How to Analyse Incident and Near-Miss Trends: A Practical Method

Most safety teams have more incident data than they use. The register is full, the monthly report has a bar chart, and nobody can say with confidence where the next injury is coming from. This guide is the method for getting from a register to an answer. It works in a spreadsheet, in a safety platform, or by asking an AI assistant, and the steps are the same in each.

Step 1: Decide what you are counting

Before you chart anything, settle the definitions, because a trend made of inconsistent categories is a trend in how people fill in forms. Agree on:

• Event types. Near miss, first aid, medical treatment, lost time, notifiable. Use the definitions from your WHS or HSWA reporting, so the categories match what you report to the regulator.
• Categories. Mechanism (manual handling, struck by, fall from height, vehicle), body part, task, plant. Keep the list short enough that people pick the right one.
• Severity. Actual and potential. A near miss with a crane load is a different thing from a near miss with a stapler, and your analysis should know that.

Step 2: Normalise by hours worked

Raw counts mislead. A site with 200 people will have more incidents than a site with 20, and that tells you nothing. Divide by hours worked and compare rates.

• LTIFR: lost time injuries per million hours worked. Use our LTIFR calculator.
• TRIFR: total recordable injuries per million hours.
• Near-miss rate: near misses per million hours, or per 100 workers if hours are hard to get. This is the one most teams skip and the one that tells you most about reporting culture.

If you cannot get hours, use headcount. It is rougher but it is better than raw counts.

Step 3: Segment, then segment again

A company-wide rate hides everything. Cut the data by:

• Site or project
• Category or mechanism
• Task or activity
• Time of day and day of week
• Crew, shift or contractor
• Tenure, if you have it. New starters and the first month on a new site are a known risk.

Then do the Pareto check: which three categories make up most of the total? In most workplaces, two or three mechanisms account for well over half of the injuries. That is your shortlist.

Step 4: Separate real change from noise

Here is where most monthly reports go wrong. Four incidents last month and seven this month is not a 75 percent increase in risk. It might be. It is more likely the normal swing.

• Look at twelve months, not two.
• Work out the usual month-to-month range. If the last twelve months ran between three and eight, seven is inside the range.
• Treat three or more months moving the same way as a signal. Treat a category that doubles, or a cluster at one site or one task, as a signal.
• If you have only a handful of incidents a month, use rolling three-month totals. Single months are too jumpy to read.

Near misses deserve a separate read. Rising near-miss reports with flat or falling injuries usually means reporting is improving, which is good news. Falling near-miss reports with rising injuries means people have stopped telling you things. Our near miss reporting guide covers how to fix that.

Step 5: Add the leading indicators

Incidents are what already happened. Put them next to what you are doing to prevent them, and the picture gets sharper.

• Inspection completion rate by site
• Open and overdue corrective actions
• Hazard reports per site
• Training currency and expiring licences
• Toolbox talk attendance

A site with rising near misses and falling inspection completion is telling you something a bar chart of injuries never will. The predictive safety analytics guide goes further into leading indicators.

Step 6: Ask why, then act

Analysis ends with a decision, not a chart. For each item on your shortlist, go back to the individual incidents and read them. Patterns in the descriptions, the same plant, the same time of day, the same crew, point at causes that the category codes miss. Then pick one or two problems, assign actions with owners and dates, and set a check for next month. If you need a structure for the why, see root cause analysis formats.

Doing this with AI

Every step above is a question you can ask an assistant, if the assistant can see your data. Safety Space AI answers these in plain English, in Safety Space or in ChatGPT, Claude or Microsoft Copilot through the Safety Space MCP connector:

• "Show me near misses per million hours by site for the last 12 months as a chart."
• "Which three mechanisms account for most of our recordable injuries this year?"
• "Is the rise in manual handling incidents at Site B bigger than the normal month-to-month swing?"
• "List every incident at Site B in the last quarter involving the pallet jack, with the time of day."
• "Which sites have rising near-miss reports and falling inspection completion?"

The method does not change. The assistant runs it faster and returns the chart, and you still decide what to do. If you are pasting data into a public chatbot to do this, read what happens to your safety data first.

A monthly routine that works

1. Update hours worked and pull the month's incidents and near misses.
2. Rates by site. Flag any site outside its twelve-month range.
3. Pareto by mechanism. Has the top three changed?
4. Near-miss rate versus injury rate. Which way are they moving?
5. Leading indicators next to the sites flagged in step 2.
6. Read the incident descriptions for the flagged items.
7. One or two actions, owners, dates. Check them next month.

That takes an hour a month by hand and a few minutes with a platform that holds the data. Either way, the hour is worth it.

Frequently asked questions

What is a good near miss to incident ratio?

There is no correct ratio, and the old 300 to 1 figure is a rule of thumb, not a target. Watch the direction. Rising near-miss reports with flat injuries is a good sign. Rare near-miss reports next to regular injuries means people are not reporting.

How do I know if a trend is real?

Compare it with the usual twelve-month swing. One month above average is noise. Three months in the same direction, a category that doubles, or a cluster at one site or task is a signal.

Can AI analyse incident trends?

Yes, if it can see the data. Safety Space AI answers trend questions with charts, inside the platform and in ChatGPT, Claude or Copilot.

How often should I do this?

Monthly for the routine above. Quarterly for the longer read, where you look at the year so far and decide whether the annual focus areas still make sense.