Data for this study are from a cross-sectional survey of patients in Los Angeles County aged 18–64 years with type 2 diabetes and Medicaid health insurance. Sampling was conducted in two stages. First, primary care physicians (i.e., family medicine, internal medicine, and general practice) were identified through publicly available network data for one of the largest Medicaid plans in the county. From a list of 471 eligible physicians, they were randomly drawn and recruited until we reached our target sample size of 100 (n = 104, 75 % response rate).
Second, all physicians were asked to refer a minimum of 10 patients with type 2 diabetes meeting the study criteria, using one of two methods: retrospective or prospective referrals. The retrospective referral method involved referring all eligible patients who had visited the office, working from most recent up to the past six months. Before referral, the office staff called the patients and, used a study-provided script offer them the chance to opt out of the referral. The prospective method was similar. Physicians referred all eligible patients scheduled to visit the office for up to the next six months, or until a minimum of 10 referrals were made. In this case, a flier was provided to each patient about the study by the office staff, giving the patient a chance to opt out of the referral.
This study was approved and given an IRB waiver of HIPAA authorization by the USC Office for the Protection of Research Subjects. The potential for physicians to hand-select patients was reduced by the requirement that “all” patients meeting the criteria were to be referred. A total of 1.8 % of patients opted out of being the referral. We then randomly selected patients until we reached a goal of about five patients per physician. There were no differences in opt-out rates, response rates, or demographics between methods.
Medical Home Performance
We used the Johns Hopkins’ Primary Care Assessment Tool–Adult Expanded version (PCAT) to assess patient-reported indicators of medical home quality. While definitions vary across organizations, there is consensus on at least four core features--accessibility, longitudinality (or continuity), comprehensiveness, and coordination15–18--all of which are measured by the PCAT. The PCAT consists of 96 questions that evaluate the four core features above and three others that are often included in definitions of a medical home: 1) community-oriented care, 2) family-centered care, and 3) cultural competence.
First-contact care refers to the concept that care is available and first sought from the medical home when a new health or medical need arises. Longitudinality refers to the use of a regular source of care over time and the relationship that develops. Comprehensiveness is the range of services available to and received by the patient. Coordination refers to the linking of health services so that patients receive appropriate care for all of their health problems. Community-oriented care refers to the concept that providers strive to be aware of, and oriented to, the health needs of a community. Family-centered care is the recognition of the family as a participant in the diagnosis, treatment, and recovery of patients. Cultural competence refers to care that respects the language, beliefs, and attitudes of people.
Each question is scored using a Likert-type response scale as follows: “definitely not” (1 point), “probably not” (2 points), “probably” (3 points), and “definitely” (4 points). The PCAT is scored such that each feature is calculated as the average of the responses to the questions comprising the feature. A total medical home score is the average of the seven feature scores such that each feature is equally weighted. Previous work has shown that the overall PCAT has good reliability and validity.19
20 The PCAT has good construct validity, with the seven features accounting for 88 % of the total 96-item variance. Internal consistency reliability for the features has ranged from 64 % to 96 %. The PCAT has been successfully used to detect differences in primary care across providers and delivery systems.21–23
The EQ-5D is a preference-based global, rather than disease-specific, measure of HRQOL.10 It consists of five questions about mobility, self-care, usual activities, pain, and anxiety. Each dimension has three levels describing the patient's health state on that day—no problems, moderate problems, or severe problems—leading to 243 unique health states. The health states are expressed in an index using a preference-based set of weights for the U.S. population.24 This index is an expression of these states, taking into account the desirability or undesirability of each state. After weighting, the index is anchored by scores 1.0 (full health) and 0.0 (dead), with some states being worse than death (<0). The EQ-5D index scores in this study ranged from −0.04 (severe problems) to 1.0 (no problems). For two patients in our analysis with EQ-5D index scores below zero, both reported being “confined to bed” (mobility) and experiencing “extreme pain or discomfort” (pain). We also presented the EQ-5D dimensions using the three ordinal response levels.
Status: Self-reported general health status is the response to the question, “In general, would you say your health is excellent, very good, good, fair, or poor?” We re-coded the responses into three ordinal categories: 1) fair/poor, 2) good, and 3) excellent/very good.
Age, gender, race/ethnicity, education, employment, marital status, the length of time with diabetes (i.e., duration), and patient-reported insulin use were accounted for in the analysis. Ethnicity was categorized into Hispanic vs. non-Hispanic. Education included less than high school vs. high school graduate or equivalent (GED) or higher. Marital status was coded as married vs. not-married. Employment status was coded as employed vs. unemployed. Duration of diabetes was included because of its negative association with HRQOL that likely indicates disease progression or severity.11 Patient-reported insulin use was included because some studies have observed that insulin use is associated with HRQOL25–27 and because it may further help account for disease severity.
All analyses were completed with the patient as the unit of analysis. Each patient had his/her own medical home score, even if they shared a physician. This approach reflects the concept that the medical home experience can validly vary even among patients of the same physician or clinic. Because patients were nested within physicians, we estimated our sample size to account for clustering, and examined the necessity of accounting for clustering in our analysis using the intraclass correlation coefficient (ICC) of the total medical home score. The ICC describes the extent to which responses are correlated among patients seeing the same physicians. In this case, the ICC was 0.20, which is a moderate correlation, suggesting that patients of the same physician reported primary care similarly (although not identically), so we adjusted for clustering using survey procedures in Stata13.
Descriptive statistics were obtained for the independent measures and the dependent measures. Next, we examined differences in medical home scores and demographics among categories of the EQ-5D score, dichotomized at the 50th percentile, and the ordinal three-level EQ-5D domains and general health status. To test significance, t tests were used to compare medical home mean scores across the dichotomized EQ-5D index, and post hoc means tests (and F-statistics) were used to compare the mean medical home scores across the ordinal EQ-5D domains and general health status. Chi-squared tests were used to test the relationship between all health measures and the demographics.
Next, two multivariable linear regressions of medical home on the EQ-5D index were performed—one for the total medical home score and then one for the seven medical home features entered simultaneously—both while controlling for study covariates. A set of ordered logistic regressions were performed for the ordinal EQ-5D dimensions and general health status. The ordered odds ratio (OR) is interpreted as the average likelihood of moving between the levels of an ordinal dependent measure. An ordered OR greater than 1.0 would reflect the average odds of moving up one response category of the three (e.g., from no problem to moderate, or from moderate to severe).
Finally, because the sample was heavily female and Hispanic, and because these groups may use health care differently than their counterparts, we examined the potential for the relationship between the total medical home score and HRQOL to vary according to these factors. Where interaction terms were significant (gender), separate regressions for men and women were run. Figure 1 shows the plot of the interaction effect using a graph of fitted values (i.e., the slope of the relationship between the medical home score and EQ-5D index, adjusted for our covariates).