; PP = Peripheral Perception; TAVT = Tachistoscopic Traffic Perception Test; COG = Cognitrone; WRBTV = Vienna Risk-Taking Test Traffic.

As can be seen in Table 7, “complex choice reaction (DT),” “simple choice reaction (RT),” “perceptual speed,” “subjectively accepted level of risk (WBRTV),” and “social responsibility” contribute to the prediction of patients' actual fitness to drive in all three samples. In the total sample, “time between the injury and participation in this study” also contributes to the validity of the predictive model, which was not the case for the subsample of stroke and TBI patients. In the subsample of TBI patients, the variable “selective attention” also contributed to the prediction of patients' actual fitness to drive.

Generalization of the Predictive Models across Both Subsamples

Although the relative relevance of the predictor variables obtained for the two specific subsamples are similar in magnitude—at least with regard to the results obtained with an artificial neural network—this does not mean that the structural relations between these predictors and the criterion are identical in stroke and TBI patients. This is particularly relevant in the case of the analyses carried out by means of an artificial neural network since the relative relevance of a predictor represents the sum of the direct-linear relation as well as the interaction and mediator effects. In order to test the degree of similarity of the predictive model in a more straightforward manner, several cross validations are calculated to investigate whether the models calculated on the basis of either stroke or TBI patients are generalize to the other subsample. In order to accomplish this aim, we applied the predictive models obtained for one specific subsample to the other. In the case of the unitary model, we restricted the predictive models calculated on the basis of one of the two subsamples to the predictor variables that proofed to be relevant in the analysis carried out with the total sample. The resulting models were later on applied to the remaining subsample to investigate the generalizability of the unitary model across both subsamples. In case, the unitary model generalizes across the two subsamples it can be concluded that even though there might be qualitative differences in the structural relations between the predictor variables and the criterion measure, these differences can be adequately accounted for by the more complex unitary model.

The results obtained for predictive models calculated by means of a logistic regression indicate that neither the more specific predictive model calculated purely on the basis of the stroke patients (stroke→TBI: r = .31, CR = 60.8%) nor the more general unitary model (stroke→TBI: r = .39, CR = 65.2%; TBI→stroke: r = .33, CR = 67.0%) can be assumed to be generalizable since neither the validity coefficient nor the classification rate was within the boundaries of the confidence intervals obtained in the bootstrap validation of the respective predictive model (cf. Table 4). In the case of the predictive model for TBI patients, this analysis was not possible due to the lack of a significant predictive model.

The results obtained by means of an artificial neural network were more promising. Although the validity coefficients and classification rates were outside of the confidence interval obtained in the bootstrap validation (cf. Table 6) when generalizing the predictive model optimized for each of the subsamples to the other subsample (stroke→TBI: r = .51, CR = 72.3%; TBI→stroke: r = .41, CR = 69.7%), the more general unitary model seemed to generalize rather well across both subsamples (stroke→TBI: r = .91, CR = 93.2%; TBI→stroke: r = .71, CR = 83.8%).

Discussion

This article investigated the contribution of driving-related ability and personality traits to the prediction of fitness to drive in patients suffering from TBI or strokes. Owing to the lack of theoretical models of driving that enable the structural relationships between measures of patients' actual fitness to drive and driving-related personality and ability traits to be precisely specified in advance, the contribution of these predictor variables was investigated using a logistic regression analysis and an artificial neural network. The results obtained with an artificial neural network outperformed those obtained by a logistic regression in the case of the predictive models for TBI patients and the unitary model. However, this was not the case for the subsample of stroke patients. For this specific subsample, the relation between the predictor and criterion measures is modeled equally well by a logistic regression and an artificial neural network. Most interestingly, the unitary model calculated by means of an artificial neural network captured the different structural relations between the criterion measure and the predictor variables observed for the two subgroups comparably well as indicated by the results obtained in the cross-validation and by the lack of significant differences in the distribution of stroke and TBI patients across those correctly and incorrectly classified by the unitary model. This was not the case when a logistic regression analysis was used to calculate a unitary predictive model. The comparison of the results obtained with the more specific predictive models and their generalizability indicates that there are indeed differences in the structural relation of the predictor variables and the criterion measure between stroke and TBI patients as suggested by previous studies (cf. Mazer et al., 1998; McKenna et al., 2004; Söderstrom et al., 2006). However, the more complex unitary model obtained by means of an artificial neural network seems to be able to take these differences into account; this is most likely due to the moderation of the structural relations by means of the variable “time between the injury and participation in this study” and the addition of a third hidden layer unit.

In line with previous studies, the results obtained with an artificial neural network indicate that complex choice reaction speed and perceptual speed contribute significantly to the prediction of fitness to drive in the total sample as well as the two subsamples (cf. complex choice reaction: Burgard, 2005; Hannen et al., 1998; Hartje et al., 1991; Lundquist et al., 1997, 2000; perceptual speed: Bouillon et al., 2006; Brooke et al., 1992; Burgard, 2005; Coleman et al., 2002; Galski et al., 1992; Hartje et al., 1991; Korteling & Kaptein, 1996; Marshall et al., 2007; Mazer et al., 1998; Niemann & Döhner, 1999; van Zomeren et al., 1988). In addition, in our study, simple choice reaction time contributes to the separation of drivers judged to be more or less likely to pass a standardized driving test. The results further indicate that “social responsibility” and “subjectively accepted level of risk” (Wilde, 1994) contribute significantly to the prediction of patients' fitness to drive. Besides these two driving-related personality traits, none of the other driving-related personality traits contributed significantly to the prediction of patients' actual fitness to drive. For instance, our measure of sensation-seeking turned out to be unrelated to patients' actual fitness to drive. This was also observed in samples of healthy adults (cf. Rim & Berg, 1999; Sommer, Herle, Häusler, Risser, et al., 2008). An inspection of the incremental validity reveals that driving-related personality traits accounted for 29.8% of the predictive validity compared with the total of 44.9% of the predictive validity explained by driving-related ability traits. In general, the finding that cognitive ability measures are more important in predicting patients' fitness to drive than driving-related personality traits is in line with results obtained in samples of healthy adults and elderly drivers (e.g., Anstey et al., 2005; Sommer, Herle, Häusler, Risser, et al., 2008).

Although the results obtained in this study are promising, there is one limitation that needs to be discussed in order to evaluate the present results. Owing to the exclusion criteria we employed in the present study, the level of severity of the brain injuries our patients were suffering at the time of their participation in this study is restricted. Even though the exclusion criteria are in line with legal regulations and recommendations on fitness to drive after TBI or strokes in Austria and Germany, the resulting sample cannot be regarded as representative of the larger population of stroke and TBI patients. It would thus be interesting to investigate whether the predictive model obtained in the present study generalizes to a much broader population of stroke and TBI patients in terms of type of diagnosis and severity of the injury. This kind of research would enrich our knowledge of the personality and ability determinants of fitness to drive in TBI and stroke patients in general.

Funding

This work was supported by the SCHUHFRIED GmbH, Mödling, Austria.

Conflict of Interest

Sommer, Heidinger and Häusler worked at the SCHUHFRIED GmbH, who financially supported the present study. Schauer and Schmitz-Gielsdorf were responsible for the data collection for which they received financial support from the SCHUHFRIED GmbH. There are no conflicts of interests for Arendasy.

References

Adams
H. P.
Bendixen
B. H.
Kappelle
L. J.
Biller
J.
Love
B. B.
Gordon
D. L.
, et al.  . 
Classification of subtype of acute ischemic stroke. Definitions for use in a multicenter clinic trial. TOAST. Trial of Org 10172 in Acute Stroke Treatment
Stroke
 , 
1993
, vol. 
24
 (pg. 
35
-
41
)
Akinwuntan
A. E.
Feys
H.
De Weerdt
W.
Pauwels
J.
Baten
G.
Strypstein
E.
Determinants of driving after stroke. A retrospective study
Archives of Physical Medicine and Rehabilitation
 , 
2002
, vol. 
83
 (pg. 
334
-
341
)
Anderson
J. A.
Rosenfeld
E.
Neurocomputing: Foundations of research.
 , 
1988
Cambridge, MA
MIT Press
Anstey
K. J.
Wood
J.
Lord
S.
Walker
J. G.
Cognitive, sensory and physical factors enabling driving safety in older adults
Clinical Psychology Review
 , 
2005
, vol. 
25
 (pg. 
45
-
65
)
Biehl
B.
Manual Tachistoscopic Traffic Perception Test (TAVTMB).
 , 
1996
Mödling
SCHUHFRIED GmbH
Bishop
C. M.
Neural networks for pattern recognition
 , 
1995
Oxford, England
Oxford University Press
Bortz
J.
Statistik für Sozialwissenschaftler [Statistics for social scientists].
 , 
1999
Berlin
Springer
Bouillon
L.
Mazer
B.
Gelinas
I.
Validity of the cognitive behavioral driver's inventory in predicting driving outcome
American Journal of Occupational Therapy
 , 
2006
, vol. 
60
 
4
(pg. 
420
-
427
)
Bridle
J. S.
Fogleman Soulie
F.
Herault
J.
Probabilistic interpretation of feedforward classification network outputs, with relationships to statistical pattern recognition
Neurocomputing: Algorithms, architectures and applications
 , 
1990
Berlin
Springer
(pg. 
227
-
236
)
Brooke
M. M.
Questad
K. A.
Patterson
D. R.
Valois
T. A.
Driving evaluation after traumatic brain injury
American Journal of Psychical Medicine and Rehabilitation
 , 
1992
, vol. 
71
 (pg. 
177
-
182
)
Brown
M. T.
Wicker
L. R.
Tinsley
H. E. A.
Brown
S. D.
Discriminant analysis
Handbook of applied multivariate statistics and mathematical modeling
 , 
2000
San Diego, CA
Academic Press
(pg. 
209
-
234
)
Bukasa
B.
Wenninger
U.
Brandstätter
C.
Validierung verkehrspsychologischer Testverfahren [Validation of traffic psychological test batteries]
 , 
1990
Wien
Literas
Burgard
E.
Fahrkompetenz im Alter: Die Aussagekraft diagnostischer Instrumente bei Senioren und neurologischen Patienten [Driving competence at advanced age: The validity of psychological tests in elder and neurological impaired respondents]
 , 
2005
Dissertation LMU München, Medizinische Fakultät
Chaloupka
Ch.
Risser
R.
Don't wait for accidents—possibilities to assess risk in traffic by applying the “Wiener Fahrprobe”
Safety Science
 , 
1995
, vol. 
19
 (pg. 
137
-
147
)
Chapman
P.
Underwood
G.
Forgetting near-accidents: The roles of severity, culpability and experience in the poor recall of dangerous driving situations
Applied Cognitive Psychology
 , 
2000
, vol. 
14
 (pg. 
31
-
44
)
Coleman
R. D.
Rapport
L. J.
Ergh
T. E.
Hanks
R. A.
Ricker
J. H.
Millis
S. R.
Predictors of driving outcome after traumatic brain injury
Archives of Physical Medicine and Rehabilitation
 , 
2002
, vol. 
83
 (pg. 
1415
-
1422
)
Evans
L.
Traffic safety and the driver
 , 
1991
New York
Van Nostrand Reinhold
Fahlman
S. E.
Faster-learning variations on back-propagation: An empirical study. Proceedings of the connectionist models summer school
 , 
1988
Los Altos
Morgan-Kaufmann
Fox
G. K.
Bowden
S. C.
Smith
D. S.
On-road assessment of driving competence after brain impairment: Review of current practice and recommendations for a standardised examination
Archives of Physical Medicine and Rehabilitation
 , 
1998
, vol. 
79
 (pg. 
1288
-
1296
)
Galski
T.
Bruno
L.
Ehle
H. T.
Driving after cerebral damage, a model with implications for evaluation
American Journal of Occupational Therapy
 , 
1992
, vol. 
46
 (pg. 
324
-
332
)
Golz
D.
Huchler
S.
Jörg
A.
Küst
J.
Beurteilung der Fahreignung [Assessment of driving ability]
Zeitschrift für Neuropsychologie
 , 
2004
, vol. 
15
 (pg. 
157
-
168
)
Groeger
J. A.
Risser
R.
Can psychological tests predict driving ability?
Assessing the driver
 , 
1997
Braunschweig
Rot-Gelb-Grün
(pg. 
59
-
72
)
Groeger
J. A.
Understanding driving. Applying cognitive psychology to a complex everyday task.
 , 
2000
East Sussex
Psychology Press
Hannen
P.
Hartje
W.
Skeczek
W.
Beurteilung der Fahreignung nach Hirnschädigung [Assessing fitness to drive after brain damage]
Nervenarzt
 , 
1998
, vol. 
69
 (pg. 
864
-
872
)
Hartje
W.
Dettmers
Ch.
Weiller
C.
Die neuropsychologische Testdiagnostik hat nur begrenzte Aussagekraft für die Fahreignung [Psychometric neuropsychological assessment has limited predictive validity regarding the evaluation of respondents' fitness to drive]
Fahreignung bei neurologischen Erkrankungen
 , 
2004
Bad Honnef
Hippocampus Verlag
(pg. 
19
-
22
)
Hartje
W.
Pach
R.
Willmes
K.
Hannen
P.
Weber
E.
Fahreignung hirngeschädigter Patienten [Fitness to drive in brain-damaged patients]
Zeitschrift für Neuropsychologie
 , 
1991
, vol. 
2
 (pg. 
100
-
114
)
Hatakka
M.
Keskinen
E.
Gregersen
N. P.
Glad
A.
Hernetkoski
K.
From control of the vehicle to personal self-control; broadening the perspectives to driver education
Transportation Research Part F
 , 
2002
, vol. 
5
 (pg. 
201
-
215
)
Häusler
J.
Sommer
M.
Vergleichende Simulationsstudien zu Methoden der statistischen Urteilsbildung [Comparative simulation studies of various approaches to statistical judgment formation]
 , 
2005
Paper presented at the 7th conference of the Task Force Psychological Methods and Evaluation, 28 September–1 October, Münster, Deutschland
Hergovich
A.
Bognar
B.
Arendasy
M.
Sommer
M.
Manual Vienna Risk-taking Test Traffic (WRBTV)
 , 
2005
Mödling
SCHUHFRIED GmbH
Herle
M.
Sommer
M.
Wenzl
M.
Litzenberger
M.
Manual Inventory of Driving-related Personality Traits (IVPE)
 , 
2004
Mödling
SCHUHFRIED GmbH
Hjälmdahl
M.
Várhelyi
A.
Validation of in-car observations, a method for driver assessment
Transportation Research Part A
 , 
2004
, vol. 
38
 (pg. 
127
-
142
)
Hornke
L. F.
Etzel
S.
Rettig
K.
Manual Adaptive Matrices Test (AMT).
 , 
2003
Mödling
SCHUHFRIED GmbH
Kinnebrock
W.
Neuronale Netze. Grundlagen, Anwendungen, Beispiele [Neural networks: Foundations, applications and examples].
 , 
1992
München
Oldenbourg
Klebelsberg
D.
Verkehrspsychologie [Traffic psychology].
 , 
1982
Berlin
Springer
Korner-Bitensky
N. A.
Mazer
B. L.
Sofer
S.
Gelinas
I.
Meyer
M. B.
Morrison
C.
Visual testing for readiness to drive after stroke: A multicenter study
American Journal of Physical Medicine and Rehabilitation
 , 
2000
, vol. 
79
 
3
(pg. 
253
-
259
)
Korteling
J. E.
Kaptein
M. A.
Neuropsychological driving fitness tests for brain-damaged subjects
Archives of Physical Medicine and Rehabilitation
 , 
1996
, vol. 
77
 (pg. 
138
-
146
)
Lezak
M. D.
Neuropsychological assessment
1995
3rd ed.
New York
Oxford University Press
Lundquist
A.
Alm
J.
Gerdle
B.
Levander
S.
Rönnberg
J.
Neuropsychological aspects of driving after brain lesions: Simulator study and on-road driving
Applied Neuropsychology
 , 
1997
, vol. 
4
 (pg. 
220
-
230
)
Lundquist
A.
Gerdle
B.
Rönnberg
J.
Neuropsychological aspects of driving after stroke—in the simulator and on the road
Applied Cognitive Psychology
 , 
2000
, vol. 
14
 (pg. 
135
-
150
)
Mahoney
F. I.
Barthel
D. W.
Functional evaluation: The Barthel Index
Maryland State Medical Journal
 , 
1965
, vol. 
14
 (pg. 
61
-
65
)
Marshall
S. C.
Molnar
F.
Man-Son-Hing
M.
Blair
R.
Brosseau
L.
Finestone
H. M.
, et al.  . 
Predictors of driving ability following stroke: A systematic review
Topics in Stroke Rehabilitation
 , 
2007
, vol. 
14
 (pg. 
98
-
114
)
Masters
T.
Advanced algorithms for neural networks
 , 
1995
New York
John Wiley and Sons
Mazer
B. L.
Korner-Bitensky
N.
Sofer
S.
Predicting ability to drive after strokes
Archives of Physical Medicine and Rehabilitation
 , 
1998
, vol. 
79
 (pg. 
743
-
750
)
McKenna
P.
Jefferies
L.
Dobson
A.
Frude
N.
The use of a cognitive battery to predict who will fail an on-road driving test
British Journal of Clinical Psychology
 , 
2004
, vol. 
43
 (pg. 
325
-
336
)
Michon
J. A.
Evans
L.
Schwing
R.
A critical review of driver behaviour models. What do we know, what should we do?
Human behaviour and traffic safety
 , 
1985
New York
Plenum Press
(pg. 
485
-
520
)
Mielke
A.
Neuronale Netze [Neural networks] [Online]
 , 
2001
 
Retrieved April 11, 2006, from http://www.andreas-mielke.de/nn.html
Niemann
H.
Döhner
A.
Schlag
B.
Fahreignung von Patienten mit Schädigungen des ZNS [Fitness to drive of patients with injuries of the central nervous system]
Empirische Verkehrspsychologie
 , 
1999
Lengerich
Pabst Science Publisher
(pg. 
69
-
90
)
Nouri
F. M.
Lincoln
N. B.
Validation of a cognitive assessment: Predicting driving performance after stroke
Clinical Rehabilitation
 , 
1992
, vol. 
6
 (pg. 
275
-
281
)
Nouri
F. M.
Tinson
D. J.
A comparison of driving simulator and road test in the assessment of driving ability after a stroke
Clinical Rehabilitation
 , 
1988
, vol. 
2
 (pg. 
99
-
105
)
Owsley
C.
Ball
K.
Sloane
M. E.
Roenker
D. L.
Bruni
J. R.
Visual/cognitive correlates of vehicle accidents in older drivers
Psychology and Aging
 , 
1991
, vol. 
6
 (pg. 
403
-
415
)
Radford
K. A.
Lincoln
N. B.
Murray-Leslie
C.
Validation of the stroke drivers screening assessment for people with traumatic brain injury
Brain Injury
 , 
2004
, vol. 
18
 (pg. 
775
-
786
)
Rakes
T.
Research in the neural networks/expert system area
Decision Line
 , 
1991
, vol. 
22
 (pg. 
11
-
14
)
Ranney
T. A.
Models of driving behavior. A review of their evolution
Accident Analysis and Prevention
 , 
1994
, vol. 
26
 (pg. 
733
-
750
)
Rim
P.-A.
Berg
L.
On the distinction between violations and errors: Sensation seeking associations
Transportation Research Part F
 , 
1999
, vol. 
2
 (pg. 
151
-
166
)
Risser
R.
Risser
R.
Validity of driver's diagnostics and selection
Assessing the driver
 , 
1997
Braunschweig
Rot-Gelb-Grün
(pg. 
89
-
101
)
Risser
R.
Brandstätter
Ch.
Die Wiener Fahrprobe [Vienna Driving Test]
 , 
1985
Wien
Literas
 
Kleine Fachbuchreihe des Kuratoriums für Verkehrssicherheit, 21
Risser
R.
Chaloupka
Ch.
Grundler
W.
Sommer
M.
Häusler
J.
Kaufmann
C.
Using non-linear methods to investigate the criterion validity of traffic psychological test batteries
Accident Analysis and Prevention
 , 
2008
, vol. 
40
 (pg. 
149
-
157
)
Rojas
R.
Neuronal networks. A systematic introduction
 , 
2000
Heidelberg
Springer
Scheiblechner
H.
Embretson
S. E.
Psychometric models for speed-test construction: The linear exponential model
Test design: Developments in psychology and psychometrics
 , 
1985
New York
Academic Press
(pg. 
219
-
244
)
Schuhfried
G.
Manual Determination Test (DT).
 , 
1998
Mödling
SCHUHFRIED GmbH
Schuhfried
G.
Manual Expert System Traffic (XPSV).
 , 
2005
Mödling
SCHUHFRIED GmbH
Schuhfried
G.
Prieler
J.
Manual Reaction Test (RT).
 , 
1997
Mödling
SCHUHFRIED GmbH
Schuhfried
G.
Prieler
J.
Bauer
W.
Manual Peripheral Perception (PP)
 , 
2002
Mödling
SCHUHFRIED GmbH
Schwarz
G.
Estimating the dimension of a model
Annals of Statistics
 , 
1978
, vol. 
6
 (pg. 
461
-
464
)
Söderstrom
S. T.
Petterson
R. P.
Leppert
J.
Prediction of driving ability after stroke and the effect of behind-the-wheel training
Scandinavian Journal of Psychology
 , 
2006
, vol. 
47
 (pg. 
419
-
429
)
Somers
M. J.
Thinking differently: Assessing nonlinearities in the relationship between work attitudes and job performance using Bayesian neural networks
Journal of Occupational and Organizational Psychology
 , 
2001
, vol. 
74
 (pg. 
47
-
61
)
Sommer
M.
Häusler
J.
Dorn
L.
Non-linear methods for the identification of drivers at risk to cause accidents
Driver behaviour and training
 , 
2005
, vol. 
Vol. II
 
Burlington
Ashgate
(pg. 
425
-
436
)
Sommer
M.
Herle
M.
Häusler
J.
Factorial and criterion validity of the Bad Tölzer Driving Test in samples of neurological patients and healthy adults: Result report
2008
Mödling
SCHUHFRIED GmbH
Sommer
M.
Herle
M.
Häusler
J.
Arendasy
M.
Ebner
G.
Fleck
G.
Von TAVTMB zu ATAVT: Eine Anwendung der Automatisierten Itemgenerierung unter einschränkenden Rahmenbedingungen
Zweites Österreichisches Symposium für Psychologie im Militär
 , 
2008
Wien
Schriftereihe der Landesverteidigungsakademie
Sommer
M.
Herle
M.
Häusler
J.
Risser
R.
Schützhofer
B.
Chaloupka
Ch.
Cognitive and personality determinants of fitness to drive
Transportation Research Part F
 , 
2008
, vol. 
11
 (pg. 
362
-
375
)
Stuss
D. T.
Shallice
T.
Alexander
M. P.
Picton
T. W.
A multidisciplinary approach to anterior attentional functions
Annals of the New York Academy of Sciences
 , 
1995
, vol. 
769
 (pg. 
191
-
211
)
Szlyk
J. P.
Seiple
W.
Viana
M.
Relative effects of age and compromised vision on driving performance
Human Factors
 , 
1995
, vol. 
37
 (pg. 
430
-
436
)
van Zomeren
A. H.
Brouwer
W. H.
Clinical neuropsychology of attention
 , 
1994
New York
Oxford University Press
van Zomeren
A. H.
Brouwer
W. H.
Rothengatter
J. A.
Snoek
J. W.
Fitness to drive a car after recovery from severe head injury
Archives of Physical Medicine and Rehabilitation
 , 
1988
, vol. 
69
 (pg. 
90
-
96
)
Venter
A.
Scott
E. M.
Tinsley
H. E. A.
Brown
S. D.
Issues in the use and application of multiple regression analysis
Handbook of applied multivariate statistics and mathematical modeling
 , 
2000
San Diego, CA
Academic Press
(pg. 
152
-
182
)
Wagner
M.
Karner
T.
Manual Cognitrone (COG).
 , 
2001
Mödling
SCHUHFRIED GmbH
Wilde
G.
Target Risk. Dealing with danger of death, disease and damage in everyday decisions
1994
Toronto
PDE Publications