EMPIRICAL TESTING OF THE REGRET-BASED THRESHOLD MODEL IN END OF LIFE CARE
Purpose: The threshold model represents one of the most important advances in medical decision-making but it has never been empirically tested in real-life setting. We aimed to empirically test the regret-based threshold model in end of life care where patients face choices between hospice and curative treatment.
Method: According to the regret-based threshold model there must be some probability of death (pDeath) at which patients should be indifferent (Pt) between hospice care (Hospice) and continuing treatment targeted at their disease (Rx). The model predicts that if pDeath>Pt, patients should choose hospice; if pDeath<Pt, they should opt for Rx. We tested these predictions by interviewing 134 terminally ill patients facing Rx vs. Hospice decisions. We determined Pt by eliciting regret of omission (i.e. losing benefits of hospice care) and regret of commission (i.e. incurring harms from unnecessary treatment) using a dual visual analogue scale1. We estimated pDeath over 6-months using the Palliative Performance Scale (PPS) and adjusted PPS prognostic models. We compared the regret-based threshold model recommendation to the patients' choice at two different time frames: immediately after the interview and one month after the interview to study the patients' preferences and actual choice of care. We used Cramer V (effect size) to calculate the strength of agreement between the model recommendations and the patients' preferences and actual choice, respectively.
Result: We observed statistically significant agreement between the model recommendations and the patients' stated preferences (p<0.0001). Out of 134 patients 111 (83%) agreed with the model recommendations immediately after the interview, 6 patients (4%) disagreed, and 17 (13%) were unsure about their preferences (figure). This converts into very large effect size (0.84). 111/134 patients were approached one month after the interview to determine what type of care the patients actually chose: 59 (53%) chose according to the model recommendations; 39 (35%) chose a different option than the model's recommendation; and 13 (12%) patients remained unsure. While the association remains statistically significant (p=0.0067), the effect size dropped to 0.21 indicating medium effect.
Conclusion: The regret-based threshold model strongly predicts what patients think they would want (preferences) and moderately predicts the patients' actual choice. This is the first empirical study testing the threshold model in a real-life setting.
Agreement between patient preferences and recommendation of regret threshold model.