Data set

The IHPS has been a cross-sectional survey repeated annually by the Commonwealth Fund.[14] Questionnaires were administered to understand health related issues in different countries.[14] The questionnaires were revised from time to time.[14] Some variables were removed and new ones might be introduced. Specifically, the 2014 IHPS surveyed 25530 adults aged 55 years and over in 11 countries: Australia, Canada, France, Germany, the Netherlands, New Zealand, Norway, Sweden, Switzerland, the United Kingdom, and the United States.[14, 15] The samples were randomly selected in each country to obtain nationally representative statistics. Adults aged 55 years or over were called for telephone interview.[14] The measures in this survey included health status, healthcare use, healthcare costs and access, timeliness of care, care coordination and safety, doctor-patient relationship, health promotion, end-of-life planning, patient engagement, management of chronic conditions and caregiving.[14] The sampling frames, the rationales and theories of the included measures were reported by the Commonwealth Fund researchers.[14] The variables related to patient engagement and activation were first introduced to the IHPS in 2014 survey.[14] The theories for the patient engagement measures were discussed in the next section.

There were several important features in the 2014 IHPS data set. The first was that many questions were country-specific. This was because the health systems differed from country to country and several countries chose to add questions pertinent to their healthcare system.[15] For example, a section of the questionnaire was specifically used to ask the status of health insurance and insurance purchase. These questions were very useful to understand health insurance coverage in the United States, but not useful for other countries where healthcare insurance was administered by governments and provided to almost all citizens, such as the United Kingdom and Canada. For the levels of education, the titles of the secondary school diploma differed and the methods to report the types of degrees varied. The use of country-specific questions led to high proportion of ineligibility in some of the variables (see S1 Appendix for the proportions of missing data).

Also there was variability of exposure to healthcare systems among interviewees. Interviewees’ experiences in health care were asked in questions about primary care, specialist care, chronic illness care, hospitalization, emergency room use and end-of-life planning. Interviewees were mostly eligible to questions about primary care and the response rates were higher than those about hospitalization and emergency room use (see S1 Appendix for details). This also contributed to high proportions of ineligibility in some variables.

In addition to ineligibility to certain questions, this survey also had other limitations.[14] First, the response rates varied significantly according to countries and sampling frames (random-digit dialling or population registry) and the direction of bias was unknown.[14] The other limitation was that institutionalized adults were not sampled in the survey and so countries with higher rates of institutionalization might seem healthier than those with lower rates.[14]

The measures of patient engagement or support in the 2014 IHPS

Patient engagement or patient support was measured by eight variables based on two theories, four related to chronic conditions and four related to specialist care in the 2014 IHPS.[14] At the time of variable creation, two important patient engagement theories were determined fit for the survey and selected by the Commonwealth Fund researchers, Patient Activation Measure and a team-based comprehensive model.[14] According to the Patient Activation Measure, a validated tool, patient engagement was related to patients’ knowledge, skills, beliefs, leadership and autonomy about the health care and treatment.[16, 17] In the team-based comprehensive care model, the continuity of care, communication with patients, patients’ involvement with providers, and access to physicians, were important factors.[18] Based on these two patient engagement or activation theories, the exact variables used in the 2014 IHPS included 1) how often specialists provided patients with treatment choices (variable short name: SP_Rx_choice); 2) how often specialists involved patients as much as patients wanted (SP_Involvement); 3) how often specialists asked what mattered to patients (SP_What_matters, only in the Netherlands); 4) how often specialists gave patients’ family and relatives the opportunity to involve in care as much as patients wanted (SP_Family_involved only in Sweden); 5) whether patients had treatment plans that they could carry out in their daily life (CC_Feasible_plan); 6) whether the treatment plans helped control or manage conditions (CC_Helpful_plan); 7) whether there were health care professionals who contacted patients to see how things were going between doctor visits (CC_contacted); and 8) whether there were healthcare professionals that patients could easily contact to ask questions or get advice between doctor visits (CC_DR_for_question) in Table 1.[14] To link these measures to existing patient engagement literature, we classified the variables into three core components of patient participation: shared decision-making, self-care and autonomy, and patient having critical self-knowledge based on the Patient Activation Measure and the team-based comprehensive care model[9, 16–18] in Table 1. The variables of different patient engagement categories were designated with different colors in Table 1 and in Figs 1 and 2.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Flow chart of variable selection and network modeling with the 2014 Commonwealth Fund International Health Policy Survey of Older Adults.

https://doi.org/10.1371/journal.pone.0201355.g001

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Networks of patient engagement, countries, and experience in health care.

(a) The largest network consisting of eight variables, out of 83 variables. Note: The thicknesses of arcs are proportional to the absolute values of Spearman’s correlation coefficients between variables. See variable definitions in Table 1. (b) The network of the measures of patient engagement or support in specialist care. (c) The network of patient engagement in chronic illness care. Note: See Table 1 for variable definitions.

https://doi.org/10.1371/journal.pone.0201355.g002

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. The measures of patient engagement or support and related variables in the Commonwealth 2014 International Health Policy Survey.

https://doi.org/10.1371/journal.pone.0201355.t001

Bayesian network modeling

To understand the overall relationships between the variables in the 2014 IHPS, we used a probabilistic graphical model, known as Bayesian networks for the analysis.[13, 19, 20] Bayesian network analysis adopted Bayesian statistics for network analysis.[19] In Bayesian networks, variables were presented as nodes and the arcs were used to present the relationships between variables in terms of conditional probabilities.[13, 19, 20] Bayesian statistics built on Bayes’ theorem that described the relationships between prior probabilities and posterior probabilities based on given events were used to propagate information between nodes or variables.[19] In other words, the probability distribution of one variable was related to its parent variable. If two variables were related only because they had the same parent variable, they were interpreted as conditionally independent in Bayesian network modeling.[13, 19, 21] Bayesian network modeling could be applied to identify the parent variables and the strengths of the linkages to parent variables in terms of conditional probabilities.[13, 19, 21]

In general, the advantages of using Bayesian network include the ability of handling incomplete data, structural learning, integrating information from diverse sources, presenting variable relationships with visual clues, and fast responding to post-analysis queries.[22] Sometimes causal interpretation can be inferred from the arcs in the networks.[20] The Bayesian network had been used in many disciplines, ranging from health systems research,[21] medicine,[23, 24] biology,[25] and sociology[26] to study the conditional dependencies between variables. The modeling process was shown in Fig 1.

Specific to the research on patient engagement measures and their interactions with contextual factors, the advantage of using Bayesian network modeling included the potential to discover latent structures of the relationships between variables.[13, 19, 20] The identified latent structure could be visualised in networks.[13, 19, 20] In contrast, regression models required researchers to treat one single variable as outcome and others as predictors.[27, 28] For regression models, it was assumed that the effects of the predictors were estimated simultaneously to one single outcome. This led to difficulties in understanding the interactions between patient engagement measures and contextual factors with conventional regression models. The assumptions of regression models might not be applicable to patient engagement variables. We considered the application of Bayesian network modeling important for the exploration of the interactions between patient engagement measures and other contextual factors.

Model development process and expert opinion

The Bayesian network models were built and reviewed in a way similar to the development process previously proposed.[21, 29] After data cleaning and missing data assessment,[29] the redundant variables or administrative variables were searched and not included for analysis (see Reasons for exclusion column in S1 Appendix). The variables that were derived by the Commonwealth Fund and were not labeled with clear definitions or not matched to the questions in the questionnaire were excluded from analysis. The Bayesian network model implementation was supervised by an expert panel based on the following steps.[30] The experts included a patient engagement expert (AB), an epidemiologist (MD), a data analyst (YSC), and a Bayesian network specialist (MS), who created the bnlearn package for the implementation of Bayesian network analysis within R environment.[13, 20] Three of them were also medical doctors with experiences in medical practices (AB, MD, and YSC). MD and MS were invited by AB and YSC for their expertise in Bayesian network modeling and epidemiological investigation respectively. The first step was that all variables were used for data-driven model building and assessed for adequacy (see S2 Appendix for original model output). Second, the expert-driven perspectives were introduced to 1) select the necessary variables for further model selection, and 2) identify essential links between variables. After discussion and model re-specification, the conclusion by the expert panel was as follows. The country-specific variables were determined inadequate and not used for the final model. Because the amount of health care spending could not be harmonized across countries, this variable was not selected for analysis. The arcs to demographic characteristics and national health care systems were blacklisted (see the next paragraph about layering or blacklisting). There were no arcs deemed obligatory or whitelisted.[13] In other words, the expert opinions helped to identify ineligible variables only and did not impose assumptions about the network structure (see S1 Appendix for the variables included for the final model and reasons for exclusion).

Layers of variables

There were restrictions on the directions between variables. First, the variables other than age (variable names available in S1 Appendix: q715), sex (q725), some college education (college), countries (q600), community types (q615), regions in countries (q630), linguistic regions (q640), and Swiss languages (q726), were not allowed to be directed to the above-mentioned variables. In other words, the other variables could not influence the probability distributions of above-mentioned demographic characteristics or health systems (layer-one variables in S1 Appendix). This technique was called blacklisting[13] or layering[24] to ensure there were no arcs directed from other variables to socio-demographic characteristics.

Correlations between variables and cross-group comparisons

In addition to Bayesian network modeling, the relationships between pairs of variables were also quantified by the correlation coefficients from Spearman’s correlation tests. The differences in continuous and categorical variables across countries or parent variables were also tested with one-way analysis of variance and Chi-square test, respectively. The p-values less than 0.05, two-tailed, were considered statistically significant. This study adopted R[31] (v. 3.10) and RStudio[32] for data analysis.

Ethics review

This secondary data analysis was approved by the ethics review committee at the Centre Hospitalier de l’Université de Montréal.

Networks of patient engagement and other variables

Based on Bayesian networks modeling, networks were drawn to describe the connectedness between variables. For the 2014 IHPS data, it was possible to identify networks that could be generalized to the data from 11 countries. Not all variables were connected in one single network. There were 19 separate networks identified and 66 variables were included in these networks. The 66 variables were related to countries of residence, sex, age, health status, healthcare utilization, patient engagement, health behavior, and experiences in primary, specialist and chronic illness care (see S1 Appendix for details). Ten networks consisted of two or three variables. Eight networks consisted of four variables. The largest network consisted of eight variables, two of which were patient engagement variables about chronic illness care in Fig 2A. The patient engagement variables were included in three networks, except for the variable, “whether specialists involved patients’ family or relatives in their care” that did not connect to other variables.

Networks of patient engagement by specialist and chronic illness care

The three networks consisted of any patient engagement variables were shown in Fig 2. There were two characteristics about the patient engagement variables. First, the patient engagement variables about specialist care tended not to connect with the engagement variables about chronic illness care or primary care (see S2 to S4 Appendices for all networks). The other was that patient engagement variables were only connected to two variables that were unrelated to patient engagement. The first variable connecting to patient engagement variables was “whether professionals provided written plans for your chronic condition” (not a patient engagement variable), which was connected to “feasible plan for your chronic condition” in Fig 2A. This connection was found in the largest network with eight variables. The other was “how confident about chronic condition management” (not a patient engagement variable), which was lined to a patient engagement variable, “how often specialists asked what mattered” in Fig 2B.

Contexts and potential measures for intervention

Implications

There are several research implications. Firstly, the measures of patient engagement or support do not necessarily interact with each other. The interactions are subject to the type of healthcare and professionals’ practices. The findings show that at least two contexts or settings could be identified for intervention: specialist care or chronic illness care. These pathways are worth further exploring to form engagement strategies that have the potential to improve health outcomes and healthcare quality. Secondly, this network approach is a method that policy makers can use to assess the extent of influence between patient engagement measures or to understand potential conflicts between organizational priorities and engagement activities. Competing priorities can be found in organizations or systems[12] and we think the understanding of the interdependencies or connectedness between variables through network models would be helpful to assess some of the concerns. With enough resources, the identified interactions between engagement measures can be verified with randomized trials or other studies.

For policy makers, the implementation strategies will rely on the policy objectives and resources. For example, providing clear instructions is directly linked to patient engagement in both chronic illness and mental health care. If the policy aims to involve patients as much as possible in chronic illness care, the formation of treatment plans may be the target of intervention. This is to focus on the immediate factors, if causation confirmed. In contrast, the opposite approach is to focus on the factors that are linked to many potential outcomes through several intermediate variables.[40] These non-immediate or distant factors may possess more extensive influence over patient engagement through multiple channels. In the same example, a policy focusing on how health professionals give patients instruction has the potential to influence engagement in chronic care and mental health.

Strengths and limitations

The strengths of this study include the use of comparable interview methods across countries, a large sample size, and models adequate for network analysis. The inclusion of patient engagement measures enables us to understand the potential role of engagement in health improvement. However, there are limitations to this study. This is a secondary data analysis using observational and cross-sectional data. The directions between nodes or variables cannot be seen as causal[42]. Second, the data cleaning policies are consistent with a previous publication[14] and require assumptions. For example, the “don’t know” category is recoded as missing. The importance of these assumptions needs to be tested in the future. Third, the 2014 IHPS was the first one to have patient engagement variables.[14] It is not possible to understand the trends in patient engagement in the included countries. Fourth, there are other patient engagement theories or research approaches that have been developed and reviewed recently.[4, 6, 7] The theoretic framework of the 2014 IHPS may not be satisfying when this study is published. Fifth, there are very limited numbers of health outcomes for study in the 2014 IHPS, including mental problems and self-rated health status. Lastly, there are other dimensions of patient engagement or outcomes that are not surveyed in the 2014 IHPS, such as health literacy, collaboration between patients and physicians, and self-management skills.[37, 43] We are not certain about the role of these unmeasured dimensions.