On Data Science in Accident Prevention

What would you say, if we could focus on 10% of locations, to prevent half of the accidents?

Image for post
Image for post

The US Occupational Safety and Health Administration agency (OSHA) estimates US companies spend around $150 billion on workplace safety each year.

Is this money being used effectively? Does it produce the intended results? Which initiatives are the best?

The results are being “measured” on a broad scale in terms of total injuries and lost time rates. Currently there is no measurement or estimation of the impact of individual initiatives. The programs and initiatives are being put in place based on the judgement of senior staff, hoping to reduce the trend over the coming years. If it doesn’t work out, there is usually one answer — do more.

Recent advances in Data Science and Analytics enable us to take a look under the hood, understand the drivers of injury rates, and determine the impact of each individual workplace safety initiative.

This information can be used in the following ways:

  • Run safety initiatives at their optimal level
  • In a budget constraint situation, decide which initiatives can be reduced without loss in performance
  • Focus on most effective initiatives for increased impact
  • Identify business units and job scenarios which are not using the safety initiatives effectively
  • Determine quantified risk (probability of injury) of a specific business unit or job scenario
  • Quantify the safety engagement of groups or individual employees
  • Propose specific actions to reduce the quantified risk and increase engagement
Image for post
Image for post

A data driven methodology was used at a leading Oil and Gas service company to understand and reduce driving accidents in the US.

A valid statistical model was developed which quantified the impact of each initiative intended to reduce driving accidents. Based on this, an optimal level of engagement was determined, highlighting what can be done better to reduce driving accidents over the next year, without adding more controls.

The statistical model was further used to determine the risk of each operational location specifically for the next month, highlighting that half of the accidents will happen at only 15% of the locations. We were able to determine precisely which locations that would be for the coming month.

Now the safety function can focus on a fraction of the locations in a targeted approach.

The insights in some cases confirmed what senior employees “knew” by intuition and proved it with real numbers. In other cases, it shed a new light on the processes that nobody had expected.

Image for post
Image for post
We were able to show diminishing returns in STOP reports. There is negligible improvement in accident rates when employees do more than 10 driving-related STOP reports per year. While there is little return on the effort beyond this quota, several locations could benefit greatly from this safety initiative if they are below this quota.
Image for post
Image for post
It was further shown that Coaching Events, another global safety control initiative at the company, with yearly quotas for every employee, are not very helpful, unless done by senior employees.
Image for post
Image for post
Another interesting insight that was derived from the historical performance is that safety training is way more effective if the conformance is above 90%.
Image for post
Image for post
As expected, a better Site Audit Score indicates better conditions, but now we can show exactly how much better. Locations with an audit score of 3 typically have 16% more accidents than locations with score 1.
Image for post
Image for post
Several Job Types are shown to have, controlled for all other effects, including driving volume and location risk, a statistically significantly higher occurrence of accidents than others. Safety personnel should focus the driving safety initiatives on these job-types.
Image for post
Image for post
What has been underestimated is the importance of Near-Miss reports, taken under the flair of “nothing actually happened, carry on.” In fact, they should be watched closely just like actual accidents. If two locations have exactly the same conditions, and one has 10 near-miss occurrences in the year (roughly one per month), the location would typically have 50% more actual accidents.

At the beginning of 2017, the statistical model was used to determine the probability of an accident at each operational location. At the end of the year, locations classified as high-risk ended up having over half of the accidents throughout the year.

Image for post
Image for post

The concentration of accidents under the high-probability flag allows the company’s safety personnel to focus on a manageable sub-set of locations.

Furthermore, having knowledge of the specific impact of each safety initiative, we can propose specific actions, and know how much improvement can be expected.

Image for post
Image for post

In this case, a step-change of 10% was identified as possible just by tweaking existing controls to their optimal levels, which were determined in the first part of this study.

While other controls are in place which could not be included due to data availability, this approach clearly shows the potential, but also the limitations of those considered.

The same approach was used to determine the quantified risk of each location each month. Performance from the previous month was used to calculate the risk level of the next month.

Image for post
Image for post

The results were similarly astonishing, we were able to identify 64% of the accidents concentrated within a small sub-set of locations.

Image for post
Image for post

We were able to quantify the impact of individual safety initiatives, determine the optimal level of each initiative, quantify the risk of each location on a yearly and monthly basis, suggest specific actions to reduce the accident rate, and identify accident locations ahead of time.

A similar approach can be used on other accident types and job scenarios, as well as to quantify job risk and business risk overall.

Data analytics is proving to be a powerful tool for answering “non-technical” questions and providing reliable recommendations for confident, data-driven business decisions.

Following projects would be possible with the same methodology:

  • Optimal Engagement — How are safety and job quality initiatives done most effectively (by whom, how often, when, etc.)
  • Targeting Audits — Probability of a site-audit uncovering non-conformances to reduce unnecessary checks and “pencil-whipping” of mostly positive reviews
  • Job Intervention — Real-Time probability of undesirable operational events like accidents or quality non-conformance (similar to credit card fraud detection upon each card-swipe)
  • Measuring Fatigue — Effects of long work hours and periods with little rest on performance and incident rates

www.linkedin.com/in/marekdanis

Disclaimer: All numbers presented are for illustration purposes only and encoded with preserved ratios.

Get the Medium app

A button that says 'Download on the App Store', and if clicked it will lead you to the iOS App store
A button that says 'Get it on, Google Play', and if clicked it will lead you to the Google Play store