Consistency of Hemoglobin A1c Testing and Cardiovascular Outcomes in Medicare Patients With Diabetes

We used the Medicare Physician and Supplier file as well as the Medicare Denominator file, in the years 2002–2009, to identify all patients with diagnosis codes indicative of the presence of diabetes mellitus during that time period. Each patient was followed forward in time for a minimum of 3 calendar years, through the year 2012. Patients were required to have a diagnosis codes for diabetes mellitus in 2 of 3 consecutive years for inclusion in the cohort. We used the year in which the patient entered our cohort to establish the patient's comorbidities using the Charlson score. Because having diabetes mellitus was a requirement for cohort inclusion, we did not include diabetes mellitus with or without end organ damage within the overall Charlson score calculation.¹⁷

We excluded patients less than 65 years of age, greater than 99 years of age, and those not enrolled in fee‐for‐service Medicare plans. Further information was obtained using the denominator file, which contains information about Medicare and Medicaid eligibility, age, sex, race, and disability. We recorded patient ZIP code and the hospital referral region of residence, as described by the Dartmouth Atlas of Health Care.¹⁸ We also linked zip code to the American Community Survey (2006–2010 aggregation) to identify local area median income and poverty status. We used county‐level data from countyhealthrankings.org to obtain measures of area level health: healthy days, smoking, and obesity as described in previous work.¹⁹ Patients left the cohort when they died or ceased enrollment in Medicare's Part A or Part B programs, such as in those who joined a Medicare HMO program such as Medicare Advantage.

Hemoglobin A1c testing is recommended at least annually for patients with diabetes mellitus.²⁰ Within our cohort, we examined whether or not patients had ever undergone hemoglobin A1c testing. We used the CPT codes available for this laboratory test (Data S1). This variable has been used in previous studies using administrative datasets.^{20, 21, 22}

To determine our exposure variable, consistency in hemoglobin A1c testing, we examined how consistently hemoglobin A1c testing was performed for each patient during the first 3 years they were followed in our cohort. Testing consistency was categorized as low (testing in none or 1 year of the first 3 years), medium (testing in any 2 years of the first 3 years), and high (testing in all 3 of the first 3 years). Our analysis considered only patients who had at least 1 physician visit per year during the first 3 years, given that a physician visit would allow for an opportunity for patients to receive hemoglobin A1C testing. Sensitivity analyses where none or 1 test during the first 3 years were analyzed independently were performed, and our findings were similar to those presented herein.

We excluded any patients who died within the first 3 years. Sensitivity analysis were performed, which required a minimum of 4 physician visits per year during the first 3 years. We also performed analyses that considered testing consistency during the entire time each patients appeared in fee‐for‐service Medicare, rather than using the first 3 years of testing as an exposure variable and examining outcomes using survival analysis thereafter. Because findings were similar between these analyses and our outcomes reported herein, we present only the latter strategy in this article.

After using the first 3 years in the cohort to measure the exposure variable, we used all remaining years patients appeared in Medicare claims to measure major adverse cardiovascular events, beginning on the first day of the fourth year. We searched for evidence of death, as well as any of the following cardiovascular events: myocardial infarction, stroke, amputation, and need for a lower‐extremity vascular procedure (codes are shown in Data S1). Death was assessed using the Denominator file. Myocardial infarction was defined using International Classification of Diseases, Ninth Revision (ICD‐9) diagnosis codes as in previous reports.²³ Stroke was defined using ICD‐9 codes, as published previously. We used a 1‐year look back to exclude patients in whom any of the cardiovascular events occurred in the past year.^{23, 24} We recorded the occurrence of a major lower‐extremity amputation at the patient level, using current procedural terminology codes indicative of above‐ or below‐knee amputation. We excluded toe and forefoot amputations, and traumatic amputations, although in sensitivity analyses, our results remained similar when we included toe or forefoot amputations in our analysis. Leg revascularization procedures were also measured using diagnosis and procedure codes reported in our previous work^{20, 25} A composite outcome of a major adverse cardiac event (MACE) was analyzed as well, defined as the occurrence of any of the following: death, myocardial infarction, stroke, major leg amputation, or need for lower‐extremity revascularization.

All cardiovascular outcomes were assessed using time‐to‐event analyses, with the initial time period beginning on the first day of the fourth year; the first 3 years in the cohort were used to assign the exposure. Death was allowed in the fourth in the cohort and thereafter. Any major adverse cardiovascular events occurring in the first 3 years were excluded. If patients left fee‐for‐service Medicare claims for Medicare Advantage or other non‐fee‐for‐service programs during this interval, they were censored on the date they ceased to appear in the fee‐for‐service Medicare program.

We began by examining the consistency in hemoglobin A1c testing among Medicare patients with diabetes mellitus between 2002 and 2012. We created Cox survival models to understand associations between the consistency of testing and cardiovascular outcomes. These models were adjusted for age, sex, race, Medicaid eligibility, disability status, and Charlson score as well as regional variables indicative of health status, income, and poverty.

Crude results were examined using linear analyses before examining our 3 consistency categories and over time. Because patient characteristics differed across the categories of consistency in hemoglobin A1c testing (Table 1), we used multilevel inverse propensity weighting to develop a matched cohort of patients, based on the patient's likelihood to receive low‐, medium‐, and high‐consistency testing.^{26, 27} We developed multinomial logistic models that identified patient and structural factors associated with each category of consistency in hemoglobin A1c testing. We then used the inverse of these probabilities to weight patients and balance the testing groups. Models were run using both baseline information only, as well as allowing covariates, including the year of testing, to change annually as the patient progressed through each year in the study. Both efforts produced similar results, and therefore the baseline‐adjusted models are presented herein. P values are reported across all three categories.

All analyses were performed using SAS (SAS Institute Inc., Cary, NC) and STATA software (StataCorp LP, College Station, TX). The Geisel School of Medicine's Center for the Protection of Human Subjects approved our study. Informed consent was waived as part of the study, because it involved only secondary data‐set analyses.

Between 2002 and 2009, we identified 1 574 415 individual Medicare patients with diabetes mellitus. These patients were followed for a mean of 6.3 years in our cohort, with a range from 3 to 11 years. Overall, 70.2% of patients received high‐consistency testing, 17.6% of patients received medium‐consistency testing, and 12.2% of patients received low‐consistency testing. Testing consistency in the first 3 years was reflected in testing in later years. For example, those with high‐consistency testing in the first 3 years had testing in 88% of later years, and those with low‐consistency testing during the first 3 years had testing in 49% of later years.

Patients were who received low‐consistency testing were older than patients receiving high‐consistency testing when they entered the cohort; the mean age was nearly 2 years older for those receiving low‐consistency testing when compared to those receiving high‐consistency testing (76.3 vs 74.6 years; P<0.001; Table 1). Differences by race were evident as well, as 14.3% of low‐consistency testing patients were black, whereas 10.0% of high‐consistency testing patients were black (P<0.0001). Finally, a larger proportion of patients getting low‐consistency testing were on disability when compared to high‐consistency testing (15.5% vs 11.9%; P<0.0001).

Because a patient clinic visit is an opportunity for a physician to order hemoglobin A1c testing, we examined patient visit patterns in our cohort. Overall, patients had an average of 17 physician visits in each year during the study period. Of these visits, an average of 8 were with a primary care physician. Patients were seen often by physicians; only 9% of Medicare enrollees in the cohort had fewer than 4 physician visits annually. Differences in the number of visits were evident across categories of testing consistency (Table 1). For example, patients who received low‐consistency testing had a higher number of physician visits than those receiving high‐consistency testing (19.7 vs 16.3 visits; P<0.001), but were also more likely to have fewer than 4 physician visits per year (11.1% vs 8.6%; P<0.001).

In unweighted analyses, we found 62.3% of patients treated with low‐consistency testing experienced death or a major adverse cardiovascular event within 7 years of follow‐up (Figure A). The rate of death or an adverse cardiovascular event was 13.2% lower, in absolute terms, for patients treated with high‐consistency testing (49.0%), a difference that was highly significant across testing consistency categories (log rank, P<0.001). Similar trends were observed for all individual components of our adverse cardiovascular outcomes, including death, myocardial infarction, stroke, lower‐extremity vascular procedures, and leg amputation (FigureB througF). We calculated unweighted hazard ratios, with surrounding 95% CIs, for each outcome, across consistency categories. These demonstrated an inverse relationship between testing consistency and the risk of death or a major adverse cardiovascular event, as well as each of its individual components (Table 2).

Because of the differences in patient characteristics across categories of testing consistency, we used inverse propensity weighting to generate 3 groups, which were similar across patient characteristics (Table 1). Many of these differences remained statistically significant given our large sample size, but all patient demographic characteristics in the inverse propensity weighted analyses varied by less than 1% across testing consistency categories.

As with our unweighted findings, Cox proportional hazards models derived from the inverse propensity‐weighted cohort again demonstrated that low‐consistency testing remained associated with worse cardiovascular outcomes, both in our composite outcome (MACE) and its individual components (Table 2). For example, patients receiving low‐consistency testing were 16% more likely to experience a cardiovascular adverse event than those who had high‐consistency testing (adjusted hazard ratio [HR]=1.16; 95% CI, 1.15–1.17; P<0.001). These findings were again similar for each of the components of our composite outcome.

Last, to ensure we accounted for differences in patient visit type and frequency, we repeated these analyses, but limited to patients who were seen by physicians at least 4 times per year to ensure that physicians had several opportunities to order hemoglobin A1c testing. As with our unweighted and inverse propensity‐weighted analyses, we again found that low‐consistency testing was associated with a higher risk of death and major adverse cardiovascular events (Table 2), with little change in the effect size evident in this sensitivity analysis.