In my previous post Saving Lives with Statistics we have quantified the impact of workplace safety initiatives, and were able to forecast locations of accident occurrence one month ahead of time.
To the second point, forecasting accident locations ahead of time, I often got feedback in the flair of “We don’t need this, we know where to expect accidents, it’s where we have the most activity.”
Sounds logical so it must be true, right? Well we have the data, so let’s see if this heuristic is correct.
Since we were initially focusing our study on driving accidents, let’s look at the distribution of locations according to their monthly miles driven.
One might expect that driving accidents would happen predominantly at locations with high mileage and therefore we should focus our safety initiatives on those locations. One could set an arbitrary cut-off point at maybe 25,000 miles per month.
Next we have the distribution of locations with accidents. Following our assumption we would expect a top-heavy distribution with most accidents at locations with high mileage.
Doesn’t seem like it’s the case.
The median is at ~12,500 miles per month. That means, 50% of the locations with accidents drive less than 12,500 miles per month.
We have roughly 85% of locations that drive less than 12,500 miles per month. Doesn’t seem so obvious now where we should expect the accidents. This is the proverbial haystack where we’re looking for the needle!
However, because our high-risk forecast is using a multi-parametric model, we are able to highlight even locations with a low driving volume as at-risk.
The distribution naturally follows the distribution of accidents.
It’s a beautiful thing.
In conclusion, we are able to focus our workplace safety initiatives on a small sub-set of locations to target most of the accidents, including locations with low activity where accidents are traditionally not expected.
This approach proves useful to avoid pitfalls of pagan heuristics and anecdotal “evidence,” and provides solid, data-driven decisions.
Disclaimer: All numbers presented are for illustration purposes only and encoded with preserved ratios.