The TAP Lab has always placed emphasis on laboratory studies, longitudinal work, and novel measures to advance knowledge about self-injurious thoughts and behaviors. Early on (2010-2014), this took the form of experiments with painful stimuli and psychophysiological measures, and longitudinal studies with implicit measures. You can access our papers on this work by clicking here. More recently, this has taken the form of virtual reality experiements and a range of machine learning studies. We briefly discuss some of our previous and ongoing work with these methods below.
Virtual reality experiments
Why do we think that virtual reality experiments may represent a game-changer in suicide research? Well, first, it is important to understand why experimental designs can be valuable. As noted in our Summarizing Knowledge Section, something called the 'counterfactual dependence test of causation' is the best way to infer a causal relationship. Although most researchers are unfamiliar with this metaphysical position on cause, much of science is built on this philosophical foundation. In essence, this view of causation proposes that we can infer cause when:
X -----> Y (i.e., the presence of X leads to Y; aka 'the fact')
No-X -----> No-Y (i.e., the absence of X does not lead to Y; aka 'the counterfactual')
Technically, a true counterfactual is what would have happened if X did not exist or occur. But we can never really know this because we only live in one universe and the arrow of time only points forward. So how can we estimate what might have happened if X never existed or occurred? The answer is some kind of control group (between-participants design) or control condition (within-participant A-B-A-B design). This is obviously imperfect because two people/groups are never exactly alike, and because time/experience mean that a given person is not exactly the same from one moment to the next. But the closer a control group/condition is to a true counterfactual, the stronger the causal inferences that can be made. This is why larger groups, randomized groups, active control manipulations (e.g., a control drug with similar side effects to the experimental drug, vs. an inert control drug), and similar design features all allow for stronger causal inferences. In traditional psychological research design terms, the closer the control group/condition is to a counterfactual, the stronger the internal validity. The ability of experimental designs to permit causal inferences is why these kinds of designs are a staple of science.
But experimental studies seeking to understand the causes of suicide have not been a major part of the suicide research tradition. Meaningful experiments require meaningful outcomes. For example, meaningful experiments on anger include actual anger as their outcome. But obviously the nature of suicide precludes using actual suicidality as an outcome in experiments that seek to understand what causes suicide. Moreover, given the extreme nature of suicidality (e.g., high places, dangerous weapons), it is difficult to approximate suicidal situations and behaviors in the lab. As a result, it traditionally has been very difficult to experimentally investigate suicidality. The field has accordingly focused on cross-sectional and longitudinal studies, which unfortunately cannot provide much useful information about cause (i.e., these designs do not attempt to approximate counterfactuals). This has made it difficult to evaluate theories (which are primarily proposals about what causes suicidality) and to develop effective interventions (which must target the causes of suicidality).
And this is why we are particularly excited about virtual reality experiments. They have the potential to safely-but-meaningfully approximate suicidal situations and behaviors in the laboratory. Virtual reality can realistically simulate a wide range of potential suicidal situations and provide environmental affordances for a wide range of suicidal behaviors. These behaviors can be used as outcome measures in experiments that seek to test ideas about the causes of suicidal behavior. Virtual suicide is obviously not isomorphic with actual suicide (thankfully), but we believe that it is close enough to provide meaningful (though tentative) information about suicide causes. As technology continues to progress and realism is intensified, virtual suicidal behavior will become an even closer approximation of actual suicidal behavior, further improving the meaningfulness of these kinds of experiments. For the past two years we have been developing and testing a viable approach to virtual reality experiments on suicide. Below, we briefly describe this work:
Initial feasibility and validity studies. As we started our work in this area, we realized that there were many reasonable approaches to experiments on virtual reality suicide. It was not clear which potential approaches were best, so we conducted a lot of initial pilot work that ultimately led us into a fruitful direction. We then conducted three studies (total N = 498) to further investigate several basic aspects of this new approach. These studies produced four important findings:
1. This method is safe and feasible. No participants reported enduring distress, experienced nausea, or indicated suicide risk. We also did not encounter any notable technical problems. This approach also seems to lend itself to widespread use and replicability. A state-of-the-art virtual reality system (including computers) only costs around $2,000, and used virtual reality scenarios that anyone can use for less than $20.
2. The VR suicide rate approximates the rate of actual suicidal behavior. Under neutral conditions, participants engage in virtual reality suicide on approximately 5% of their opportunities to do so. This is what one might expect from a reasonable approximation of actual suicidal situations. Bolstering this, participants rated these scenarios as strongly suicide-relevant and realistic. The modal participant reported that they would never engage in virtual suicidal behaviors under any circumstances.
3. The predictors of virtual reality suicide are similar to the predictors of actual suicidal behavior. This includes things like suicidal desire, suicidal capability, agitation, identifying as male, and a history of suicidal behaviors. Interestingly, the reasons that participants gave for not engaging in virtual suicide (e.g., identity, fear, moral/religious reasons) resemble the reasons that people give for not engaging in actual suicide.
4. Virtual reality suicide is sensitive to experimental manipulations. As a proof of concept, we tested whether minor rewards or punishments would increase the virtual reality suicide rate. Both of these manipulations raised the rate from 5% to 25%. This paves the way for future studies that test more theoretically-relevant manipulations' effects on the virtual suicide rate.
Changing conceptualizations to cause (and prevent) virtual reality suicide. We are currently working on a large study using this method to evaluate our central hypothesis about suicide: when someone conceptualizes suicide as the most sensible thing to do in a given situation, they will do it; if they don't conceptualize suicide as the most sensible thing to do, they will not do it. This study uses a wide range of experimental manipulations to systematically alter participants' conceptualizations, leading some participants to conceptualize virtual reality suicide as highly sensible and others to conceptualize it as extremely non-sensible.
Machine Learning Work
Machine learning is a method that fits well with our view of suicidality predictors/causes as complex and indeterminate (see our Summarizing Knowledge Section for more information on this view). This has clear implications for prediction and, indeed, machine learning greatly improves predictive accuracy compared to traditional methods. Led by our collaborator Dr. Jessica Ribeiro at FSU, we have several previous and ongoing projects aimed at using machine learning methods to better predict and understand suicide risk. See here for Dr. Ribeiro's work.
Because machine learning is a general computational approach (i.e., it is not specific to prediction), we are also applying it to address other types of questions about the nature of suicidality. Below, we note a few of our ongoing projects in this area:
Is the difference between suicide ideators and attempters complex and indeterminate (or simple and determinate)? Several recent theories have proposed that particular factors or small sets of factors should distinguish ideators from attempters. But none of these specific factors seems to distinguish between ideators and attempters much better than random chance. As an alternative consistent with our Protean Paradigm, we hypothesize that the difference between ideators and attempters is complex and indeterminate. In other words, to distinguish between ideators and attempters, it will require the consideration of a relatively large number of factors combined in a complicated way, and there is no single complicated algorithm or recipe for this distinction. We are using both unsupervised and supervised machine learning to investigate this possibility. Across five large data sets, preliminary results strongly support the complex and indeterminate hypothesis.
Are there meaningful differences between suicidal civilians and suicidal military service members? Inspired by the temporary rise in the military suicide rate (compared to the civilian rate) approximately 10 years ago, many have proposed that there may be important differences between suicidal civilians and suicidal military service members. Unfortunately, none of the proposed distinguishing factors seemed to differentiate between these two groups much better than random chance. Another possibility is that there are differences between these groups, but they are complex and indeterminate. We are currently using supervised machine learning methods to investigate this possibility.
Are suicidality phenomena (and related phenomena) natural kinds or non-essentialized categories with blurry boundaries? Traditionally, many researchers have assumed that suicidality phenomena are natural kinds (i.e., each phenomenon has an essence and firm boundaries between it and all other phenomena). Some hold a strong version of this view, proposing many very specific subtypes of suicidality and essences for each; others hold a weaker version of this view, proposing a few strong subtypes (e.g., ideation vs. attempt) with murkier subtypes in between. Based on our Protean Paradigm, we hypothesize that all subtypes of suicidality (including ideation vs. attempt) are non-essentialized. In other words, none of these categories has a distinguishing feature that separates it from all others, and each has great overlap with all others. We hypothesize that this overlap extends to non-suicidal phenomena as well, including nonsuicidal self-injury, drug use, and eating disorder symptoms. We believe that traditional statistical methods and unsupervised machine learning will do a poor job of distinguishing among these categories, and that supervised machine learning will do a much better (though still far from perfect) job of distinguishing among these categories.