The facilitators and barriers associated with implementation of a patient-centered medical home in VHA

Independent variables were obtained from a national, web-based survey fielded to all VHA primary care personnel in May and June 2012. The survey link was distributed via an email from VHA central office leadership to regional network and facility leaders with instructions to distribute to primary care personnel. Because the survey was anonymous and voluntary, the precise denominator is unknown; however, the estimated response rate was 25 %. Respondents to the VHA PCMH survey had very similar demographics with primary care respondents from a general employee survey that was fielded the month prior to the VHA PCMH survey. The general employee survey achieved a 62 % response rate. Tables comparing demographic variables have previously been published, and show that the PCMH sample had a slightly higher proportion of supervisors and slightly fewer African Americans but was not significantly different in terms of employee age, tenure, gender, or other racial or ethnic groups [11]. The present analysis focuses on the 96 clinics with survey responses (n = 532) from among the 164 clinics in the top and bottom deciles for VHA PCMH implementation, described further below. Analyses of middle decile clinics are included in the Appendices 1, 2, and 3. The survey methodology has previously been reported [11].

The VHA PCMH survey included items assessing barriers to the delivery of optimal patient-centered care and facilitators of PCMH implementation. The items were developed in collaboration with five regional Demonstration Laboratories, which were established to evaluate the impact of the PCMH initiative in VHA and test methods of improvement [12]. The survey also included respondent characteristics such as team role, proportion of time spent working in primary care, and supervisory level.

The independent variables were 19 items assessing the barriers and 10 items assessing the facilitators. The question stem for the 19 barrier items was, “How much, if at all, does each of the following factors limit your ability to provide optimal, patient-centered care for your patients?” The barrier items were preceded by a definition of patient-centered care, “An approach to healthcare that prioritizes the Veteran and their values and partners with them to create a personalized strategy to optimize their health, healing and well-being.” Respondents scored each barrier on a 3-point ordinal scale from “Does not limit,” to “Limits somewhat,” to “Limits a great deal.” There was also an “N/A or Don’t Know” response option. Barriers included factors such as difficulty accessing specialists, recruiting and retaining providers, and high-volume of Computerized Patient Record System (CPRS) automated alerts (e.g., for contraindicated medications).

The question stem for the 10 facilitator items was, “How helpful are the following [PCMH]-related activities or resources?” Respondents scored each facilitator on a four-point ordinal scale from “Very Helpful,” “Somewhat Helpful,” “Not Helpful,” to “Not Available/Not Involved.” They could also respond “N/A or Don’t Know.” Facilitators included local education sessions (e.g., a facility or team level in-service on PCMH to help define team roles and responsibilities of the different team members); measurement tools (e.g., use of patient data to evaluate improvement benchmarks); and information systems (e.g., the Primary Care Management Module is a suite of software tools that can be used by primary care teams to assign patients to teams and generate reports). The dependent variable was a validated composite measure of PCMH implementation, the Patient Aligned Care Team Implementation Progress Index (Pi²), which assesses progress in implementing PCMH within VHA, and incorporates data from surveys of primary care personnel, patient surveys, and administrative data [13]. The Pi² includes 53 measures mapped to 8 domains representing goals of the initiative related to improving (1) access; (2) continuity; (3) care coordination; (4) comprehensive care; (5) self-management support; (6) patient-centered care and communication; (7) shared decision-making; and (8) team-based care. The Pi² score is calculated as the number of domains in which the clinic appeared in the top and bottom quartiles for the domain scores, ranging from 8 (all domain scores in the top quartile) to −8 (all domain scores in the bottom quartile). The Pi² has previously been validated by testing the association of clinics in the highest decile of the Pi² (a score of +5 to +8) versus those in the lowest decile of the Pi² (a score of −8 to −5) using measures of quality of care and outcomes that were expected to be favorably associated with PCMH implementation, including higher patient satisfaction, lower staff burnout, higher quality of care, and lower emergency department visits and hospitalizations for ambulatory care-sensitive conditions [13]. For the present analysis, we compared clinics in the lowest decile (−8 to −5, n = 87) to the highest decile (+5 to +8, n = 77); 96 of the 164 clinics in the high and low groups had one or more survey respondents.

Covariates included respondent characteristics, clinic workload and capacity, and setting. Respondent characteristics included VHA tenure, supervisory level, percentage of work hours spent in primary care teams, and occupation. Clinic workload and capacity included four variables from administrative data: (1) overall clinic average panel size, which was adjusted for type of provider (physician vs. nurse practitioner and physician assistant) and FTE; (2) the degree to which the clinic was over paneled relative to a 1200 patients per provider target panel size (0 if panel size ≤1200, the percent that average panel size exceeded 1200 if panel size >1200) [12]; and (3) clinical “intensity” of the patients in the panel, which was the average Diagnostic Cost Group score. Diagnostic Cost Group score is based on inpatient and outpatient diagnoses using the risk adjustment condition categories used by the Center for Medicare and Medicaid Services [14]. For setting, we adjusted for whether the clinic was community based or located with a VHA medical center, the logarithm of total annual visits, and whether the clinic had an academic-affiliation agreement indicating a teaching facility. The variables were selected because we expected they would reflect variation in local patient demand and workforce availability.

We conducted respondent-level multinomial logistic regressions, in which the dependent variable was the respondent’s clinic’s Pi² score, categorized as high (top decile), medium (middle eight deciles), and low (lowest decile), as a function of barriers and facilitators. We ran separate models for each barrier and facilitator and adjusted for respondent characteristics and clinic-level workload and staffing measures in each model, described above.

For facilitators, we constructed two sets of models. In the first, we assessed the associations of facilitator “availability” with Pi² score (high vs. low). A facilitator was defined as unavailable if the respondent rated it Not Available/Not Involved; otherwise, it was defined as available. We constructed a second set of models based upon respondents’ categorization of the facilitators as, Very Helpful versus Not or Somewhat helpful.

For primary analyses, responses that were NA/Don’t Know were categorized as missing and excluded from the relevant individual model using a standard listwise deletion approach, eliminating observations with missing values from a given model but including those observations in any models where values were not missing. Because NA/Don’t Know responses might actually indicate the absence of a barrier or facilitator (rather than truly missing data), we also conducted secondary analysis where NA/Don’t Know values were re-coded as a negative response rather than missing (i.e., not a barrier; facilitator not available; or facilitator not helpful). Because frequent NA/Don’t Know responses might be indicative of a respondent who either truly did not know or was passively filling out the survey, we also ran the secondary models excluding respondents with 70 % or more NA/Don’t Know on items.

In total, we constructed 39 models: 10 models on facilitators available/not available; 10 models on facilitator helpfulness; 19 models on barriers. For reporting purposes, in facilitator models, low implementation clinics were the referent and in barrier models, high implementation clinics were the referent. To simplify the presentation of data, we focus on findings comparing high versus low Pi² clinics which are presented in the main results, but also include findings for comparisons of medium Pi² clinics versus low (referent) for facilitators, and low Pi² clinics versus medium (referent) for barriers in the appendices. Models were cluster-adjusted at the clinic level to allow for correlated responses within clinics. We report adjusted odds ratios and 95 % confidence intervals.

In order to assess potential non-response bias, we compared clinics without survey response to clinics with survey responses on covariates. We also assessed the level of inter-rater agreement on barriers and facilitators within site and within occupation.

Analyses were conducted using SAS Enterprise Guide 6.1.

The availability of five facilitators was associated with higher odds of a respondent’s clinic being in the highest versus lowest decile for Pi² scores (Table 3): teamlet huddles (OR = 3.91, CI 1.34–11.44), measurement tools (OR = 3.47, CI 1.47–8.16), regular team meetings (OR = 2.88, CI 1.31–6.30), information systems (OR = 2.42, CI 1.21–4.82), and disease registries (OR = 2.01, CI 1.11–2.80).

The helpfulness of four facilitators was associated with higher odds of a respondent’s clinic being in the highest versus lowest decile for Pi² scores (Table 4): local education sessions (OR = 2.02, CI 1.13–3.63), information systems (OR = 1.97, CI 1.19–3.24), disease registries (OR = 1.90, CI 1.22–2.96), and regular team meetings (OR = 1.68, CI 1.05–2.70).

For six of the 19 barrier items, reporting that delivery of patient-centered care was limited a great deal by the barrier was associated with higher odds of a respondent’s clinic being in the lowest versus highest decile for Pi² scores (Table 5): difficulty recruiting and retaining providers (OR = 2.37, CI 1.29–4.33); difficulty recruiting and retaining non-provider clinicians (OR = 2.17, CI 1.17–4.00); lack of support from clinical leadership (OR = 2.15, CI 1.21–3.82); lack of control over one’s schedule (OR = 1.98, CI 1.15–3.43); inadequate support for patient behavioral change needs (OR = 2.15, CI 1.36–3.38); and recruiting and retaining non-clinicians (OR = 1.96, CI 1.09–3.53).

We also found that the availability of three facilitators was associated with higher odds of the respondent’s clinic being in the middle eight deciles versus the lowest, and that the reporting of seven barriers was associated with higher odds of the respondent’s clinic being in the lowest Pi² decile versus the middle eight deciles (Appendix 1). In all cases, these were also significant associations in the comparison of respondents in clinics with the highest versus lowest decile for Pi² score, and in all cases, the associations were smaller but in the expected direction. There were no significant associations between helpfulness of facilitators and the odds of a respondent’s clinic being in the middle eight deciles for Pi² scores versus in the lowest decile.

In the secondary analyses testing how robust the results were to different assumptions about NA/Don’t Know responses, the results were virtually identical.

This was a cross-sectional analysis of observational data. This design is susceptible to unobserved confounding, and we cannot assume that the observed associations are causal. We found potential selection bias in terms of the clinics excluded due to lack of survey data. These clinics tended to be smaller, community-based clinics, and were more likely to be a low-Pi² clinic. This type of selection bias would most likely increase our risk of a type 2 error, and it is possible that we would have found additional significant associations with a more representative sample of clinics. It is also possible that our findings are due to inferential bias, in which respondents at high-implementation clinics infer that they encountered fewer barriers and received more helpful resources because they observed greater implementation of the VHA PCMH model. However, this type of bias, called representativeness heuristic, usually arises when someone infers an answer to a more complex question (e.g., how much progress have you made on VHA PCMH) from a simpler, specific observation (e.g., have you been able to hire staff), which is the opposite of the present analysis [20]. It is also of greater concern when the respondent is rating both the outcome and the explanatory variables. In this case, the outcome is a clinic-level index incorporating data from a range of sources and respondents completed the survey without knowledge of their clinic’s Pi² score.

We tested 39 associations, and it is possible that significant findings are just chance associations. However, we found a total of 15 significant associations when we would expect to find approximately two by chance (39 * 0.05 = 1.95). Moreover, the associations were all in the expected direction, and we found a similar but weaker pattern of associations when comparing clinics in the lowest decile for Pi² score with clinics in the middle eight deciles.

Finally, we used a discrete set of structured measures of barriers and facilitators of PCMH implementation that may have omitted important factors. Some factors we may have omitted, such as strategies for integrating additional clinic members, for example, pharmacists [21], could play an important role in implementation of the VHA PCMH model. Other potentially omitted factors, such as the effects of healthcare reform and conflicting criteria for PCMH designation [22], are likely less salient for VHA clinics, but may be vitally important in other settings.