David and Drew discuss a research paper published this year in the Journal of Safety Research entitled “Exploring Bias in Incident Investigations: An Empirical Examination Using Construction Case Studies” by Sreeja Thallapureddy et al.
You’ll hear David and Drew delve into the often overlooked role of bias in accident investigations. They explore the potential pitfalls of data collection, particularly confirmation bias, and discuss the impacts of other biases such as anchoring bias and hindsight bias. Findings from the paper are examined, revealing insights into confirmation bias and its prevalence in interviews. Strategies for enhancing the quality of incident investigations are also discussed, emphasizing the need to shift focus from blaming individuals to investigating organizational causes. The episode concludes with the introduction of Safety Exchange, a platform for global safety community collaboration.
Discussion Points:
Quotes:
"If we actually don't understand how to get a good data collection process, then it really doesn't matter what happens after that." - David
"The trick is recognizing our biases and separating ourselves from prior experiences to view each incident with fresh eyes." - Drew
"I have heard people in the industry say this to me, that there's no new problems in safety, we've seen them all before." - David
"In talking with people in the industry around this topic, incident investigation and incident investigation quality, 80% of the conversation is around that causal classification taxonomy." - David
Resources:
The Safety of Work on LinkedIn
Drew: You're listening to The Safety of Work Podcast episode 112. Today, we're asking the question, how biased are incident investigators? Let's get started.
Hey, everybody. My name is Drew Rae. I'm here with Dave Provan. We're from the Safety Science Innovation Lab at Griffith University in Australia. Welcome to The Safety of Work Podcast.
In each episode, we ask an important question in relation to the safety of work or the work of safety, and we have a look at the evidence surrounding it. David, it's been a little while, but we've got some good papers stacked up starting with this one. Do you want to tell us something about it?
David: Drew, you said this one through to me. I don't know what it is, Drew. You might enlighten our listeners, but you seem to have some kind of affinity with incident investigation, generally. I know it was the Disaster Cast podcast that got me originally really excited by podcasts. It seems like you're literally excited when you see a new piece of research around incident investigation. Is that what's happening here?
Drew: David, I guess I've just always been fascinated by the epistemological status of accident investigations as data about safety. It's one of the most fundamental important things to our understanding of safety, but it's also really quite problematic and difficult to unpack how good evidence of what's gone wrong is useful for what we should be doing. I'm always interested when someone takes a little bit more sophisticated look at it.
David: A little bit of background on incident investigation as it relates to this paper. This information comes out of the literature review of the paper itself, and then we'll do as we always do, introduce the research and the conclusions. Like you said there, Drew, the incident investigation is a foundational tool of safety management and for a long time was the way that we developed our theories of how organizations can create safety.
Really, what we're about is we're about having a process, where we collect information about something that was unplanned, unexpected, or undesirable in the organization. We collect information about it after the fact. We then determine the contributing factors, we develop some corrective measures, and we have some mechanisms to communicate and implement change in our organization. That's, I guess, what our listeners would understand to be a fairly standard incident investigation process.
Drew, I guess there's a lot of commentary around incident investigations. We, probably five or six times on the podcast, talked about investigations in a certain way. But for all that history around incident investigation and centrality in safety management, I think there's some fairly commonly accepted challenges that organizations face when it comes to good quality investigations and good quality learning. Do you want to talk a little bit about your thoughts around, I guess, the state of incident investigation across industry broadly?
Drew: Sure. I guess there's a couple of points that are worth making. The first is that the idea of investigation quality can hide so many different interpretations of what quality means. Often, when people talk about wanting to improve the quality of investigations within their organization, they're talking about wanting to improve the standardization of reporting of incidents. Whereas I would personally see quality coming not from standardization, but quality coming from degree of surprise and usefulness of information that is gathered.
The second thing, which I guess is not going to bode well in an episode that we've titled how biased our instant investigators are, is that I don't think bias is an appropriate word at all in this space. All humans are subjective and have subjective experiences. Bias always implies that there is an objectively right answer.
When we're trying to describe something as complex as the causes of an accident, I don't think there are objectively right answers. But there are some answers which are fuller than others and more complete and more useful than other answers. Usually, when we talk about bias, we're saying they've got it wrong. I don’t quite think that’s the right way of looking at it.
David: I think for some biased insights, visceral reaction is a little bit like using the word error. I guess in this podcast, we'll talk about biases. I think it's also useful as a framing point for thinking about how we might achieve those two things that you mentioned that I quite liked there. What information surprised us and what information is useful through the investigation process?
I guess some of the other challenges we know that exist in organizations that are trying to improve the usefulness and the insights from their investigation processes are the culture of trust, the time and production constraints around performing investigations, and investigations being single loop learning processes with no feedback or double loop learnings.
Drew, you've got a better understanding of the incident investigation, literature, or body of literature than me. This paper claims that a lot of the studies into incident investigation that talks about a number of these limitations then go on to propose new analytical techniques and methodologies as a way of trying to say, we can get better outcomes if we've got a better analytical process.
I know a lot of people in industry asked me this very question, which is, what do you think is the best analytical method for incident investigation? These researchers are somewhat surprised and maybe suggested an important understudied area of incident investigation processes, the initial data collection phase of the investigation. What are your thoughts on the current literature body?
Drew: I think they're even, perhaps, overselling the quality of the current body of literature. There is very, very little empirical study of accident investigations. There's a huge amount of work that uses accident reports as data, or that makes commentary on poor quality of investigations or on distortions in investigations.
One of the challenges in interpreting accident reports is we don't really know from a scientific point of view how those reports are produced. Like many safety practices, most people have only got experience within one organization. They're trying to map their understanding of the very small number of investigations they've seen to all these other investigation reports.
In the case of academics, academics often are involved in hardly any investigations at all. All they've got other reports to go on and their own reinterpretation of those reports without really a good understanding of what goes on to produce those reports. It's an area that is very ripe for any research to contribute to our understanding of what actually goes on when investigators investigate.
David: It’s a great point there about future research opportunities. It's one of the papers that when we're preparing for this episode, I thought, gee, I'd like to do more research. I'd like to see the outcomes of more research in this area.
Really, these researchers are about what's happening at the front end of the investigation and particularly focused on the subjective and qualitative data collection process. Not so much about photos and training records and other forensic type evidence that might be gathered, but more about the process of interviewing people involved in the incident or witnesses to the incident. Drew, should we introduce the paper?
Drew: Yeah, let's talk about the paper itself. David, I'm going to get you to actually introduce the authors though because you know some of them and you're far better at pronouncing these names than I am.
David: Not so much. Five authors for this paper, Sreeja Thallapureddy, Fred Sherratt, Siddharth Bhandari, Matthew Hallowell, and Hayley Hansen. I don't know all of these authors. However, I do know Fred and Matthew's work reasonably well, at least their recent work in the last four or five years.
Fred Sherratt, she authored the paper, The Zero Paradox. We reviewed this in episode 12 when we talked about zero harm. For a long time, I had a look at the stats today because I wanted to validate. I was going to claim it was our most listened to episode and it was for a very long time.
The episode based on your research, episode 95, has about 20% more downloads than our second most listened to episode. Some of our listeners have been doing a lot of sharing on your Take-5 research. I don't know if you want to comment about that.
Drew: I love the fact that our two most downloaded episodes are on probably the two topics where we have the strongest negative takes of the stuff that we're talking about.
David: Let's just hope it’s part of an organization decluttering process. Fred, she was a professor based in the UK. I only just worked out in preparing for this episode that she's relocated to join the Construction Safety Research Alliance in the US, so congrats on that move, Fred.
Matt Hallowell, I think he's the head or the director of the Construction Safety Research Alliance at the University of Colorado in Boulder. He was the lead author of the paper, The statistical invalidity of recordable injury rates, which we reviewed in episode 55. Not as many downloads as I'd expect on that one, Drew, so maybe people aren't quite ready to drop their recordable injury reporting.
This is a very credible research center with very credible professors and researchers doing investigation in the construction industry, which is the industry that this alliance has been set up to research in. As we'll see, in the methodological design of this research, some very good research is coming out of that center.
Drew: Yeah, a lot of playing around with novel research methods, as well as very highly industry informed academic research, which I think is exactly what we wanted to be seeing in Safety Science.
David: The paper is titled Exploring Bias In Incident Investigations: An Empirical Examination Using Construction Case Studies. It was published in the Journal of Safety Research. That's one of the journals that we published in during my PhD, a US based journal. Quite credible, I think, and published online on the 31st of July 2023. It hasn't yet gone to print, but it's been available for the last two months or so. It's open access, so we can link to the paper in the show notes.
It is a really readable article. We're seeing this with very industry focused research centers that, at least in my experience, more recent experience, that academic journal articles seem to be getting a little bit easier and easier for people who are non academics to review or make sense of.
Drew: There's been a real push in some fields to rather than have paper abstracts, have industry ready and public ready summaries of papers as part of the publishing process. We haven't quite gotten that far in Safety Science yet, but I always do appreciate when papers are written to be read rather than written to be published.
I think the immediate next thing we've got, though, is a paragraph you've extracted from their method, which says, fundamentally, this research adopts a realist ontology and post positivistic epistemological position. That's written squarely for academics or possibly for the peer reviewers. I don't know if it was in the original version of the paper.
It's rather interesting in that they've sidestepped a lot of the complications around layers of reality when it comes to instant investigations by their method where they've created an artificial reality through a simulation. Post-positivistic basically just means that they're assuming that there is actually an objective truth out there, but that objective truth is not fully knowable and includes some social reality. You see throughout the papers they make like direct comments, like bias, which implies that there is a right answer.
In some places, they actually say, the interviewer was incorrect. That's the thing that you need to believe that there is a correct version of reality before you can make those claims. It's good to have the researchers recognize that philosophical position they're taking. I personally don't think it's quite the right approach for talking about accident investigation. That will come with a little bit into the interpretation of the findings. But it's nice and explicit that they're very clear about what they're doing while they're taking the position that they have.
David: I think I'm a little bit more of a social constructivist when it comes to incident investigations that we construct that reality. I think for the purpose of setting up the simulation, like you said, I think that that is written very much for the review as someone who's tried to publish qualitative research. Probably our qualitative research that we did publish in the Journal of Safety Research, they seem to accept that but not all reviewers, without making it sound a lot more empirical or quantitative by using those words.
Drew: Let's get into the actual method of what they did because this is really fun. This is a role play simulation. All of the research participants, so these are the people being studied, are industry practitioners with experience in incident investigation. They bring those along to the research, where they're faced with firstly a short video summary, which is meant to simulate the phone call you'd get telling you that an incident has happened and giving you basic facts. They then get the participants to write down their initial assumptions and thoughts about the incident. And then they go in to conduct interviews.
They are faced with two people that they can interview. One of the people is a direct victim of the accident, and the other person is a witness for the accident. In fact, these victims and witnesses are students who are acting, but they've been prepared both with scripts and with background information that they can improvise around to provide consistent stories to each participant. They've got a couple of features built into the method to ensure that consistency, even if the people ask questions that aren't expected.
The idea is that they're studying what questions these practitioners ask. How they conduct the interviews, what questions they start with, what they focus on, and then they're going to use that information from the interview to figure out whether the participants are displaying various types of bias. We'll talk a little bit about some of the biases that they were looking for and that they found when we get into the results. David, you've often been a bit of a fan of simulation research, but you seem to be a little bit less of a fan of this one?
David: I think I'm not so much of a fan. I think there's an opportunity to do real world ethnography here in working with a number of organizations in their incident investigation space. The researchers talk about maybe ethics and access might be a challenge, but I'm less convinced.
I think it would make a really helpful embedded ethnography to put researchers inside investigation teams for the purpose of observing and understanding how these investigations get carried out in real organizations. I think we've learned a lot and created a lot of opportunities for improvement. I think this research we're talking about today is a really great pilot study, but I don't think it's a big step to take what we're talking about today and do it in the real world.
Drew: I think it's really interesting as a pilot. I would almost have preferred them to go the other way around to do the ethnography first, and then come at this simulation with a little bit more of a specific concrete question to try to ask. The beauty of the simulation, and they display that in this study, is that you can directly compare a whole heap of research participants but keep everything else the same. You can tease apart.
The problem with real world investigations is each one only happens once. Two different investigators might be investigating different facts. Of course, they're going to come up with two different answers. But if you hold the facts constant, then you can look at the behavior of the investigators.
The sacrifice you make is you lose all of the contextual validity. As far as I know, most real world investigations don't involve a solo person being told, here are the two people you're going to interview. You have to interview each one once, and then the investigation is over. There's a lot more communication and interaction and choices being made in a real investigation that I think feeds directly into some of these questions of bias that we're missing out on, which is why we'd like a more focused question that we could take advantage of this ability to prepare this study gives us.
David: I think the institutional influences as well are really important. The social and the political influences within the broader organizational context where the investigations taking place get isolated. The investigator in the simulation gets an opportunity to perform an investigation, how they think it should be performed. In some ways, these results that we're going to talk about, some of the challenges that we learned through the simulation may even be far, far worse in the real world when you add in a lot of other challenges that might make some of these biases even show up even more.
Drew: I guess one of the great things about this as a pilot study is it gives a lot of ammunition to the standard objections. No one can say, oh, just one investigator did this, because they had 34 practitioners. No one can say it was just one company, because they came from a bunch of different companies, a bunch of different sectors, presumably different training, different experiences. No one can say, oh, it's just lack of training or lack of experience, because on average, these people had 20 years of incident investigation experience.
This isn't taking 34 students and getting them to run the interviews and then talk about how biased they are. These are real investigators who are used to doing real investigations. Presumably, in this simple scenario they're given, this should just be bread and butter, easy stuff to do really well. Any negative behaviors they show in the simulation, we can guarantee it's probably worse in the real world rather than better out in the real world.
David: A couple of comments on the method there just to land a few points, and we'll get on to the finding, you have your 34 experienced investigators, average of 20 years of investigation experience. It was a purpose sampling technique. The researchers intentionally went out to find experienced industry based investigators as a way of attempting to make the results more generalizable to the investigator population.
There were two incident stories. Those 34 participants were randomly assigned to one of the two case studies. The same students, like you said, Drew, were doing the acting in all of these. They had their script, they had their background information, so they could answer any question that the investigator threw at them. If the investigator threw a question, and they had to make up an answer on the spot, then after that interview, that answer was recorded. If the same question came up another time with a subsequent participant, then the same answer was given.
They continue to build the script out through the course to make sure that the same answers were being provided in the same way to the same questions. It sounds like they're able to go back and forward. If they interviewed the witness, they could then go back and interview the injured person again. If they learned something else, they could then go back and interview the witness. It looked like they had the opportunity to go and do that if they wanted to.
There were two scenarios. One was called the staircase incident, and one was called the concrete form incident. It's probably not worth going through these in total detail, Drew, but one involved a staircase collapsing in a residential home and one involved a laborer on a construction project who was basically slicing cardboard, stripping around a light pole using a utility knife, and they injured their hand, I think.
Drew: Actually, the supervisor stopped them using the knife and got them to use a hammer instead and accidently hit the supervisor with the hammer.
David: Got it. You got these two different scenarios going randomly assigned. Like Drew said at the start, the participants got a two minute video summary, and then they got an opportunity to basically just ask questions of the two participants. At the end, they were debriefed by the lead researcher about basically what they think caused the incident, or what were the contributing factors to the incident.
Drew: In terms of analysis, basically, they were specifically looking for known biases. They didn't come into this just asking these random questions about what's happening. They're looking for biases, but then they're looking for specific examples, patterns, and details of those broad categories of bias.
It shouldn't surprise you that they found bias. That was exactly what they were looking for. But what's interesting is the number of people who displayed these biases and how those biases manifested. David, should we go through the broad things that they've found?
David: I'm just going to read the method there. What I actually quite like is they just talked about the limitations of the research and what they tried to do in the design of the research to limit the negative impacts of those limitations. I think that it's always good when the research team reflects on obviously what the research can and can't tell us due to the design. Let's talk about the themes. There are six or seven themes out of the research, so do you want to get us started on that?
Drew: Okay. The first one is the idea of what you look for is what you find or the technical name for that is confirmation bias. To measure this, they asked each participant before they did the interviews, what contributing factors they thought were likely to be in the accident. What they noticed is a lot of the interviewers asked questions that were trying to confirm their initial guess at what the factors were.
For example, if they thought the problem was that the formwork wasn't stable, then they would ask questions like, was the formwork stable? They're looking for confirmation of what they suspected was there, rather than broadening out and looking for new information about causes that they weren't already suspecting.
What actually surprised me was how few interviews this applied to. They said 30% of interviews were structured around those initial assumptions. I'm actually surprised that 70% said, oh, here are my initial assumptions, but then we're able to completely put those aside and ask more open questions in the interviews. We'd say like, you wouldn't say it was a rare mistake, but it certainly wasn't the majority of people who's shown that confirmation bias.
David: How I interpreted that is that 30% of the interviewers basically structured their interview and in some cases, their whole interview around their assumptions. I suspect in other cases, maybe it showed up from time to time, but maybe they ask more open questions at the start, like, can you give me a summary of what happened in your opinion? Maybe they didn't structure their interview around their assumptions.
One of the interesting things that I also saw here was that most participants, and they didn't tell us how many, did not ask any questions to challenge their personal notions of what may have happened. This idea that we seek out disconfirming evidence, or we actually try to find things that invalidate the way that we're currently thinking about the problem. Most investigators didn't do that, and I think that was the telling point in this part of the data for me that I don't think it's natural for incident investigators to go and actually try to disprove their own assumptions.
Drew: David, to be honest, that's one where I'm a little bit skeptical. I don't fully trust the researchers enough here. I want to see the raw data, because I know that when I do research interviews myself, if I'm trying to find disconfirming evidence, I come at it sideways. It won't be explicit in my question that I'm looking for disconfirming evidence. I'm asking a question that opens up a space that if that disconfirming evidence is there, then the person I'm interviewing will be likely to say it.
That's one way you really want to see the original data from the interviews to know how they were interpreting and coding some of these things that the investigators were doing. I absolutely believe the answer. I absolutely believe them that that is what was going on, but I would like to actually see the evidence myself for that one. I think it's actually really quite subtle how you go about looking for disconfirming evidence in a real interview.
David: I think for each of these six themes or so we're going to talk about, they probably deserve a deeper analysis for exactly how this theme shows up and in the ways that you said, because you can now take this confirmation bias theme and go back into the at least 68 transcripts from 34 participants and specifically try to go really deep on just confirmation bias and maybe answer those questions. Drew, if you really want me to reach out and see if I can get your copies of the transcripts, I'm sure maybe we could arrange it in your spare time.
Drew: In a simulation, what I'd really like to do there would be ethnomethodological analysis, which is have them take me back through the tape recording of their interview and explain to me while we watched the tape together, what they were trying to do at each stage. I think that would be very interesting. That's beyond this pilot study that they've done. That's not a criticism of what they did. That will be the next step.
This one is non controversial. They said, 10% of the participants asked leading questions. That's not a question of interpretation. It's really easy to tell an interview transcript what's a leading question. It's real life training no no in interviews.
That one both doesn't surprise me, but it's also like a big red flag for confirmation bias that so many people were asking leading questions. That's what police do when they want to get evidence of you confessing. It's not what investigators do when they want to find out what happened and disconfirm their previous assumptions.
David: During the staircase case study, there was a very explicit production pressure and resource constraint, where two of the workers have been fired for safety breaches, one of the workers was on their own, and the client was coming to inspect the completion of the work that afternoon. Questions from the investigators are like, so you are rushing to complete the stairs, do you feel that that pressure affected you? The only scripted answer there is to just say yes.
Drew, theme number two. I really liked the way that they labeled these biases. That's why they did that, this fundamental attribution error. The headline finding is that 2/3 or 21 out of the 34 interviewers stated in their post interview that they felt that the personal characteristics of the people involved and their resulting behaviors contributed to the incident. The interviewers relied on character to make judgments about what led to the incident occurring. We are looking at, what was it about the person that made them do that?
Drew: Which says a lot about the quality of the acting of the students in this simulation. These hardened industrial practitioners went away making character judgments against the characters who were being played by the students. This one's interesting, and this comes back to the post positivist approach. David, I would guess that neither you nor I would be very comfortable with incident investigations, which blame the individuals for their personal characteristics and their behaviors. That's not a question of absolute truth.
That's the old structure versus agency debate, where both structure and agency are real end matter. It's just a question of which you focus on. I don't think it's absolutely a systematic, cognitive tendency. But calling it an error, I think, is a little bit severe here. Fair to say, yes, most of them did focus on personal characteristics rather than structural things.
David: I guess we see this in investigations in the industry, where in this case, the researchers saw investigators attributing behavior to personality rather than considering the situational context of the work and their surroundings. One example you might think is the person putting their hand in there, because they're an impulsive or arrogant person, as opposed to the person putting their hand in there just like anyone else would, because of the explicit time pressure to perform the task.
I think that the red flag, if you like, here was the fact that it was 2/3. When asked at the debrief by the researchers, what do you think contributed to the accidents, 2/3 of people started to list judgments about the characters involved in the event.
Drew: I guess, particularly because these two examples that they've picked out are filled with systemic and structural factors that could have been focused on instead. They're not stories, where there was no explanation available other than someone misbehaved.
David: Theme number three, nothing surprises me anymore, this is this past experience bias. You mentioned at the start of the episode about investigation quality being a little bit about, what did we find out that surprised us? Do you want to talk a little bit about this theme?
Drew: The idea here is something that has been observed with incident investigations before, which is the investigators are interpreting the incidents in light of their previous similar experience to the point where they're almost guaranteed not to find anything new. They're saying, okay, I've seen this happen before. When it happened before, this is what I thought the cause was. Therefore, that must be the cause this time as well. The investigation reproduces their prior assumptions from prior investigations.
There was pretty clear evidence of this happening, particularly with the debriefing of the people after the exercise. Also, some hints from actual questions that people asked in the interviews, it's almost like the investigator is trying to explain to the people they're interviewing what the cause is rather than finding out from the people who are interviewing what the cause is. Again, not surprising, but worrying when you think of what makes a good investigation.
David: In that language, starting a question in the interview with the statement, I've seen this before, here's how it can happen, can you tell me how you think it happened. I think one of the challenges here is maybe the more experienced you get at incident investigations, the more incidents you see, maybe the stronger that this gets. I have heard people in the industry say this to me, that there's no new problems in safety, we've seen them all before. I think that's a really interesting one for us to think about how we separate ourselves. How do we look at every incident with a fresh set of eyes is a real challenge.
Drew: Theme four, they called getting stuck somewhere, which they've linked to the idea of anchoring bias. There are actually a couple of different forms of anchoring bias. The one they're talking about here is akin more to tunnel vision, which is where you find one thing to focus on, and then you just stay stuck focusing on that one thing without being able to set it aside and move on to something else.
David: Yeah, Drew. I guess a statement in here that there's an infinite number of things to consider when we're thinking about what may have contributed to an incident in a complex work situation, even such as these staircase falling down, or someone getting hit with a hammer, so many human process, technical aspects of the work, all that necessitate sense making processes around. At this point in the research, 18 of the 34, so more than half fixated on just one aspect of the incident. They're in the researchers' opinion to the detriment of the process as a whole.
I found one thing. There's structural instability of the stairs. That's clearly the most important contributing factor, and that's all we're going to talk about during the interview.
Drew: This is another one, where I would really like to see the actual recordings of the interviews, because the examples they give of anchoring bias seem equally explainable by lack of interview skill, where the person giving the interview is stuck at a point in the interview, doesn't know where to go from there, and doesn't have the interview tools to step back from their current line of questioning to move the conversation on to somewhere else, which is a real social skill that a lot of untrained interviewers lack, because it's quite an artificial thing to do.
It may be that outside the artificial constraints of the simulation, they're better at that, although what we see from investigation reports absolutely confirms this idea of anchoring bias that entire organizations get anchored on single explanations across multiple investigations. Even though there are some suspicions there about exactly how they got this conclusion, I don't doubt the conclusion that it’s a real thing, it’s happening.
David: I think around this confirmation bias and anchoring bias themes is where I think in real organizations with real investigations, there are a lot more organizational aspects to anchor around. I'd love to know the extent to which the known problems in the organization or the known strategic improvement areas in the organization then show up in investigation reports after they've been determined to be a really important strategy for the organization as a reinforcement of the weaknesses in the organization, the validation of the strategy, and so on. I think there'd be some really interesting ways that these bias show up not just in the investigators' mindset, but in the organization's safety narrative at that point in time.
Drew: Yup. Theme five is it was preventable. This is the idea of hindsight bias. They say this showed up in nine of the participants out of 34, which is not a lot. But remember, we're just talking about the interviews here, we're not talking about the reports. In a real investigation, there will be many more opportunities to display hindsight bias than they had here.
This is nine people just from the interviews and the conversation with the researcher afterwards, are already establishing these counterfactuals and talking about what they would have done and how they would have prevented the accident. Presumably, on their mind as they're interviewing, is this alternative picture of reality that is shaping the way they're looking at what happened?
David: This idea that this incident is preventable, the investigator forms a view about how it could be prevented, I guess it leads them to not fully explore the situation around the incident to just direct conversations towards the solutions that they've got for the problem. I guess that for every complex problem, there's a simple solution that's wrong. I think that's the point that the researchers are making here, that when we form the view that an incident is preventable, then we stop looking for broader influences on the event.
Drew: It's really interesting that we can see this at a really early stage of the investigation. You can imagine this final report, which says, okay, this accident was caused by failing to completely shut off the staircase. I recommend that we put in place procedures that staircases be shut off with signs across them saying that they're not usable.
That looks very naturally like it comes out of the investigation and out of what happened. This is revealing that the investigator, at the very start, that's what they would have done. They've reproduced through the interview in the investigation, their own way of doing things that they knew about before the accident even happened.
David: Okay. Theme six, taking sides and sticking to them. This is conservatism in belief revision. I guess this is really the ability of the person to change their mind during any investigation process and also to accept that there are different truths with different people in the investigation.
People have their stories. There were two people interviewed, a witness and an injured person, for each of these scenarios. They might or might not have had different views about certain aspects. Essentially, we might expect the role of the investigator to remain objective to those involved.
What we did see in 23 of the 34 is that the investigator took sides. Where there was a discrepancy between something the injured person and the witness was saying, the investigator would take aside and favor either one of those two interviewees and therefore, just basically made conclusions and judgments based on which side they wanted to take.
Drew: This is even more extreme than it sounds. I should mention, they suddenly seated contradictions into these scenarios. The two witnesses have slightly different versions of facts. It's not that the investigators were saying, this person said this, and that other person said this, and I believe this person, it's that they were talking about one person's version as if it was the fact.
For example, in one case, one of the witnesses said that they had told something to the other person, and the other person said they had not been told. The person was just talking as if it was obvious that the person had been told. They just fully taken on board one version of reality. It's interesting that it was fairly consistently the version of reality that they encountered first. Once they've formed that opinion, contradictory evidence was just seen as not real.
David: I think we go back 25 years, or maybe even further, you can tell me. Richard Cook and Dave Woods wrote a paper titled A Tale of Two Stories. We started introducing this idea of multiple truths in incident investigations. I think it was the early 2000s when that paper was drafted, if not late 1990s.
The last conclusion before we might talk about some practical takeaways was that, they titled it theme seven, which was bigger than the sum of the parts, which was about biases working together. They referred to some work in 2006 around a theory of combined cognitive bias hypotheses. Basically, this is that some of these biases can combine and work together in ways that actually makes the total effect greater than any of them individually.
If you become anchored early on in the conversation, then you go and seek out information that confirms confirmation bias around that, and then you've got this conservatism in belief revision, and a tendency to favor the person who aligns with your perspective on the situation. You can see how these biases can work together to, I guess, really just drive an outcome of the investigation towards the investigators' initial position.
Drew: Yup. David, before we move directly on to takeaways, I want to mention a little bit about the relationship between this research and the common solutions that organizations try to have to improve their investigations. Very often, people are looking for investigation techniques, frameworks, or models. That's often what academics love providing as well, new extant models, human error classification models, or human factors models, that give us a bunch of factors.
What this research point intends very, very clearly is, there's almost no opportunity for that model to fix the fundamental problems in the process. If the model gets applied after these initial interviews, it's way too late, because your data has already been heavily, heavily shaped by the cognitive processes, which includes what questions were asked, so literally, what data was collected to go into the model.
If you have your model before you go into the interview, this just shows how people already latch on to particular parts of that model. It then guides all your investigation again. You might get something that is a little bit more standardized in its reporting if you start with a model, that it's absolutely not going to fix any of these cognitive problems. You're populating the model with that data.
David: Yeah, it's a great point. In talking with people in the industry around this topic, incident investigation and incident investigation quality, 80% of the conversation is around that causal classification taxonomy. Whether it's ICAM, essential factors, TapRooT, [ICMAP 00:40:44], STAMP, or whatever it is, we feel like methodology is going to solve these problems.
The researchers in the introduction to this paper talk about the processes garbage in, garbage out. If we actually don't understand how to get a good data collection process, then it really doesn't matter what happens after that, because the outcomes to the investigation have already been shaped. If we have this attribution error to the individual based on personal characteristics, then of course, if we apply [ICMAP 00:41:14] for the framework, we're going to get fixated on the people factors and so on.
I think it's a really good point. Do we know the problem we're trying to fix to improve our quality of investigations, which takes us nicely into practical ideas?
Drew: Takeaway number one, go for it.
David: I thought one of the initial ones is we talk a lot when we talk about risk assessment, I think in one of our episodes on risk assessment, you were quite strong that we need to document all of our assumptions in a risk assessment, what frame the decisions that we made through the risk assessment process.
I think it's really important before the commencement of data collection that the investigator or the investigation team does the process that was done in this research, which actually writes down what their initial assumptions and causal ideas are for the investigation and really documents those. It gives us the opportunity, (1) for the investigation team to be mindful of their preconceived ideas through the process, and (2) also it gives us the opportunity at the end of the investigation to do what the researchers did here, which is look at what the investigators went in with and if they come out with exactly the same thing. I really liked that.
Drew: It's a common misunderstanding to believe that the way to avoid bias is to try to be unbiased. I bet we've got listeners who are saying, oh, but when I do investigations, I go in with an open mind. We've got lots and lots of evidence that that doesn't work. This is one of the things that gets hammered in the social sciences. Why should all accident investigators have arts degrees, not engineering degrees? It's the idea that you avoid bias by being open, transparent, and understanding of your own position.
No one is free of subjectivity, so you've got to be reflexive. You've got to know where you're coming from, and you've got to work with that. You're claiming that you've got an open book, you've got an open mind, that's just deceiving yourself.
Understanding, oh, my initial assumption here is that the person probably stuffed up. Okay, I can recognize that, I can write it down. I can realize, okay, if I come out of this investigation, just believing the person stuffed up, I've got moved nowhere. That's something I've got to be conscious of my own beliefs. I've got to test that, challenge it, and look for other explanations. You can only do that if you want to stop fronting about what your preconceptions already are.
David: I like that, Drew. Thanks for explaining that. You had made during my PhD, because I was a safety professional and quite an experienced safety professional researching the role of safety professionals in organization, you have me right, which was published in my final thesis, a reflective piece about my experience as a safety professional, what I believed about the profession, where I saw the value of the profession, what I saw as the challenges, and really all of the things that I was researching.
But before I even did any of the research, I wrote pages and pages, which is actually quite cathartic really, of what I initially felt so that we could check in whether or not I was actually looking at the data or just looking at what I already thought.
Drew: The second one you've written down here, David, which I agree with absolutely is to focus on local rationality. To go into the investigation, with the presumption that it made sense for everyone to do what they did, no one was insane, no one was stupid, no one was malicious. Our job is to find out why it made sense to them.
Even if that is factually wrong, even if there are stupid people out there, even if there are malicious people out there, even if there are crazy people out there, that's not actually helpful in an investigation to reach those conclusions, because it doesn't lead to any organizational action. Methodologically, let's just assume it made sense to people, because that's what's going to help the next person in that same situation.
David: You said to me when I've asked you previously about how you would think about the quality of investigation, you said to me where the investigator can advocate that it made perfect sense for everyone to do everything exactly as they did it. I'm reminded here of Diane Vaughan's book, The Challenger Launch Decision, which was based on an almost decade long historical ethnography into the challenger launch decision.
In that, she makes the point that there was a lot going on in there, but everyone acted in ways that made exact sense to them. Anyone else in the position of all of the people in that room at the time would have made decisions in exactly the same way based on the situation. This idea here is if someone else in that situation can do the same thing, then there's no point trying to fix the person.
Drew: Obviously, there's a limitation to that approach, which is, you do it to the extent that Vaughan did it, or that I do it. You come to believe that the entire universe is inevitable and unchanging, that no one could have prevented the accident, because everything everyone did made sense in context.
I think the reality is safety people, particularly, have got such an instinct to change, fix things, and to problematize things. You don't really need to worry about accidentally not finding problems in fixable things. The thing to worry about is at the other extreme. We're pushing you to one extreme, but we're pushing you probably against where most people’s tendencies lie.
David: Drew, do you want to go into the third one?
Drew: Okay. The third one I've got here is, and this is the point they make throughout the paper, a lot of these problems don't require some fundamental rewiring of the brain to stop investigators being biased, or to stop them having cognitive distortions. A lot of them come down to interview techniques that are conscious of the way people think and conscious of getting around some of those problems.
I think we do a lot to train people in safety and safety models, but we don't do a lot of actual training in how to do interviews. When people do have training, they've often come in from environments like the police, where it's a totally different type of interview with different objectives. The skills are actually quite counterintuitive.
One example there is, how do you avoid confirmation bias? You structure your interviews, so you start with open questions leading gradually towards closed clarifying questions. That's a skill that you can learn. You can look in an interview, see whether someone has done that, give them advice for next time, and say, look, the reason you got this answer is because you asked it as a closed question upfront. If you'd asked a more open question, you wouldn't have been locked in.
Another basic one is just don't ask explicit leading questions. That's an interview skill thing. It takes training, it takes practice, it takes good understanding of exactly what is and isn't a leading question. But it's trainable.
Another one that gets around a lot of the things here is just not to ask questions. Really good interviewers don't ask nearly as many questions as you might think they do. That is a hard skill, but it is a learnable and trainable skill, to train people in not asking questions. The less you ask questions, the less you have the opportunity to accidentally shape the interview through any of these biases, because you're not leading the interview. The person you're interviewing is leading with their experiences, which is what you're trying to capture.
David: I then came up with a list of six questions for these six different themes. I haven't run this by you, so good to get a live check on what you think. I guess I just thought that, if you're looking at instant investigation, you get an opportunity to have a conversation with the investigator. There might be six questions that you could ask them to think about how much their own biases impacted the investigation outcome.
The first is a question like, what did you learn that you didn't know before the investigation? What did you learn? Did you just confirm everything you already knew, or did you learn something new?
Drew: That is the first law of accident investigation, David, that you've reproduced there, which is, in order for the organization to learn from an investigation, the investigator themselves must increase their knowledge. If the investigator hasn’t learnt something, the investigation can’t learn something.
David: Hopefully the investigator can surprise you with all these amazing things that they've learned about the organization. The second question is simply, why did it make sense for the person to do what they did? Can the investigator advocate for the decisions and the actions of the people involved from that attribution error bias?
The third one, you mentioned at the very start, I had in here before you mentioned that, just for our listeners, the third was, what surprised you during the investigation? This is the extent to which the investigator was genuinely surprised. I know you like that one, Drew, so I'll move on.
Number four, what other possible things could have contributed to the incident? If we're anchoring around the things that we think did asking the investigator, say that turned out to be wrong, what other things did you dismiss during the course of the incident investigation and see how long their list is of other things that they considered? Thoughts?
Drew: I love that question. That's just a deliberate invitation to broaden out your explanations. I'd even love that as an interview question for accidents. What else might we have contributed that you haven’t told me.
David: The fifth one about not having simple solutions to complex problems is asking the investigator, what would make avoiding this incident in the future extremely difficult? This idea that if we just sign the stairs, it'd be fine. It actually have them go, no, no, what would make avoiding this genuinely, really difficult?
I think the last thing is really just to ask that question about the multiple stories, asking investigators, what were the differences and the discrepancies in the opinions? How did you make sense through that process? Like you said, it is possible to train people in doing good data collection. I think this research is really interesting and useful framing for real organizations. I think you can strengthen your process without throwing out all of your existing methodology.
Drew: I really liked those six questions you've got there. That was really great just for a feedback cycle around any investigation process. Let people do what they're currently doing. Ask them that question. Where they struggle, they will naturally be prompted next time to come up with a better answer for you, because I know you're going to ask that question again next time. Just building that into the feedback cycle, I think, would improve the organization's investigations.
David: Drew, anything you want to add before we wrap up if you miss a topic of yours?
Drew: No, I'm pretty comfortable. You've in fact handed it to me. I get to say, the question we asked this week was, how biased are incident investigators?
David: The answers I guess is very, and we all are as human beings. It does mean that we should probably worry more about the data collection phase of our investigations more than the causal analysis methodology and taxonomy that we concern ourselves with.
Drew: That's it for this week. We hope you found this episode thought provoking and ultimately useful in shaping the safety of work in your own organization. We normally say, send any comments, questions or ideas for future episodes to feedback@safetyofwork.com. But David, you have a new platform that we can talk and discuss podcast episodes on. Do you want to say a quick word?
David: Yeah, we do. Safety Exchange was launched last week or two weeks ago at the time that this episode will go out. You can go to safety.exchange, or you can download the Safety Exchange app. What we've tried to put together is a space for the global safety community to engage and collaborate.
We know that LinkedIn and other social media platforms are becoming increasingly less psychologically safe for the community to learn from each other. It's a mass social experiment, Drew. But yes, there is a space for The Safety of Work Podcast, and each episode will be going up there. Drew and I are going to host a Safety Science space on the platform. I'd love you to come across and join us there.