Driver's use of mobile phones and road safety: a case- crossover study

Martin Dawes takes you through a case-crossover study, which aims to find out if talking on the phone while driving poses a greater risk of crashing

This month's paper is McEvoy S, Stevenson M, McCartt A, Woodward M, Haworth C, Palamara P, et al. Role of mobile phones in motor vehicle crashes resulting in hospital attendance: a case-crossover study BMJ 2005;331:428-30. You can read it by going to studentbmj.com and clicking on the link.

Hazard (control) interval=up to 10 minutes. Comparison=before crash versus yesterday.
Exposed=Used a mobile phone during the interval of interest (crash or control). Not exposed-Did not use a mobile phone during the interval of interest

Why do the study?
On the face of it, anything that detracts a driver from concentrating on the driving of a motor vehicle, be it a mobile telephone or drinking a cup of coffee, is more likely to lead to a crash. Motor vehicle crashes are a major cause of death in people under the age of 45 and are a considerable cause of morbidity in this group. A previous study in 699 people using a case-crossover design compared mobile phone use during the crash with mobile
phone use during the previous week.' They found that the risk of having a crash was four times higher during the time when mobile phones were used than when they were not being used.

This new study set out to check that finding. Although studies are powered, which means that they are designed with big enough samples (giving enough statistical "power") to show a real difference between groups and to ensure that the finding was not likely to occur by chance, we often need three or four studies (and sometimes a lot more) before really believing the data.

What is the study design?
Ideally, we would like a randomised controlled trial ensuring somehow that one group use their phones when driving and the other group are unable to do so and then observing the number of crashes in the two groups. This is clearly impossible for ethical and practical reasons. The next method is to identify people who have had a crash and determine their use of a mobile phone when driving and compare that with a group of similar aged individuals who have not had a crash and assess their rate of mobile phone use when driving. Two problems arise with this case-control design. The first is that it is difficult to accurately and retrospectively determine mobile phone use when driving when there is no critical incident to stimulate recall. Secondly, there may be inherent differences between the two groups that confound the results. For example, people who use the phone when driving may be more prepared to take risks generally than people who do not. What is needed is to determine if in those people who would use a phone when driving, the use of mobile phones leads to more crashes. This then leads to the design used in this study, called the case-crossover design.

What are the details of the study?
Drivers involved in car crashes between April 2002 and July 2004 in Perth, Western Australia, who were seen in one of the emergency departments, were the participants. The researchers accessed mobile phone records to identify phone use two hours before and after the crash. In addition, researchers identified mobile phone use during three control periods (for that individual) at 24 hours, 72 hours, and seven days before the crash. Mobile phone use was defined as calls made or received and text messages sent.

The primary aim was to compare mobile phone use in the 10 minutes before the crashes with mobile phone use (by the same driver) in three similar periods of driving 24 hours, 72 hours, and one week previously when the drivers could confirm they had been driving.

What were the results?
The researchers identified 1625 drivers, of whom 72°/o owned mobile phones. After various exclusion criteria this left 941 subjects. The researchers were able to get mobile phone data from 744 of these drivers, of whom 456 had at least one verifiable control interval. So we have three groups, the last of which contains the study participants.

In any research, it is normal to reduce the numbers using exclusion and other criteria. What is important is to ensure that by doing so you do not alter the composition of the individuals with respect to characteristics relevant to the study. In this case, the authors have clearly described in two tables the driving, crash, weather, trip, and phone characteristics of the groups as well as other factors, such as age. There are no statistical comparisons in this table because they are unnecessary. This is an important point when studying the differences of two groups in trials. Often authors feel the need to put in a statistical test here when in fact it is the clinical difference, not the statistical difference, that is important. For example, if there was a statistical difference in age of one month between the two groups this would not be relevant. The difference, even though statistically significant, is not clinically significant in terms of leading to a different pattern of behaviour of the two groups. So, in this table, we are not looking for a statistical difference, but a clinically significant difference. No clinically meaningful differences, rather than statistically meaningful, were shown, so we can say that the final study group was similar in all important respects to the initial large group.

Of the final study group, 28% of these drivers said they did not use mobile phones when driving. Seven per cent of drivers said that they had used the phone during the trip. Mobile phone records show that 9°/o of participants used their phones in the 10 minute interval before the crash. Mobile phones were used during 3% of the multiple control intervals.

How to understand odds
Few people understand odds. The main message to take away is that odds in large studies are similar to probabilities. If you don't like the idea of exploring this right now, you should maybe skip the next part

Ideally, we would like to know the probability (or percentage) of drivers using a phone and having a crash out of all drivers having a crash. We would then look at drivers who did not have a crash and calculate the probability of those who were using the phone at a similar time out of all the drivers who might have used a phone. The latter figure is clearly impossible to collect. But, for example, if the probability in the first group was 12% and in the second 3% we could say that the probability of phone use was four times higher in the crash group. We do not have that data, however, but odds match probability closely if the numbers studied are quite large. If the number of bee stings per year is 12 in 100, bee keepers would have a probability of being stung of 12% (12/100). The odds, in contrast to the probability, is calculated by dividing the number, not by the total, but by the number who did not have a sting (12/88=0.136). This is not so far from 0.12 (or 12%).

From this example, you can see that we can calculate the odds of using the phone when having an accident by looking at the crash group (phone use/total phone users), and the odds of using the phone and not having an accident in the control period. One divided by the other gives the odds ratio, which, for practical purposes, is similar to saying "it is four times more likely." To further complicate things statistically, a matched case control study requires a different way of calculating the odds ratio from that of calculating an odds ratio in an unmatched case control study, because everyone in the trial had the bad outcome or is a case.

The odds ratio in a matched case-control study is defined as the ratio of the number of pairs a case was exposed and the control was not (A), to the number of pairs the control was exposed and the case was not (B).

The author kindly provided me with the data for the following calculation from the trial in the table. Authors are often incredibly helpful, as in this case, and will share data, give explanations, and generally help people who want to know more about their research. This data comes from die 24 hours before the crash, and there were 248 matched pairs of data.

Remember, all these drivers were in a crash. So, of the people who crashed who were using the phone during the crash interval (n=26), 22 were not using the phone during the control period. Of the people who crashed who were not using a phone during the crash (n=222), six were using the phone during the control period. The odds ratio in this trial of paired data is the first number (n=22) divided by the second (n=6). OR = 22/6 = 3.7.

DAVID MCNEW/ GETTY IMAGES

Martin Dawes, chair of family medicine
Department of Family Medicine, McGill University,
515 Avenue des Pins, Montreal, Quebec, H2W 154, Canada
martin.dawes@mcgill.ca

1. Redclmeier DA,Tibshirani RJ. Association between cellular-telephone calls and motor vehicle collisions. N Engl J Med 1997;336:453-8