Confronting Data Bias to Improve Safety Outcomes

In safety management, data is often treated as objective truth. Leaders use incident rates, near-miss reports, injury trends and predictive models to guide them as they prioritize risk and allocate organizational resources.

Yet data can quietly mislead us, particularly when bias is embedded in what we collect and our measurement and interpretation methods. Effective, ethical safety leaders continuously work to recognize and address these distortions.

Exploring Various Biases
A widely cited World War II-era example illustrates the dangers of biased data. During the war, Allied forces studied returning aircraft to determine where additional armor was needed. Analysts initially recommended reinforcing areas with the most visible bullet holes. Statistician Abraham Wald challenged this reasoning, highlighting what is now known as survivorship bias. He observed that the only aircraft analyzed were those that survived their missions. Aircraft that failed to return home had likely sustained catastrophic damage to areas where no bullet holes were observed on the surviving planes. Wald’s insight suggested that undamaged areas required reinforcement, not the visibly damaged ones.

Survivorship bias remains a powerful warning for leaders whose safety programs rely on incomplete or filtered data. However, it is only one source of potential distortion.

Selection bias occurs when data is drawn from an unrepresentative sample. In a utility environment, this could happen when organizations heavily rely on information from crews or regions with strong reporting cultures while underestimating risk in areas where incidents and near misses are less likely to be reported. Leaders may inadvertently prioritize the wrong hazards when the dataset does not reflect the entire population.

Even when data is broadly collected, confirmation bias can still emerge (i.e., leaders subconsciously favor data that supports their existing beliefs or assumptions). For example, if management believes a particular work practice is safe, near-miss data that challenges their belief may be discounted or dismissed as anomalous. Over time, selective interpretation reinforces blind spots and weakens organizational learning.

Measurement bias can be introduced at the point of data capture, resulting in inconsistently defined or poorly standardized safety data. Metrics that depend on subjective judgment – such as what qualifies as a safety observation or near miss – can vary widely among supervisors, crews and contractors. When measurement practices differ, trends become unreliable and comparisons across departments or time periods lose meaning.

Historical bias arises when data reflects outdated assumptions, norms or exclusions that no longer align with today’s workforce or operating environments. Caroline Criado Perez’s book “Invisible Women: Data Bias in a World Designed for Men” highlights how systems built on incomplete data can overlook entire populations. In safety-critical industries, this could appear in PPE design, equipment ergonomics or training materials developed for a narrow segment of the workforce, leaving others at elevated risk.

More recently, algorithmic bias has emerged as organizations increasingly adopt predictive analytics and other safety tools driven by artificial intelligence, which can inherit and amplify patterns embedded in historical data. Any algorithms trained using past incident data that underrepresents certain hazards, job roles or worker groups may consistently underestimate risk in those areas. Since algorithmic outputs often appear objective, this bias can be difficult to detect and challenge without deliberate oversight.

Overcoming Vulnerabilities
Embedded bias distorts safety intelligence and can create organizational vulnerabilities. Resources may be misdirected. Early warning signs could be missed. Emerging hazards might remain invisible until a serious incident occurs. Overreliance on lagging indicators like recordable injury rates could create a false sense of security, especially in high-risk utility operations.

Biased data can also further erode trust. Reporting declines when frontline workers witness leadership decisions that conflict with their lived experiences, deepening the data gap.

Despite these risks, high-quality data remains indispensable to effective safety management, enabling organizations to identify trends, prioritize controls, evaluate interventions, and shift from reactive responses to proactive prevention. Decisions made without data are often driven by anecdotes and intuition.

The challenge, therefore, is not whether to use data but how to use it thoughtfully and with full awareness of its limitations.

Recognizing bias is the first step. Leaders should routinely ask, who is missing from this dataset? What assumptions shaped these metrics? What risks could be hidden? A questioning approach encourages more accurate, proactive, ethical decision-making. Leaders who understand bias are more likely to consult multiple data sources, blending quantitative indicators with qualitative insights from job observations, worker feedback and learning teams. Most importantly, confronting data bias helps to ensure that safety systems are designed to protect all workers, not just those most visible in the data.

Intentional effort is required to turn awareness into action. Organizations must routinely audit safety data for gaps and inconsistencies, standardize definitions and measurement practices, and foster psychologically safe reporting environments. As use of predictive analytics and other AI tools expands, transparency and human oversight are essential. Leaders must treat model outputs as decision aids – not decision-makers – and be accountable for how data-driven insights are applied in the field.

Conclusion
Numbers carry authority, shaping organizational budgets, priorities and narratives. However, as Abraham Wald demonstrated decades ago, some of our greatest threats may never appear in the data we see. Safety leaders who understand and deliberately question, test and correct for biases ultimately position their organizations to more effectively mitigate risk.

About the Author: Gina Vanderlin, CSP, CHMM, CIT, CUSP, is the customer operations health and safety program manager at PSEG Long Island. With over 15 years of experience leading EHS initiatives in high-reliability industries, she remains passionate about elevating safety from a compliance function to a strategic driver of culture, engagement and operational excellence. Reach Vanderlin at gina.vanderlin@psegliny.com.