Mitigating Case Mix Factors by Choice of Glycemic Control Performance Measure Threshold Free

Data for this cross-sectional analysis come from the 1999–2002 National Health and Nutrition Examination Survey (NHANES). This ongoing survey conducted by the National Center for Health Statistics uses a stratified, multistage, probability cluster design that results in a nationally representative sample of the noninstitutionalized, civilian U.S. population. The sampling design oversamples non-Hispanic blacks, Mexican Americans, individuals aged ≥60 years, and low-income individuals to provide more precise estimates for these specific groups. We included 843 adults aged ≥20 years with valid measures of A1C who were not pregnant and who reported during an interview that a health care professional had told them they have diabetes. The NHANES 1999–2002 underwent institutional review board approval, and written informed consent was obtained from participants.

Sociodemographic characteristics and information on diabetes, health, and insurance status were ascertained during interviews. Age-groups were based on approximate quintiles of the analytic sample. Race/ethnicity was based on self-identified race and Hispanic origin. Income-to-poverty ratio was defined as the ratio of reported total family income to the U.S. Census Bureau poverty threshold. Diabetes medication was categorized as use of insulin (with or without oral agents), use of oral agents only, and no use.

Weight and height were measured during a physical examination. BMI was calculated as weight in kilograms divided by the square of height in meters, and weight status was categorized as underweight or normal weight, overweight, obese, and severely obese with BMI cut points of <25, 25−29.9, 30–39.9, and ≥40 kg/m² (17). A1C was assessed by boronate affinity high-performance liquid chromatography using Primus CLC330 and Primus CLC 385 (Primus, Kansas City, MO), and values were standardized according to the method of the Diabetes Control and Complications Trial (18,19), yielding interassay coefficients of variation of <3.0%.

The mean difference provides a measure of disparity among different groups within a population (20). We used the mean difference to assess differences in A1C threshold attainment by various population groups within potential case mix factors. For each factor, we calculated a mean difference to provide a summary of the difference in proportion of people with diabetes whose A1C was >7, >8, and >9%. The mean difference is expressed in percentage points, across any number, p, of population groups. The referent for each threshold, r_ref, was selected as the proportion exceeding the threshold among the overall population. These referent proportions are constant at a given threshold and provide an intuitive standard from which to gauge differences across groups. For each population group, j, we calculated the proportion, r_j, exceeding the threshold, and the absolute difference in that proportion from that of the referent, r_ref. For each factor, the mean difference was calculated using Equation 1.

All calculations were made with Wesvar (version 5; Westat, Rockville, MD) using the 4-year examination sample weights to account for the sample design and nonresponse. Sampling variance was calculated using the jackknife procedure, appropriate for the complex, multistage sample design (21), and 95% confidence limits were calculated by taking the natural logarithm of the mean difference, estimating the standard error of the transformed estimate, deriving confidence limits, and back-transforming the confidence limits via exponentiation. This transformation results in asymmetric confidence limits but improves their validity (22). Student's t tests were used to test whether the differences between the mean differences for the 7 and 8% thresholds relative to the 9% threshold were equal to 0.

The proportion of individuals whose A1C concentrations were >9, >8, and >7% are presented by various patient factors (Table 1). Frequency distributions for each factor across the overall population are provided for reference. Of the overall (unstratified) population, 18.4% exceeded an A1C threshold of 9%, 32.8% exceeded an A1C threshold of 8%, and 52.6% exceeded the ADA recommended target A1C of 7%. These proportions provide the referent to which subgroup threshold attainment is compared. We first present differences in threshold attainment within factors and then describe the effects of lowering the A1C threshold on these differences.

For sociodemographic factors, important differences existed across groups of age, race/ethnicity, and insurance status at each of the A1C thresholds, whereas differences across sex and education were minor. The proportion with A1C >9% declined steadily with increasing age, such that 32.7% of those aged 20–44 years but only 13.4% of those aged 65–74 years and only 5.8% of those aged ≥75 years surpassed the threshold. The age gradient was not explained by a shorter reported duration of diabetes or a lower proportion of insulin use among elderly individuals (data not shown). In comparing each of the five age-groups to the overall population, the mean difference in the proportion with A1C >9% was 7.8 (95% confidence limits 4.4–13.9) percentage points. The difference across race/ethnic groups was of a magnitude similar to that for the age difference at this threshold. Among race/ethnic groups, the proportion exceeding any threshold was much lower for non-Hispanic whites than for any minority race/ethnic group. Almost one-third of those most economically disadvantaged (income-to-poverty ratio ≤0.8) had A1C >9%, but the proportions of individuals in the other income-to-poverty ratio categories with A1C >9% were much lower and ranged from 14.9 to 17.5%, resulting in a mean difference of 5.4 (3.0–9.8) percentage points. Uninsured individuals were more likely to have A1C >9% than those with any insurance. The lower proportion with A1C >9% among those with public insurance (and among those with additional private insurance) appears to be explained by the Medicare eligibility of individuals aged ≥65 years, who have a lower probability of poor control (data not shown).

Lowering the threshold to A1C >8 or >7% did not significantly affect most sociodemographic disparities. The age gradient persisted, and mean differences across age and insurance status were modestly reduced as the threshold was lowered from >9 to >7%. Mean differences across race/ethnic groups at the lower thresholds remained substantial. Lowering the threshold to A1C >8% did not significantly reduce the disparity across income-to-poverty groups, but lowering the threshold to A1C >7% reduced the mean difference to 2.5 (95% confidence limits 0.7–9.0) percentage points, a significant decline from the estimate at >9% (P < 0.05). Although our analysis was framed in terms of the effect on the mean difference across various groups of lowering the threshold for poor control, the results are similar if framed in terms of raising the A1C threshold for good control.

Among health status factors, the largest difference at the >9% threshold was found across diabetes medication categories. Compared with 18.4% overall, more than one-quarter of individuals taking insulin, 16% of those taking oral medications, and 13% of those using no medications had concentrations above this level, yielding a mean difference of 5.7 (95% confidence limits 2.3–13.8) percentage points. The relation of poor glycemic control by diabetes duration was U-shaped, as a lower proportion among those with duration of diabetes <5 and at least 20 years had A1C >9%; the mean difference was 4.9 (2.2–11.1) percentage points. The mean difference for weight status was small at 3.3 (1.0–11.3) percentage points. Only 9.8% of individuals who reported good or excellent health had A1C >9%, whereas 20% of those reporting good or fair to poor health had concentrations above the threshold, yielding a mean difference of 4.2 (2.0–9.1) percentage points for self-reported health. At A1C >9%, mean differences across categories of number of prescription medications and number of health care interactions and hospitalization in the last year were small, at 3.1 (0.8–11.9), 4.1 (1.5–11.1), and 2.1 (0.1–34.5) percentage points, respectively.

Lowering the threshold to A1C >8 or >7% had a large and significant effect on the mean difference across categories of diabetes medication. Compared with 32.8% of the overall population, nearly half of those who used insulin had A1C >8%, and compared with 52.6% overall, >70% of insulin users had A1C >7%. Mean differences were 10.1 (95% confidence limits 6.5–15.7) percentage points at A1C >8% and 14.7 (10.8–20.1) percentage points at A1C >7% (P < 0.05 and P < 0.001, respectively, compared with A1C >9%). The increase in the mean difference for diabetes medication categories at sequentially lower A1C thresholds reflects a differential increase by type of medication in the proportion exceeding the lower thresholds (Fig. 1). When the threshold is lowered to A1C >8%, the proportion of insulin users exceeding the threshold increased by 20 percentage points (to include those with A1C >8–9%), compared with increases of 13 and 9 percentage points for those taking oral agents and those not using medication, respectively. When the threshold is lowered to A1C >7%, the increase in the proportion above this threshold was 23 percentage points for the insulin users and those taking oral agents (after including those with A1C >7–8%), but only 6 percentage points for those using no medication. The mean differences for the remaining health status factors were not significantly affected by lowering the A1C to >8 or >7%.

In Table 2, patient factors were ranked by mean differences at each A1C threshold. Considering a mean difference of ≥5 percentage points as noteworthy, important disparities were consistently observed across all three A1C thresholds for age, race/ethnicity, insurance, and diabetes medication, but their rank varied by threshold. At A1C >9%, insurance, age, and race/ethnicity contributed the greatest mean differences in threshold attainment. At lower thresholds, diabetes medication exhibited the greatest disparity with mean differences of 10 and almost 15 percentage points. The average mean difference across all factors was consistent (4.7–4.8) among all thresholds.

This study demonstrates the extent to which disparities in glycemic control by patient subgroups are mitigated or exacerbated by choice of A1C thresholds of >9, >8, and >7%. For 11 of the 13 factors examined, choice of A1C threshold did not significantly affect the magnitude of population subgroup differences in glycemic control. For example, we observed pronounced race/ethnicity differences in achievement of glycemic control at all three thresholds, a finding consistent with previous research (23). We found no evidence, however, that establishing a lower A1C cutoff as a performance measure would exacerbate this disparity or that selecting a less stringent threshold would obviate the disparity.

Diabetes medication is the major patient factor for which differences in threshold attainment would be exacerbated by lowering the current NDQIA and NQF glycemic control performance measure of A1C >9%. The importance of diabetes treatment for good (A1C <7%) glycemic control has been shown to be independent of sociodemographic characteristics, abdominal obesity, and health care access and utilization (15). The magnitude of this difference at the lower thresholds has implications for pay for performance quality improvement models being considered by the Centers for Medicare & Medicaid Services and other payers. Practices that include a greater proportion of patients receiving insulin may have proportionately more patients exceeding A1C thresholds. A quality improvement system that includes an incentive to achieve a performance measure for glycemic control may not recognize appropriate care given by physicians whose practice includes a larger proportion of people needing insulin for glycemic control, unless the system accounts for this factor. Reporting performance measure attainment specific to diabetes medication use, and insulin use in particular, might be one means to account for this difference in case mix, although implementing this recommendation may not currently be practical under most systems. As health systems move forward with electronic medical records, however, capturing information on medication use will be more feasible. Although automated review of electronic medical records may not consistently capture data on prescription medications (24), performance measures stratified by prescription medication may be a viable solution as long as measures receive adequate field testing (25).

Differences across other patient health factors examined (self-reported health, number of prescription medications taken, and history of hospitalization) were not greatly influenced by the choice of glycemic control performance measure threshold. Of the 13 factors studied, only socioeconomic status (as measured by the income-to-poverty ratio) had a reduced disparity in threshold attainment with a lower A1C threshold.

We observed a sizable difference in threshold attainment by age, with improved glycemic control across increasing age-groups. The effect may be one of survivorship bias whereby life-threatening complications are reduced among individuals with well-controlled diabetes or may reflect milder disease with more β-cell function in elderly individuals. The poorer control in younger patients is ominous as they will have a longer duration of diabetes and thus greater risk of complications. Our findings also suggest that the current NDQIA and NQF performance measure of A1C >9% is unlikely to drive major improvements in glycemic control in the Medicare population as only 13.4% of those aged 65–74 and 5.8% of those aged ≥75 had A1C >9%.

Our analysis of these national data offers a response to a specific health policy question being debated among groups committed to establishing national health care quality measures. Nonetheless, we acknowledge several limitations. Recent evidence suggests that differences in A1C by race may reflect physiological differences such as rates of glycation and erythrocyte survival (26). Fair implementation of threshold-based performance measures will depend on determining whether small differences in A1C threshold attainment truly indicate differences in quality diabetes care after adjustment for case mix. Diagnosed diabetes status was assessed with a validated question that nonetheless has a sensitivity of ∼81% and a positive predictive value of ∼89% (27). Another limitation is that our analysis relied heavily on self-reported case mix factors. The NHANES surveys are designed to result in a representative sample of the noninstitutionalized U.S. population but may underrepresent noninstitutionalized persons with mental health and substance abuse conditions, a group shown to have poor glycemic control (28).

We conclude that implementing an A1C performance measure to indicate poor control that is more stringent than the current measure of A1C >9% would not increase disparities across most patient factors including age, race/ethnic group, and socioeconomic status. To account for a case mix that includes more insulin users, however, stakeholders considering the selection of a performance measure threshold for glycemic control that is ≤8% should consider reporting of threshold attainment specific to diabetes medication use. Stratified reporting would allow more rigorous goals for glycemic therapy in groups for whom a benefit has been established. In so doing, lowering the A1C performance measure threshold may advance the overall quality of diabetes care, without magnifying overall differences among population groups and without discouraging physician acceptance of insulin-requiring patients whose diabetes is more difficult to manage.

This work was supported by the National Institute of Diabetes and Digestive and Kidney Diseases (N001-DK-1-2478 and HHSN267200700001G).

Parts of this study were presented in abstract form at the 66th annual meeting of the American Diabetes Association, Washington, DC, 9–13 June 2006, and a small number of the estimates were presented at the 2007 Joint Statistical Meetings, Salt Lake City, Utah, 29 July–2 August 2007.

We thank Valerie Hsu, BS, from Westat for statistical support and Danita Byrd-Holt, BBA, from Social & Scientific Systems for programming support.