Researchers at the University of Colorado Boulder’s Construction Safety Research Alliance have published a test of Check Yourself, a web-based prompt tool meant to keep incident investigators from steering interviews toward the answer they already expect. Fourteen of 20 experienced safety professionals showed fewer signs of bias after using it.
The paper appeared Oct. 2 in the journal Safety. The authors call it exploratory, and they mean it.
How the incident investigation bias tool was tested
Each participant ran two simulated post-incident interviews over Zoom with trained actors, one built around a staircase incident and one around a concrete form incident. Check Yourself was introduced between the two rounds. The researchers then coded the transcripts for observable indicators of six biases: confirmation, anchoring, experience, hindsight, conservatism in revising beliefs, and the fundamental attribution error, which is the habit of blaming the person before the conditions.
The 20 participants averaged about 12 years of investigation experience. They were recruited from oil and gas, service, utility and heavy civil work.
On the authors’ scoring, the group mean moved from 3.45 in the first round to 4.85 in the second, a 1.40-point gain (paired t-test p = 0.0018, Cohen’s d = 0.76). Four participants didn’t change. Two got worse. In a follow-up survey, 70 percent said they were more aware of their own biases and 80 percent called the tool relevant to real investigations.
What the construction safety study can’t show
The limitations section is longer than most, and worth reading before anyone rewrites an investigation procedure.
- No control group. Everyone did a second interview after a first, so practice alone could explain part of the gain.
- The second round included a structured reflection step the first didn’t, which the authors say limits direct comparison of the scores.
- Actors on a video call aren’t a crew the morning after someone got hurt.
- Independent double-coding covered only five of the 20 participants (90 percent agreement, kappa 0.78).
- No follow-up measured whether the effect lasts.
One more point the paper discloses but doesn’t dwell on: the Construction Safety Research Alliance funded the study, and its researchers built the tool. That’s normal for applied research. It’s also a reason to want a replication from somebody else.
Why investigator bias matters to contractors
An incident report is only as good as the interviews under it. If the investigator arrives believing the worker skipped a step, the questions go looking for the skipped step, and the corrective action becomes retraining. The design, the schedule pressure and the missing equipment never make the page.
That’s the case for a cheap, checklist-style intervention even on thin evidence. The downside is a few minutes of reflective questions. A 20-person sample with a medium-to-large effect is a reason to pilot the idea. It isn’t a reason to cite it in a safety manual as proven.
The same instinct runs through other recent work on jobsite blind spots, from cross-trade hazard training to North Carolina State’s study of why crews resist Type II helmets. The hazard is often in what people assume.
The authors want field trials on real incidents with long-term follow-up next. Until then, 14 of 20 is the number. So is 6 of 20.