Method

Participants were 101 adults who had reported active suicidal thoughts in the previous week. Participants were recruited with online advertisements on Reddit and Craigslist. The history of suicidal thoughts and behaviours in the sample was 64% lifetime suicide attempt and 38% past year suicide attempt; the median number of lifetime days with active suicidal thoughts was 1825 (range 30–8000 days).

Each participant downloaded a smartphone-based survey app that sent them three types of survey over a 6-week period: daily surveys (1 time per day), momentary surveys (5 times per day) and burst surveys (6 times/h, 2 per day, 4 days/week). Four days a week, participants received 18 surveys per day (1 daily, 5 momentary and 12 burst surveys) and three days a week, participants received 6 surveys per day (1 daily and 5 momentary surveys). Each survey assessed the desire, urge and intent to kill oneself on a 0 (not at all) to 10 (very much) scale, and scores were combined to create a suicidal thinking severity score. At the end of every momentary and daily survey, participants were provided with a list of hotlines and resources.

The authors assert that all procedures contributing to this work comply with the ethical standards of the relevant national and institutional committees on human experimentation and with the Helsinki Declaration of 1975, as revised in 2008. All study procedures were approved by the Harvard University-Area Institutional Review Board (IRB no. 19-1819; ‘High-resolution real-time capture of suicidal thoughts and urges’). Written informed consent was obtained from all participants.

For question (a), we ran a three-level ordinal model in which responses were nested within participants and burst episodes (total number of surveys n = 10 479 over 2890 episodes). The predictor was time (i.e. minutes between the first completed burst survey and the last burst survey). If frequent assessments were iatrogenic, we would expect to see a positive association between the severity of suicidal thinking over time during a burst episode. To address the possibility that those with increased suicidal thoughts stopped completing surveys, we ran a multilevel survival analysis on missing data. The outcome was time (i.e. minutes between the first and the last completed burst surveys) with a binary variable (suicidal thinking increased versus no change/decreased) as the predictor. If assessments were iatrogenic, participants with increased suicidal thinking would stop assessments earlier. The survival analysis was estimated using the ‘coxme’ package.^{Reference Therneau8}

For question (b), we ran a multilevel ordinal model in which the outcome was daily reports of the severity of suicidal thinking (total number of surveys n = 2314) and the predictor was a day with burst surveys (maximum n = 18 surveys sent) or without burst surveys (maximum n = 6 surveys sent). As a sensitivity analysis, we examined whether the effect of number of surveys per day differed by suicide attempt history (present versus absent). If high-risk participants were more affected by the assessments, we would expect there to be a significant interaction between number of surveys per day and suicide attempt history.

All multilevel models were estimated within a Bayesian framework with the ‘brms’ package.^{Reference Bürkner9} Results were interpreted by summarising the 95% highest density interval (HDI) around the median beta values. If the 95% HDI includes zero, it is indicative of a null effect.