Comparability of patient-reported health status: multi-country analysis of EQ-5D responses in patients with type 2 diabetes

Transcription

1 Comparability of patient-reported health status: multi-country analysis of EQ-5D responses in patients with type 2 diabetes Joshua A Salomon, Anushka Patel, Bruce Neal, Paul Glasziou, Diederick E. Grobbee, John Chalmers, Philip M. Clarke, On behalf of the ADVANCE Collaborative Group Text S1. Technical appendix In this appendix we provide additional details on the methods used to compute predicted probabilities from the regression models, multivariable regression results and sensitivity analyses. Computing predicted probabilities To decompose observed differences in EQ-5D responses across regions into those due to actual variation in health status vs. those due to differential item functioning, we developed three different sets of predictions for the probabilities of reporting problems in specific domains. The starting point for these predictions in a given domain was the set of coefficients estimated in the multivariable logistic regression model, which took the following general form: where p is the probability of reporting problems, ASIA is a dummy variable for patients living in Asia, EEUR is a dummy variable for patients living in Eastern Europe, and X k is an element in the covariate vector X. Parameters to be estimated in the model included the vector of coefficients on the covariates other than region,!; " 0, which may be interpreted as the log odds for a patient in Established Market Economies (EME), the reference region, with all other covariates set at 0; and " 1 and " 2, which are the log odds ratios for patients living in Asia or Eastern Europe, respectively, compared to a patient with the same covariate values living in EME. The three sets of predictions pertained to the following scenarios: (A) no standardization, which preserved differences in both true functional health status and response style; (B) standardized for demographics, risk factors and event history, which isolated the effects of differential item functioning; and (C) standardized for response style, which isolated the effects of demographics, risk factors and event history. Predictions for regions were made by first computing predicted probabilities for individuals; and then computing the average probability over a particular sample of individuals. Table S1 summarizes the prediction equations and the sets of individuals used to define each regional prediction under each of the three scenarios. 1

2 Table S2. Computation of predicted probabilities by region and scenario. Scenario Region EME Asia Eastern Europe A Equation " 0 +!!X " 0 + " 1 +!!X " 0 + " 2 +!!X Sample EME Asia Eastern Europe B Equation " 0 +!!X " 0 + " 1 +!!X " 0 + " 2 +!!X Sample All patients All patients All patients C Equation " 0 + " * +!!X " 0 + " * +!!X " 0 + " * +!!X Notes: Sample EME Asia Eastern Europe EME = Established Market Economies Equations give predicted log-odds. If predicted log-odds equals y, then predicted probability equals exp(y) / [1 + exp(y)]. " * is a standardized region effect computed as v 1 " 1 + v 2 " 2, where v 1 and v 1 are the proportions of the total patient sample in the Asia and Eastern Europe regions, respectively. Because the EME region effect is implicitly 0 in the model, this formulation yields a standardized effect that may be understood as the average response style in the sample. In the base-case analysis, all predicted probabilities were computed with the random effect fixed at 0. As a sensitivity analysis, we computed alternative predictions that included draws from the estimated random effect to account for individual differences not explained by the covariates in the regression. The approaches yield different predictions because the random effect term is additive and has mean 0 in log-odds space, which yields a non-mean-preserving effect (specifically, a shift toward 0.5) once predicted logodds are back-transformed into predicted probabilities. The two approaches invite subtle differences in interpretation of the predictions: for the base case they are predicted probabilities for an average patient having a particular set of covariates, whereas for the alternative approach they are predictions of the average probability across patients who share a particular set of covariates. Notwithstanding the shift in all predicted probabilities toward 0.5 when draws from the random effect were included, the conclusions drawn from comparing regional predictions under the three scenarios were unaffected by the choice of approach to handling the random effect term. 2

3 Multivariable regression results Results from the full multivariable models are reported in Table S2. In the overall summary and in all domains, both patient age and duration since diagnosis of diabetes had significant associations with worse reported health status. Males were significantly less likely than females to report problems in all domains, with odds ratios ranging from 0.42 (95% CI 0.38, 0.48) for anxiety/depression to 0.73 (0.60, 0.88) for self-care. Higher education was associated with significantly lower odds of reporting problems in most domains, but the effects were small. Differences associated with blood pressure and cholesterol were statistically insignificant for most domains, although LDL cholesterol had a significant association with pain/discomfort and anxiety/depression. Higher body mass index was significantly associated with higher reporting of problems in every domain except anxiety/depression, while regular exercise was associated with significant reductions in reporting problems in all domains. Smoking was significantly associated with problems in mobility and anxiety/depression, while weekly alcohol consumption was associated with significant reductions in problems in self-care, usual activities and anxiety/depression. A Mini- Mental-State Examination score meeting the criterion for dementia was associated with a significant increase in odds of reporting problems in all domains except pain/discomfort. Death within the next 90 days was strongly and significantly preceded by increased problems in all domains. A pre-trial history of stroke or heart failure was significantly associated with problems in all domains, with odds ratios >3 for mobility, self-care and usual activities in the case of stroke. Several other variables indicating history of complications at baseline were significantly associated with worse reported health in some but not all domains, including angina, peripheral revascularization, proliferative retinopathy and blindness. The remaining pre-trial history variables failed to attain significance for any of the five domains. Likewise, the importance of within-trial events varied, with stroke, coronary events and peripheral vascular disease having significant associations with at least four domains, but other types of events showing less consistent associations. We evaluated the potential for multicollinearity by computing variance inflation factors for each explanatory variable. All of the factors were less than 2, which is well below the threshold of 10 that is often suggested as a rule of thumb for detecting multicollinearity problems. 3

4 Table S2. Multivariable logistic regression results for reporting problems in EQ-5D Domains. * Variable Mobility Self-care Usual activities Pain/discomfort Anxiety/depression All domains Region EME (reference) Asia 0.26 (0.22,0.31) 0.70 (0.55,0.89) 0.30 (0.25,0.36) 0.98 (0.87,1.10) 1.17 (1.02,1.35) 0.77 (0.68,0.87) Eastern Europe 1.98 (1.69,2.33) 3.67 (2.90,4.64) 1.26 (1.06,1.49) 1.45 (1.27,1.65) 3.48 (2.99,4.06) 1.82 (1.58,2.10) Demographics Sex (male) 0.57 (0.50,0.65) 0.73 (0.60,0.88) 0.57 (0.50,0.66) 0.46 (0.42,0.51) 0.42 (0.38,0.48) 0.44 (0.39,0.48) Age 1.12 (1.11,1.13) 1.11 (1.09,1.12) 1.09 (1.08,1.10) 1.03 (1.03,1.04) 0.99 (0.98,1.00) 1.05 (1.04,1.06) Education 0.98 (0.98,0.99) 0.98 (0.97,0.99) 0.98 (0.97,0.99) 0.99 (0.98,0.99) 0.99 (0.98,1.00) 0.98 (0.98,0.99) Risk factors Duration of diabetes 1.02 (1.01,1.03) 1.02 (1.00,1.03) 1.02 (1.01,1.03) 1.02 (1.01,1.02) 1.02 (1.01,1.03) 1.02 (1.01,1.02) SBP 1.00 (1.00,1.01) 1.00 (0.99,1.00) 1.00 (1.00,1.00) 1.00 (1.00,1.00) 1.00 (1.00,1.00) 1.00 (1.00,1.00) LDL 1.00 (0.95,1.04) 1.03 (0.96,1.10) 0.99 (0.95,1.04) 1.06 (1.03,1.10) 1.05 (1.01,1.09) 1.08 (1.04,1.12) HDL 0.87 (0.76,0.99) 0.99 (0.81,1.21) 1.01 (0.87,1.16) 1.03 (0.93,1.14) 1.01 (0.90,1.14) 0.99 (0.89,1.10) HbA1c 1.03 (1.00,1.07) 1.02 (0.97,1.07) 1.00 (0.96,1.03) 1.03 (1.00,1.05) 1.03 (1.00,1.06) 1.05 (1.02,1.08) Body mass index 1.13 (1.12,1.15) 1.09 (1.07,1.11) 1.09 (1.07,1.10) 1.07 (1.06,1.08) 1.00 (0.99,1.02) 1.07 (1.06,1.08) Exercise 0.41 (0.37,0.45) 0.30 (0.26,0.35) 0.38 (0.34,0.42) 0.67 (0.62,0.73) 0.68 (0.61,0.74) 0.58 (0.53,0.64) Smoking 1.32 (1.13,1.55) 0.97 (0.75,1.26) 1.15 (0.97,1.36) 1.04 (0.92,1.17) 1.23 (1.07,1.42) 1.18 (1.04,1.34) Alcohol 0.92 (0.82,1.03) 0.80 (0.66,0.96) 0.80 (0.71,0.90) 1.04 (0.95,1.13) 0.88 (0.79,0.97) 1.00 (0.91,1.10) MMSE 1.48 (1.13,1.93) 2.97 (2.14,4.13) 2.65 (2.04,3.46) 1.06 (0.84,1.33) 1.73 (1.36,2.21) 1.70 (1.30,2.22) Death < 90 days 2.42 (1.95,3.01) 2.60 (1.94,3.47) 2.42 (1.94,3.02) 1.39 (1.16,1.66) 1.24 (1.01,1.52) 1.56 (1.28,1.89) Event history, baseline Angina 1.26 (1.03,1.53) 1.15 (0.87,1.52) 1.50 (1.23,1.83) 1.37 (1.18,1.59) 1.37 (1.15,1.64) 1.37 (1.16,1.61) TIA 0.77 (0.58,1.02) 0.96 (0.66,1.41) 0.97 (0.74,1.29) 0.90 (0.73,1.11) 1.04 (0.82,1.33) 0.85 (0.68,1.06) MI 1.13 (0.94,1.37) 1.20 (0.92,1.58) 1.21 (0.99,1.47) 0.97 (0.84,1.13) 0.98 (0.82,1.17) 1.00 (0.85,1.18) CABG / PTCA 1.13 (0.90,1.43) 0.85 (0.60,1.19) 1.17 (0.92,1.48) 1.17 (0.98,1.40) 1.03 (0.83,1.28) 1.19 (0.98,1.45) Stroke 3.91 (3.20,4.78) 5.50 (4.22,7.16) 4.01 (3.28,4.91) 1.34 (1.15,1.57) 1.41 (1.18,1.69) 1.90 (1.60,2.25) Heart failure 1.65 (1.19,2.29) 2.16 (1.43,3.26) 1.91 (1.38,2.64) 1.72 (1.32,2.25) 1.60 (1.19,2.16) 1.48 (1.10,1.99) 4

5 Revascularization 3.85 (2.57,5.76) 1.68 (0.96,2.92) 1.58 (1.05,2.38) 1.74 (1.25,2.42) 1.42 (0.97,2.07) 2.36 (1.63,3.42) Amputation 1.53 (0.72,3.23) 1.26 (0.44,3.56) 0.84 (0.38,1.83) 1.16 (0.64,2.08) 0.69 (0.34,1.39) 1.04 (0.55,2.00) Macular oedema 0.90 (0.56,1.45) 0.59 (0.29,1.22) 0.93 (0.57,1.52) 0.81 (0.56,1.17) 0.97 (0.63,1.49) 0.74 (0.50,1.09) Proliferative retinopathy 1.55 (1.10,2.20) 1.50 (0.91,2.45) 1.44 (1.01,2.06) 1.14 (0.88,1.48) 1.18 (0.87,1.60) 1.27 (0.96,1.68) Photocoagulation 1.01 (0.72,1.41) 1.08 (0.68,1.73) 0.85 (0.60,1.19) 0.88 (0.68,1.14) 1.05 (0.77,1.42) 0.97 (0.74,1.29) Blindness 1.87 (1.05,3.32) 1.41 (0.64,3.08) 2.70 (1.53,4.76) 1.93 (1.22,3.04) 1.43 (0.85,2.40) 2.05 (1.23,3.41) Events in trial Coronary 1.60 (1.30,1.96) 1.36 (0.99,1.85) 1.58 (1.28,1.96) 1.40 (1.18,1.67) 1.50 (1.23,1.83) 1.67 (1.37,2.02) Cerebrovascular 3.34 (2.62,4.26) 6.80 (4.97,9.32) 4.23 (3.31,5.40) 1.51 (1.24,1.85) 1.99 (1.60,2.48) 2.06 (1.64,2.59) Heart failure 2.67 (1.89,3.78) 2.92 (1.93,4.40) 3.25 (2.32,4.56) 1.21 (0.90,1.62) 1.16 (0.84,1.59) 1.85 (1.30,2.61) Peripheral vascular 1.56 (1.25,1.95) 1.66 (1.22,2.26) 1.60 (1.28,2.01) 1.36 (1.13,1.63) 1.12 (0.91,1.38) 1.38 (1.13,1.70) Nephropathy 1.26 (0.96,1.64) 2.34 (1.60,3.42) 1.87 (1.42,2.48) 1.09 (0.87,1.37) 1.76 (1.37,2.25) 1.32 (1.03,1.68) Retinopathy 1.33 (1.06,1.68) 2.01 (1.45,2.80) 1.88 (1.48,2.40) 0.95 (0.80,1.14) 1.11 (0.90,1.36) 1.07 (0.88,1.30) * Regression results shown as odds ratios of reporting problems (with 95% confidence intervals). Bold figures indicate significance at p<0.05. Regressions were run on a total sample of 40,968 EQ-5D responses from 11,118 individuals with complete data. Education was missing for 10 participants, duration of diabetes for 2 participants, and EQ-5D scores for 10 participants. Abbreviations: EME, Established Market Economies; SBP, systolic blood pressure; LDL, low-density lipoprotein cholesterol; HDL, high-density lipoprotein cholesterol; HbA1c, hemoglobin A1c; MMSE, Mini Mental State Examination; TIA, transient ischemic attack; MI, myocardial infarction; CABG / PTCA, coronary artery bypass graft or percutaneous transluminal coronary angioplasty. 5

6 Sensitivity analyses Our base-case model was fit to multiple observations per study participant, with a random effect to account for correlations in errors between repeated measures on the same individuals. As a sensitivity analysis, we fit analogous models to the baseline observations only, with and without adjustment for demographic characteristics, risk factors and event history. The adjusted model specification for the baseline-only data differed from the main specification used in the paper by omitting terms for events during the trial, which were all 0 at baseline by definition. Results from this model are reported in Table S3. Differences from the base-case results were minimal. Region effect sizes tended to be smaller in the baseline-only model than in the full model, although this was not uniformly true. One odds ratio coefficient for Asia in the unadjusted model and two in the adjusted model lost significance in the baseline-only analysis. Otherwise, all coefficients in the baseline model had the same direction of effect as in the full model and continued to be significant. We also conducted a sensitivity analysis in which we specified a more parsimonious model. In this alternative specification, we excluded from all models the six baseline event history variables that were not statistically significant at p<0.05 for any domain in the full models. These included transient ischemic attack, myocardial infarction, coronary artery bypass graft or percutaneous transluminal coronary angioplasty, amputation secondary to vascular disease, macular oedema and retinal photocoagulation therapy. The more parsimonious specification had virtually no effect on the estimated region coefficients, with none of the estimated odds ratios for region changing by more than 4 percent in the alternative model compared to the base-case results (results not shown). 6

7 Table S3. Results from logistic regression analyses of reported problems in domains of EQ-5D, by region, with and without adjustment for demographic variables, risk factors and clinical history, baseline measures only. Domain Unadjusted odds ratios Adjusted odds ratios Asia Eastern Europe Asia Eastern Europe Mobility 0.37 (0.33 to 0.41) 1.76 (1.58 to 1.95) 0.50 (0.43 to 0.57) 1.56 (1.38 to 1.77) Self-care 0.48 (0.38 to 0.60) 2.83 (2.37 to 3.38) 0.55 (0.41 to 0.72) 2.38 (1.93 to 2.95) Usual activities 0.36 (0.32 to 0.41) 1.42 (1.26 to 1.60) 0.42 (0.36 to 0.49) 1.19 (1.03 to 1.37) Pain/discomfort 0.93 (0.86 to 1.01) 1.68 (1.52 to 1.87) 1.04 (0.87 to 1.25) 1.47 (1.31 to 1.65) Anxiety/depression 1.21 (1.10 to 1.34) 2.79 (2.50 to 3.11) 1.02 (0.90 to 1.16) 2.23 (1.97 to 2.52) All domains 0.87 (0.80 to 0.94) 2.08 (1.85 to 2.33) 1.00 (0.89 to 1.12) 1.80 (1.59 to 2.05) * Results show odds ratios (with 95% confidence intervals) for Asia and Eastern Europe, compared to Established Market Economies. Bold figures indicate significance at p<0.05. Regressions were run on a total sample of 11,118 individuals with complete data. Education was missing for 10 participants, duration of diabetes for 2 participants, and EQ-5D scores for 10 participants. Adjusted odds ratios controlled for sex, age, education, duration of diabetes, systolic blood pressure, LDL and HDL cholesterol, hemoglobin A1c, body mass index, exercise, smoking, alcohol use, Mini Mental State Examination results, death within the 90 days following the survey, and baseline history for a range of different microvascular and macrovascular events at baseline. 7