Paper Plus: Driver sleepiness and risk of serious injury to car occupants

This month, John Fletcher looks at a population based case control study about driver sleepiness and the risk of injury to car occupants

This month's paper is Connor J, Norton R, Ameratunga S, Robinson E, Civil I, Dunn R, et al. Driver sleepiness and risk of serious injury to car occupants: population based case control study. BMJ 2002;324:1125. You can read the paper by going to studentbmj.com and clicking on the link..

Why do the study?

At first glance, you might think this is silly research: it stands to reason that falling asleep at the wheel of a car is hazardous, so why go to all this trouble to prove it? Fortunately the authors tackle this straight away in the introduction when they tell us that what is not known is just how important sleepiness is as a cause of crashes. Existing studies suggest that dozing drivers account for anything between 1% and 33% of crashes, which leaves some room for disagreement about how vigorously to tackle the problem.

Is this paper worth reading?

Accidents are the commonest cause of death in younger people, and cars account for most of these deaths. So this is an important topic, though it will be of most interest to those who think prevention is better than cure.

By looking at the title and the last sentence of the introduction, the authors are clearly focusing on sleepiness as a possible cause of car crashes that result in death or serious injury. They used a case-control study design, which is appropriate for assessing risk factors for a rare harmful event. Obviously a clinical trial of sleep deprivation cannot be used, because it would be unethical to cause death and injury.

The alternative observational study design would be a cohort study, in which researchers classify people according to their risk factors-such as sleepiness, age, sex, and so on-at the start of the study. Then the researchers monitor them over time and count and compare all outcomes, such as crashes or deaths, between categories of risk factor. Since crashes are fairly unusual, a cohort study would involve monitoring a huge number of people for a long time, simply to observe a non-event in most of them and a crash in a very few; this is expensive and time consuming.

In this study, the researchers selected and classified the subjects on the basis of the outcome-that is, having had a crash, with a suitable number of controls who have not had a crash. They then look back into the past and assess the risk factors to which each group were exposed to count and compare between case and control categories. Since they start with people who have had a crash, they need only find suitable controls, and the whole study is much less wasteful and more efficient.

Is the study likely to be reliable?

Case-control studies provide seemingly endless possibilities for being misled but, only three things matter-bias, confounding, and chance. By keeping these three sources of error in mind while reading the paper, an informed opinion about the paper's reliability is possible.

Bias is most likely to occur during the selection of subjects and the gathering of information about their risk factors, or "exposures" as they are often referred to. Ideally, cases should be clearly defined according to criteria that do not allow researchers any leeway and all eligible cases should be included in the study. In this study, cases were the drivers of any car involved in a crash that resulted in death or treatment in hospital of any of the car's occupants, including the driver. Also, the crash must have occurred in a defined geographical region in New Zealand. They went to the hospital records and to the coroner's office to find their cases, and this looks fairly clear cut and comprehensive. If the case definition is in any way subjective, there is a danger that, unwittingly or not, especially where the decision is borderline, people may be more likely to be classified as cases if they have an exposure that is considered harmful. This kind of selection bias builds the association into the study and the results become self fulfilling-and wrong.

Selecting controls

Controls are often harder to select than cases. They should come from the same population as the cases, and they should represent that population's exposure to the risk factors of interest. These researchers stood by the roadside at random times and places and stopped cars to interview drivers. They did their best to balance the sampling according to time of day and traffic volume. It is reassuring that they took the numbers of drivers that did not stop and attempted to trace them and conduct a telephone interview.

Avoiding information bias

To avoid information bias, data should be collected from cases and controls in exactly the same way, if possible. There were some differences between cases and controls in the interview techniques used. Researchers interviewed all the controls by telephone up to a few days after being selected at the roadside. In contrast, many cases were interviewed one to one in hospital, and someone else gave information if the driver had died. Researchers got information on alcohol concentration by roadside breath test for all controls but by a mixture of methods, or not at all, for cases. How much these differences matter depends on whether they are likely to affect the answers given to the questions about sleep in particular.

As well as different methods of collecting data, differing memories or accounts of events may introduce bias. In the last sentence of the "Data collection" paragraph, the researchers imply that the questions about sleep were a small part of the interview, perhaps to suggest that subjects would answer them without thinking. If you believe that drivers who have had a crash are likely to be as honest about their sleep patterns as drivers who have not had a crash, then this possible source of recall bias will not concern you too much.

Confounding factors

Confounding occurs when a factor exists that is strongly related to both the exposure and the outcome of interest. For true confounders this relationship is also independent but this is sometimes hard to unpick. The questions to ask of a study are: have they identified the important potential confounders, and have they made appropriate adjustments in the analysis? To answer the second question, look to see if the researchers mention the words "adjusted, regression, multivariate, stratified, or modelling" in the analysis section. If they use any of these words, they have probably used a statistical adjustment technique. It is probably easiest to trust the researchers and the journal's referees at this point and hope that they have done a good job. To answer the first question, note down the things you think might be potential confounders and then look at the list of things the authors considered to see how well they agree. This paper uses a fairly sophisticated logistic regression model to account for potential confounding and they give a long list of things they have considered both in the text and in table 1. Should they have asked about drugs and make of car? If you think drugs are often taken and can affect the probability of a crash as well as symptoms of sleep, this may worry you. Similarly, if you think some cars both keep you awake and make injury to occupants less likely-for example, heavy noisy four by fours-this will make you cautious in interpreting the results.

What are the results?

Using the result in the abstract is the easiest way to find a number to quote. It is also usually, but not always, the number chosen by the authors to have the highest impact. The authors present an odds ratio of 8.2 for those with acute sleepiness versus those who are less sleepy. This means acute sleepiness was seen more frequently in cases than controls. Since this is a case control study the odds ratio is interpreted in the same way as the relative risk would be from the equivalent cohort study. That is to say, acute sleepiness makes an injury causing crash 8.2 times more likely. The researchers report a similar odds ratio of 5.6 for driving at night. This suggests it really is sleepiness that is a risk and not just how people answered a questionnaire. They put this in context by suggesting that the population attributable risk of sleepiness is 19%. In other words, if the sleepiness of drivers was a real cause of crashes and was eliminated entirely, then 19% of injuries would be prevented.

Looking at the results quoted in the tables is a good check, to ensure a balanced view. Table 2 contains the necessary information. There seems to be a fairly clearly increased risk no matter what measure of acute sleepiness is used and whether the drivers who consumed alcohol are included in the analysis or not. This consistency of association suggests the association is real. By contrast, the measures of chronic sleep disturbance do not show a strong or consistent relationship with crash risk.

If you have considered bias and confounding and think that the study is probably OK, the last question to ask is could it all be due to chance? Look for P values or a 95% confidence intervals around the main result. In Table 2 and the abstract, the odds ratio for the whole population for a score on the Stanford sleepiness scale of 4 and above versus 3 and below was 8.2, and the 95% confidence interval is from 3.4 to 19.7. Since the lower and the upper limit both suggest that sleepiness is associated with a crash, it is unlikely that this result is due solely to chance. Lack of at least one night of full sleep in the previous week was associated with a relative risk of around 1.5. But the lower 95% confidence limit is 0.9 suggesting the possibility that lack of sleep may be safer and the upper 95% confidence limit is 2.4 suggesting lack of sleep may be less safe. Here the play of chance means the result is not sufficiently precise to say for sure if chronic sleep deprivation is a risk factor for a crash, though if it is, it is not a strong one.

Before rushing out and doing something, reflect on whether there is anything special about the study setting that limits its wider application. This study was done in an urban area of New Zealand, and it is hard to see why it would not have a message for other temperate Western countries.

How important is a relative risk of 8 or so? Using the data in table 1, working out an unadjusted odds ratio for any of the listed variables is possible. One of the strengths of a case-control study is just this, that the risk associated with any exposure may be estimated, though of course there is only one outcome measure. The odds ratio of a crash causing injury for reporting any alcohol consumption is (418/129) divided by (537/41) which is 4.0. In the same way the odds ratio for any measured alcohol consumption is 22.5. It looks like driver sleepiness may be about as important as drink driving as a cause of serious accidents.