Variation in First-line Type 2 Diabetes Treatment due to eGFR and Provider Preferences: A Novel Statistical Analysis

Christina X. Ji, Saul Blecker, Michael Oberst, Ming‐Chieh Shih, Leora I. Horwitz, David SontagmedRxiv 2024 · DOI

Abstract

The selection of first-line treatment for Type 2 Diabetes (T2D)—typically between metformin and alternatives like DPP-4 inhibitors or sulfonylureas—is influenced by clinical factors such as kidney function (eGFR) and provider-specific preferences. This study analyzes treatment variation across a cohort of 10,643 patients using a novel statistical framework. By comparing models with and without provider-specific random effects using a Generalized Likelihood Ratio Test (GLRT), we found significant variation in treatment decisions attributable to providers (p < 0.0001), particularly at lower eGFR levels where clinical guidelines are less definitive. These findings suggest that provider preference plays a substantial role in T2D management, highlighting the need for further research into the appropriateness of this variation.

Introduction

Metformin is widely regarded as the gold-standard first-line therapy for Type 2 Diabetes (T2D) due to its efficacy, low cost, and favorable safety profile. However, its use is historically restricted in patients with significant renal impairment, as measured by the estimated glomerular filtration rate (eGFR), due to concerns regarding lactic acidosis. While the American Diabetes Association (ADA) and FDA updated guidelines in 2016 to allow metformin use in patients with moderate renal impairment (eGFR 30–60 mL/min/1.73 m²), clinical adoption of these updates can be inconsistent.

This study investigates the drivers behind first-line treatment selection, specifically examining the extent to which treatment variation is dictated by patient clinical characteristics versus individual provider preferences.

Methods

The researchers utilized health insurance claims data from a large provider in the Northeast United States, spanning the years 2012 to 2021.

Cohort Selection

The initial cohort consisted of 10,643 patients who were observed prior to their first-line treatment, had available eGFR results, and no history of Type 1 or gestational diabetes. For the provider-specific analysis, the cohort was restricted to 2,271 patients treated by 173 providers who had managed at least 10 patients each.

Statistical Framework

To isolate the impact of provider preference, the study employed a novel statistical approach comparing two families of generalized linear models:

  1. Model 0 (Patient-only): Predicted the probability of receiving metformin based on patient features, including age, sex, eGFR, history of heart failure, and the date of treatment.
  2. Model 1 (Patient + Provider): Incorporated the same patient features but added provider-specific random effects (random intercepts and random slopes for eGFR). This allowed the model to account for different baseline tendencies and sensitivities to kidney function among doctors.

The researchers then performed a Generalized Likelihood Ratio Test (GLRT) to determine if the inclusion of provider-specific data significantly improved the model's ability to explain the observed treatment decisions.

Results

The analysis confirmed that while eGFR is a primary driver of treatment choice, provider preference is a statistically significant independent factor.

The Impact of Kidney Function

As expected, metformin prescription rates were strongly correlated with eGFR. Among patients with an eGFR ≥ 90, 88.8% received metformin. This dropped to 14.1% for patients with an eGFR < 30, where the drug is strictly contraindicated.

Provider-Level Variation

The GLRT yielded a G-statistic of 387.3 (p < 0.0001), indicating that provider identity significantly influences the choice of first-line therapy even after controlling for all available patient-level clinical data. Interestingly, the study did not identify specific "outlier" providers; rather, the variation was distributed across the entire population of physicians.

The "Gray Zone"

The most pronounced variation between providers occurred in the "gray zone" of renal function—specifically eGFR levels between 30 and 60. In this range, the probability of a patient receiving metformin varied significantly depending on their doctor. For a hypothetical patient with an eGFR of 50, some providers had a nearly 100% probability of prescribing metformin, while others dropped to near 0%, preferring DPP-4 inhibitors or sulfonylureas instead.

Discussion

The findings suggest a "spectrum of practice" across the medical community. In clinical scenarios where guidelines are less rigid or have recently changed, individual "practice styles" become a dominant factor in decision-making. This suggests that the "who" of medical care is often as influential as the patient's clinical data.

Limitations

The study's reliance on claims data means certain nuances, such as patient-reported gastrointestinal intolerance to metformin or specific cost-sharing concerns, were not captured. Additionally, the data reflects a specific geographic region and may not be generalizable to all healthcare systems.

Conclusion

This research demonstrates that provider preference is a major determinant in Type 2 Diabetes treatment, particularly when clinical guidelines allow for interpretation. The novel statistical approach used here provides a robust framework for evaluating provider variation in other areas of medicine. Future research should focus on whether this variation impacts long-term patient outcomes, which could help refine clinical guidelines and reduce unnecessary treatment disparities.

Summary updated

Abstract

Abstract Introduction The decision between metformin and a DPP-4 inhibitor or sulfonylurea for first-line type 2 diabetes treatment relies on many factors, including estimated glomerular filtration rate (eGFR), history of heart failure, age, sex, and even provider preferences. This study evaluates variation in this treatment decision across two factors: eGFR and provider preferences. Research Design and Methods Using health insurance claims data, we defined a cohort based on observation prior to first-line treatment, availability of eGFR results, and no type 1 or gestational diabetes (n=10,643). We performed a chi-squared test to verify the association between eGFR and treatment. The cohort was then restricted to providers with at least 10 patients (n=2,271 patients). We conducted a novel statistical analysis to assess variation across providers. We fitted two models to predict treatment—one using only patient characteristics (age, eGFR, sex, history of heart failure, and treatment date) and another using both patient characteristics and provider-specific random effects. With these models, we performed a generalized likelihood ratio test (GLRT) to assess whether including provider-specific random effects improved fit. Results The chi-squared test confirmed significant association between treatment and eGFR (p &lt; 0.0001). The GLRT in our novel statistical analysis found significant variation existed across providers even after accounting for patient characteristics (p &lt; 0.0001). Visualizations of the observed treatment decisions and treatment policy models show that most of this variation across providers occurred at low eGFR levels, where the level of kidney damage at which metformin should be contraindicated is unclear. Conclusions While some variation in first-line type 2 diabetes treatment was associated with eGFR, some variation may be due to provider preferences that cannot be explained by treatment guidelines. Further studies can elucidate whether such variation across providers is appropriate. Our approach can be applied to other treatment decisions to improve diabetes management. Key Messages What is already known on this topic Guidelines for first-line type 2 diabetes treatments recommend metformin unless there are contraindications, such as kidney damage indicated by low estimated glomerular filtration rate (eGFR). What this study adds This study uses a health insurance claims dataset to verify that first-line treatment is significantly associated with eGFR levels. Then, we propose a novel statistical analysis to assess whether significant variation exists across providers even after accounting for patient age, eGFR, sex, history of heart failure, and treatment date. By fitting two random effects models—one with only patient characteristics and one that also utilizes provider-specific random effects—and comparing the likelihoods of the observed treatment decisions under the two models, we find that the treatment decisions can be explained significantly better when accounting for differences among providers in treatment preferences and eGFR considerations. How this study might affect research, practice, or policy Our results suggest future studies about whether the significant variation across providers found in our analysis is appropriate may help improve first-line type 2 diabetes treatment decisions, and our novel statistical approach can be applied to evaluate variation across providers throughout the diabetes management process.
BibTeX
@misc{ji2024variation,
  author = {Christina X. Ji and Saul Blecker and Michael Oberst and Ming‐Chieh Shih and Leora I. Horwitz and David Sontag},
  title = {{Variation in First-line Type 2 Diabetes Treatment due to eGFR and Provider Preferences: A Novel Statistical Analysis}},
  year = {2024},
  doi = {10.1101/2024.09.19.24313155},
  howpublished = {medRxiv},
}
Figure