Using previous H&S data to predict future incidents

The whole point of looking at data using a statistical process control approach (horrible name), we prefer the term process performance prediction, is to:

• Know when a change made to processes leads to an improvement
• Identify “signals” (in the data) from “noise”
• Predict the future

Yes, you can with quite small amounts of data, just seven points, have sufficient information to predict the future! 

In effect, Shewhart, when he invented this approach, was looking to understand when it was “economic” (read profitable) to investigate a particular data point and when it was not. 

He called the data points that required investigation “special cause” variation. These are points outside the green limit lines shown in the charts below. All points within the green lines were not worthy of investigation individually as they belong to the realm of chance and he called these “common cause” variation.

Whilst data points attributable to common cause variation are not worth investigation individually, the complete data can, and should, be looked at in order to identify issues that repeat regularly throughout the data. Pareto analysis (the 80/20 rule) is an excellent tool for looking at data in this way as this separates the “vital few” issues from the “trivial many”. If these vital few issues can be located, addressed and eliminated, it will bring the average down and reduce the range of variation for future events.

Accident and incident data

The data we are looking at here is a variety of H&S data from an engineering services company.

The data above shows the number of incidents being recorded per month over a 3 and a half year period, where there is a stable incident process from May 2016 to April 2017; where, in some months, the processes employed are capable of delivering no incidents but that in others there could be an expected maximum of just over 4 incidents per month.

However, in May 2017 the process changed, and since which time, the processes employed have been generating an average of just over 4 incidents a month, but up to 12 incidents a month (the top green line) would be within the bounds of expectation.

If nothing changes the on-going number of incidents a month would be expected to be within this range of 0-12.

…but… there is a “suspicion” that something may have already changed…

If the chart is split at January 2019 it reveals that the average has gone up from 3.8 accidents a month to 6.8. But this means, if the process has changed, the company could now expect up to 17 incidents reported per month. However, we only have 5 data points since that time and until we get 7 data points we cannot be totally sure.

Obviously, the total number of incidents is greater than the number of incidents where time has been lost. Looking at the lost time incidents above, the process has remained stable over the whole period for which data is available.

However, because the data point from July 17 is so close to the limit it might be worth investigating what happened at that point…assuming any reasons have not been lost in the mists of time.

Again, if there is no change to the process it would be expected that lost time incidents would continue to range between zero and three per month.

It was commented that over the last few years there had been a “lot of work” done with staff to promote the importance of capturing near misses, as a result, it would actually be hoped that the near miss rate would actually improve as a result of increased awareness. This is seen from the data and the number of reports now being made.

There are two real points of interest April 2019 when staff were targeted to record near misses which saw the number of report rise dramatically, but also Sept 2018…What happened here?

Rolling Data

More recently the company has adopted a programme of rolling data over a 12 month average. This has a couple of effects:

• The data profile is “smoothed”
• Critical data points end up being hidden

When data is analysed using a “rolling” approach there can be big surprises at the end of the 12 month period if, along the way, an extreme data point has been included in the data set and at the end of the period it is no longer part of the data set.

Conclusion

Processes prediction charting is a great way of being able to:

• Know when a change made to processes leads to an improvement
• Identify “signals” (in the data) from “noise”
• Predict the future

Process prediction charting allows for better decisions to be made from better data; only special causes are "investigated" and all other data is collated so that common causes can be identified and removed creating an ever safer environment.

Next steps

There are still things that could and should be done to both improve the above data set and to improve the analysis ...In view of the data that is now available the following recommendations are made:

• Consider refraining from the rolling 12-month approach it “covers up” information and can leave you with surprises
• Consider “normalising” the data in order to reduce the effects of skewing results. That is if you get more work you are likely to get more incidents and accidents, so normalise against 1) Sales or turnover 2) Hours worked or number of contracts 3) Number of people doing the work
• Ensure that the whole data is reviewed over a longer period (as above) rather than on an annual basis where the analysis starts afresh every January
• Consider looking at the “time between accidents” as this might reveal new information
• Consider recording the affected areas of the body where accidents are encountered as this can drive the need for PPE. i.e. very simplistically, if there are a lot of eye incidents make goggles compulsory
• Consider analysing accidents and incidents in relation to the activity undertaken at the time 

#HealthAndSafety #HealthAndSafetyData #StatisticalProcessControl #SPC #PredictingTheFuture #Accidents #Incidents #NearMiss #Improvement #SignalsFromNoise