Study Design and Data Source
We conducted a cohort study using data from the Longitudinal Cohort of Diabetes Patients database (LHDB) from the National Health Insurance (NHI) Research Database of Taiwan. The NHI system covers >99% of Taiwan's population and has been in operation since 1995.[7,15] The LHDB is a sub–data set comprising a randomly sampled cohort of deidentified patients with incident diabetes from the NHI Research Database, which is constructed and maintained by the National Health Research Institutes, Taiwan. The LHDB randomly sampled 120,000 patients with incident diabetes each year. For the present study, we analyzed data from patients with incident diabetes from 2004 through 2010. We obtained their claims data, including inpatient records, outpatient records, registry for contracted medical facilities, registries for beneficiaries (including scrambled identification number, birthday, sex, coverage period, location of residence, occupation, and income), and registries for patients with catastrophic illness (copayments are waived for patients receiving medical treatments related to the registered diseases).
The Institutional Review Board of the National Taiwan University Hospital approved this study and waived the requirement for informed consent because the database had only deidentified information, and linkage to other databases was not allowed.
Adult patients with incident CKD who had diabetes were enrolled. A patient was defined as having CKD if the in- or outpatient diagnosis code included one of the International Classification of Diseases, Ninth Revision, Clinical Modification (ICD-9-CM) codes listed in Table S1 (provided as online supplementary material). The algorithm by which ICD-9-CM codes were used to identify CKD in the NHI Research Database has a high positive predictive value of 88%. The incidence date of CKD was defined as the date of the first appearance of a diagnosis code for CKD from: (1) the date of the outpatient record or (2) the date of discharge from the hospital in the inpatient record, whichever occurred first. The index date was defined as day 180 after the incidence date of CKD. Patients were not enrolled if they were younger than 18 years or CKD had been diagnosed before diabetes. We also excluded patients who underwent long-term dialysis therapy or died before the index date and those who received care at medical facilities taking care of fewer than 15 incident patients with CKD.
Quality Scores of CKD Care. To measure the quality of CKD care, we applied the quality indicators we developed based on claims data using the RAND-modified Delphi method.[5,18] The RAND-modified Delphi method is a standard method to achieve expert consensus on quality indicators. From the original set of 11 quality indicators, 3 were both applicable to patients with diabetes and also had data available from the LHDB: prescription of renin-angiotensin system (RAS) inhibitors, testing for proteinuria, and nutritional guidance (Table S2). The observation period for the quality indicators was the 180-day period between the incidence of CKD and the index date. For each patient and each quality indicator, we assessed the quality indicator status (either 0 or 1) and then summed the scores on the 3 quality indicators into an overall quality score (range, 0–3).[10,19]
Potential Confounders. We obtained patient data from the LHDB. Those data included age, sex, comorbid conditions, location of residence, occupation type, income per month in New Taiwan Dollars, time from diagnosis of diabetes to diagnosis of CKD, and the medical facility at which CKD was first diagnosed. We defined patients' comorbid conditions by diagnosis codes from in- and outpatient records during the 180-day period between the incidence of CKD and the index date. Charlson Comorbidity Index (CCI) scores were calculated based on comorbid conditions.
Primary and Secondary Outcomes. The primary outcome was long-term dialysis therapy. Secondary outcomes were hospitalization due to AKI and all-cause death. The observation period started on the index date and ended on the date of the outcome or December 31, 2011, whichever occurred earlier. Because results of laboratory examinations were not recorded in the original claims data, we defined outcomes by specific diagnostic codes in the registries for patients with catastrophic illness and the inpatient records. Individuals requiring long-term dialysis therapy were registered as catastrophic illness patients, with a specific ICD-9-CM code (which is 585), and were exempted from copayment for dialysis. The date on which long-term dialysis therapy began was defined by the registration date for this catastrophic illness in the NHI system.[21,22]
The date of hospitalization due to AKI was defined as the admission date of the first hospitalization with the ICD-9-CM code of 584. Because the LHDB does not provide data for mortality and linkage to other administrative databases (including those with death registrations) is not allowed, we identified patients who died and their date of death by using either: (1) the hospital discharge date if the patient's record indicated that the patient had died in the hospital or (2) the date of withdrawal from the NHI system if the record indicated that the patient had chosen to die at home and therefore left the hospital against medical advice. Because Taiwan's NHI is a compulsory single-payer program, the only reason for withdrawal after a patient leaves the hospital under such a condition would be death.
Mean values ± standard deviations are shown for continuous data. Frequencies and percentages are shown for categorical data. Univariable analyses were done with the independent-samples t test, Wilcoxon signed rank test, Pearson χ 2 test, Spearman rank correlation coefficient, or 1-way analysis of variance, as appropriate.
To examine the association between overall quality scores and the outcomes, we first used univariable Cox proportional hazards models to estimate unadjusted hazard ratios (HRs) and their 95% confidence intervals (CIs) for each outcome. Second, we used multivariable Cox proportional hazards models to estimate adjusted HRs for each outcome, after adjusting for age, sex, CCI scores, location of residence, occupation type, monthly income, and the time from diagnosis of diabetes to diagnosis of CKD. Third, we used an instrumental variable approach, by fitting 2-stage residual inclusion estimation models, to adjust for residual confounding due to unmeasured variables, such as CKD stage, glomerular filtration rate, and proteinuria (more information about the instrumental variable approach is given in Item S1).[25,26] Two-stage residual inclusion is the framework for extending instrumental variable analysis to nonlinear models, including Cox proportional hazards models and Weibull models. We treated death as a censoring event for analyses of long-term dialysis therapy or hospitalization due to AKI.
A good instrumental variable must be strongly associated with the exposure of interest, and its only association with the outcome should be by that exposure. For each quality indicator, we looked at the data from the previous patient at the same facility and took the previous patient's status with regard to that quality indicator (ie, 0 or 1) as an indicator of the facility-level preference for that treatment, which was the instrumental variable.[27,28] Because facility-level preference could vary widely among facilities and could change over time within a single facility, we obtained the instrumental variable after stratifying patients by the calendar year of the index date and the category of time from diagnosis of diabetes to diagnosis of CKD. We excluded the first patient treated at each facility in each calendar year because no instrumental variable could be defined for them. We excluded patients who received care at facilities taking care of fewer than 15 incident patients with CKD because we did not expect a facility to have a clear consistent preference (the instrumental variable) if it had so little experience taking care of patients with CKD.
In the first stage of the 2-stage residual inclusion, we used linear regression to model associations between instrumental variables and overall quality scores, adjusting for all patient-level covariates. From the first-stage model, we estimated the raw residual for each patient. The raw residual was defined as the difference between the model-predicted estimates of overall quality scores and the actual overall quality scores. In the second stage of survival models with Cox regression assuming Weibull distribution, we examined the association between overall quality scores and the outcomes, adjusting for all patient-level covariates and for the residuals. To assess the validity of the instrumental variables, we estimated associations between the overall quality score of each patient and the quality indicator status of the previous patient (F statistics in the first-stage model). A good instrumental variable should have an F statistic > 10. We assessed the covariate balance across the previous patients' quality indicator status and examined associations between each quality indicator and each instrumental variable using a logistic regression model. In sensitivity analysis to examine the robustness of our results to different categorizations of facility size, we changed the cutoff values of facility size into those taking care of at least 10, 20, 30, 40, or 50 incident patients with CKD. We also used the same statistical models to examine the associations between each quality indicator and the outcomes.
For statistical analyses, we used Stata (version 14.1; StataCorp LP) and SAS (version 9.3; SAS Institute Inc) software.
Am J Kidney Dis. 2017;70(5):666-674. © 2017 The National Kidney Foundation
The National Kidney Foundation
Published by Elsevier