Measurement theory and provider profiling Timothy P. Hofer MD Dept. Medicine, University of Michigan VA Ann Arbor HSR&D Center of Excellence The measurement problem construct indicator Quality ei(jk) Levels of care Site (clinic, hospital) Provider (physician) Patient i i i Indicators Implications of the measurement model The indicator is a fallible measure of the construct Some indicators are less precise than others Quality indicators are very imprecise for a variety of reasons You need to account for the measurement error The location of the construct variability can suggest different causes, interventions and measurement procedures Intra-class correlation(=reliability) Ability to distinguish between physicians (or sites) single observation under a specified set of conditions of measurement. physician Rx 2 2 2 site physician patient 2 Vol. 281 No. 22, pp. 2065-2160, June 9, 1999 MD laboratory utilization profiles Deviations of MD profiles from mean laboratory utilization ($) Unadjusted for reliability 60 20 -20 -60 Adjusted for reliability Table 3: Amount of variation in hospitalization and outpatient visit rates attributable to a physician practice style effect Age-Sex Adjusted Casemix Adjusted 13% 10% 7% 4% 0.51 0.40 Unadjusted for reliability (R2) * 8% 8% Reliability adjusted (ICC) † 2% 1% 0.24 0.17 Variable Visits % of variation associated w/ physician Unadjusted for reliability(R2) * Reliability adjusted (ICC) † Reliability‡ Hospitalizations % of variation associated w/ physician Reliability‡ VA Network 11 Diabetes Care Project Resources available VA Diabetes Registry Project (1998-2001) Automated Clinical Databases Data warehouse (VA Healthcare and analysis group) Database Components Encounter records (OPC/PTF ) Outpatient Pharmacy Lab primary care provider database (PCMM () Vitals Cohort identification procedure Data quality and measure validation Kerr EA , et al. Journal on Quality Improvement 2002; 28(10):555-65. Selected Measures: Resource Use Cost of hypoglycemic medications Cost of home glucose monitoring for patients not on insulin Cost of calcium channel blockers Quality Outcomes Intermediate Outcomes Processes Selected Measures : Intermediate Outcomes Last A1c value A1c 9.5% Last LDL value LDL 3.6 mmol/L (140mg/dl) Quality Outcomes Intermediate Outcomes Processes Selected Measures: Process Measures Hemoglobin A1c obtained LDL-C obtained Lipid profile obtained Quality Outcomes Intermediate Outcomes Processes Selected Measures: Mixed or Linked Measure LDL 3.6 mmol/L (140mg/dl) or on a statin Are there differences between physicians? What are the sources of variation? Noise Unmeasured differences Physician effects Noise Clinic or group effects Physician Health System/payor effects Clinic System/Payor Outcomes Diabetes Care Indicator Level of Care Facility Team PCP 1% 0 2% Cost of home glucose monitoring for patients not on insulin 18% 3% 3% Cost of home glucose monitoring for patients on insulin 8% 2% 1% Cost of Calcium Channel Blockers 1% 0 0 Resource Use Cost of hypoglycemic medications Intermediate outcomes Diabetes Care Indicator Level of Care Facility Team PCP 12% 0 1% Last HbA1c value 9.5% 16% 0 0 Last LDL-C value‡ 7% – 1% Last LDL-C value 3.6 mmol/L (140mg/dL)‡ 2% 1% 1% Last LDL-C value < 3.6 mmol/L (140 mg/dL) or on a statin (%)‡ 2% 2% 5% Intermediate Outcomes Last HbA1c value Process measures Diabetes Care Indicator Process Measures† Hemoglobin A1c (HbA1c) obtained Low Density Lipoprotein Cholesterol (LDL-C) obtained‡ Lipid profile obtained Level of Care Facility Team PCP - 1% 8% – 2% 8% 7% 2% 9% Physician effect size Physician effect size Negligible PCP Effect 200 Last LDL-C Value (1%) Panel size 150 100 Small Moderate Cost of home glucose monitoring for patients not on Insulin Last LDL-C value <3.6 mmol/L or on a statin (5%) Hemoglobin A1c obtained (8%) 50 Median PCP Panel size in study sample 0 .02 .04 .08 Variance attributable to level of care .10 Implicit chart review – site level .8 1 (to detect differences between sites) .6 Site level reliability .4 HTN 0.17 DM 0.10 COPD 0.11 .2 How many reviews are needed? 0 Trained physician reviewers 621 records 26 clinical sites Reliability of site level quality score 0 10 20 30 cases 40 50 Conclusions Measurement models are fundamentally important to measuring and profiling quality. There is often little reason or capability to profile at the physician level. Profiles that ignore measurement error Misrepresent the variability in quality Are difficult (or impossible) to validate Example – the imprecise thermometer Budget cuts inspire innovation in the clinic 85 90 95 100 105 Observed temperature observed 85 90 95 100 105 Observed vs. true temperature true observed 85 Body temperature(F) 90 95 100 105 Strength in numbers true observed average 0 .2 .4 .6 .8 1 Scale transformation 1 2 3 4 Rating of preventability 5 6 Reliability “A person with one watch knows what time it is” “A person with two watches is never quite sure” Effect of gaming Outlier Physicians (1991) Deviations of MD profiles from mean Hgb A1c levels 0.4 0.2 0.0 These physicians eliminate from their 1992 panels the patients who in 1991 had HgbA1c levels -0.2 in the top 5%. -0.4 Non-outlier Physicians (1991) 0.4 0.2 These physicians have the same patient panels in 1992 as in 1991. 0.0 -0.2 -0.4 1991 1992