“Low-value” glycemic outcomes among older adults with diabetes cared for by primary care nurse practitioners or physicians: a retrospective cohort study

Abstract

Background:

“Low-value” healthcare is care without benefit to patients. Overly intensive glycemic control (i.e., HgbA₁C < 7%) can cause harm to patients at high risk of hypoglycemia, particularly among older adults with co-morbidities. It is unknown whether overly intensive glycemic control differs among patients with diabetes and at high-risk of hypoglycemia cared for by primary care nurse practitioners versus physicians.

Objective:

This study examined patients with diabetes at high risk of hypoglycemia receiving primary care between Jan 2010-Jan 2012, comparing patients reassigned to nurse practitioners to those reassigned to physicians after their previous physician separated from practice in an integrated United States health system.

Design:

This was a retrospective cohort study. Study outcomes were collected at two years after reassignment to a new primary care provider. Outcomes were predicted probabilities of HgbA₁C < 7% using two-stage residual inclusion instrumental variable models, controlling for baseline confounders.

Setting:

Primary care clinics within the United States Veterans Health Administration.

Participants:

38,543 patients with diabetes at increased risk for hypoglycemia (age ≥ 65 years with renal disease, dementia, or cognitive impairment), who had their primary care physician leave the Veterans Health Administration and who were reassigned to a new primary care provider in the following year.

Results:

Cohort patients were on average 76y and 99% men. Of these, 33,700 were reassigned to physicians and 4,843 to nurse practitioners. After two years with their new provider, in adjusted models, patients reassigned to nurse practitioners had a −20.4 percentage-point [95% CI −37.9 to −2.8] lower probability of two-year HgbA₁C < 7%.

Conclusions:

Aligned with prior studies on care quality, rates of overly intensive glycemic control may be appropriately lower among older patients with diabetes at high-risk of hypoglycemia, cared for by nurse practitioners than physicians.

Tweetable abstract:

Primary care nurse practitioners deliver equivalent or better rates of low-value diabetes care for older patients, compared to physicians.

Background

“Low-value” care is considered poor quality healthcare, and occurs when treatment offers no clinical benefit or has potential harms that outweigh the benefits.¹ Among patients with diabetes, overly intensive glycemic management can increase the frequency of adverse health outcomes and becomes low value among patients at highest risk for hypoglycemia. Intensive glycemic management (i.e., glycated hemoglobin test or HgbA₁C < 7%) in older adults has no macrovascular and limited microvascular benefits, but triples the risk of severe hypoglycemia.^2,3 Increasing clinical evidence about the harms from glycemic overtreatment, including increased mortality, was shown across several large trials in 2008.^3–5 Recommendations to ease glycemic control (e.g., an HgbA₁C < 8%) to avoid hypoglycemia in higher-risk adults (e.g., older adults or those with greater co-morbidity burdens) subsequently followed in national clinical practice guidelines from the Veterans Health Administration/Department of Defense (2010)⁶ and American Diabetes Association (2012),⁷ and amplified in international campaigns (for example, Choosing Wisely in the United States (US)).^8,9

Primary care providers are the frontline of care for patients with diabetes. Nurse practitioners, one group of advanced practice registered nurses, represent a growing proportion of primary care providers worldwide, and made up one-fifth of primary care providers in the US in 2010.¹⁰ It has been well documented that nurse practitioners and physicians provide comparable clinical care quality for chronic diseases,^11,12 and nurse practitioner-provided care has equivalent or superior outcomes for mortality, patient satisfaction, and patient communication.^13,14 However, in describing ambulatory low-value care, a few studies have found no differences between outpatient nurse practitioners and physicians in rates of low-value imaging for uncomplicated back pain or headache, or in antibiotic prescriptions for viral respiratory infection - or related referrals for these conditions.^15,16 No studies have yet compared low-value care for diabetes between these clinician groups. Identifying practice differences in delivery of low-value diabetic care has widespread relevance to health systems seeking to determine equivalence between nurse practitioners and physicians for delivery of high-quality care, identifying opportunities for tailored provider interventions to reduce low-value practices, and to patients choosing primary care providers. Care comparability, including low-value practices, is increasingly important as nurse practitioners are a growing solution to primary care workforce shortages and to expanding demand for primary care as populations with chronic disease age globally.^10,17,18

The Veterans Health Administration is one of the largest employers of nurse practitioners in the US, and is an integrated health system utilizing a patient-centered medical home model of care delivery across over 900 outpatient clinics nationally.¹⁹ Patients are empaneled to primary care providers providing ongoing primary care and care coordination. In this team-based model, nurse practitioners or physicians (and less frequently, Physician’s Assistants) serve as primary care providers. Nurse practitioners in the Veterans Health Administration have full practice authority as primary care providers, and do not require the supervision of a physician except within state licensure restrictions on controlled substances.^20–22 primary care providers are supported by a registered nurse care manager, clinical associate, and an administrative assistant. With primary care provider role equivalence between physicians and nurse practitioners, this system provides an opportunity to examine practice differences with less bias from nurse practitioner roles without independent practice privileges, such as Medicare or fee-for-service practice environments that require physician supervision.²³ With an objective to understand differences in low-value glycemic control for diabetes, this study leveraged a natural quasi-experimental condition of primary care patients who experienced reassignment to either a physician or a nurse practitioner after their original primary care provider left the Veterans Health Administration. To improve causal inference using observational data, an instrumental variable approach was used based on existing variation in facility staffing levels of nurse practitioners to reduce endogenous confounding from differences in patient- and facility-level characteristics. Clarifying these provider practice differences in low-value diabetes care provides evidence for policy makers and health systems interested in establishing if nurse practitioners provide comparable care to physicians – knowledge with international utility as workforce shortages compound and for justifying scope-of-practice policy expansions of independent practice authority.

Methods

1. Study overview

This research examines the difference in glycemic overtreatment of patients with diabetes at risk of hypoglycemia, comparing patients assigned to primary care physicians or nurse practitioners in an integrated health system in the US. This study conceptually built from published validated recommendations on ambulatory practices to de-intensify (i.e., reduce). We selected a definition of diabetes overtreatment for patients at risk of hypoglycemia as HgbA₁C < 7% as established by these validated recommendations and as a conservative threshold for overtreatment; other clinical practice guidelines further liberalize HgbA₁C targets goals for high-risk populations.^24–26 These recommendations define patients at highest risk of hypoglycemia as adults ≥ 65 years with a diagnosis of chronic kidney disease stage 3 or higher, cognitive impairment, or dementia.²⁴ Empirical analyses were conducted using detailed Veterans Health Administration administrative data capturing information on national clinical encounters. To estimate differences in glycemic overtreatment, the study compared patients who were reassigned to primary care physicians and nurse practitioners, respectively, after being managed by a primary care physician who left the practice. To address potential selectivity in the reassignment of patients to physicians compared to nurse practitioners, an instrumental variable approach was applied, which seeks to isolate the component of provider reassignment attributable to pseudo-random variation in facility staffing.

2. Data sources and patient sample

The primary data source was the national Veterans Health Administration Corporate Data Warehouse.²⁷ Within this, patients were identified using the Primary Care Management Module, a comprehensive longitudinal record tracking the assignment of patients to providers in primary care. The sample for this study was derived from a previously defined cohort of Veterans Health Administration-enrolled patients whose primary care was reassigned to the management of either a new primary care provider as a nurse practitioner or a physician (due to low proportions, Physician’s Assistants were not included as primary care providers the original dataset). The original study cohort has been previously described.²⁸ In brief, the initial cohort (n=806,434) was formed by first identifying physician primary care providers who departed primary care between Jan 1, 2010 and Jan 1, 2012, then identifying their assigned patients who had been subsequently reassigned to a new primary care provider (either a physician or nurse practitioner).

From this original cohort, patients at risk for hypoglycemia were identified. Patients were included (n=117,133) if they were ≥ 65 years old, enrolled in primary care, and had a diagnosis of type II diabetes mellitus in the baseline year (the year prior to primary care provider reassignment). Patients were excluded if without a high-risk baseline year comorbidity for hypoglycemia (n=78,590), defined as ≥ 1 of: a) renal failure; b) dementia; c) cerebrovascular disease; or d) neurologic disorders. Comorbidities were defined by the International Classification of Diseases-9 codes, based on the validated Gagne comorbidity index.²⁹ The final cohort was 38,543 patients. Outcomes were collected at two years (24 months) after primary care provider re-assignment.

3. Instrumental variable models

Instrumental variable models were applied to account for unobserved confounding affecting patient reassignment to either a new physician or nurse practitioner primary care provider. An instrumental variable approach aims to isolate the component of patient reassignment that is due to pseudo-random variation from an external variable.³⁰ Consistent with prior instrumental variable modeling approaches using these data, pseudo-random variation in nurse practitioner staffing at primary care facilities in 2010 was used as a binary variable above or below the median proportion of primary care provider nurse practitioners, divided by total nurse practitioners + physicians in primary care at the facility.²⁸

Instrumental variable approaches rely on two key requirements. First, the instrumental variable must be highly correlated with the exposure (i.e., primary care provider reassignment). Second, the instrumental variable must not directly relate to the outcome of interest (i.e., HgbA₁C < 7%). The study hypothesized that the instrumental variable (facility nurse practitioner staffing) would meet the first requirement because, all other aspects equal, greater availability of facility nurse practitioners is likely to increase the probability of patient reassignment to a nurse practitioner. The study hypothesized the instrumental variable would satisfy the second requirement because nurse practitioner availability is a function of natural variation in state scope-of-practice laws (i.e., more nurse practitioners practice where state law is less restrictive) and location (rural locations have a greater proportional supply of nurse practitioners by county, while urban practices tend to have more proportional physician providers).^17,18 After controlling for observed confounders, nurse practitioner assignment therefore should be related to differences in availability of nurse practitioners at a facility – represented by the instrumental variable – but should not relate to variation in underlying patient- or facility-level unobserved factors that might otherwise affect patient glycemic outcomes.³⁰ To help validate these two instrumental variable requirements, instrumental variable strength was confirmed using the partial F-statistic using a previously established minimum threshold of 10 and the balance of covariates levels was examined across the instrumental variable.³¹

There are regional geographic patterns to state scope-of-practice laws (e.g., proportionately more full-practice states are in the west, while more reduced/restricted states are in the southern US).¹⁴ Historically, more proportionate nurse practitioners than physicians serve rural patients.^17,18 In contrast, more total nurse practitioners overall serve at hospital-affiliated medical centers, which are more often urban, than often rural community-based outpatient clinics.^28,32 These structural characteristics may result in differences between the respective groups of patients reassigned to physicians or nurse practitioners across those facilities. However, importantly, after adjustment these facility characteristics are unlikely to be systematically associated with the individual patient characteristics related to glycemic control that would influence provider group assignment, other than through the instrumental variable mechanism.

4. Outcome and predictors

The primary predictor was reassignment to a new nurse practitioner vs. physician primary care provider. Primary care physicians who left the health system from 2010–2012 were identified as those without a patient panel for two consecutive quarters. Their empaneled patients were then identified if reassigned to a nurse practitioner or physician primary care provider via taxonomy codes.

The primary outcome was the proportion of patients with HgbA₁C < 7 % in the second year (i.e., 12–24 months) after the index date of reassignment to their new primary care provider. For multiple lab measures in a single year, the most recent (furthest from reassignment date) was used.

5. Covariates

Models were adjusted for characteristics that may influence the outcome and provider reassignment. Patient baseline covariates included age, sex, marital status (married vs. other), race and ethnicity (Black non-Hispanic, Hispanic, other/missing, or White non-Hispanic), non-Veterans Health Administration insurance (any Medicaid/Medicare vs. Veterans Health Administration-only), new primary care provider relationship length (≤ 90, 91–180, 181–270, 271–365, or > 365 days), primary care copay exemption status (exempt due to disability, low-income, or nonexempt), primary care visit count, if average HgbA₁C was < 7.0% (as binary), high-risk comorbidity diagnosis, high total comorbidity score (Gagne score ≥ 3), and drive distance to their primary care clinic (< 40 vs. ≥ 40 miles). Clinic covariates included US Census region of the primary care clinic, facility-level hospital or community-outpatient clinic affiliation, and geographic location (urban or rural).

6. Statistical analysis

Unadjusted and multivariable regression models assumed logit links with a binomial family based on goodness-of-fit tests (i.e., modified Park, Pregibon’s link, Pearson’s correlation, and Hosmer-Lemeshow tests). Models accounted for patients clustering in facilities (heteroskedastic cluster-robust standard errors) with the Iteratively Reweighted Least Squares algorithm.

The primary analysis was an instrumental variable approach. Balancing was assessed by standardized mean differences with correlation matrices to confirm non-collinearity between selected covariates before model inclusion. The analysis then applied a two-stage residual inclusion (i.e., 2SRI) estimator. The first stage estimated the probability of nurse practitioner reassignment as a function of the instrumental variable (facility-level nurse practitioner staffing) and covariates using a generalized linear model with logit link and binomial family. Using the fitted first stage model, generalized residuals were calculated using a logistic density function that served as explanatory variables in the second stage. Two-stage residual inclusion has been favored for nonlinear probability models with binary outcomes.^33,34 Prior work also demonstrated two-stage residual inclusion instrumental variable methods with generalized residuals minimize bias in estimating average treatment effects for non-rare outcomes (probabilities > 5%).³⁵ Goodness-of-fit tests were also conducted to support model fit. Patients with missing data were excluded, except if levels could be recoded as missing. Standard errors were estimated using a bootstrap procedure (1,000 repetitions) and accounted for patient clustering by facility.

7. Sensitivity analysis

To improve the theoretical link between provider active treatment practices and glycemic outcomes, a sensitivity analysis was conducted among patients receiving prescriptions for diabetic medications from the Veterans Health Administration. For this, patients who were not prescribed either an oral or injectable diabetic medication that lowers blood sugar (a hypoglycemic) starting ≥ 90 days after the switch to a new primary care provider were excluded. This subgroup was n=1,025. As the sample size was reduced with fewer patients per facility, facility-level variance was accounted for by including facilities as fixed effects, rather than clustering by facility. Goodness-of-fit, instrumental variable tests, estimator selection, and standard errors were otherwise as above.

8. Ethical approval

This study was approved by the Veterans Health Administration Puget Sound Institutional Review Board (MIRB #00765). Individual informed consent was waived for this retrospective data analysis.

Results

1. Description of cohort

A total of 38,543 patients with diabetes ≥ 65 years at risk for hypoglycemia were reassigned to a physician versus nurse practitioner primary care provider after their initial primary care provider left (Table 1). On average, these patients were 75.9 years old, mostly male (98.7%) and non-Hispanic White (74.7%). Among all patients, 87.4% (n=33,700) were reassigned to a physician primary care provider and 12.6% (4,843) were reassigned to a nurse practitioner primary care provider. Those reassigned to a new nurse practitioner primary care provider versus a physician were slightly older (76.4 vs. 75.8 years), more proportionately non-Hispanic White (77.1% vs. 74.3%) and more rural residents (26.2% vs. 16.2%) living further from their primary care clinics (22.7% vs. 19.3% with distance to clinic ≥ 40 miles).

Table 1.

Patients with diabetes ≥ 65 years at risk for hypoglycemia reassigned to a new primary care physician vs. nurse practitioner (NP). Mean (SD) shown, except as noted.

	Total N=38,543	Reassigned to physician n=33,700	Reassigned to NP n=4,843	SMD
Age, years	75.9 (7.6)	75.8 (7.6)	76.4 (7.6)	0.08
Female, n (%)	500 (1.3)	426 (1.3)	74 (1.5)	0.02
Race, n (%)
Black, non-Hispanic	5,138 (13.3)	4,581 (13.6)	557 (11.5)	0.06
Hispanic	2,012 (5.2)	1,830 (5.4)	182 (3.8)	0.08
Other/missing	2,605 (6.8)	2,236 (6.6)	369 (7.6)	0.04
White, non-Hispanic	28,788 (74.7)	25,053 (74.3)	3,735 (77.1)	0.06
Married, n (%)	25,463 (66.1)	22,270 (66.1)	3,193 (65.9)	0.003
Rural, n (%)	6,722 (17.4)	5,453 (16.2)	1,269 (26.2)	0.27
Copay exemption, n (%)
Due to disability	15,297 (39.7)	13,427 (39.8)	1,870 (38.6)	0.03
Due to income	15,111 (39.2)	13,200 (39.2)	1,911 (39.5)	0.01
Nonexempt	8,135 (21.1)	7,073 (21.0)	1,062 (21.9)	0.02
Other insurance, n (%)	34,530 (89.6)	30,123 (89.4)	4,407 (91.0)	0.05
Relationship length, days				0.08
0 – 90	8,121 (21.1)	7,129 (21.2)	992 (20.5)	0.02
91 – 180	4,367 (11.3)	3,652 (10.8)	715 (14.8)	0.12
181 – 270	3,936 (10.2)	3,357 (10.0)	579 (12.0)	0.06
271 – 365	3,578 (9.3)	3,141 (9.3)	437 (9.0)	0.01
> 365	18,541 (48.1)	16,421 (48.7)	2,120 (43.8)	0.10
Community clinic, n (%)	16,809 (43.6)	14,725 (43.7)	2,084 (43.0)	0.01
United States Region, n (%)
West	7,549 (19.6)	6,484 (19.2)	1,065 (22.0)	0.07
Midwest	8,367 (21.7)	6,837 (20.3)	1,530 (31.6)	0.26
South	19,130 (49.6)	17,394 (51.6)	1,736 (35.9)	0.32
Northeast	3,497 (9.1)	2,985 (8.9)	512 (10.6)	0.06
Distance to clinic ≥ 40 miles, n (%)	7,612 (19.8)	6,513 (19.3)	1,099 (22.7)	0.08
Baseline PC visits	6.1 (6.5)	6.1 (6.3)	5.9 (7.2)	0.04
High total comorbidity, n (%)	22,539 (58.5)	19,683 (58.4)	2,856 (59.0)	0.01
Renal failure, n (%)	21,109 (54.8)	18,498 (54.9)	2,611 (53.9)	0.02
Dementia, n (%)	2,602 (6.8)	2,243 (6.7)	359 (7.4)	0.03
Neurologic disease, n (%)	10,061 (26.1)	8,721 (25.9)	1,340 (27.7)	0.04
Cerebrovascular disease, n (%)	17,274 (44.8)	15,132 (44.9)	2,142 (44.2)	0.01
Baseline HgbA1C	7.16 (1.29)	7.16 (1.28)	7.15 (1.31)	0.01
Baseline HgbA1C < 7%, n (%)	16,563 (51.1)	14,530 (51.1)	2,033 (51.1)	0.001
2-year HgbA1C	7.11 (1.24)	7.11 (1.24)	7.11 (1.27)	0.01
2-year HgbA1C < 7%, n (%)	16,901 (54.2)	14,858 (54.2)	2,043 (53.8)	0.01

PC = primary care. SMD = Standardized Mean Difference.

At baseline, the unadjusted mean proportion of patients with HgbA₁C < 7% was 51.1% (n=15,852 with HgbA₁C ≥ 7%; n=16,563 with HgbA₁C < 7%). The mean baseline HgbA₁C was 7.16 (SD=1.29, IQR 6.3–7.8). For those reassigned to a physician primary care provider, 51.1% (n=14,530) had a baseline HgbA₁C < 7% [mean 7.16, SD=1.28]; among those reassigned to a nurse practitioner primary care provider, 51.1% (n=2,033) had a baseline HgbA₁C < 7% [mean 7.15, SD=1.31].

After two years of follow-up, the mean proportion of patients with a HgbA₁C < 7% was 54.2%, with an average HgbA₁C of 7.11 (SD=1.24, IQR 6.3–7.7). Of these, 54.2% (n=14,858) of patients reassigned to physicians had a follow-up HgbA₁C < 7% [mean A₁C=7.11 (SD=1.24)], and 53.8% (n=2,043) nurse practitioner patients had a follow-up HgbA₁C < 7% [mean A₁C=7.11, (SD=1.27)].

2. Logistic Regression Models

Unadjusted, the predicted probability of having a low HgbA₁C (< 7%) at two years of follow-up was not statistically significantly different between those reassigned to a physician versus nurse practitioner primary care provider (difference, −0.004 percentage points, 95% CI −0.02 to +0.01, P=0.65) (Table 2). After adjusting for observed patient and clinic characteristics, there was also no statistically significant difference in the groups at two years of follow-up (difference −0.004 percentage points (−0.02 to + 0.01), P=0.62).

Table 2.

Estimated predicted probability of HgbA₁C < 7% after 2 years in older patients with diabetes reassigned to primary care physicians (MD) vs. nurse practitioners (NP), adjusting for clinic and patient baseline differences.

	MD	NP	Difference (95% CI)	P
Logistic regression	53.26	52.81	+0.45 (−2.19 to +1.30)	0.62
2SRI instrumental variable	55.56	35.59	−20.38 (−37.92 to −2.83)	0.02
2SRI, among patients on medications only	50.55	28.75	−23.85 (−92.60 to +44.91%)	0.50

2SRI = Two-stage residual inclusion. MD = Medical Doctor. NP = Nurse practitioner.

3. Instrumental variable results

In the first stage model, the selected instrumental variable had an F-statistic of 17.9, above the threshold defining a strong instrumental variable (> 10).³¹ Balancing by the instrumental variable was also assessed (Supplemental Appendix 1). Standardized mean differences between patient characteristics were mostly < 0.1, indicating similarly between instrumental variable levels. Some differences remained after rebalancing; more patients seen in facilities with a higher proportion of nurse practitioners were hospital-affiliated (63% vs. 49%), and facilities with a higher proportion of nurse practitioners were less common in the South (38% vs. 61%). First stage model results are shown in Supplemental Appendix 2.

Table 2 shows main model results. Those who were reassigned to a nurse practitioner had a 20.38 percentage-point lower average predicted probability (95% CI −37.92 to −2.83, P=0.02) of having a HgbA₁C < 7% at two years than those reassigned to a physician.

Using the adjusted instrumental variable approach among those prescribed a hypoglycemic medication for diabetes at ≥90 days after primary care provider reassignment, the average predicted probability of a HgbA₁C < 7% at two-year follow-up was 23.85 percentage-points lower among those reassigned a nurse practitioner than a physician (95% CI, −92.60 to +44.91 percentage-points, P=0.50).

Discussion

This study examined a cohort of older patients with diabetes and medical comorbidities to compare if reassignment to a new nurse practitioner primary care provider resulted in greater risk of overly intensive glycemic control, relative to patients reassigned to new physician primary care providers. Leveraging an instrumental variable approach allowed further accounting for potential unobserved factors influencing patient primary care provider reassignment. At two years, those reassigned to a nurse practitioner primary care provider had an over 20% lower probability of a HgbA₁C < 7.0%. While not statistically significant, among patients on hypoglycemic prescriptions at follow-up, those assigned to a nurse practitioner had a similar 24% lower probability of HgbA₁C < 7%.

There are several explanations for these findings. These study data were collected at a time of increasing awareness of the threat of hypoglycemia and growing consensus on less-restrictive glycemic targets in select populations.^6–8,36,37 Older patients at risk for hypoglycemia assigned to nurse practitioners may have experienced greater liberalization of glycemic targets, or conversely, physicians may have been able to safely achieve lower glycemic ranges in these patients. While a comparison of prescribing differences between physicians and nurse practitioners was beyond this study’s scope, the sensitivity analysis of patients with system-prescribed hypoglycemic medications supports that active treatment differences may account for findings. Knowledge of clinical practice guidelines or recognition of organizational efforts to de-escalate glycemic control may have been different between nurse practitioners and physicians.^8,34,35 Alternatively, nurse practitioners may have systematic practice patterns beyond prescribing that lead to higher HgbA₁Cs. This may occur due to differences in panel sizes or workload from physicians. This study was unable to adjust for patient panel size among providers in this dataset; however, Veterans Health Administration panel sizes are approximately 15–25% larger for physicians than nurse practitioners.²⁸ Appointment length and other organizational components were identical, but smaller nurse practitioner panel sizes may reduce the workload from other demands and permit nurse practitioners to individualize glycemic goals or better accommodate patient needs around treatment regimens. Relatedly, physicians with comparably larger panel sizes may be more likely to suffer from competing demands for time or perpetuate prior treatments (even when not appropriate).³⁸ Supporting this possibility, prior literature suggests patients seeing nurse practitioners for primary care have better satisfaction and perceive longer visits with providers.^39,40

Residual differences after balancing by instrumental variable remained in facility and geographic distribution where this cohort sought care, which may influence findings. Patient characteristics were generally well balanced, as expected with the pseudo-randomized instrumental variable approach. However, as anticipated, the instrumental variable had artifactual residue from a) state scope-of-practice laws affecting local hiring availability with fewer nurse practitioners in facilities in restricted-practice states, and b) staffing patterns at Veterans Health Administration hospital-affiliated clinics, as more commonly urban-based and with proportionately more nurse practitioners per facility than rural locations. These imbalances are inherent to the instrumental variable, but are unlikely to systematically affect the assignment of an individual patient to an available nurse practitioner or a physician at a facility after controlling for facility characteristics, other than through the instrumental variable mechanism.

These findings add to the literature on practice patterns between primary care physicians versus nurse practitioners, with increased relevance given implications for diabetic care quality. Low HgbA₁C < 7% among older patients with diabetes can indicate overtreatment and increases the risk of hypoglycemic events without corresponding benefit.^2–5 As awareness of the potential harms of diabetic overtreatment has increased over the past two decades, individualizing or liberalizing HgbA₁C targets is supported by organizational guidelines with increasing understanding of the patients at highest-risk of hypoglycemia, such as older adults, those with chronic kidney disease, underlying existing cognitive impairment, or limited life expectancy.^24,41,42 Our findings suggest that after rigorously accounting for differences in patient assignment, there may be variation in diabetic care administered by nurse practitioners compared to physicians. While a handful of studies have looked at clinician differences in other areas of low-value care,^15,16 this may be the first examining diabetes.

These findings have clinical and policy implications. As of early 2023, 27 US states have provided full practice authority to nurse practitioners. However, scope-of-practice debates involving the care provided by nurse practitioners versus physicians are ongoing.^43,44 This study adds to the current evidence demonstrating nurse practitioners have equivalent, or better, chronic disease management quality to physicians for some populations.^11,12,45 This study in particular also supports that nurse practitioners may provide more individualized care among older patients with diabetes at highest risk for hypoglycemia. Given the increasingly importance of nurse practitioners to augmenting primary care capacity,¹⁰ findings reinforce state expansion of nurse practitioner scope-of-practice laws. At the clinic level, panel composition and size (“caseload” or workload) determinants are complex, but rarely incorporate evidence-based patient and system outcomes associated with nurse practitioner care.^46,47 This study supports the ability of nurse practitioners to provide quality care of older, complex patients with diabetes. Our work adds to evidence suggesting health systems, including both non-Veterans Health Administration systems in the US and internationally, may improve the flexibility of their workforce by removing precedents on the limited size (e.g., 75% of a physician panel size is used in the Veterans Health Administration), composition, and complexity (e.g., allocation of more complex patients to physicians in some health centers) of nurse practitioner patient panels.^46–48 Given the broad implications to the primary care workforce, patient safety, and clinical care quality, future research should extend to non-Veterans Health Administration settings and additional health conditions.

Limitations:

This study has several limitations. This dataset, while historic, has unique strengths and disadvantages. These data provide exceptional insight into care patterns, leveraging the natural experiment of following a patient after their primary care provider left the system, when combined with an instrumental variable approach, improves causal inference within an observational study design. Inferences from these data retain relevance to contemporary policy debate about scope-of-practice laws,^43,44 yet predate the significant confounding from the Veterans Access, Choice and Accountability Act (i.e., “Choice Act”, 2014) and Maintaining Internal Systems and Strengthening Integrated Outside Networks (i.e., “MISSION Act”, 2018), which led to growth in the proportion of care delivered by non-Veterans Health Administration providers, care fragmentation, and risk of differential care patterns for higher-risk and geographically remote (rural, highly rural) subgroups.^49–51 We note the instrumental variable approach has drawbacks, such as the imbalances inherent to the instrumental variable, as discussed above. The discrepancy between the multivariate regression model findings, adjusting only for observed differences, and the more rigorous instrumental variable findings, illustrates the importance of controlling for endogenous confounding.

Several other limitations exist. First, the limited sample of patients receiving Veterans Health Administration-administered prescriptions diminished the power of the sensitivity analysis and ability to clarify the role of prescribing in the findings. This small number of patients prescribed system medications may be secondary to patients filling prescriptions outside the system (e.g., via Medicare). Secondly, this study occurred in the Veterans Health Administration, which has several differences compared to outside settings in the US and globally. For example, this system provides integrated care only to veterans of US military service. However, prior research has found Veterans Health Administration enrollees are highly comparable to Medicare enrollees within the US.⁵² In addition, the Veterans Health Administration is one of the largest employers of nurse practitioners and provides an unrestricted licensure allowing nurse practitioners to practice at any system facility.^20,21 Differences from non-veteran, and non-US settings should be considered when applying findings from this study. Finally, the study used validated comorbidity indices based on administrative data as hypoglycemia risk markers, which are potentially overly conservative for identifying patients. For example, Gagne-defined chronic kidney disease is limited to stage 5 kidney disease/renal failure.²⁹

Conclusions

In a population of older adults with diabetes receiving primary care from the Veterans Health Administration, assignment to a nurse practitioner led to lower rates of diabetes overtreatment in patients at greater risk of hypoglycemia. These findings suggest at least equivalent (and potentially superior) care outcomes versus physicians for specific older patients with diabetes, with policy implications supporting expanded state scope-of-practice regulations in context of growing presence of nurse practitioners among the primary care workforce and for promoting beneficial diabetes care practices among high-risk patients.

What is known:

• Nurse practitioners (advanced practice registered nurses) comprising a growing part of the United States’ primary care workforce and have independent practice authority in the integrated Veterans Health Administration.

• Overly intensive glycemic control among patients with diabetes at high risk of hypoglycemia (e.g., older adults) may cause harms that outweighing benefits.

• It is not known how primary care nurse practitioners, compared to physicians, vary in delivery of overly intensive glycemic outcomes for patients with diabetes at high risk of hypoglycemia.

What this paper adds:

• After accounting for differences between the patient populations, patients at high risk of hypoglycemia assigned to primary care nurse practitioners had appropriately lower rates of overly intensive glycemic control, compared to those assigned to physicians.

• This study contributes to understanding of practice differences between physicians and nurse practitioners related to low-value care delivery for diabetes.