Estimating the Completeness of Physician Billing Claims for Diabetes Case Ascertainment Lisa M. Lix, Xue Yao, George Kephart, Khokan Sikdar, Hude Quan, J. Paul Kuwornu, Wilfrid Kouokam, Mark Smith IHDLN Meeting Vancouver April 27-30, 2014
Overview of Presentation Background Objective Methods Results Conclusions Key Assumptions & Future Research
Background Physicians have traditionally been remunerated by the fee-for-service (FFS) method Physicians submit billing claims to their ministry of health for all services provided to patients These claims are captured in electronic administrative databases An increasing number of physicians are remunerated by non-fee-for-service (NFFS) methods, such as salaries and contracts Shadow billing is the practice by which NFFS physicians submit claims, using the same procedures as FFS physicians Not all jurisdictions require physicians to shadow bill
Background Gaps in administrative databases associated with a lack of shadow billing may result in biased estimates of: Healthcare use Chronic disease prevalence and incidence rates Existing statistical models could be used to estimate the bias in disease prevalence and incidence rates. Examples of these models include: Missing data models Capture-recapture models Prediction models
Objective To develop a population-based model to predict prevalent diabetes cases from FFS physician billing claims and apply it to estimate missed diabetes cases seen by NFFS physicians for whom billing claims data are incomplete. We focus on diabetes because administrative health data have good sensitivity and specificity to ascertain diabetes cases.
Overview of Model Development Identify Diabetes Case Cohort Link Diabetes Case Cohort to Physician Registry Model the Number of Diabetes Cases per FFS Physician Predict the Number of Diabetes Cases per NFFS Physician Sum Observed and Predicted Cases to Estimate Prevalence
Prediction Model Data Newfoundland and Labrador administrative data Physicians do not shadow bill The province has a high proportion of NFFS physicians Specific data sources Physician billing claims, hospital discharge abstracts, physician registry records, insured resident registry records Fiscal years 2002/03 to 2003/04
Constructing the Diabetes Case and Physician Cohorts Diabetes Case Cohort: 20+ years Case definition: one hospitalization or two physician billing claims within a 730-day period Diagnosis codes: ICD-9 250 or ICD-10-CA E10 to E14 Physician Cohort: Individuals in the physician registry who had submitted a billing claim for at least one individual in the Diabetes Case Cohort Physicians were assigned to cases in the Diabetes Case Cohort using the physician identification number on the billing claim for the case-qualifying diagnosis
Selecting Diabetes Cases for the Prediction Model Diabetes cases were classified into three mutually exclusive groups Group 1: cases ascertained only from hospital discharge abstracts Used to Build the Prediction Model Group 2: cases ascertained from physician billing claims for whom the case-qualifying diagnosis was from a FFS physician Group 3: cases ascertained from physician billing claims for whom the case-qualifying diagnosis was from a NFFS physician In Newfoundland & Labrador, a small number of NFFS physicians also receive some remuneration by the FFS method
Prediction Model Generalized linear regression model with gamma distribution Dependent variable: number of cases in the Diabetes Case Cohort per FFS physician in the Physician Cohort Covariates years since specialty licensure (quartiles; reference = lowest quartile) sex (reference = female) region of practice (reference = Labrador, a remote rural region) specialty vs. general practitioner (reference = specialist) External validation Same model applied to data from Manitoba Many physicians shadow bill This province has a lower proportion of NFFS physicians than Newfoundland & Labrador
Characteristics of Diabetes Case Cohort 100% 90% Percent of Total Diabetes Cases 80% 70% 60% 50% 40% 30% 20% Group 2 N=28,89 8 Group 2 N=42,93 3 10% 0% Newfoundland & Labrador Manitoba Group 1: Cases from Hospital Data Group 2: Cases from FFS Physician Data Group 3: Cases from NFFS Physician Data
Number of Diabetes Cases per FFS Physician in Newfoundland & Labrador Physician Mean (SD) Median characteristics Overall 75.5 (84.6) 49.0 Specialty General practitioner 79.0 (66.2) 66.0 Specialist 61.0 (136.8) 9.0 Sex Male 89.3 (94.1) 75.0 Female 41.5 (37.2) 32.5 Age group < 40 years 54.9 (64.8) 32.5 40 64 years 99.9 (98.3) 91.0 65+ years 63.8 (68.6) 34.5 Health region Eastern 67.8 (73.1) 42.0 Central 87.8 (87.7) 59.0 Western 108.1 (129.5) 86.0 Labrador 47.6 (47.9) 38.5
Regression Parameter Estimates for NL Prediction Model (Gamma Distribution) Variable Estimate & p-value Intercept 0.0279 Sex p <.0001 Male -0.0098 * Female Reference Health region p =.0711 Eastern 0.0001 Western -0.0037 Central -0.0033 Labrador Reference Physician specialty p =.0883 Specialist 0.0033 General practitioner Reference Years since specialty licensure p =.0006 Q1: 0 9 years 0.0090 * Q2: 10-17 years 0.0008 * Q3: 18-26 years -0.0011 * Q4: 27-35 years Reference * Indicates a statistically significant parameter estimate, α =.05
Number of Diabetes Cases per FFS Physician in Manitoba Physician Mean (SD) Median characteristics Overall 43.4 (74.2) 25.0 Specialty General practitioner 45.1 (45.7) 35.0 Specialist 37.6 (132.8) 3.0 Sex Male 47.7 (76.0) 33.0 Female 30.5 (67.2) 17.0 Age group <40 years 25.5 (34.2) 14.5 40-64 years 52.1 (90.4) 34.0 65+ years 47.1 (49.8) 34.5 Health region Interlake-Eastern 45.9 (37.8) 48.0 Northern 49.4 (59.4) 20.0 Prairie Mountain 42.4 (39.7) 35.0 Southern 35.1 (29.1) 28.0 Winnipeg 44.3 (86.7) 20.0
Observed and Predicted Diabetes Cases, 2002/03 2003/04 Percent of Total Diabetes Cases 100% 90% 80% 70% 60% 50% 40% 30% 20% 10% Predicted Cases: 18,546 Observed Cases: 31,303 Predicted Cases: 8,829 Observed Cases: 45,183 0% Newfoundland & Labrador Manitoba Observed cases: Hospital discharge abstracts Observed cases: Billing claims for FFS physicians Predicted cases: NFFS physicians
Comparing Observed and Predicted Diabetes Cases in the MB External Validation Dataset FFS NFFS Observed Predicted Observed Predicted Overall 43.4 43.8 21.7 32.7 Health region Interlake-Eastern 45.9 45.5 20.7 43.5 Northern 49.4 49.4 15.1 26.8 Prairie Mountain 42.4 42.2 16.0 41.2 Southern 35.1 34.5 17.1 24.6 Winnipeg 44.3 44.8 39.5 36.5 Note: Average number of cases per physician
Conclusions Substantial under-estimation of the number of prevalent diabetes cases was observed for one Canadian province in which there is no shadow billing and a high proportion of NFFS physicians External validation of the prediction model using data from a second Canadian province in which NFFS physicians shadow bill revealed a smaller amount of under-estimation The magnitude of under-estimation had face validity: Smallest in an urban health region with few NFFS physicians Largest in rural health regions with a greater proportion of NFFS physicians
Key Points and Further Research The prediction model assumes that physician characteristics (age, sex, specialty, region of practice, and years since specialty licensure) will have the same association with the number of prevalent diabetes cases in FFS and NFFS physician populations If this assumption is not tenable, then our predictions will probably represent an upper bound on the number of missed cases External validation is currently underway with other chronic diseases, including: Hypertension: cases are primarily seen by general practitioners Inflammatory bowel disease: cases are primarily seen by specialists
Acknowledgements This research was supported by funding from the Manitoba Health Research Council and the Canadian Institutes of Health Research (#275189). The authors are indebted to the Newfoundland and Labrador Centre for Health Information (NLCHI) and Manitoba Health (HIPC 2012/2013-04) for the provision of data. The results and conclusions are those of the authors, and no official endorsement by NLCHI or Manitoba Health is intended or should be inferred.