The Effect of Depression and Rurality on Diabetes Control

Data

The Minnesota Department of Health contracts with MN Community Measurement (MNCM), a nonprofit organization, to manage the quality reporting.¹⁵ For 2017 encounters, the quality reporting for diabetes was done through electronic data submission from 604 primary care and endocrinology practices in 100 parent medical groups. MNCM provided the patient-level diabetes data collected from primary care practices for use in our study. Our data contained HbA1c values for 1,697,173 patient-year records collected from 2010 to 2017, an average of 212,000 patients per year. This sample of adults aged 18 to 75 years includes the majority of Minnesotans with diabetes in this age range (currently estimated to be 330,000 across all ages).² Because the encryption scheme for patient IDs changed annually, we could not group together records that were repeated observations of the same patient over different years. Therefore, we treated this as repeated cross-sectional data; the statistical implications of this treatment are discussed below. This study was reviewed and approved by the University of Minnesota’s Institutional Review Board.

Our dependent variable was glycemic control, defined as HbA1c < 8%. Our independent variables of interest were depression status and patient’s residential location. Practices submitting these data provided an indicator of patient’s current depression status. MNCM’s data submission guidelines recommend the Major Depression or Dysthymia (DEP-01) Value Set to define depression, but accept any documentation of a new or existing diagnosis of depression during the measurement period. Patient ZIP codes were mapped to Rural-Urban Commuting Areas (RUCA), and summarized to a 3-category variable of residential location describing where the patient lived 1) urban area, 2) large rural town, or 3) small rural town.¹⁶ RUCA coding examines both population density and commuting patterns to determine classification.

Patient-level covariates included: age, sex, diagnosis of ischemic vascular disease (IVD), diabetes type, and type of insurance coverage (commercial, Medicare, Medicaid, dual Medicare/Medicaid, self-pay, and unknown coverage). The indicator of IVD is provided by the practice. MNCM’s data submission guidelines recommend the Ischemic Vascular Disease Value Set to define an IVD diagnosis, but accept any documentation of a new or existing diagnosis in the measurement or prior years. We formatted patient age as a categorical variable to allow for nonlinear effects (18 to 44 years, 45 to 54 years, 55 to 64 years, 65 to 75 years). Because the data from MNCM do not include patient socioeconomic variables, we used 5-year average American Community Survey (ACS) data (2011 to 2015), mapped to patient ZIP codes, to describe the neighborhood in which the patient lived. From these data we created a neighborhood race/ethnicity variable (percent of population non-Hispanic White), a neighborhood poverty variable (percent of households with income under the federal poverty level) and the distribution of educational achievement in the neighborhood (percent of adults age 25 or greater with less than high school diploma, high school diploma or GED but no 4-year college degree, or a 4-year college degree).

Practice-level covariates included an indicator of Minnesota certification as a patient-centered medical home. In addition, we included the size of the medical group that owned the primary care practice. Size categories included single-site practices, practices in small medical groups (owning 2 to 11 primary care practices) and practices in large medical groups (owning 12 or more primary care practices).

We excluded records for patients with care managed by an endocrinology practice (85,058 records) because MNCM provided data only for locations reporting primary care providers onsite, in keeping with the goals of the UNITED project. Thus, we could not assume we had a representative sample of endocrinology practices in the state. In addition, endocrinology practices are rare in rural areas, so including patients reported by endocrinology practices might bias our results. We also excluded records for patients with missing data (19,333 records). Missing data were almost exclusively from incomplete matches with RUCA and ACS data. Together, eliminating data reported by endocrinology practices and records with missing data reduced our 8-year sample from 1,801,564 to 1,697,173 records, a 5.8% reduction in sample size. Because we restricted our sample to those with care managed by a primary care practice, patients in omitted records tended to be younger, with an associated higher prevalence of type 1 diabetes and less Medicare coverage.

Statistical Analysis

We modeled achievement of glycemic control (HbA1c < 8%) in a logit framework as a function of depression and residential location, and the interaction between having depression and living in a rural area. We estimated a baseline model without this interaction to provide a direct comparison with the existing literature that tested association between glycemic control and depression, or glycemic control and rurality. The logit regression models were adjusted for annual trend as a continuous variable, and the patient-, practice- and neighborhood-level characteristics described above. We included practice-level random effects to capture unobserved time-invariant practice and patient characteristics. We conducted the statistical analysis with Stata 15 statistical software using α = 0.05 as the cutoff for statistical significance.¹⁷

A positive coefficient from our logit regressions indicates an increase in the probability of glycemic control as that coefficient’s variable increases, but does not tell the reader how large that increase in probability is. For this reason, logit results are often presented as odds ratios to give the reader a more intuitive understanding of the change in likelihood of the outcome. But in our results, we have the added complexity of interaction terms, a complexity that is not addressed by a simple transformation from coefficients to odds ratios. Therefore, to illustrate the impact of a change in depression status on the probability of glycemic control, we estimated the average percentage-point difference in the probability of glycemic control with and without depression, presented below in Figure 1. In other words, the models were used to compute the average expected change in the probability that HbA1c < 8% when depression status changes from “no depression” to “depression.” This was done multiple times, holding residential location fixed at each of the 3 possible categories (urban area, large rural town, small rural town), to demonstrate how these changes in probability were affected by residential location. All other patient, practice and neighborhood characteristics were held at their true value. A negative percentage-point change in probability would indicate that glycemic control was predicted to get worse with depression diagnosis.

• Download figure
• Open in new tab

Figure 1.

Predicted change in the probability of achieving HbA1c  < 8% associated with change in depression status by residential location. Note: Predicted changes in probabilities are based on the multivariate logit models. The models are used to compute the average expected change in the probability that HbA1c < 8% when patient depression status changes from “no depression” to “depression,” assuming the residential location is fixed at the level described and all other patient characteristics are held at their true value. The negative changes in probability indicate that glycemic control worsens when a patient is depressed, though the level of this decline may vary by patient location. Error bars indicate the 95% CIs for the estimated changes in probability.