1. The Matched Sampling Technique

Although road safety researchers focus primarily on driver behavior, vehicular defects and road design, there is general agreement that environmental factors, such as weather and darkness, also affect accident risk. Research on weather-related hazards, especially precipitation, has made extensive use of matched sampling.

This application of matched sampling first requires weather data that can be used to identify precipitation events. Each event is then matched with a suitable control period (i.e. a period of 'nice' weather). For example, a Monday afternoon rain shower lasting from 1 p.m. until 4 p.m. would be paired with the same three hour period on a Monday afternoon just one or two weeks prior to or following the event. The absence of any kind of adverse weather during the control period is an essential feature of this method. Events without a suitable control are deleted from the sample.

Accident frequencies (and sometimes characteristics) during events are then compared with those during control periods. Usually, these comparisons take the form of relative risk ratios. For example, if there were 12 collisions during the event and 8 during the control, then the relative risk ratio for this particular event-control pair would be 12 / 8 = 1.5. Typically risk estimates are based on the combined data from many event-control pairs.

The main reason for choosing this quasi-experimental design is related to the degree of control that is implicitly incorporated in the study, i.e. the fact that many variables that have nothing to do with weather, but which do affect accident frequency, such as traffic volume, road conditions and the incidence of impaired driving, are controlled rather effectively. As a result, it seems reasonable to suggest that any deviations from a risk ratio of 1.0 are largely attributable to weather.

Matched sampling has been applied in many case studies of weather hazards. Codling (1974) pioneered the first use of matched sampling, estimating relative risk for wet weather days for 1968-69 in the United Kingdom. Later, Sherretz and Farhar (1978), Bertness (1980) and Smith (1982) applied this technique to the cities of St. Louis, Chicago and Glasgow, respectively.

2. Case Study

A Ph.D. thesis by Andrey (1989) provides a Canadian application of the matched-sample approach. Her study was based on a sample of more than 25,000 accidents and nearly 1,000 event-control pairs that occurred in seven different cities of Alberta : Calgary, Edmonton, Medicine Hat, Lethbridge, Red Deer, Fort McMurray, and Grande Prairie during the 1979-1983 time period. Click here to see a map of Alberta. Selected results for rainfall events are summarized below :

Accident Frequency	During Events	During Controls
Minimum Value	0	0
Maximum Value	124	89
Median	3	2
Mode	1	1
Mean	11	7
Standard Deviation	19	11
Sum	6066	3634

• As evident in the above table, accident frequencies during events and controls are highly variable : frequencies range from 0 to more than 100 and the standard deviation is greater than the mean. This is partly because the cities vary in population (i.e. Calgary and Edmonton both had populations between 500,000 and 600,000 and the other 5 cities had populations below 50,000). Another reason is that the events vary in length, from two-hour rain showers to one situation where rainfall occurred intermittently for 35 hours.
• Overall, accident frequencies are greater during events than during controls. The relative risk ratio is 1.67 (6066/3634). This tells us that, on average, there are 67% more collisions when rainfall occurs compared to when it does not. "On average" are important words, however. As shown in the following table, even the relative risk ratios for individual event-control pairs are highly variable :

Relative Risk Ratio	# Event-Control Pairs	% Event-Control Pairs
<= 1.000	199	36.7
1.001 to 2.000	132	24.4
2.001 to 3.000	115	21.2
3.001 to 4.000	19	3.5
> 4.000	77	14.2
All	542	100

• Usually, accident frequencies are greater during events than during controls, but not always. In fact in 36.7% of the observations, the reverse was true.
• In some cases, however, accident rates increase dramatically during rainfall events, as shown by the 14.2% of observations where risk ratios were greater than 4.0.