Characteristics of international primary care practices and physicians related to advance care planning: a cross-sectional survey study

Data from the 2015 Commonwealth Fund International Health Policy Survey of Primary Care Physicians in 10 Nations was used to perform analyses. This survey consists of physician responses from nationally representative random samples of 12,049 PCPs in 11 countries: Australia, Canada, Germany, France, Netherlands, New Zealand, Norway, Sweden, Switzerland, United States, and United Kingdom. The Commonwealth Fund survey is an International Health Policy survey collecting nationally representative data in Organization for Economic Cooperation and Development (OECD) countries to compare features of health system performance. The sample included general practitioners, family physicians, internists, and pediatricians. Responses were collected online, by mail, or by phone. Complete survey data and methods are available [17]. This study was approved by the Colorado Multiple Institutional Review Board.

Physicians were asked “Do you have conversations with older or sicker patients about the health care treatment they want or do not want in the event they become very ill or injured, and cannot make decisions for themselves?” The primary ACP outcome, an indicator for ACP conversations, was created by grouping responses as “yes, routinely” versus other responses. In secondary analyses we created a binary outcome to reflect whether PCPs who routinely engage patients in ACP also routinely document their patients’ preferences resulting from these conversations.

Independent variables included physician characteristics and primary care practice characteristics. Physician variables included the following: age (younger than 45 years vs. 45 years or greater), sex, and whether the physician considered their work to be high stress (“Extremely” and “Very” vs. “Somewhat”, “Not too”, or “Not at all” stressful).

Practice characteristics included location (rural, small town, suburb, urban), and how many full-time equivalent (FTE) physicians were employed at the practice, with categories based on quartiles of the distribution of responses (1-1.45 FTE, 1.5–2.95 FTE, 3-5.95 FTE, and 6-100 FTE). One survey question asked how many hours per week physicians worked, and we created a four-category variable for hours worked, with cut points at each quartile of the distribution of hours worked (0–34 h/week, 35–40 h/week, 40–49 h/week, 50–80 h/week). Finally, we also created a four-category variable using a question from the survey regarding how long the PCP’s typical visits lasted, with cut points based on quartiles of the distribution. Length of visit categories included 1–11 min, 12–14 min, 15–19 min, and 20–240 min.

To assess the impact of PCMH-like qualities that influence ACP in this sample, we used a validated, 41-item PCMH index which we previously developed using the sample of US physician respondents from this same international dataset [9]. According to the Bodenheimer model there are 10 building blocks to high-performing, advanced, primary care practices: engaged leadership, data-driven improvement, empanelment, team-based care, patient-team partnership, population management, continuity of care, prompt access to care, comprehensiveness and care coordination, and a template of the future to allow for fewer and longer in-person visits [12]. Briefly, to create the index we used descriptive items reported about the practice from the survey. From these practice descriptors items to be considered for the PCMH-index were selected through a 2-round Delphi process. Panelists for the Delphi process were primary care researchers familiar with logic models of PCMH practices. After 2 rounds of the Delphi process items were removed if they did not meet an item-total cutoff of 0.20 for inclusion in the index. The index had a Cronbach’s alpha value of 0.879. This index was then validated against a survey question asked of US respondents to the survey: whether the physician practiced in a PCMH. Because the data used to create this index were secondary data, this index does not include items for every single building block of the Bodenheimer model. Our team and Delphi panelists felt that this index has question items specific to 5 of the 10 building blocks with components of questions reflecting 8 of the building blocks.

An unadjusted logistic regression was performed with the index as the predictor and binary PCMH question as the outcome. The results from this model found an odds ratio of 1.11 (95% CI: [1.09, 1.13], p < 0.001), indicating that every additional item in the index that a respondent endorsed was associated with an 11% increase in the odds of the practice being a PCMH. Each item in the PCMH index represents a practice feature characteristic of practices that have PCMH designation. Further details for the derivation of this index can be found in prior work [9]. Higher index scores (ranging from 0 to 41) mean that the practice has more features characteristic of and associated with a US PCMH. The average index score in US practices from which the index was derived was 23.7 (SD 7.7).

Bivariate analyses were performed using chi-square tests. For the PCMH scale, a continuous variable, we performed a two-sample t-test. Missing values were excluded from these analyses. Next, single random imputation was used for items missing less than 10% of responses. These were randomly identified based on the range of responses to each item. We then performed a series of mixed-effects logistic regression models using the PCMH index to predict the ACP outcome. We allowed for a random intercept for country to account for within-country clustering. First, we used an unadjusted model including only the PCMH index to predict ACP. In the survey, respondents from two countries (Sweden and Switzerland) were not asked for their age or sex. Therefore, two adjusted models were performed. First, an adjusted model with all covariates except for physician age and physician sex was performed, including respondents from Sweden and Switzerland. Next, a model with all covariates, including physician age and sex was used, excluding respondents from Sweden and Switzerland. Adjusted models included the imputed independent variables. Because the relationship between ACP and APCMs has been established in the US sample from these data, [9] we performed sensitivity analyses including all countries except the US.

Finally, among the subgroup of PCPs who endorsed routine ACP, we performed an analysis focusing on whether patient preferences are recorded in their medical records. We used both unadjusted and adjusted logistic regressions with random a random intercept for country to investigate the effect of the PCMH scale on documentation of preferences.

All models were performed using R version 5.3.1 and the glmer function from the lme4 R package [18]. Two-sided p-values less than 0.05 were considered statistically significant. Age group and gender had high rates of missing responses because Swedish and Swiss respondents were not asked these questions.