An instrumental variable approach finds no associated harm or benefit with early dialysis initiation in the United States

Abstract

The estimated glomerular filtration rate (eGFR) at dialysis initiation has been rising. Observational studies suggest harm, but may be confounded by unmeasured factors. As instrumental variable methods may be less biased, we performed a retrospective cohort study of 310,932 patients who started dialysis between 2006 and 2008 and were registered in the United States Renal Data System in order to describe geographic variation in eGFR at dialysis initiation and determine its association with mortality. Patients were grouped into 804 health service areas (HSAs) by zip code. Individual eGFR at dialysis initiation averaged 10.8 ml/min per 1.73 m(2) but varied geographically. Only 11% of the variation in mean HSA-level eGFR at dialysis initiation was accounted for by patient characteristics. We calculated demographic-adjusted mean eGFR at dialysis initiation in the HSAs using the 2006 and 2007 incident cohort as our instrument and estimated the association between individual eGFR at dialysis initiation and mortality in the 2008 incident cohort using the two-stage residual inclusion method. Among 89,547 patients starting dialysis in 2008 with eGFR 5-20 ml/min per 1.73 m(2), eGFR at initiation was not associated with mortality over a median of 15.5 months (hazard ratio, 1.025 per 1 ml/min per 1.73 m(2) for eGFR 5-14 ml/min per 1.73 m(2); and 0.973 per 1 ml/min per 1.73 m(2) for eGFR 14-20 ml/min per 1.73 m(2)). Thus, there was no associated harm or benefit with early dialysis initiation in the United States.

Figure 1. Participant flow diagram

Medical evidence form refers to Centers for Medicare and Medicaid Services (CMS) Medical Evidence Form 2728; eGFR, estimated glomerular filtration rate; HSA, health service area. Differences are displayed relative to the overall national mean eGFR at dialysis initiation (10.8 ml/min per 1.73 m2)

Figure 2. Variation in the timing of dialysis initiation by Health Service Area and Region

(A), mean estimated glomerular filtration rate (eGFR) at dialysis initiation smoothed by a Bayesian model. (B), Difference between mean eGFR in the region and the national average. Differences were tested without adjustment (blue square) and with adjustment for age, sex, race, Hispanic ethnicity, comorbidity index, diabetes, congestive heart failure, median income in the patient’s zip code, insurance status, modality/vascular access at dialysis initiation, and RUCA code of the patient’s residence (red circle). Bars represent 95% confidence interval.

Figure 2. Variation in the timing of dialysis initiation by Health Service Area and Region

(A), mean estimated glomerular filtration rate (eGFR) at dialysis initiation smoothed by a Bayesian model. (B), Difference between mean eGFR in the region and the national average. Differences were tested without adjustment (blue square) and with adjustment for age, sex, race, Hispanic ethnicity, comorbidity index, diabetes, congestive heart failure, median income in the patient’s zip code, insurance status, modality/vascular access at dialysis initiation, and RUCA code of the patient’s residence (red circle). Bars represent 95% confidence interval.

Figure 3. Variation in the timing of dialysis initiation by Health Service Area after demographic adjustment

Map depicts mean estimated glomerular filtration rate at dialysis initiation by Health Service Area (HSA) after adjusting for age, sex, race and Hispanic ethnicity using a random effects model and followed by Bayesian smoothing.

Figure 4. Difference in patients’ estimated glomerular filtration rate (eGFR) at dialysis initiation according to health service area (HSA)-level characteristics

Estimates are adjusted for patient age, sex, race, Hispanic ethnicity, comorbidity index, diabetes, congestive heart failure, median income in the patient’s zip code, insurance status, modality/vascular access at dialysis initiation, and rural urban commuting area (RUCA) code of the patient’s residence. Herfindahl index is a measure of market competition calculated as the sum of squares of the market share of each provider type. % for-profit is referring to the percentage of prevalent hemodialysis patients treated in for-profit dialysis facilities in the HSA. % pre-emptive transplant referred to the percentage of incident end-stage renal disease patients treated with a kidney transplant prior to any dialysis modality. % peritoneal dialysis refers to the percentage of prevalent dialysis patients in the HSA treated with peritoneal dialysis. Hemodialysis patient-bed ratio refers to the total number of prevalent in-center hemodialysis patients relative to the number of available dialysis beds in all facilities in the HSA.

Figure 5. Relationship between patients’ estimated glomerular filtration rate (eGFR) at dialysis initiation and log hazard of mortality incorporating a penalized spline

(A), Instrumental variable analysis result obtained using the two-stage residual inclusion model with demographic-adjusted mean health service area (HSA)-level eGFR at dialysis initiation as the instrumental variable. Model is adjusted for patient-level variables (age, sex, race, Hispanic ethnicity, comorbidity index, diabetes, congestive heart failure, median income and rural urban commuting area (RUCA) code in the patient’s zip code, insurance status, modality and vascular access at dialysis initiation, primary cause of end-stage renal disease, serum albumin, hemoglobin and body mass index). Reference is set at the mean eGFR and log hazard ratio is indicated by the red line with 95% confidence intervals (95% CI) indicated by dashed yellow lines. Hazard ratios for mortality per 1 ml/min/1.73m² higher eGFR at dialysis initiation obtained from piecewise linear spline analysis with a knot at eGFR of 14 ml/min/1.73m² are reported above the corresponding line segment. (B), Cox proportional hazard model result with adjustment for patient-level variables (age, sex, race, Hispanic ethnicity, comorbidity index, diabetes, congestive heart failure, median income and rural urban commuting area (RUCA) code in the patient’s zip code, insurance status, modality and vascular access at dialysis initiation, primary cause of end-stage renal disease, serum albumin, hemoglobin and body mass index). Reference is set at the mean eGFR and log hazard ratio is indicated by the red line with 95% CI indicated by dashed yellow lines. Hazard ratios for mortality per 1 ml/min/1.73m² higher eGFR at dialysis initiation obtained from a simplified linear model is reported above the spline function. A piecewise linear spline is not used for this estimate because a knot at eGFR=14 ml/min/1.73m² was not significant (p=0.7).

Figure 5. Relationship between patients’ estimated glomerular filtration rate (eGFR) at dialysis initiation and log hazard of mortality incorporating a penalized spline

(A), Instrumental variable analysis result obtained using the two-stage residual inclusion model with demographic-adjusted mean health service area (HSA)-level eGFR at dialysis initiation as the instrumental variable. Model is adjusted for patient-level variables (age, sex, race, Hispanic ethnicity, comorbidity index, diabetes, congestive heart failure, median income and rural urban commuting area (RUCA) code in the patient’s zip code, insurance status, modality and vascular access at dialysis initiation, primary cause of end-stage renal disease, serum albumin, hemoglobin and body mass index). Reference is set at the mean eGFR and log hazard ratio is indicated by the red line with 95% confidence intervals (95% CI) indicated by dashed yellow lines. Hazard ratios for mortality per 1 ml/min/1.73m² higher eGFR at dialysis initiation obtained from piecewise linear spline analysis with a knot at eGFR of 14 ml/min/1.73m² are reported above the corresponding line segment. (B), Cox proportional hazard model result with adjustment for patient-level variables (age, sex, race, Hispanic ethnicity, comorbidity index, diabetes, congestive heart failure, median income and rural urban commuting area (RUCA) code in the patient’s zip code, insurance status, modality and vascular access at dialysis initiation, primary cause of end-stage renal disease, serum albumin, hemoglobin and body mass index). Reference is set at the mean eGFR and log hazard ratio is indicated by the red line with 95% CI indicated by dashed yellow lines. Hazard ratios for mortality per 1 ml/min/1.73m² higher eGFR at dialysis initiation obtained from a simplified linear model is reported above the spline function. A piecewise linear spline is not used for this estimate because a knot at eGFR=14 ml/min/1.73m² was not significant (p=0.7).