Objective: To determine whether groups of surrogates for severe acute brain injury (SABI) patients with poor prognosis can be identified based on their prioritization of goals-of-care (GOC) decisional concerns, an online survey of 1588 adults recruited via a probability-based panel representative of the US population was conducted.  

Methods: Participants acted as a surrogate for a GOC decision for a hypothetical SABI patient and were randomized to one of two prognostic scenarios: the patient likely being left with a range of severe functional disability (SD) or remaining in a vegetative state (VS). Participants prioritized a list of 12 decisional concerns via Best-Worst Scaling. Latent class analysis (LCA) was used to discover decisional groups.

Results: The completion rate was 44.6%; data weighting was conducted to mitigate nonresponse bias. For 792 SD respondents, LCA revealed four groups. All groups shared concerns regarding respecting patient wishes and minimizing suffering. The four groups were otherwise distinguished by unique concerns that their members highlighted: an older adult remaining severely disabled (34.4%), family consensus (26.4%), doubt regarding prognostic accuracy (20.7%), and cost of long-term care (18.6%).  For the 796 VS respondents, LCA revealed five groups. Four of the five groups had similar concern profiles to the four SD groups. The largest (29.0%) expressed the most prognostic doubt. An additional group (15.8%) prioritized religious concerns.

Conclusions: While surrogate decision makers for patients with SABI are concerned with respecting patient wishes and minimizing suffering, certain groups highly prioritize other specific decisional factors.  These data can help inform future interventions for supporting decision makers. 

The survey versions for this study were developed in 2015, with the main data for this study then collected between February 18 and 29, 2016, from a nationally representative online survey of US adults 30 years and older. The following sections outline the survey development process first; followed by descriptions of the sample, data collection, and statistical design for the main study.

 

Standard protocol approval

The study was approved by the Yale Human Investigation Committee (protocols #1406014207 and #1505015893).

 

Survey instrument development

Two versions of the survey instrument were drafted by the authors that both asked a participant to consider an 85-year-old, previously healthy, close family member admitted to an ICU following a severe intracerebral hemorrhage (ICH) and remaining intubated and poorly responsive despite two weeks of care without limitations (full surveys available in Appendix e-1). Both versions of the survey then asked the participant to imagine that he or she was the patient’s surrogate decision maker for possible tracheostomy and gastrostomy placement (i.e., LST).

The two survey versions that were developed differed with regards to the long-term poor prognosis presented. One version—the more common real-life scenario—stated that the patient would most likely remain unresponsive, with a small but unlikely chance that the patient could gradually recover the ability to speak, eat, recognize others, and/or travel in a wheelchair with time (the “severe disability” survey). The other version stated a less common but more severe scenario that the patient would remain bedbound and unresponsive long-term with certainty (the “vegetative state” survey). Both survey versions specified that the patient would be ventilator-free following tracheostomy and that the patient did not have an advance directive.

For both versions of the survey, a list of concerns that a surrogate could prioritize when deciding on LST versus CMO for an intubated patient with SABI was also developed (Table e-1). To review the drafted hypothetical scenarios and to create the list of decision making concerns, semi-structured interviews were conducted with 11 family members of former patients in the Yale New Haven Hospital (YNHH) Neuroscience ICU (six of whom had requested the ICU make their relatives CMO). Family members who provided input were recruited by telephone within 8 months of their neuroscience ICU experience. Detailed notes were taken as the family members (1) provided feedback on the wording of the drafted hypothetical scenarios and (2) listed what their own top decisional concerns would be if acting as a surrogate decision maker. They were also asked if they agreed or disagreed with concerns that other families had expressed in prior published qualitative research regarding similar ICU goals-of-care decisions.

Interviews with family members were conducted until thematic saturation regarding the decision making concerns was achieved, with a total of 10 concerns attained. Fifteen national experts—in the fields of neurology, critical care, palliative care, geriatrics, ethics, and medical decision making with publication track records on the topic of ICU GOC decision making—were then contacted by e-mail, given revised copies of the hypothetical scenarios and the list of 10 concerns elicited from the family interviews, and asked to provide additional feedback (Table e-1).

The final list of 12 decision making concerns in Table 1 were used to create a 12-question Best-Worst Scaling (BWS) exercise utilized in both survey versions to allow participants to prioritize their concerns based on the hypothetical scenario presented. A BWS question asks respondents to choose the “best” (or most concerning) and the “worst” (or least concerning) item from a series of sets containing different combinations of items from a master list. An example is provided in Appendix e-1. BWS questions are easy to understand, and respondents from diverse socioeconomic backgrounds have been shown to provide reliable data. The format avoids scale-related response bias and is thus more efficient than rating scales. BWS has been employed to assess patients’ preferences in a wide variety of health care settings.

The question design used in the BWS application was based on a Youden balanced incomplete block design. Each of the 12 questions presented four potential decision making concerns, with each concern presented an equal number of times in each version of the survey, and each concern presented with each of the other 11 concerns an equal number of times.

Following the BWS exercise, both survey versions asked participants about their decision regarding requesting LST versus CMO for the hypothetical patient, as well as their certainty about the decision on a 4-point scale.

Surveys were designed using Sawtooth Software, Inc. (SSI) Web v8.2. Both versions of the survey underwent cognitive testing among a total of 11 ethnically diverse subjects recruited from the local community around YNHH. The final survey versions incorporated feedback and were both expected to take at least eight minutes to complete based on testing.

 

Sample

Following survey instrument development, participants for the main study were recruited from the GfK (now Ipsos) KnowledgePanel, a probability-based online panel designed to be representative of all non-institutionalized civilian adults living in US households. The 55,000 active panel members cover 97% of US households, and the panel sample is updated every three months through address-based sampling, to include those that are served only by cellphones or have no telephone service. Selected panel members who do not have computers or internet service are provided with both.

Because the minimum age of real-life surrogate decision makers in prior studies in Neuro ICUs was approximately 30 years old, this age was used as the minimum age for recruitment of participants from the survey panel.

 

Data collection

Following a two-day operational test period with 36 participants, the KnowledgePanel members accessed the surveys using a link sent by the company and were randomized to one of the two survey versions, with their differing prognostic scenarios. Any respondent completing a survey in less than eight minutes was excluded. The KnowledgePanel provided the following basic demographic information regarding participants: age, gender, race, education, household internet access, region of residence, marital status, employment status, and income. All other demographic variables were self-reported by respondents directly in response to additional survey questions.

 

Statistical analysis

Data from the severe disability and vegetative state survey versions were analyzed separately. To minimize bias due to differential nonresponse in the KnowledgePanel recruitment process as well as nonresponse during the field period, iterative proportional fitting was used to weight the survey data and geodemographically match the U.S. population age 30 and older, based on the 2015 Current Population Survey.

Latent class analysis was performed on the data from the surveys’ BWS exercise using Sawtooth Software’s Latent Class Segmentation Module v4 to estimate part-worth utilities (i.e., importance scores for each of the concerns listed in the surveys), scaled from zero to 100, and to determine previously unmeasured class membership among participants. Class solutions were replicated 20 times from random starting seeds, with a maximum of five groups to allow for possible heterogeneity of survey responses.

For both versions of the survey, the Bayesian Information Criterion (BIC) and Akaike Information Criterion (AIC) decreased as the number of groups increased, suggesting improving fit. However, for the severe disability survey, two of the groups in the five-group solution had utility distributions that were nearly identical (Figure e-1), which made four groups the best practical solution.

For both survey versions, associations between participant characteristics and group membership were explored by comparing univariate data and subsequently performing multinomial logistic regression including only those surveys with all questions answered. Analyses were conducted using SAS software v14 (SAS Institute Inc., Cary, NC).

 

Sample size

Factors affecting the performance of latent class analyses include the number of items in the BWS exercise, the manner in which different concerns cluster among participants, and the eventual size of each group in the best solutions. Meaningful segmentation analyses have been performed in the medical literature with as few as 183 respondents for a survey that contained as many as 18 items to prioritize. In comparison, recruitment for this study obtained nearly 800 respondents for each of two 12-item BWS exercises.

 

Data availability

Anonymized data from this study can be shared by request by any qualified investigator for the purposes of replicated procedures and results.