High-Dimensional Fixed Effects Profiling Models and Applications in End-Stage Kidney Disease Patients: Current State and Future Directions

Abstract

Profiling analysis aims to evaluate health care providers, including hospitals, nursing homes, or dialysis facilities among others with respect to a patient outcome, such as 30-day unplanned hospital readmission or mortality. Fixed effects (FE) profiling models have been developed over the last decade, motivated by the overall need to (a) improve accurate identification or "flagging" of under-performing providers, (b) relax assumptions inherent in random effects (RE) profiling models, and (c) take into consideration the unique disease characteristics and care/treatment processes of end-stage kidney disease (ESKD) patients on dialysis. In this paper, we review the current state of FE methodologies and their rationale in the ESKD population and illustrate applications in four key areas: profiling dialysis facilities for (1) patient hospitalizations over time (longitudinally) using standardized dynamic readmission ratio (SDRR), (2) identification of dialysis facility characteristics (e.g., staffing level) that contribute to hospital readmission, and (3) adverse recurrent events using standardized event ratio (SER). Also, we examine the operating characteristics with a focus on FE profiling models. Throughout these areas of applications to the ESKD population, we identify challenges for future research in both methodology and clinical studies.

Keywords: Dialysis facility staffing; Poisson regression; United States Renal Data System; end-stage kidney disease; fixed effects; generalized linear mixed model; high-dimensional parameters; multilevel varying coefficient model; propensity score; random effects.

Figure 1:

Selected main literature on profiling methodology development and applications over past two decades for random effects (RE) profiling model (top panel – green) and fixed effects (FE) profiling models focused on/tailored to unique aspects of end-stage kidney disease (ESKD) patients. Key works for RE models include Norman, Glickman and Gatsonis (1997) [19], Ash et al. (2012) [24], and Centers for Medicare & Medicaid Services (CMS) launch of “Hospital Compare” in ~2009 and for FE models this include the seminal works of Kalbfleisch and Wolfe (2013) [30] and He et al. (2013) [38] at the University of Michigan Kidney Epidemiology and Cost Center (UM-KECC). Our works also focused on profiling dialysis facilities and are highlighted in the red box and described in this paper.

Figure 2:

Estimated standardized dynamic readmission ratio $\left(\mathit{\text{SDRR}}\right(\mathit{t}\left)\right)$ as a function of time $t$ (days since transition to dialysis); displayed are five dialysis facilities found to have readmission rates significantly (a) worse, (b) better, and (d) not different relative to the national norm. Shown in (c) are five dialysis facilities with $\mathit{\text{SDRR}}\left(\mathit{t}\right)$ that significantly vary over time with some time periods worse, better, or not different (mixed) compared to the national norm. Adapted from [35].

Figure 3:

Outline of two-stage analysis: (1) profiling modeling and (2) creating matched-sets of facilities of significantly worse (SW) standardized readmission ratio (SRR) and facilities with SRR not significantly (NS) different relative to the national norm/average rate for comparative analysis of facility characteristics.

Figure 4:

Patient demographic, risk factors and facility size before and after propensity score matching. Shown are results for the year 2012.

Figure 5:

Differences in percent nurse-to-patient staff and patient-to-nurse ratios in dialysis facilities with significant worse than expected 30-day readmission compared to facilities with 30-readmissions not different relative to the national norm. Adapted from [37].

Figure 6:

(a) Dialysis facilities flagged for recurrent adverse events (RAEs: anemic events) using Poisson and negative binomial models for RAEs outcome. (b) Impact of overdispersion on flagging facilities as significantly worse (sensitivity – worse) and not different (specificity) relative to a reference norm in simulation. Adapted from [34].

Figure 7:

Sensitivity or rate of correctly flagging truly worse providers for fixed effects (FE) and random effects (RE) profiling model for increasing case-mix complexity (correlation among patient risk adjustment variables.) Adapted from [31].

Figure 8:

Instability in estimation of provider effects (left) and corrected estimates (right). Adapted from [45].

Figure 9:

Sensitivity (Sen) of correctly flagging “better” (B) providers and false negative classification of B providers as not different (ND) and as worse (W) providers as a function of outcome sparsity (outcome event rate: 5% to 20%). Adapted from [45].