If you think the title is provocative, you are right. This blogpost aims to discuss a different way of looking at tailings dam failures.
Biases occur if one uses statistical samples from a failure database to evaluate tailings dams’ likelihood of failure. That is because one considers the sub-sample of failed dam as opposed to the overall inventory. Indeed, if one ignores the base inventory, one could draw misleading conclusions.
Berkson’s Paradox Applied to Tailings Dams
Berkson’s paradox is a selection bias. It applies, for instance, to studies examining the potential root cause of failures from a statistical sample of post-failure records. We can immediately draw a connection with tailings dams.
Studies using samples from post-failure records, rather than from the general inventory, may result in an illegitimate association between failures and the alleged root cause.
Let’s assume a study looks at dam collapses in a post-failure database. The query bears, for example, on the dams’ foundation failures caused by earthquakes. The record is obviously biased because it contains, by design, only failed dams. Because the sample is from that biased database, results are more likely to show earthquakes generating foundation failure than in the general unknown portfolio. Furthermore, one will obtain the result regardless of whether there is any association between earthquakes and foundation failure in the general inventory.
Tailings Dam Failure Records
Let’s take a hypothetical example where from 2000 to 2018, we have 11 upstream dam failures, four downstream dam failures and two centerline dam failures.
From the list above, one can easily assert that upstream tailings dams are more likely to fail than downstream dams. In addition, downstream dam failures are twice as likely as centerline ones.
Is that correct? Not really! If, for example, the world inventory encompassed two downstream dams for every centerline dam, the rate of failure for centerline dams would actually be equal to the downstream dams, not double as suggested by the first look at the sample above.
Similarly, let’s suppose the worldwide inventory of tailings dams had a 3:1 ratio of upstream dams vs downstream dams, meaning that for every downstream dam, there are more than three upstream dams around the world. As a result, the initial conclusion that upstream dams are more likely to fail would be incorrect.
Conclusion 1: Using only failed dams to define general rules is fallacious if the general inventory is not known.
The Devil is in the Details
So, the question is, how do we even count the number of dams? Let’s note that we must properly define the term “dam” first. Does a ring dyke impoundment count as one or several dams? When a client asks us to deploy ORE2_Tailings™, we always state that our analysis covers homogeneous segments of a dam. As a result, we may split a ring dyke impoundment into six different homogeneous segments of tailings dams. Similarly, we may split a very long linear dam into several segments.
Then, if we look at the alleged causes for the failures of our hypothetical case, let’s assume that during the same period, five failures were allegedly due to earthquakes and 10 were allegedly due to foundation failure. Two were allegedly due to both foundation failure and earthquakes.
Again, are we drawing the right conclusion or are we simply stating that many dams were built in a seismic area? How many dams in a seismic area survived?
Conclusion 2: Building general causalities or failure mode “rules” can be fallacious.
Conclusion 3: Drawing correlations, even if they seem mathematically sound, can be incorrect.
Let’s Go Back to that Famous 3,500 Active Dams Portfolio Estimate
We were very pleased to read Franks et al.’s 2021 article titled “Tailings facility disclosures reveal stability risks.” In our opinion, this research is a lot closer to depicting the tailings dam world inventory than others. The article states the following: “1,743 facilities, 725 of which are currently active, representing an average of 36% of contemporary global commodity production,” have been sampled.
Let’s do some “napkin” math. If 725 currently active tailings dams represent 36% of world production, then 100% of world production should be covered by 2,013 active tailings dams. Then considering the above numbers, one could infer there is a total of 4,860 noteworthy facilities around the world! Of course, these values bear uncertainties as the distribution of tailings volume likely follows a Pareto distribution than a uniform one. Nevertheless, let’s note that 2,013 and 4,860 have as an average the “magic” number of 3,500 used earlier by various authors, including us.
