Prevalent network screening methods rely on a set of analytical models known as Safety Performance Functions (SPFs), that are developed on the basis of a database of historical data (i.e. reported collisions and site characteristics such as geometry and exposure). The SPFs are used to predict the number of collisions (by severity) at each site, which is then combined, using an Empirical Bayes (EB) approach, with the observed number of collisions at the site, to obtain the expected number of collisions. SPF-based network screening methods in the HSM include:

• Expected Average Crash Frequency with Empirical Bayes (EB) Adjustment
• Equivalent Property Damage Only (EPDO) Average Crash Frequency with EB Adjustments
• Excess Expected Average Crash Frequency Using SPFs with EB Adjustments

There are four main challenges to carrying out network screening using these methods:

1) Methods that rely solely on historical collisions suffer from limitations such as bias due to low mean and small sample size and regression-to-the-mean (RTM). Furthermore, these methods are reactive, rather than proactive, as they can only identify hazardous sites after collisions have occurred.

2) SPFs must be developed using local data and must be regularly updated (re-developed). This activity is resource intensive, requiring extensive data and specialized statistical and road safety modelling expertise, and therefore posing a barrier for small and medium sized municipalities in Canada. Further work is needed to (i) develop tools to enable municipalities to determine in an objective way if an existing SPF needs to be updated; and (ii) develop methods that permit transferability of an SPF that has been developed in another municipality to a local jurisdiction with minimal recalibration effort.

3) Conventional SPFs consist of analytical forms such as the Negative Binomial model which make assumptions about the distribution of the collision data that are frequently invalid. Emerging AI techniques provide an opportunity to develop models that can better represent collision data and other forms of data that reflect safety.

4) Typically, there is substantial error in the exposure data for motorized vehicles (normally expressed as AADT and estimated from a small number of count stations) and little or no data to quantify exposure of vulnerable road users such as pedestrians and cyclists. This results in both error and bias in the network screening outcomes. Methods are required to improve the accuracy of exposure data for vehicles and VRUs and for NS screening methods that are less sensitive to errors in exposure data.