Approach and objective

While analyses at the national level add critical insight, policymakers and key stakeholders can use state-specific data [14] to inform policy-relevant planning and forecasting at more local levels. Such information is valuable in estimating future expenditures and responding to anticipated needs. Further state-based analyses can be especially relevant in identifying high-priority subgroups that may be more at-risk within a given state. For example, states with a relatively larger share of rural residing individuals may benefit from analyses that highlight potential variation in costs by rurality. Such studies that tailor detailed cost-related analyses more closely aligned with state priorities, may provide highly relevant and actionable data that can guide multi-sector planning throughout a given state and be updated to reflect changing priorities over time.

As a specific example, Texas, home to a diverse population residing in a wide range of geographic areas (e.g., remote rural areas, border cities, large metropolitan areas), provides a unique setting for the current study. Diagnosed diabetes costs an estimated $26 billion annually in direct medical expenses and other indirect costs [15]. Given this large diabetes burden, efforts are needed to better understand the drivers of unmanaged diabetes and diabetes-related costs among population subsets. Thus, we sought to investigate potential variation in medical costs associated with diagnosed type 2 diabetes across clinical-related characteristics (e.g., A1c, comorbidities), sociodemographic characteristics (e.g., sex, age), and geographic location of a person’s residence (e.g., rurality) using a Texas-wide commercial insurance dataset. The specific objectives of this study were to: (a) describe the characteristics of commercially insured among adults diagnosed with type 2 diabetes between the ages of 18 and 64 residing in Texas; (b) identify factors associated with their higher A1c values; and (c) identify factors, including A1c values, associated with higher diabetes-related costs. This study adds updated analyses of both clinical indicators such as A1c, cost-related factors, and associated cost estimates for a large southern state.

Data source and sample selection

We used 2018–2019 commercial claims data provided to us by a commercial insurer with a significant presence as the largest private health insurer in Texas. We restricted our analysis to commercially insured individuals 18–64 years of age residing in Texas with an ICD type 2 diabetes diagnostic code during the study period. We restricted the analysis to individuals with valid A1c values between 4 and 14 reported during the study period and provide a supplemental table to show that this restriction eliminated less than 1% of reported A1c values. To identify individuals with type 2 diabetes we used the International Classification of Diseases, Tenth Revision (ICD10), code E11 [16]. Fig 1 illustrates the final analysis sample after all the exclusions resulting in 133,581 individuals with type 2 diabetes.

Download:

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

Fig 1. Consort diagram of patient flow.

https://doi.org/10.1371/journal.pone.0289491.g001

Theoretical framework

Our study goals are influenced by the desire to document and ameliorate health disparities in chronic disease care. The National Institute on Aging Health Disparities Framework (https://www.nia.nih.gov/research/osp/framework) designates priority populations and illustrates different analytical levels associated with illuminating disparities research: environmental, sociocultural, behavioral, and biological. Our research explicitly seeks to understand health care in rural populations and focuses on the interplay of environmental, social and biological factors. While we recognize the importance of behavioral factors in health and health care, given the current data set we were not able to include in-depth health behaviors in our model. However, the available variables fit nicely into the Anderson model [17] with sex as Predisposing, rurality as Enabling, and age, comorbidity, and diabetes complications as Need factors.

Measures

Statistical analyses

The raw data were in a longitudinal format per quarter. To generate individual-level summary statistics, we calculated summary statistics from the 8 quarters in 2018 and 2019 for each participant, including the baseline and median A1Cvalues, median age, highest CCI over the two years, short- and long-term diabetes-related complications over the two years, rurality, and sex. Descriptive statistics including median with interquartile range for continuous variables, and frequency with percentage for discrete variables were provided for all participants. Variables were compared by complication type (i.e., no complications, short-term complications only, long-term complications only, and both short and long-term complications) and rurality (i.e., metro, non-metro). Kruskal-Wallis test and Pearson’s Chi-square test were used for testing continuous and categorical variables respectively.

Distributions of the two outcome variables (i.e., A1c and type 2 diabetes-related cost) were found to be highly skewed to the right. Different transformations were attempted, including log, square root, and box cox transformations. We used the log transformation because it is sufficiently operationalizable, and in the plot of A1c and cost, both variables showed bell-shaped distributions. The raw A1c was between 4 and 14, and the raw cost value was greater or equal to 0. To regulate the data on the log scale, we set the log transformation such that: (1) if the cost was 0 in the raw scale, the cost was also 0 on the log scale (i.e., log of cost = log(cost+1)), and (2) if A1c was 4 in the raw scale, the A1c was 0 on the log scale (i.e., log of A1c = log(A1c-3)).

Longitudinal A1c measurements were modeled using age, sex, rurality, CCI, and diabetes-related complications in generalized linear longitudinal regression models adjusting the observation time, which was one of the 8 quarters in 2018 and 2019. We accounted for the cluster effect of observations of the same individual using generalized estimating equations for longitudinal models [22]. Developed by Liang and colleagues [23], a generalized estimating equation can estimate the parameters of a generalized linear model with possible correlations between observations from different timepoints in longitudinal data. The type 2 diabetes-related costs were modeled in both univariable and multivariable generalized linear longitudinal regression models adjusting the observation time by calendar quarters and covariates. Among the covariates, A1c, CCI, and complications are time-dependent, while other covariates including age, sex, and rurality are not significantly time-dependent. Rurality and diabetes-related complications were also tested in univariable models of type 2 diabetes-related cost. To account for a possible non-linear relationship between A1c and cost, we tested polynomial regression terms of A1c, including the linear and quadratic terms, as well as the interaction between A1c and time. A P-value ≤ of 0.05 was considered statistically significant. All analyses were conducted using SAS software, version 9.4 (SAS Institute, Cary, NC).

Human subjects

This study consists of a secondary data analysis of data from health claims records. It was ruled as “non-human subjects” research by the Texas A&M Institutional Review Board since the data was deidentified. However, we recognized that the primary data set did involve humans.

Limitations

The current analyses included one state, limiting generalizations to other states nationally. Further, the analysis sample includes commercially insured individuals from one insurance company, which further limits its external validity. However, it is important to note that we do have data from the largest private health insurer in Texas, which serves over 6 million members in all 254 TX counties. This plan also offers coverage across multiple lines of business, including employer group programs and Affordable Care Act (ACA) programs. We believe that makes our results applicable not only to TX commercial population, but also to comparable commercial populations in other states with large rural areas. We do concede that the implications of our results for publicly insured individuals (Medicare and Medicaid) are less precise as the sample excluded those 65 and older given medical costs were likely to be attributed elsewhere (e.g., Medicare). However, at the same time, a major strength was the use of statewide data that can inform key stakeholders and provide actionable information with which to use in planning efforts by multiple relevant sectors.

While administrative claims data provide information on a large population group, there are nevertheless inherent issues in this type of study relative to smaller clinically based studies which have their own limitations. For example, claims data do not document when diagnoses are initially made, potentially blurring the relationship between treatment and outcomes. Further, the plateauing of costs at higher A1C levels may reflect better management among those with commercial insurance coverage or, alternatively, that those in poorer control were using less care and that higher cost implications may appear at a later date.

Another potential limitation is the relatively short two-year time frame for this analysis. Although we are examining eight quarters of services, a longer time frame would be desirable to track changes in A1c values over time and their association with health care utilization, complications and costs. The shorter time frame was advantageous in minimizing geographic shifts from baseline that occurred in less than 1% of the population.

Additionally, the sample had more missing data than anticipated, specifically for the primary outcome variable A1c value. Missing lab values can be common in claim-based data even if a lab test was performed [32]. It was impossible to tell if missing data reflected that many enrollees with type 2 diabetes were not getting the recommended minimum of two professional A1c checks annually, or if these tests were done but simply not recorded in the claims database. The latter is a likely explanation since reporting A1c values was not required or reimbursed. A closer analysis of available A1c values indicated that 80% of the participants had two or less A1c values recorded in their administrative claims. Future research is needed to examine the overall health care utilization and costs associated with the frequency of A1c checks, controlling for other risk factors.

Further, information on race, ethnicity, education, and other relevant sociodemographic variables was not available in the data, a limiting factor, but not uncommon when using claims data. To counter this gap, we strongly recommend that administrative databases routinely collect data that can be used to assess social determinants of health. This is critical given the key influence of social and structural factors on inequities [33] and their strong relationship to morbidity and mortality outcomes [34]. Further studies should concentrate on adding Medicare, Medicaid, and uninsured populations to see how their outcomes compare to commercially insured populations. In addition, extending the study period to the post-COVID-19 era would help to analyze the impacts of the pandemic on healthcare costs given an increase in unmanaged diabetes as well as complications such as diabetic ketoacidosis [35].