Categorical variables are available in the CDW for veteran related demographic characteristics, rurality, and VHA priority group for each FY, which are described previously [8]. For this evaluation, VA administrative data provides an adequate sample size for subgroup analysis by mean age, marital status (i.e., divorced, married, not married, or separated), race (i.e., Black or African American, White or Caucasian, other races), male or female sex, Hispanic or not Hispanic ethnicity, and rurality used by VA to categorize veteran’s residences as rural or urban using VA administrative data and VHA priority group status. VHA priority group status is healthcare eligibility defined by group numbers and descriptions for the inclusion criteria for each group. While a general description is provided here, more details can be found elsewhere [18]. Priority 1–4 are veterans who have a VA service-connected disability, receive VA aid and attendance, receive VA homebound services, or who are determined to be catastrophically disabled; Priority 5–6 are veterans who do not have a service-connected disability, have a non-compensable service-connected disability that VA rated as 0% disabling and have an annual income level that is below VA adjusted income limits (based on resident zip code), receive VA pension benefits, or are eligible for Medicaid programs (group 5) or have a non-compensable service-connected disability that VA rated as 0% disabling and meet criteria of certain military related exposures or war zones. Priority 7–8 are veterans whose income is above (group 8) or below (group 7) VA income limits and geographically adjusted income limits for where they live and who agree to copayments. See Table 1 for the description of the total sample.
Latent class indicators
We used seven items of ACP self-efficacy derived from the ACP-GV Participant Worksheet used in ACP-GV. The participant worksheet was created by the National ACP-GV Program provided within VHA to all facilities with special attention devoted to veterans residing in rural areas. The first seven questions are reviewed and answered by participants at the start of the group session (e.g., pre-intervention). Knowledge is measured with one question, written as “How knowledgeable are you about advance care planning?” with a five-item Likert response option: “not at all, a little bit, moderately, quite a bit, and extremely.” Comfort is assessed with two questions: “I have thought about what I would want if I were hurt, injured or sick and could not communicate” and “I have thought about my treatment preferences if I could not communicate them during a mental health episode.” Response options for these and subsequent questions were “Yes” or “No.” Confidence is assessed with four items: “I have talked with someone I trust to make health care decisions for me; I have named someone to make health care decisions for me; I have discussed these topics with someone on my health care team (such as a doctor, nurse, or social worker); I have filled out an advance directive (also known as ‘living will’) to guide those I trust to make health care decisions for me.” We used the first seven items in the LCA to identify veteran subgroups and assign class membership to each veteran in our sample. Table 2 displays the question, response options, and patterns for each of the seven latent class indicators.
Additional ACP self-efficacy measures and outcome expectancy items
Near the end of the ACP-GV discussion, participants return to the worksheet and write responses to a final set of questions (Questions 8–11). A post-intervention knowledge gain question is queried first with “How much has your knowledge increased about advance care planning?” with a five-item Likert response option (i.e., not at all, a little bit, moderately, quite a bit, extremely). For outcome expectancies, the participants are asked, “If you are ready to take a next step” and instructed to enter their next step in the blank text box, if applicable. To ease administrative burden for providers, the National ACP-GV Program scripted a range of possible categorical response options within the electronic health record documentation to include: [1] Check current AD status and update as needed [2], Complete AD [3], Discuss with family [4], Discuss with health care provider [5], Discuss with non-family member [6], File a copy of an existing AD [7], Goals not clear/unsure [8], Learn more about ACP [9], Update existing AD [10], I need time to think and reflect on my values and wishes, and [11] Other that are used to document the next step response. Responses are documented and referred to in a follow up phone contact with the veteran to engage them in their next step.
The worksheet then provides four text boxes to reflect on the development of a goal if a next step was identified previously. The question prompt reads “If you are ready to take a next step, please write what your next step will be,” followed by four questions with their own respective text box, “When will you do this?”, Who will you do this with?”, How will you do this?”, Where will you do this?” The final question related to the likelihood of taking the next step, or behavioral activation, is queried with “How likely are you to take this next step?” with a five-item Likert response (i.e., definitely will not, probably will not, not sure, probably, definitely will). See Table 2 for a description of the types of next steps and the distribution of the types of next steps by the resulting class membership.
Statistical analysisLatent class analysis
To identify subgroups of veterans’ knowledge, confidence, and comfort with ACP, we conducted a series of LCA using robust maximum likelihood estimation with the MPlus version 8.8 software [19]. We used a combination of criteria to determine the number of latent classes, including (1) examination of fit indices (e.g., Akaike Information Criteria [AIC] and Bayesian Information Criteria [BIC], etc.) of which we weighed the values for the BIC (smaller is better) and the sample adjusted BIC as most accurate given its superior performance for LCA models, (2) Vuong-Lo-Mendell-Rubin Likelihood Ratio Test (VLMR-LRT) where significant p-value indicates the (K-1) class model is rejected in favor of a model with at least k-classes, (3) entropy values (larger is better), and (4) clinical judgment regarding the practical utility of classes for intervention efforts [20]. Consistent with the views of Nylund et al. (2007) [20], we first identify the point where our model fit indices start to plateau across the different LCA models we executed. This flattening effect suggests minimal or no improvement in model fit with the inclusion of additional classes. To decide whether to include additional classes after fit indices values start to plateau, we consider the heuristic and clinical value of adding additional classes and weighed this against the value of using more parsimonious solutions. We note that veterans did not provide responses to all items of the ACP self-efficacy questionnaire; therefore, the LCA analysis was based on case-wise deletion.
Comparison across subgroups
After identifying the best fitting LCA model and assigning each veteran to the classes with highest membership probability, descriptive measures, including frequency and percentages, describe veteran demographic characteristics overall and are stratified by latent classes. Additionally, demographic characteristics across the latent classes are compared using chi-square tests (see Table 1). For each next step response category, separate logistic regression models access the association between the latent class member and types of next steps taken. In addition, multivariable logistic regression accounts for demographic characteristics including age, marital status, race, ethnicity, urban/rural, sex, and VHA priority group. We report both unadjusted and adjusted odds ratios (OR) along with their respective 95% confidence intervals (CI).