Baseline comparison of three health utility measures and the feeling thermometer among participants in the action to control cardiovascular risk in diabetes trial

The SF-36 was converted into the SF-6D HU score, using techniques previously described [32]. The HUI was scored according to both the HUI2 and the HUI3 algorithms, resulting in two HU scores [3]. The FT represents how the patient feels on a 0 to 100 scale. We converted the FT values to 0.00 to 1.00 (dividing by 100) because values range from below 0.00 to 1.00 for the other instruments in the study, Demographics, physiologic and laboratory measures, complications, and comorbidities were obtained from baseline case report forms. All data have been edited according to data capture and verification procedures established for the ACCORD trial. For all statistical analyses, a two-sided alpha was set at 0.05.

For our first objective, we calculated pair-wise intra-class correlation coefficients (ICCs) using one-way random models to determine level of agreement between HU instrument scores as well as FT values [33]. We also divided the scores into quartiles, and determined whether values from each HU or the FT, fell within the same quartiles. We analyzed the level of agreement between the pairs of instruments using chi square tests. We also calculated the mean differences per patient (and 95% confidence intervals) between pairs of instruments to determine the extent of variation between the HUs.

For the second objective, we developed multiple regression models with HUs or FT values as dependent variables and demographics, clinical measures, and diabetes complications as independent variables. Categorical variables were entered into the model as indicator variables (for example, presence of characteristics = 1, absence = 0). We used a stepwise, forward selection process with p < 0.5 and report the variables that were significant contributors to the model at p < 0.05 All data computation and analyses were performed using SAS Software, V9.1 (SAS Institute). We tested normality of the distributions of HUs and FT using Kolmogorov-Smirnov tests and assessed Goodness of Fit for the models based on the significant F values.

We note that each sample of patients are dissimilar and thus are not directly comparable to our results, however we did identify similarities to previous studies. The range of mean HU values across the instruments (0.684 to 0.818) in our study were similar to those reported in Action in Diabetes and Vascular Disease study (ADVANCE) which used the SF-6D and the EQ5D (0.678 to 0.801) [35]. The mean FT value in our study (0.748) is similar to the mean value reported among participants of United Kingdom Progressive Diabetes Study (0.74) [36] and among patients with obesity (0.751) [27]. However, our mean FT value was higher than the value reported in the Cost of Diabetes in Europe -Type 2 (CODE-2) (0.628, converted from 62.8 to have similar decimal places to our results) [37] and the FT value found by Matza, et al. among patients with diabetes (0.623) [38]. The differences may reflect study instrument administration techniques as well as differences in study samples. In a study designed to determine the impact of hypothetical diabetes medication outcomes, the patients’ FT scores from diabetes were lower than ACCORD participants, who had higher rates of hypertension (85.4% vs. 37.2%), while body mass index (BMI) was similar (32.2 vs. 31.3) [38]. A study using time-tradeoff measures found a mean HU of 0.76 for conventional glucose control [18], which is similar the FT HU obtained from ACCORD participants at baseline.

The cumulative distributions (Figure 1) help elucidate differences between the methods. The FT scores were concentrated at the interval values listed on the instrument, for example, multiples of 0.05 or 0.10. This finding may suggest lower sensitivity to changes in HRQOL smaller than 0.05, which is smaller than what has been suggested as a clinically important difference in HU (0.03) [39]. The limitations of the FT when compared to HU measures have been previously described [13, 40].

HUs obtained from the SF-6D varied by the smallest range among the middle 50% of the patients (from quartiles >25% to <76%), the difference was only 0.12 points for the SF-6D versus 0.21 for FT, 0.18 for HUI2, and 0.35 for HUI3. The finding suggests that the scoring algorithm for the SF-6D may be less sensitive to differences in HUs among participants whose scores are within this range. A narrower range of scores for the SF-6D has been previously documented when compared to the EQ-5D, and is considered a potential limitation of the SF-6D [41, 42]. The narrow scoring range of the SF-6D was also demonstrated among rheumatoid arthritis participants when compared with HUI3 and EQ-5D [43] and among participants in an implantable defibrillator study when compared with the HUI3 [17]. The HUI2 scoring range was also narrow among the middle 50% of patients (Figure 1), as was previously shown among rheumatoid arthritis patients [43]. The HUI3 cumulative distribution plot is the most gradual across the mid-range scores, suggesting more differentiation between participants within the middle quartiles.

We note several differences between the HUI 2 and HUI3. The HUI3 includes scales for vision, hearing, and speech versus a sensation domain for the HUI2; has separate domains for dexterity and ambulation versus mobility for HUI2; and uses different questions for emotion and pain. Therefore the domain scores are not directly comparable. Only the domain of cognition uses the same questions, but cognition is scored differently between the 2 instruments (6 levels in HUI3 compared to 4 levels in HUI2) Furthermore, the domain scores are then entered into different scoring algorithms for the HUI2 and HUI3, resulting in different values. We note that since the HUI3 differentiates patients more broadly within the middle quartiles, it may be a better method for scoring the HUI in the ACCORD population. The ACCORD CEA sub-study planning committee selected the HUI3 a priori [30].

The ICCs found between HU and FT values, with the exception of the expected higher value between the HUI2 and HUI3 scoring algorithms, provide a summary statistic showing poor or fair agreement between the instruments. Fair agreement between SF-6D and HUI3 was found among patients in an implantable cardiac defibrillator trial (ICC = 0.45) [17] and a percutaneous coronary intervention trial (ICC = 0.40) [14]. Our results by quartile describe these discrepancies more specifically. Comparing the FT and HUI2, 6.1% (n = 123) participants would be measured as being in the highest quartile by one instrument, while scoring in the lowest quartile of the other (Figure 2). Furthermore, an additional 19.2% (n = 383) of comparisons between the FT and the HUI2 differ by two quartiles.

The mean (± standard deviations) differences per participant showed similar results, with average differences from 0.100 (HUI2/HUI3) to −0.122 (SF-6D/HUI2) and 95% confidence intervals as great as −0465 to 0.417 (SF-6D/HUI2). These large discrepancies indicate that the choice of HU instrument could impact results of the overall CEA. Specifically, one instrument might show increased HU over time while another shows a negative or no impact in the same participant. Such discrepancies between HUs have been identified previously among a primary care population in East Asia [44] and among rheumatoid arthritis patients in British Columbia [43]. A longitudinal analysis is needed to determine the full impact of these discrepancies in regard to sensitivity to changes in physiologic diabetes measurements (e.g. glycated hemoglobin or cardiovascular complications).

In multivariable analyses, we identified significant relationships between HUs and FT values. Comorbidities negatively associated with HUs were presence of CVD, current smoking, and obesity measured by BMI or waist circumference. Either waist circumference or BMI were significant for all HU instruments and both were significant for FT. When BMI was significant it may have addressed the variance in HUs associated with waist circumference and vice versa. Previously, in a study of the impact of long-term diabetic complications on HRQOL, BMI was a significant predictor in all regression analyses of SF-36 domains with the exception of mental health [45]. Similarly, in CODE-2, obesity was a significant predictor of VAS scores [37]. Relationships between VAS and obesity have also been shown among patients with obesity without a diagnosis of diabetes [46]. History of CVD was significantly associated with lower HU and FT values. Significant relationships between HU and CVD among patients with diabetes have been shown in other studies [35]‐[37, 45].

Among physiologic measures, total cholesterol and low-density lipoprotein were significantly associated with lower values for all HU instruments, but not for the FT. Glycated hemoglobin was only associated with FT values. None of the renal function measures (serum creatinine, micro- and macro-albuminuria) were significant in any models. Regarding use of blood pressure, lipid, and glycemic medications; none were significant in multivariable models. Our models were similar to a study of diabetes-related complications, which used the EQ-5D in 1143 Canadian participants [34]. Specifically, the researchers reported significant relationships between HU and duration of diabetes (negative), age and male gender (both positive), and CVD complications (myocardial infarction and stroke, negative).

Recently a simulation study was conducted to demonstrate the impact of complications on life expectancy among patients with Type 2 diabetes using data from the Fenofibrate Intervention and Event Lowering in Diabetes (FIELD) study [47]. The simulation study showed that an HU less than 1.00 at baseline was associated with increased all-cause mortality and lower quality adjusted life expectancy [47]. Occurrences of diabetic complications were associated with a mean decrease of 0.045 HU (95% CI = −0.073 to −0.017), as measured by the EQ-5D [47]. The greatest impact on HU was stroke (−0.165 HU. 95% CI = −0.246 to −0.0840). Similarly, in our multivariable models we found secondary CVD to be associated with significantly lower HU at baseline (Tables 2 and 3). The association varied by type of measure, SF- 6D (−0.015), HUI3 (−0.064), and HUI2 (−0.028).

We note that clinical and demographic factors associated with HU are similar to results of an observational trial of predictors of hypertension management [48]; in which persons with diabetes, obesity and Hispanic ethnicity were found to have decreased blood pressure control. The study found the lowest percent of patients with controlled blood pressure control (23%) among diabetic persons with obesity [48].

Adequate goal attainment of CVD risk factors continues to be illusive among persons with Type 2 diabetes. In a study of the data from National Health and Nutritional Examination Survey from 1999 to 2008, goal attainment improved significantly for low density lipoprotein (LDL), from 29.7% to 54.4%, but control of hypertension did not significantly improve (47.6% to 55.1%; P = 0.1333) even though significantly more patients were receiving antihypertensive medications (35.4% to 58.9%; P < 0.0001) [49]. Prevalence of hypertension was not significantly increased from 1999 to 2008 (66.6% to 74.2%; P = 0.3724) [49].

Another concern is under-diagnosis of diabetes among CVD patients. Researchers reviewed health records of all Danish myocardial infarction (MI) patients who were not previously diagnosed with diabetes to identify the initiation of glucose lowering medications within 1 year after discharge [50]. The rates increased from 19.6 to 27.6 per 1000 person year from 1997 to 2001, at which time the rates leveled off through 2005 [50]. These rates were much lower than expected, since other researchers had shown higher rates of abnormal glucose tolerance among MI patients. However a recent population study of screening for diabetes and CVD found no difference in HU among screened versus non-screened populations [51].

A limitation of the study is that ACCORD participants are a select group (i.e. mean age > 62, type II diabetics who met study inclusion criteria and were at risk for CVD); thus, our results are not generalizable to all other individuals with type 2 diabetes. Further research would be needed to make comparisons to other patient groups with type 2 diabetes. This analysis is limited to baseline measures only; our results do not indicate how values may be influenced by changes in diabetes severity or the study interventions over time.

We identified significant differences in HU values obtained from the SF6D, HUI2 and HUI3. Since differences in HU values could impact CEA results, the type of patient preference measure used is an important consideration in designing and interpreting CEAs. Although we found statistically significant relationships between HUs and demographic and clinical variables, the variances explained by the models were relatively small (6.1% to 7.7%).