Prevalence of long-term decision regret and associated risk factors in a large cohort of ICU surrogate decision makers

To accomplish these aims, we conducted a cohort study using prospectively collected data from the PARTNER 2 trial (NCT02445937). PARTNER 2 (PAiring Re-engineered ICU Teams with Nurse-driven Education and Relationship-building) is a stepped wedge randomized control trial of a multicomponent nurse-led family support intervention for surrogate decision makers of adult ICU patients at high risk of death conducted at five ICUs in Pittsburgh, Pennsylvania [20]. The trial enrolled 848 patients and surrogates over four years between 2015 and 2019. Patients were enrolled if they were incapacitated and critically ill with a high risk of death or severe functional impairment (Additional file 1: Table S1). Eligible surrogates were those appointed in the patient’s advanced directives or, if no advanced directive existed, identified according to the hierarchy of surrogates codified in Pennsylvania state law. Surrogates were ineligible if under the age of 21, non-English speaking, or unable to complete study questionnaires due to mental or physical limitations. For this secondary analysis, surrogates were also excluded if responses to the decision regret questionnaire were incomplete.

As per trial protocol, data on exposures of interest for this study were collected directly from participants at enrollment (surrogate demographics), through chart abstraction following hospital discharge (patient demographics and illness information), and by phone interviews with surrogates at 6 months following the index hospitalization (patient outcomes). Data on surrogate decision regret, our outcome of interest, were also collected by phone at 6-month follow-up by assessors blinded to treatment group.

Surrogates were asked to score decision regret using the Decision Regret Scale (DRS), a validated 5-item measure of regret associated with past medical decisions [14]. Surrogates were specifically asked to consider decisions about life support or end-of-life care (Fig. 1). The DRS is one of three scales used for medical decision-making and was selected due to its applicability to past decisions and validation in multiple diverse patient cohorts [21, 22]. Items are scored using a Likert scale from 1 to 5 and then transformed into a final score ranging from 0 to 100 (increasing in intervals of 5), with higher scores indicating higher levels of regret. DRS scores are often divided categorically into no regret (0), mild regret (5–25), and moderate-strong regret (30–100) [23].

First, we determined the 6-month prevalence of decision regret among surrogates and conducted nonparametric testing to assess if median regret scores differed between intervention and control groups within the trial, as well as for surrogates whose loved one died in hospital or prior to 6-month follow-up. We then used multiple linear regression to identify explanatory factors that may influence decision regret. We decided that use of a multi-level model to adjust for clustering by trial site was unnecessary as intraclass correlation values were negligible when tested.

To create our multiple regression models, we identified variables of clinical interest a priori, drawing from a conceptual model of decision regret arising from the psychological literature on decision justification theory [19, 24, 25]. Decision justification theory proposes three forms of regret: process regret, option regret, and outcome regret. In other words, regret can arise from dissatisfaction with the decision-making process, dissatisfaction with the chosen option (regardless of outcome), or dissatisfaction due to an unfavorable outcome [25]. Using this conceptual model, we identified variables in our data set that related to the decision-making process (e.g., presence of advanced directives, prior surrogate experience, trial intervention), options chosen by surrogates (e.g., withdrawal of life support, change in code status), and outcomes (e.g., death, disability, patient location at 6 months). We also included patient and surrogate characteristics thought to be potential modulators of regret. The number of events per variable was greater than 15 for all models [26].

Missing data on 6-month location and functional status due to patient death, the largest cause of missingness, were handled using single-value imputation. Patients who died were assigned a 6-month location of ‘death’ and a Katz ADL score of 0, corresponding to ‘highly disabled’ (Additional file 1: Table S2) [27]. All missing data were assumed to be missing at random, and other sources of missingness, which were minor, were handled using complete cases analysis [28].

Two models were generated for the full cohort; the first included covariates related to patient outcome, while the second focused only on covariates known during the ICU admission. The purpose of this second model was to identify surrogates in the ICU setting who might be at especially high risk of regret and who may benefit from additional supports. Variance inflation factor testing was used to assess for collinearity and to refine our models [29, 30].

We further hypothesized that factors associated with regret may differ for surrogates whose loved ones died soon after critical illness. Thus, in addition to the full regression model, we also conducted stratified subgroup analyses according to patient survival at hospital discharge and 6-month follow-up. Given the smaller sample sizes for these models, backwards elimination at an alpha level of 0.5 was used to further refine variable selection. Sensitivity analysis was then performed by comparing variable selection models to complete models with all candidate variables included. All statistical tests were deemed significant at an alpha level of 0.05. Given the exploratory nature of our analyses, no correction for familywise error rate was performed [31, 32]. We performed all statistical analyses using Stata 16 (StataCorp, Texas, USA). This manuscript was written in accordance with the STROBE recommendations for reporting observational studies [33].

Out of 848 patient-surrogate dyads enrolled in the trial, 772 surrogates completed 6-month follow-up (Additional file 1: Fig. S1). Of those, 748 had complete decision regret scores and were included in the final analysis. Patient and surrogate characteristics are listed in Tables 1 and 2. Surrogates had a mean age of 56 years, 73% were female, 85% identified as White, and 72% were a child or spouse. Half had prior surrogate decision-making experience. Patients had a mean age of 64 years, 45% were female, and 80% identified as White. Just under half (40%) had completed an advanced directive prior to their index hospitalization, and 85% were full code at ICU admission. One third of patients died in ICU, and 92% of those had limitations of life support prior to death. At 6-month follow-up, 44% of the original cohort were still alive. Of those still alive, 77% were living at home, and 60% were functionally independent (Katz ADL score of 6).

Decision regret scores for surrogates were positively skewed (Fig. 2) with a median (IQR) DRS score of 15 (0–25). Overall, 27% of surrogates reported no regret (DRS score 0), 54% reported mild regret (DRS score 5–25); and 19% reported moderate to strong regret (DRS score 30–100) (Fig. 2). Median DRS scores did not differ significantly for surrogates who received the PARTNER 2 intervention (compared to usual care) but were significantly higher for surrogates whose loved one died in the hospital (median [IQR] of 10 [0–25] for survivors versus 20 [5–25] for non-survivors; p < 0.0001) or prior to 6-month follow-up (10 [0–20] for survivors versus 20 [5–30] for non-survivors; p < 0.0001) when compared using the Wilcoxon rank-sum test.

Table 3 lists all covariates examined in our multiple regression analysis and more detailed definitions are provided in Additional file 1: Table S3. Covariates found to be significant in multiple regression analysis are shown in Table 4. For model 1 (patient outcomes included), only the composite patient outcome of death or severe disability at 6-month follow-up was significantly associated with regret (Fig. 3a). Surrogates of these patients had a 10.1-point increase in DRS scores (95% C.I. 3.2, 17.0). None of the other covariates had a significant effect on the likelihood of regret. Overall, model 1 had an adjusted R² value of 0.06 and F-statistic of 2.17 (p < 0.001).

For model 2 (patient outcomes excluded), three covariates were associated with more regret (β; 95% C.I.): palliative care consultation while in the ICU (3.0; 0.1, 5.9), limitations in life support prior to death (6.3; 3.1, 9.4), and surrogate black race (6.3; 0.3, 12.3) (Fig. 3b). Model 2 had an adjusted R² value of 0.03 and F-statistic of 1.78 (p = 0.01).

When stratified by patient survival, surrogates with a loved one who was alive but severely disabled at 6 months had more regret (+ 10.0 points; 95% C.I. 2.0, 18.0) compared to those with less severe disability or who were functionally independent. Surrogates of survivors who were full code at the time of ICU admission had less regret than those of patients who had a DNR at time of ICU admission and survived (-8.8 points; 95% C.I. − 16.8, − 0.9).

For surrogates with a loved one who did not survive to 6-month follow-up, the following covariates were associated with less decision regret (β; 95% C.I.): patient death in the hospital (− 10.1; − 19.6, − 0.5) compared to death after hospital discharge; older surrogate age (− 0.31 points per year of age; − 0.5, − 0.1); surrogate graduation from high school (− 6.0; − 12.0, − 0.1) or college (− 6.9; − 12.4, − 1.5) compared to not finishing high-school, and being a child (− 7.0; − 13.6, − 0.4) or friend (− 13.3; − 22.1, − 4.5), rather than spouse of the patient. Conversely, limitation of life support treatments prior to death was associated with more regret than not limiting treatment (10.2; 1.0, 19.3).

Results were similar when stratified by death during index hospitalization (see Table 4). For patients who survived to hospital discharge, the composite outcome of death in the post-discharge period or severe disability at 6 months significantly increased surrogate regret (11.5; 6.0, 17.1). Except for code status in the 6-month survival group, subgroup results remained robust across sensitivity analyses.

Our study has several limitations. First, the use of existing trial data did not allow us to examine all covariates that may influence decision regret, only those collected as part of the trial. Assuming missing data to be missing at random and the use of single-item imputation to handle missingness may have biased regression parameters and underestimated standard errors. We also lacked detailed information about the decisions surrogates considered when completing the DRS. Importantly, other studies have found that decision regret increased over time [34, 37]. Thus, the point at which we measured regret may have influenced our results. Serial measures of regret would help us better understand the surrogate experience. Our cohort was also enriched for patients with a high risk of poor outcomes and these results may not be generalizable to ICU surrogate decision makers in all circumstances. Surrogates who agree to participate in a clinical trial may also differ systematically from surrogates who do not participate, leading to volunteer bias. Lastly, our regression models were exploratory in nature and spurious results are a possibility. Further studies are necessary to refine our understanding of factors responsible for decision regret in ICU surrogates.

Measuring regret remains challenging. As a psychometric construct, the Decision Regret Scale focuses primarily on the option and outcome components of regret, rather than the decision process [21]. This feature of the scale may have influenced our finding that patient outcomes were most strongly associated with regret. Determining the threshold DRS score for clinically meaningful regret is also a challenge [23, 48]. A cut-off of 25 has previously been used to distinguish higher levels of regret without clear justification [23]. In a systematic review of medical decision-making studies, mean DRS score across studies was 16.5/100 and 4–20% of patients had scores over 25 [23]. Without an established gold standard measure, it is unclear if further scale calibration is needed or, less likely, if medical decision-making tends not to cause much regret. Additionally, the original DRS was not validated for use in surrogates [14]. A Decision Regret Scale for Caregivers (DRS-C) has subsequently been developed [48].

Finally, optimal measures of surrogate decision-making remain unclear [49]. Although decision regret causes psychological distress, it does not automatically follow that less surrogate regret indicates better decisions, particularly when considered from the patient’s point of view. Other measures of decision-making include decisional self-efficacy and decisional conflict [49]. Higher levels of decisional conflict have been associated with end-of-life decisions, whereas advanced directives have been associated with lower decisional conflict [5, 50, 51]. However, lower decisional conflict may also indicate a lack of engagement with hard choices, rather than better decision-making [52]. Moreover, decisional conflict is a transitional state with unclear long-term effects and an uncertain connection to decision regret [13, 53]. More work is needed to understand the links between decisional conflict, self-efficacy, and regret, as well as to determine what outcomes define success when it comes to surrogate decision-making in the ICU.