top of page

Spilling the Beans About Hazardous Liquid Incidents

Tim K. Keyes, Ph.D.

In previous editions of the Quinnehtukqut, we introduced readers to hazardous liquid incidents in the U.S., covering 2010 to April 2018, using data from the Pipeline Hazardous Materials Safety Administration (PHMSA) which is managed by the Department of Transportation (DOT). We concluded the pipeline system is inconspicuously incident-prone and highlighted the primary pipeline operators and failure causes. In this third and final installment, we delve into the results of regression analytics (a method in statistical modeling), used to uncover drivers of the likelihood of an incident having an environmental cost and drivers of the severity of that environmental cost.

Modeling Approach

  • Environmental Cost was the ‘consequence’ metric

  • Data were randomly divided into an 80 percent training and 20 percent testing sets, the latter for validation

    • We examined two stages of finding the drivers of:

    1. Odds of nonzero environmental cost, or loss “frequency,” using logistic regression

    2. (Logged) Amount of nonzero cost, or loss “severity,” using general linear modeling

Modelling Results

Figure 1: Relative Impact Drivers of a) Incident Odds, and b) Incident Severity


Key drivers are shown in Figure 1, with the relative impact levels of a factor (e.g., such as MATERIAL (Mat’l) FAILURE as one cause among eight) compared to the average impact across all levels of a factor. Factor levels with a positive impact increase the odds of nonzero environmental cost, or increase cost severity, compared to all levels (and inversely, a negative impact decreases the odds or cost severity), This implies that to mitigate risks, pipeline stakeholders can prioritize their focus accordingly:

  1. Odds: 

    1. Material Failure and Corrosion are associated with 80 percent and 60 percent odds increases, respectively, ​compared to the average across all eight causes.

    2. Odds of nonzero cost is reduced when Ignition occurs; otherwise, odds increase by 70 percent.

    3. Pipeline, including Valve Sites as a system part, contributes to 50 percent odds increase, compared to the average across system parts, of which there are five types.

    4. If an Employee Drug Test was administered as required, it is associated with 30 percent greater odds. Tests are not always done, so 'indication of impairment' (creating need for the test) may not be causal. 

    5. Altitude above sea level matters - with odds increasing slightly for each one percent elevation change.

    6. Other factors shown tend to decrease odds. Noteworthy: when there have been Nearby Previous Incidents, odds are lower, apparently the result of regional adaptation. ​

  2. Severity: ​

    1. Among all five release types, a Rupture increases cost severity by 1.6x, unlike for a Leak.​

    2. If an Employee Drug Test were administered, it is associated with a 100 percent increase in cost.  Perhaps tests are done when losses are greater, i.e., when suspicion of impairment is greater.

    3. Pipeline, including Valve Sites as a system part is associated with a 40 percent increase in cost severity compared to the average across five system part types.

    4. Other factors shown tend to decrease severity of environmental cost.


In closing, we hope that this journey through data analytics applied to PHMSA data has been illuminating, if not provocative to action for industry insiders and watchdogs alike.

Tim K. Keyes, Ph.D., Evergreen Business Analytics, LLC is a CT Chapter Sierra Club member and investigator of the CT Chapter Sierra Club report on methane regulation Gas Pipelines.

bottom of page