to validity discussed in section 4.2. The degree of control for confounding factors actually obtained by the various research designs depends on two factors:
1. The way the research design is implemented
2. The way measures intended to influence speed are introduced.
4.3.1 Commonly employed study designs
Table 4 lists some of the more commonly found versions of various study designs employed in road safety evaluation research. Randomised controlled trials
(experiments) are rarely used in road safety evaluation research. When used, such trials nearly always rely on a matched pair design. In this design, pairs of study units are formed, so that the members of each pair are as similar as possible with respect to all factors affecting road safety. Then one member of the pair is drawn at random for introduction of a measure intended to influence speed. The other member of the pair forms the control group. In general, matched pair experiments in road safety are rarely very rigorous by the standards that have developed in medical research. Features such as double blind analyses of data are rarely
implemented. Despite this, randomised controlled trials can be assumed to control for all confounding factors, at least if the number of pairs used is sufficiently large for the law of large numbers in sampling theory to hold.
Table 4: Commonly employed study designs in studies evaluating the relationship between speed and road safety
Study design Commonly found forms of the study design
A. Randomised controlled trials A.1 Matched pair design
B. Before-and-after studies B.1 With a matched comparison group
B.2 With a non-equivalent comparison group B.3 With data on some confounding factors
B.4 Without comparison groups or data on confounding factors
C. Cross-sectional studies C.1 Employing multivariate models to form homogeneous groups
C.2 Stratifying data by confounding factors C.3 Simple bivariate studies
D. Case-control studies D.1 Controlling for confounding factors by multivariate analyses
D.2 Controlling for confounding factors by stratifying data D.3 Not explicitly controlling for confounding factors Source: TØI report 740/2004
Before-and-after studies come in many different forms. In Table 4, the most commonly found forms of this design have been listed from those that best control for confounding factors to those that basically do not control for any confounding factors at all (simple before-and-after studies with no comparison group and no data on confounding factors).
The various forms of cross-sectional studies and case-control studies have also been listed from those that presumably embody the best control of confounding factors to those that do not control for confounding factors.
4.3.2 Ways of introducing speed-influencing measures
Whether a certain potentially confounding factor is likely to exert an influence on a study or not, depends to some extent on how the measure intended to influence speed has been introduced. A distinction can be made between two broad classes of cases:
1. Cases in which the measure is local, i.e. introduced on certain road sections only.
2. Cases in which the measure is system-wide, i.e. introduced on all roads in a country or state.
In the former case, the road sections where a measure is introduced are always deliberately selected from a larger set of roads. In this process of selection, a bad accident record may be one of the reasons for introducing a safety treatment on a certain road. When the measure is local, one can never rule out the possibility that regression-to-the-mean can (but not necessarily will) influence a study. A similar point of view applies when a speed limit is temporary (seasonal). The period during which it applies will then very often have been identified as a period during which there are more accidents than during the rest of the year.
In the latter case, the measure will usually involve changing speed limits on a system-wide basis. Although even the total number of accidents in a country or state is subject to random fluctuations, it is highly unlikely that these could exert a very great influence on a study. Consider, as an example, a system in which the annual expected number of injury accidents is 1,000. The 95% confidence interval for this expected number of accidents spans 938 to 1,062 accidents. Hence, the maximum conceivable size of regression-to-the-mean in this case would be amount to an accident reduction, or accident increase, of slightly more than 5%.
Regression-to-the-mean is, therefore, in general less likely to be a major confounding factor if a measure is system-wide than if it is confined in time or space.
Long-term trends can influence all before-and-after studies, but not cross- sectional studies or case-control studies. This applies irrespective of whether a measure intended to influence speed is introduced locally or on a system-wide basis.
Changes in traffic volume, or differences in it, is a threat to validity in all studies, except those that have been matched with respect to traffic volume. Effects of risk factors other than speed figures prominently as a potential source of error in cross- sectional studies and case-control studies, but is less likely to affect before-and-
after studies, in particular if these studies have employed a comparison group that reflects the effects of all risk factors influencing accidents.
In chapter 5, the assessment that has been made in this chapter will be employed in order to score studies with respect to control for confounding factors.