Clinical and Socioeconomic Determinants of Angiotensin Receptor-Neprilysin Inhibitor Prescription at Hospital Discharge in Patients With Heart Failure With Reduced Ejection Fraction

What is New?

  • Receiving an angiotensin receptor-neprilysin inhibitor (ARNI) while inpatient and taking an ARNI before hospitalization are powerful predictors of receiving an ARNI at discharge; while this pattern has been described in the prescription practices of other guideline-directed medical therapy for heart failure, this is the largest study to show this pattern with ARNIs.

  • In addition, this study demonstrates that having no medical insurance and living in the lowest quintile of economic prosperity, as measured by the distressed community index, are associated with a decreased likelihood of ARNI prescription at hospital discharge. These disparities appear to be increasing with time.

What are the Clinical Implications?

  • Initiating an ARNI during hospitalization may improve prescription rates. Additionally, ARNI therapy should be initiated in the outpatient setting whenever possible, as this is associated with inpatient use and continuation at discharge.

  • Increasing inpatient administration of ARNI may be particularly effective for patients living in distressed communities and those without insurance. The availability of inpatient resources, such as social services and pharmacists, may make this a more effective strategy than outpatient prescription in these populations.

Heart failure (HF) affects over 6 million people in the United States and was associated with 43.6 billion United States dollars (USD) in health care costs in 2020.1,2 Approximately one-half of patients with HF have reduced ejection fraction, defined as a left ventricular ejection fraction <40%.

In large randomized clinical trials of patients with HF with reduced ejection fraction, angiotensin receptor-neprilysin inhibitors (ARNIs) demonstrated a 20% reduction in cardiovascular death and hospitalizations compared to 10 mg bid of enalapril, an active ACE (angiotensin-converting enzyme) inhibitor comparator.3 Modeling suggests that >30 000 deaths per year might be prevented by switching patients from an ACE inhibitor or angiotensin receptor blocker (ARB) to an ARNI.4 Accordingly, HF guidelines now recommend ARNI as first line therapy for HF with reduced ejection fraction.5

Despite strong evidence, prescription rates for ARNIs have been low, 4% to 26% in studies performed in the past 5 years.6–8 Concerns of cost are likely tied to under-prescription, as despite multiple studies demonstrating ARNIs to be cost effective, the therapy is not covered by all insurance vendors, and out-of-pocket costs to patients can exceed 680 United States dollars per month without insurance.9,10 Contraindications for ARNIs are similar to ACE inhibitors and ARBs, and studies have showed similar rates of side effects and complications, so it is unlikely the low prescription rates are due primarily to medical reasons. Studies examining the association between hospital type and rates of ARNI prescription have found no relationship.11 Understanding the factors driving low rates of ARNI prescription is essential to developing strategies to increase use of this effective medication.

A growing body of evidence suggests that lower levels of economic well-being are associated with decreased prescription of effective medical therapy for HF.12 We hypothesized that with the substantial difference in up-front cost of ARNIs compared with ARB (just under 1400 United States dollars annual cost difference), lower economic well-being would be associated with decreased ARNI prescription.13 In this study, we investigated the impact of clinical factors and community well-being on ARNI prescription at discharge from a HF hospitalization in the Get With The Guidelines-Heart Failure (GWTG-HF) registry, leveraging socioeconomic data from the Economic Innovation Group’s distressed community index (DCI).


Study Design and Population

This is a retrospective study of adult patients enrolled in the GWTG-HF registry from 2017 to 2020. The GWTG-HF program prospectively enrolls patients hospitalized for new or worsening HF at participating hospitals and has been extensively described.14 These data are available from the American Heart Association upon reasonable request ( Patients were included in this analysis if they had an ejection fraction <40% at time of hospitalization and were alive at discharge from the hospital. Patients were excluded if they were discharged to hospice care and had a history of left ventricular assist device. Additionally, patients were excluded if they were missing data regarding prescription of an ARNI at discharge. Patients were not excluded from the study based on the presence of an ARNI contraindication as some patients with a contraindication received an ARNI at discharge despite this contraindication; instead we included the presence of a contraindication as an independent covariate in analysis. In patients with multiple hospitalizations, only the first hospitalization in the data set was included in analysis.

The primary outcome measure was prescription of an ARNI at discharge from hospitalization. The primary independent variables of interest were those related to patient demographic and socioeconomic status. Socioeconomic status data included insurance status and the DCI.

DCI data are compiled and maintained by the Economic Innovation Group (website: and are available from them for a small fee. The DCI data set contains ZIP Code–level socioeconomic data derived from the United States Census Bureau. The principal measure of interest from this data set is the distress score, which ranks communities across 7 components (no high school diploma, housing vacancy rate, adults not working, poverty rate, median income ratio, change in employment, and change in establishments) to generate a score from 0 to 100, with 0 being the most prosperous and 100 being the most distressed. Based on this score, patients are split into quintiles based on their home ZIP Code: prosperous, comfortable, mid-tier, at-risk, and distressed. The DCI data set also designates ZIP Codes as urban, suburban, small town, and rural. Data from the DCI data set were merged with the GWTG-HF data set on patient ZIP Code. Previous studies have demonstrated that increasing DCI is associated with lower health care quality and worse health outcomes.15,16

Medical comorbidities, inpatient medications, discharge vitals, discharge serum laboratory measurements, contraindications to guideline-directed medical therapy, and discharge prescriptions were included as covariates in the analysis.

Statistical Analyses

Baseline characteristics were summarized using median (interquartile range) when continuous and number (percent) when categorical. There was extensive data missingness, ranging from 0% to 70% missingness per variable. After pattern analysis, data were deemed to be missing at random, as opposed to missing completely at random and missing not at random. Patterns of missingness occurred primarily at the hospital level, with most hospitals systematically missing the same variables across all their patients. Additionally, data were found to be hierarchical in nature (patients within hospitals) based on unconditional mean modeling showing a large intraclass correlation coefficient. To take this into account, missing data were imputed via multilevel multiple imputation by fully conditional specification using the R package MICE (version 3.13.15), which has been extensively described.17–19 This particular analysis was complicated by a hierarchical data structure, so in addition to using traditional generalized linear modeling techniques in the imputation step, we leveraged random forest, a machine-learning decision tree analysis method.20,21

To investigate the impact of socioeconomic, demographic, and clinical determinants of ARNI prescription, we built an explanatory mixed-effects logistic regression model following the 3-step procedure outlined by Sommet and Morselli.22 To facilitate interpretation of the model, continuous variables were grand mean centered. We proceeded to build a multilevel logistic regression model using the lme4 package (version 1.1.21).23

A complete case analysis (an analysis using records only with complete data so that no imputation is performed) was subsequently performed as a sensitivity analysis for the imputation analysis.

To investigate changes in practice over time, a post hoc analysis was performed to explore the major determinants of ARNI prescription based on the year of patients’ discharge. The data were stratified into 2 groups early versus late based on patients’ discharge year, with early being defined as patients discharged in 2017 or 2018 and late being defined as discharged in 2019 or 2020. Multilevel logistic regression models with identical independent variables were fit to these 2 patient populations. These models were compared with identify major changes in effect sizes of determinants of ARNI prescription.

Data handling, descriptive statistics, and modeling were performed in R version 3.6.0. Our R code and an annotated R Markdown file are available on GitHub at

Computational Details

IQVIA (Parsippany, New Jersey) served as the data collection and coordination center for the GWTG-HF registry. Each participating hospital received either human research approval to enroll cases without individual patient consent under the common rule, or a waiver of authorization and exemption from subsequent review by their institutional review board. Analyses were completed using the American Heart Association Precision Medicine Platform ( on an r5.12xlarge AWS EC2 computing instance.


Population Characteristics

Of the 593 053 entries in the GWTG-HF registry, 206 145 met inclusion criteria. After removing duplicate hospitalizations and patients who met exclusion criteria, 136 144 patients remained in the data set (Figure 1). A total of 110 923 patients were missing one or more elements of data.

Figure 1.

Figure 1. Consort diagram of analysis population. Consort diagram showing sample extraction from the Get With The Guidelines-Heart Failure (GWTG-HF) registry. AMA indicates against medical advice; ARNI, angiotensin blocker-neprilysin inhibitor; EF, ejection fraction; HFrEF, heart failure with reduced ejection fraction; and LVAD, left ventricular assist device.

The baseline characteristics and data missingness of the analyzed population are presented in Table 1. Frequencies reported for categorical data do not include missing data in the denominator. The median age of the study population was 68 (interquartile range, 57–78) years. The majority of patients were male (66%) and White (57.7%). The median ejection fraction was 25% (interquartile range, 20–33). Of patients with known insurance status, 6.4% were uninsured. There were more patients living in distressed communities versus prosperous communities based on the DCI (25.0% versus 16.3% of patients, respectively). A total of 6.3% of patients were taking an ARNI before hospitalization. A total of 10.9% of patients received an ARNI while hospitalized.

Table 1. Descriptive Statistics

ARNI (no) ARNI (yes) Total
N 118 929 (87.4%) 17 215 (12.6%) 136 144
 Age* 68 (57–79) 66 (56–76) 68 (57–78)
  Missing 0
  Male 78 546 (87.2%) 11 498 (12.8%) 90 044 (66.3%)
  Female 40 105 (87.6%) 5673 (12.4%) 45 778 (33.7%)
  Missing 322 (0.2%)
  Asian 2607 (90.8%) 265 (9.2%) 2872 (2.1%)
  Black 33 369 (86.2%) 5333 (13.8%) 38 702 (28.4%)
  Hispanic 10 154 (89.4%) 1209 (10.6%) 11 363 (8.3%)
  White 68 596 (87.4%) 9911 (12.6%) 78 507 (57.7%)
  Other 4203 (89.4%) 497 (10.6%) 4700 (3.5%)
  Missing 0
Socioeconomic factors
 Distress score 57.1 (30.2–80.0) 57.4 (31.0–79.6) 57.1 (30.2–79.9)
  Missing 59 969 (44.0%)
 Patient distress score quintile
  Prosperous 10 900 (87.8%) 1515 (12.2%) 12 415 (16.3%)
  Comfortable 11 659 (88.4%) 1536 (11.6%) 13 195 (17.3%)
  Mid-tier 12 922 (87.3%) 1883 (12.7%) 14 805 (19.4%)
  At-risk 14 617 (87.4%) 2115 (12.6%) 16 732 (22.0%)
  Distressed 16 728 (87.9%) 2300 (12.1%) 19 028 (25.0%)
  Missing 59 969 (44.0%)
 ZIP Code designation
  Urban 22 258 (89.2%) 2683 (10.8%) 24 941 (32.7%)
  Suburban 23  415 (87.4%) 3379 (12.6%) 26 794 (35.2%)
  Small town 11 503 (87.3%) 1669 (12.7%) 13 172 (17.3%)
  Rural 9650 (85.6%) 1618 (14.4%) 11 268 (14.8%)
  Missing 59 969 (44.0%)
  Medicaid 22 149 (88.4%) 2913 (11.6%) 25 062 (19.7%)
  Medicare 49 574 (87.4%) 7130 (12.6%) 56 704 (44.6%)
  Other 31 866 (85.6%) 5376 (14.4%) 37 242 (29.3%)
  None 7434 (92.1%) 642 (7.9%) 8076 (6.4%)
  Missing 9060 (6.7%)
Clinical data
 Ejection fraction 25 (20–33) 23 (19–30) 25 (20–33)
  Missing 0
 Serum potassium at discharge 4.0 (3.7–4.3)
  ≤5.0 34 743 (89.1%) 4231 (10.9%) 38 974 (97.3%)
  >5.0 1019 (93.1%) 75 (6.9%) 1094 (2.7%)
  Missing 96 076 (70.6%)
 Serum creatinine at discharge§ 1.3 (1.0–1.8) 1.2 (0.95–1.5) 1.3 (1.0–1.7)
  Missing 94 145 (69.2%)
 Systolic blood pressure at discharge 117 (105–131) 113 (102–127) 116 (105–131)
  ≥90 67 374 (87.9%) 9306 (12.1%) 76 680 (98.0%)
  <90 1334 (83.5%) 263 (16.5%) 1597 (2.0%)
  Missing 57 867 (42.5%)
 Heart rate at discharge 78 (69–88) 77 (69–88) 78 (69–88)
  ≥60 63 779 (87.6%) 9041 (12.4%) 72 820 (94.4%)
  <60 3740 (87.8%) 520 (12.2%) 4260 (5.5%)
  Missing 59 064 (43.4%)
 Listed for heart transplant
  Yes 118 (86.1%) 19 (14.9%) 137 (0.1%)
  No 118 811 (87.4%) 17 196 (12.6%) 136 007 (99.9%)
  Missing 0
 Taking an ARNI before hospitalization
  Yes 632 (22.4%) 2189 (77.6%) 2821 (6.3%)
  No 39 024 (93.5%) 2711 (6.5%) 41 735 (93.7%)
  Missing 91 588 (67.3%)
 Has a contraindication to ACE inhibitor/ARB/ARNI
  Yes 54 102 (99.2%) 424 (0.8%) 54 526 (45.1%)
  No 53 858 (81.0%) 12 638 (19.0%) 66 496 (54.9%)
  Missing 15 122 (11.1%)
Inpatient medications
 Received ACE inhibitor or ARB as inpatient
  Yes 42 615 (93.8%) 2821 (6.2%) 45 436 (56.4%)
  No 27 876 (79.5%) 7185 (20.5%) 35 061 (43.6%)
  Missing 55 647 (40.9%)
 Received ARNI as inpatient
  Yes 1158 (13.2%) 7587 (86.7%) 8747 (10.9%)
  No 69 333 (96.6%) 2417 (3.4%) 71 750 (89.1%)
  Missing 55 647 (40.9%)
 Received any inotrope infusion as inpatient
  Yes 4498 (88.0%) 611 (12.0%) 5109 (3.8%)
  No 114 431 (87.3%) 16 604 (12.7%) 131 035 (96.2%)
  Missing 0
Medical comorbidities
 No prior medical history
  Yes 3098 (90.1%) 342 (9.9%) 3440 (2.7%)
  No 108 436 (87.3%) 15 764 (12.7%) 124 200 (97.3%)
  Missing 8504 (6.2%)
 History of congestive heart failure
  Yes 82 700 (86.6%) 12 820 (13.4%) 95 520 (74.8%)
  No 28 834 (89.8%) 3286 (10.2%) 32 120 (25.2%)
  Missing 8504 (6.2%)
 History of chronic kidney disease
  Yes 26 627 (91.1%) 2600 (8.9%) 29 227 (22.9%
  No 84 907 (86.3%) 13 506 (13.7%) 98 413 (77.1%)
  Missing 8504 (6.2%)
 History of end-stage renal disease
  Yes 4495 (95.6%) 209 (4.4%) 4704 (3.7%)
  No 107 039 (87.1%) 15 897 (12.9%) 122 936 (96.3%)
  Missing 8504 (6.2%)
Discharge medications/planning
 Discharged with GDMT β-blocker
  Yes 98 215 (86.4%) 15 469 (13.6%) 113 684 (85.4%)
  No 17 938 (92.3%) 1498 (7.7%) 19 436 (14.6%)
  Missing 3024 (2.2%)
 Discharged with ACE inhibitor or ARB
  Yes 71 440 (96.8%) 2325 (3.2%) 73 765 (54.4%)
  No 47 294 (76.4%) 14 572 (23.6%) 61 866 (45.6%)
  Missing 513 (0.4%)
 Discharged with MRA
  Yes 36 851 (82.5%) 7835 (17.5%) 44 686 (34.7%)
  No 75 930 (90.2%) 8262 (9.8%) 84 192 (65.3%)
  Missing 7266 (5.3%)
 Follow-up visit scheduled before discharge
  Yes 106 280 (87.3%) 15 450 (12.7%) 121 730 (94.7%)
  No 5830 (86.2%) 934 (13.8%) 6764 (5.3%)
  Missing 7650 (5.6%)
 Discharged to continued care
  Yes 32 595 (90.1%) 3532 (9.8%) 36 127 (57.7%)
  No 22 945 (86.8%) 3504 (13.2%) 26 449 (42.3%)
  Missing 73 568 (54.0%)
 Year of discharge
  2017 37 822 (91.9%) 3340 (8.1%) 41 162 (30.2%)
  2018 35 762 (88.6%) 4596 (11.4%) 40 358 (29.6%)
  2019 34 055 (83.6%) 6660 (16.4%) 40 715 (29.9%)
  2020 11 290 (81.2%) 2619 (18.8%) 13 909 (10.2%)
Hospital characteristics
 Heart transplant center
  Yes 14 073 (90.0%) 1564 (10.0%) 15 637 (11.5%)
  No 104 856 (87.0%) 15 651 (13.0%) 120 507 (88.5%)
  Missing 0
 Academic center#
  Yes 58 191 (87.8%) 8095 (12.2%) 66 286 (48.7%)
  No 60 738 (86.9%) 9120 (13.1%) 69 858 (51.3%)
  Missing 0

A total of 17 215 (12.6%) patients were prescribed an ARNI at the time of discharge. ARNI prescription rates increased yearly over the period of study, with 8.1% of patients prescribed an ARNI on discharge in 2017, 11.4% in 2018, 16.4% of patients in 2019, and 18.8% in the first half of 2020 (Table 2). In the population of 66 496 patients without documentation of an ARNI contraindication, the average rate of ARNI prescription over the study period was 19.0% (Table 3).

Table 2. ARNI Prescription Rate by Year in All Patients

Year Prescribed an ARNI Total patients included in the study
2017 3340 (8.1%) 41 162
2018 4596 (11.4%) 40 358
2019 6660 (16.4%) 40 715
2020 2619 (18.8%) 13 909
Total 17 215 (12.6%) 136 144

Table 3. ARNI Prescription Rate by Year in Patients With No Documented ARNI Contraindication

Year Prescribed an ARNI Patients with no documented ARNI contraindication
2017 2407 (11.4%) 21 056
2018 3303 (16.5%) 20 043
2019 4948 (26.2%) 18 859
2020 1980 (30.3%) 6538
Total 12 638 (19.0%) 66 496

Nearly half (45.1%) of patients were documented as having a contraindication to ACE inhibitor, ARB, or ARNI (Tables 4, 5, and 6, respectively), the most frequently cited of which was other medical reason for ACE inhibitor and ARBs and ACE inhibitor use within the last 36 hours for ARNIs. The second most common contraindication to ARNI use was patient reason, which was not specified further. Of the 57 274 patients with an explicit ARNI contraindication, 424 were prescribed an ARNI at discharge. A total of 47 294 patients did not receive an ACE inhibitor, ARB, or ARNI at discharge. Of these, 5215 (11.0%) had no documented ACE inhibitor/ARB/ARNI contraindication. The ACE inhibitor/ARB/ARNI status of 195 patients is uncertain as they were not prescribed an ARNI, and discharge ACE inhibitor/ARB prescription status was also missing.

Table 4. Reasons Given for Having a Contraindication to ACE Inhibitor

Contraindication No Yes
Hypotension 43 884 (89.3%) 5251 (10.7%)
Azotemia 39 604 (80.6%) 9531 (19.4%)
Other medical reason 17 742 (36.1%) 31 393 (63.9%)
Patient reason 44 089 (89.7%) 5046 (10.3%)
System reason 48 639 (99.0%) 496 (1.0%)

Table 5. Reasons Given for Having a Contraindication to ARB

Contraindication No Yes
Hypotension 43 129 (89.1%) 5263 (10.9%)
Azotemia 38 674 (79.9%) 9718 (20.1%)
Other medical reason 17 506 (36.2%) 30 886 (63.8%)
Patient reason 44 190 (91.3%) 4202 (8.7%)
System reason 47 734 (98.6%) 658 (1.4%)

Table 6. Reasons Given for Having a Contraindication to ARNI

Contraindication Count
ACE inhibitor use within the past 36 h 21 847
Allergy 997
Hyperkalemia 1323
Hypotension 6823
Prohibitive renal dysfunction 10 792
Other medical reason 2571
Patient reason 12 139
System reason 782
Total 57 274

Patients were treated at 560 unique sites, 11.5% of which were heart transplant centers and 48.7% academic centers based on the presence of resident physicians working at the site.

Predictors of ARNI Prescription at Hospital Discharge

Results of the mixed-effects logistic regression model are shown in Table 7 and key findings are illustrated in Figure 2. The intercept indicates that the baseline odds of being prescribed an ARNI were 0.10 (95% CI, 0.08–0.13; P<0.001) in patients with the baseline reference characteristics. For a comprehensive set of baseline characteristics, see the comparison groups in Table 7. Odds ratios (ORs) presented in Table 7 are relative to this baseline profile.

Table 7. Predictors of ARNI Prescription at Hospital Discharge

Odds ratio 95% CI Lambda P value
Lower Upper
Age* 0.988 0.985 0.991 0.256 <0.001
Sex (comparison group: male)
 Female 0.975 0.910 1.045 0.194 0.474
Race (comparison group: White)
 Asian 1.285 1.011 1.632 0.163 0.041
 Black 1.115 1.013 1.227 0.277 0.026
 Hispanic 1.264 1.096 1.460 0.235 0.001
 Other 0.942 0.778 1.140 0.136 0.536
Distress score quintile (comparison group: prosperous)
 Comfortable 0.962 0.844 1.069 0.472 0.559
 Mid-tier 0.920 0.813 1.042 0.417 0.192
 At-risk 0.880 0.769 1.007 0.404 0.064
 Distressed 0.813 0.695 0.950 0.468 0.010
ZIP Code designation (comparison group: urban)
 Rural 1.111 0.962 1.284 0441 0.153
 Small town 1.054 0.915 1.284 0.368 0.469
 Suburban 1.051 0.944 1.170 0.428 0.365
Insurance (comparison group: other nonmedicare/medicaid insurance)
 Medicaid 0.824 0.744 0.913 0.212 <0.001
 Medicare 0.968 0.890 1.051 0.198 0.437
 None 0.597 0.496 0.717 0.354 <0.001
Ejection fraction† 0.937 0.933 0.941 0.190 <0.001
Discharge serum creatinine* 0.747 0.692 0.807 0.723 <0.001
Discharge serum potassium (comparison group: K≤5)
 K>5 0.887 0.679 1.158 0.562 0.379
Discharge sBP (comparison group: sBP≥90)
 sBP<90 0.815 0.631 1.052 0.340 0.117
Discharge heart rate (comparison group: HR≥60)
 HR<60 1.091 0.927 1.282 0.403 0.295
No prior medical history (comparison group: no)
 Yes 0.870 0.699 1.083 0.150 0.213
History of congestive heart failure (comparison group: no)
 Yes 0.947 0.865 1.036 0.335 0.235
History of chronic kidney disease (comparison group: no)
 Yes 0.809 0.714 0.915 0.480 0.001
History of end-stage renal disease (comparison group: no)
 Yes 1.418 1.038 1.937 0.420 0.029
Listed for heart transplant (comparison group: no)
 Yes 0.485 0.160 1.473 0.056 0.202
Taking ARNI before hospitalization (comparison group: no)
 Yes 9.488 6.752 13.333 0.881 <0.001
Contraindication to ACE inhibitor/ARB/ARNI (comparison group: no)
 Yes 0.113 0.104 0.124 0.347 <0.001
Received inpatient ARNI (comparison group: no)
 Yes 72.091 58.301 89.142 0.846 <0.001
Received any inpatient inotrope (comparison group: no)
 Yes 0.527 0.436 0.637 0.093 <0.001
Discharged with GDMT BB (comparison group: no)
 Yes 1.818 1.634 2.022 0.237 <0.001
Discharged with ACE inhibitor or ARB (comparison group: no)
 Yes 0.098 0.088 0.109 0.551 <0.001
Discharged with MRA (comparison group: no)
 Yes 1.773 1.636 1.922 0.386 <0.001
Scheduled follow-up on discharge (comparison group: no)
 Yes 1.172 1.001 1.373 0.261 0.050
Discharged to continued care§ (comparison group: no)
 Yes 0.771 0.703 0.846 0.516 <0.001
Year of discharge (comparison group: discharged in 2017)
 2018 1.378 1.261 1.505 0.131 <0.001
 2019 1.831 1.678 1.999 0.138 <0.001
 2020 2.055 1.835 2.301 0.140 <0.001
Heart transplant center (comparison group: no)
 Yes 0.555 0.347 0.887 0.023 0.014
Academic center (comparison group: no)
 Yes 1.178 0.939 1.478 0.068 0.158
Intercept 0.100 0.075 0.134 0.298 <0.001
Figure 2.

Figure 2. Illustration of key findings. Illustrated findings of the analysis exploring socioeconomic, clinical, and institutional factors that influence the prescription of angiotensin receptor-neprilysin inhibitors (ARNI) at discharge after hospitalization for heart failure with reduced ejection fraction (HFrEF). The driving determinants are primarily clinical factors, but socioeconomic factors and practice trends over time also play a role. DC indicates discharge; and OR‚ odds ratio.

The strongest predictors of ARNI prescription at discharge were inpatient ARNI use (OR, 72 [95% CI, 58–89]; P<0.001) and taking an ARNI before hospitalization (OR, 9.5 [95% CI, 6.8–13]; P<0.001). The strongest predictors against ARNI prescription at discharge were documented contraindication to an ACE inhibitor, ARB, or ARNI (OR, 0.11 [95% CI, 0.10–0.12]; P<0.001) and being prescribed an ACE inhibitor or ARB at discharge (OR, 0.10 [95% CI, 0.09–0.11]; P<0.001).

The likelihood of ARNI prescription increased steadily from 2017 to 2020, with ORs of 1.4 (95% CI, 1.3–1.5; P<0.001), 1.8 (95% CI, 1.7–2.0; P<0.001), and 2.1 (95% CI, 1.8–2.3; P<0.001) for patients discharged in 2018, 2019, and 2020, respectively, relative to patients discharged in 2017. Post hoc analysis demonstrated no significant changes in the major determinants of ARNI prescription by year of patient’s discharge.

Examination of demographic and socioeconomic factors associated with ARNI prescription showed an inverse association with age (OR 0.99 per 1-year increase in age [95% CI, 0.99–0.99]; P<0.001). Identifying as Asian, Black, or Hispanic (as opposed to identifying as White) was associated with an increased likelihood of ARNI prescription (Asian: OR, 1.3 [95% CI, 1.0–1.6]; Black: OR, 1.1 [95% CI, 1.0–1.2]; Hispanic: OR, 1.3 [95% CI, 1.1–1.5]; P<0.05 for all 3 groups). Having no insurance or Medicaid insurance was associated with a significantly lower likelihood of ARNI prescription relative to having non-Medicare/Medicaid insurance (OR, 0.60 [95% CI, 0.50–0.72] and OR, 0.82 [95% CI, 0.74–0.91], respectively; P<0.001 for both groups). Living in a ZIP Code identified as distressed based on the DCI was associated with a lower likelihood of ARNI prescription compared with living in a prosperous community (OR, 0.81 [95% CI, 0.70–0.95]; P=0.010).

Several clinical factors were associated with ARNI prescription. Increasing ejection fraction was associated with a decreasing likelihood of ARNI prescription (OR 0.937 per 5% increase in ejection fraction [95% CI, 0.93–0.94]; P<0.001). Higher serum creatinine was also associated with decreasing likelihood of ARNI prescription (OR 0.75 per 1 mg/dL creatinine [95% CI, 0.69–0.81]; P<0.001). Receiving an inotrope infusion while hospitalized was associated with a lower likelihood of ARNI prescription (OR 0.53 [95% CI, 0.44–0.64]; P<0.001). A history of chronic kidney disease was associated with lower likelihood of ARNI prescription (OR, 0.81 [95% CI, 0.71–0.92]; P=0.001) but having end-stage renal disease was associated with higher likelihood of ARNI prescription (OR, 1.4 [95% CI, 1.0–1.9]; P=0.029). Being discharged to continued care, defined as discharge with home health care, or to a skilled nursing facility, inpatient rehabilitation, intermediate care facility, long-term acute care facility, or another acute care facility, was associated with lower likelihood of ARNI prescription (OR, 0.77 [95% CI, 0.70–0.85]; P<0.001). Having scheduled follow-up on discharge was associated with an increased likelihood of ARNI prescription (OR, 1.2 [95% CI, 1.0–1.4]; P=0.050). Serum potassium >5, systolic blood pressure <90 mm Hg, and being listed for heart transplant were not significantly associated with the likelihood of ARNI prescription.

In terms of site characteristics, being hospitalized at a heart transplant center was associated with lower likelihood of ARNI prescription (OR, 0.56 [95% CI, 0.35–0.89]; P=0.014). Being hospitalized at an academic facility was not significantly associated with likelihood of ARNI prescription (OR, 1.2 [95% CI, 0.94–1.5]; P=0.16). The adjusted intraclass correlation coefficient for the model was 0.254, indicating 25% of the total variance observed in the model could be explained by unobserved differences between sites (eg, physician practice styles, knowledge of the employed physician groups, aggressiveness of prescribers, etc).

A total of 25 221 records (18.5%) were available for complete case analysis. Comparison of the complete case analysis model with the full model demonstrated highly similar effect estimates between the 2 models.

Predictors of ARNI Prescription Stratified by Year of Discharge

Results of the stratified analysis are demonstrated in Table S1. In general, effect estimates remained consistent between the 2 stratified populations with a few notable exceptions; namely, the baseline odds of being prescribed an ARNI was higher in the late versus early population (OR, 0.25 [95% CI, 0.17–0.36] versus OR, 0.08 [95% CI, 0.05–0.12] for late versus early, respectively). Additionally, living in a ZIP Code identified as distressed based on the DCI was associated with a significantly lower odds of being prescribed an ARNI in the late population but not in the early population (OR, 0.76 [95% CI, 0.62–0.93]; P=0.009 versus OR, 0.86 [95% CI, 0.70–1.1]; P=0.157 for late versus early, respectively).


In this large analysis of over 130 000 hospitalizations for HF in the GWTG-HF registry from 2017 to 2020, we found a low rate of ARNI prescription at hospital discharge (12.6%). In 66 496 patients in whom no ARNI contraindication was explicitly documented, the rate of ARNI prescription was similarly low at 19.0%. There was a statistically significant, steady increase in the rate of ARNI prescription from 2017 to 2020, from 8.1% in 2017 to 18.8% in the first 6 months of 2020 for the entire cohort. This increase in ARNI use may be due to increasing coverage by insurance agencies, broader awareness of ARNI’s beneficial effects among providers over time, or greater comfort with the medication as its use becomes more widespread and better understood. Overall, the major determinants of ARNI prescription did not change from 2017 to 2020, based on the results of our stratified analysis.

Our explanatory model suggests that the dominant clinical determinants of ARNI prescription at discharge are receipt of an ARNI while inpatient, taking an ARNI before hospitalization, and having no contraindications to an ACE inhibitor, ARB, or ARNI. The first 2 findings are consistent with previous studies of the GWTG-HF population, which have associated inpatient initiation or continuation of guideline-directed medical therapy with persistent use following hospitalization, as well as subsequent reductions in rehospitalization and mortality.24–27 These findings reinforce the importance of initiating or continuing optimum guideline-directed medical therapy, including ARNIs, during hospitalizations. In-hospital initiation of any medication allows safe initiation in a highly monitored setting and should always be considered when a patient’s therapeutic regimen deviates from optimal guideline recommendations.

Patients with a contraindication to ACE inhibitor/ARB/ARNI constituted a large (45.1%) proportion of the population. The GWTG-HF registry contains extensive and unique data on contraindications to therapy, including reasons specific to each medication (see Table 3 for more information). Previous studies of GWTG-HF data have largely excluded patients with therapy contraindications. We opted not to exclude these patients for 3 reasons: First, there were a small handful of patients (424) prescribed an ARNI at discharge despite documentation of a contraindication. Second, we felt that some contraindications may be modifiable rather than absolute, and hence represent a strategic target to improve prescription rates. Finally, contraindications to ARNI use included patient reasons and system reasons, neither of which are specified further, and these reasons constituted over 12 000 ARNI contraindications. We were concerned that patient and system reasons might include socioeconomic factors that were not otherwise captured other than perhaps in our insurance and DCI variables. However, it was ultimately not possible to test the impact of individual contraindications due to the inability to distinguish between patients who had no ARNI contraindications from those for whom the data was merely missing. To shed further light on this issue, further detail on contraindications may be helpful in future data sets, including explicit documentation of the absence of a contraindication. Nevertheless, medical contraindications such as hyperkalemia, azotemia, and prohibitive renal dysfunction made up the vast majority of contraindications. ARNI prescription rates may be improved by further research into the extent of renal dysfunction and electrolyte abnormalities associated with adverse events on ARNI initiation, hence truly constituting a contraindication, rather than subjective cutoffs.

It is of interest that hypotension and hyperkalemia on discharge assessment were not significantly associated with decreased rates of ARNI prescription at discharge, but these effects must be interpreted in the presence of the ACEI/ARB/ARNI contraindication covariate. For example, the effect of hypotension was 0.815 with a P value of 0.117; however, this effect size is interpreted in the absence of a contraindication to ACE inhibitor/ARB/ARNI, which includes clinically significant hypotension. So, the population to whom the 0.815 OR applies is the population with a discharge blood pressure <90 mm Hg, but without clinically significant hypotension as assessed by the treating medical team. It should also be noted that discharge laboratory and vital signs values had the greatest degree of missingness, and this high degree of missing data increases the SE (uncertainty) around the estimates.

Our analysis adds new information about the impact of socioeconomic factors on prescription of ARNIs at discharge. Having no insurance or Medicaid insurance as opposed to having private insurance, and living in a distressed community as opposed to living in a prosperous community were independently associated with lower likelihood of ARNI use at time of discharge. These findings support the growing body of evidence that socioeconomic strain impacts medical decision-making during hospitalization and through time of discharge, contributing to perpetuation of health disparities. Notably, the disparity in prescription rates between distressed and prosperous communities appear to be increasing over time rather than diminishing, based on the stratification analysis. This reinforces the urgent need to ensure patients from distressed communities receive optimum guideline-based care while hospitalized and on discharge. The best way to implement such change has not been demonstrated, but we suggest early and systematic use of hospital ancillary and support staff to address financial and insurance issues that may impact prescription of more expensive but more effective medications such as ARNIs. Encouragingly, no significant disparities were identified in ARNI prescription rates amongst ethnic minority groups, with Asians, Blacks, and Hispanics actually being more likely to receive an ARNI at discharge.

Our analysis suggests that being a transplant center is associated with lower likelihood of ARNI prescription independent of ARNI contraindications and discharge labs and vitals. This hospital characteristic was not explored in previous studies investigating associations between prescription practices and hospital characteristics (Luo et al11). The lower likelihood of ARNI prescription at transplant centers may be due to the higher proportion of New York Heart Association class IV patients at these centers or higher expectations of outpatient follow-up. In any case, strategies advocating ARNI prescription should be implemented hospital-wide regardless of center expertise.

Finally, a great deal of attention was spent on proper accounting for missing data. The most common methods of handling missing data include complete case analysis (also known as listwise deletion), mean imputation, and assigning missing data its own category, but use of any of these methods causes statistical models to become biased, which is to say that effect estimates can no longer be legitimately extrapolated to represent true effect size in the general population.19,28–30 In our multiple imputation analysis, we trained the computer on the potential distribution of missing values based on similar cases in the data. These distributions were then used to generate effect estimates. These estimates are considered unbiased (ie, can be extrapolated to the general population), with uncertainty about missing values reflected in CIs. Multiple imputation offers a powerful and statistically valid method of handling missing data, and this paper and its technical appendix demonstrate how multiple imputation can leverage machine learning and traditional generalized linear modeling to use available data as fully as possible, even in a hierarchical data structure. Although each multiple imputation procedure is unique to the scientific question being asked, the R markdown file we have made publicly available on GitHub lays out a step-by-step approach that we hope can offer some guidance to investigators pursuing studies with the GWTG-HF data and other large data sets.

Study Limitations

The primary limitation of our study is the degree of missing data. Analyses leveraging multiple imputation provide estimates that ultimately converge on true effect sizes regardless of the volume of missingness, but this is not without a trade-off: uncertainty about the true value of the missing data is added into the calculation of standard errors. As a result, CIs are widened and there is increased uncertainty about the true value of the effect size in the general population.

An additional study limitation includes lack of patient-level socioeconomic data in the GWTG registry. The DCI provides socioeconomic information at the ZIP Code level only. Within that ZIP Code, individuals will have variation in their true degree of economic well-being, which is not captured by the data available in the registry.

Finally, the GWTG-HF registry provides a wealth of information regarding contraindications to therapy and is a potentially valuable source in understanding why patients do not receive optimal therapy; however, the absence of an ARNI contraindication is not recorded within the GWTG-HF data so there is no way to determine whether a patient truly has no ARNI contraindications or if that patient’s contraindications were not reported. Having an ARNI contraindication was ultimately found to be a powerful negative predictor of ARNI prescription, but we were unable to determine which contraindications (Table 3) were the primary drivers of this effect.

Additionally, the reason patients were assigned contraindications to ARNIs could not be comprehensively investigated, as only admission and discharge laboratory and vitals assessments were provided. For example, patients listed as having a contraindication to an ARNI due to renal dysfunction or hypotension might not have significantly aberrant values at time of discharge, but may have had concerning values during hospitalization. These inpatient data were unavailable, so we were not able to clearly understand why some patients were designated as having contraindications to ARNIs. In addition, the most commonly cited contraindication in the data was “other medical reason” about which no additional data are available.


The overall rate of ARNI prescription in the GWTG-HF registry from 2017 to 2020 was low at 12.6%, with inpatient administration of an ARNI and taking an ARNI before hospitalization being strong positive predictors of ARNI prescription at discharge. While the overall rate of ARNI prescription increased over the study duration, the disparity in prescription rates between patients from distressed versus prosperous communities became more pronounced over time. Even in this population of patients hospitalized for HF who are a high risk for morbidity and health care costs, socioeconomic status appears to drive disparities in medical care. This reinforces the urgent need to ensure patients from distressed communities receive optimum guideline-based care while hospitalized and on discharge.

Article Information


This project was developed through the Heart Failure Data Challenge, using the Get With The Guidelines-Heart Failure registry data to target research related to heart failure and social/structural determinants of health. The data challenge was hosted by the American Heart Association and the Association of Black Cardiologists. The American Heart Association Precision Medicine Platform ( was used for data analysis.

Supplemental Material

Table S1

Nonstandard Abbreviations and Acronyms


angiotensin-converting enzyme


angiotensin receptor blocker


angiotensin receptor-neprilysin inhibitor


distressed community index


Get With The Guidelines-Heart Failure


heart failure


odds ratio


This manuscript was sent to Dr. W. H. Wilson Tang, MD, Guest Editor, for review by expert referees, editorial decision, and final disposition.

Supplemental Material is available at

For Sources of Funding and Disclosures, see page 1047.

Correspondence to: Jeffrey S. Tran, MD, Department of Medicine, University of Arizona, 8771 N Ash Grove Ct, Tucson, AZ 85742. Email


  • 1. Virani SS, Alonso A, Benjamin EJ, Bittencourt MS, Callaway CW, Carson AP, Chamberlain AM, Chang AR, Cheng S, Delling FN, et al; American Heart Association Council on Epidemiology and Prevention Statistics Committee and Stroke Statistics Subcommittee. Heart disease and stroke statistics-2020 update: a report from the American heart association.Circulation. 2020; 141:e139–e596. doi: 10.1161/CIR.0000000000000757LinkGoogle Scholar
  • 2. Heidenreich PA, Albert NM, Allen LA, Bluemke DA, Butler J, Fonarow GC, Ikonomidis JS, Khavjou O, Konstam MA, Maddox TM, et al; American Heart Association Advocacy Coordinating Committee; Council on Arteriosclerosis, Thrombosis and Vascular Biology; Council on Cardiovascular Radiology and Intervention; Council on Clinical Cardiology; Council on Epidemiology and Prevention; Stroke Council. Forecasting the impact of heart failure in the United States: a policy statement from the American Heart Association.Circ Heart Fail. 2013; 6:606–619. doi: 10.1161/HHF.0b013e318291329aLinkGoogle Scholar
  • 3. McMurray JJ, Packer M, Desai AS, Gong J, Lefkowitz MP, Rizkala AR, Rouleau JL, Shi VC, Solomon SD, Swedberg K, et al; PARADIGM-HF Investigators and Committees. Angiotensin-neprilysin inhibition versus enalapril in heart failure.N Engl J Med. 2014; 371:993–1004. doi: 10.1056/NEJMoa1409077CrossrefMedlineGoogle Scholar
  • 4. Fonarow GC, Hernandez AF, Solomon SD, Yancy CW. Potential Mortality Reduction With Optimal Implementation of Angiotensin Receptor Neprilysin Inhibitor Therapy in Heart Failure.JAMA Cardiol. 2016; 1:714–717. doi: 10.1001/jamacardio.2016.1724CrossrefMedlineGoogle Scholar
  • 5. Yancy CW, Jessup M, Bozkurt B, Butler J, Casey DE, Colvin MM, Drazner MH, Filippatos G, Fonarow GC, Givertz MM, et al; WRITING COMMITTEE MEMBERS. 2016 ACC/AHA/HFSA Focused update on new pharmacological therapy for heart failure: an update of the 2013 ACCF/AHA guideline for the management of heart failure: a report of the American College of Cardiology/American heart association task force on clinical practice guidelines and the heart failure society of America.Circulation. 2016; 134:e282–e293. doi: 10.1161/CIR.0000000000000435LinkGoogle Scholar
  • 6. Luo N, Fonarow GC, Lippmann SJ, Mi X, Heidenreich PA, Yancy CW, Greiner MA, Hammill BG, Hardy NC, Turner SJ, et al. Early adoption of sacubitril/valsartan for patients with heart failure with reduced ejection fraction: insights from Get With The Guidelines-Heart Failure (GWTG-HF).JACC Heart Fail. 2017; 5:305–309. doi: 10.1016/j.jchf.2016.12.018CrossrefMedlineGoogle Scholar
  • 7. Mohanty AF, Levitan EB, King JB, Dodson JA, Vardeny O, Cook J, Herrick JS, He T, Patterson OV, Alba PR, et al. Sacubitril/valsartan initiation among veterans who are renin-angiotensin-aldosterone system inhibitor naïve with heart failure and reduced ejection fraction.J Am Heart Assoc. 2021; 10:e020474. doi: 10.1161/JAHA.120.020474LinkGoogle Scholar
  • 8. Ozaki AF, Krumholz HM, Mody FV, Tran TT, Le QT, Yokota M, Jackevicius CA. Prior authorization, copayments, and utilization of sacubitril/valsartan in medicare and commercial plans in patients with heart failure with reduced ejection fraction.Circ Cardiovasc Qual Outcomes. 2021; 14:e007665. doi: 10.1161/CIRCOUTCOMES.120.007665LinkGoogle Scholar
  • 9. Sandhu AT, Ollendorf DA, Chapman RH, Pearson SD, Heidenreich PA. Cost-effectiveness of sacubitril-valsartan in patients with heart failure with reduced ejection fraction.Ann Intern Med. 2016; 165:681–689. doi: 10.7326/M16-0057CrossrefMedlineGoogle Scholar
  • 10. King JB, Shah RU, Bress AP, Nelson RE, Bellows BK. Cost-effectiveness of sacubitril-valsartan combination therapy compared with enalapril for the treatment of heart failure with reduced ejection fraction.JACC Heart Fail. 2016; 4:392–402. doi: 10.1016/j.jchf.2016.02.007CrossrefMedlineGoogle Scholar
  • 11. Luo N, Lippmann SJ, Mentz RJ, Greiner MA, Hammill BG, Hardy NC, Laskey WK, Heidenreich PA, Chang CL, Hernandez AF, et al. Relationship between hospital characteristics and early adoption of angiotensin-receptor/neprilysin inhibitor among eligible patients hospitalized for heart failure.J Am Heart Assoc. 2019; 8:e010484. doi: 10.1161/JAHA.118.010484LinkGoogle Scholar
  • 12. Hung CL, Chao TF, Su CH, Liao JN, Sung KT, Yeh HI, Chiang CE. Income level and outcomes in patients with heart failure with universal health coverage.Heart. 2021; 107:208–216. doi: 10.1136/heartjnl-2020-316793CrossrefMedlineGoogle Scholar
  • 13. DeJong C, Kazi DS, Dudley RA, Chen R, Tseng CW. Assessment of national coverage and out-of-pocket costs for sacubitril/valsartan under medicare part D.JAMA Cardiol. 2019; 4:828–830. doi: 10.1001/jamacardio.2019.2223CrossrefMedlineGoogle Scholar
  • 14. Smaha LA; American Heart Association. The American Heart Association Get With The Guidelines program.Am Heart J. 2004; 148(5 Suppl):S46–S48. doi: 10.1016/j.ahj.2004.09.015CrossrefMedlineGoogle Scholar
  • 15. Weeks WB, Ouayogodé MHL, Weinstein JN. Association between a measure of community economic distress and medicare patients’ health care utilization, quality, outcomes, and costs.J Gen Intern Med. 2018; 33:1433–1435. doi: 10.1007/s11606-018-4478-7CrossrefMedlineGoogle Scholar
  • 16. Charles EJ, Mehaffey JH, Hawkins RB, Fonner CE, Yarboro LT, Quader MA, Kiser AC, Rich JB, Speir AM, Kron IL, et al; Investigators for the Virginia Cardiac Services Quality Initiative. Socioeconomic distressed communities index predicts risk-adjusted mortality after cardiac surgery.Ann Thorac Surg. 2019; 107:1706–1712. doi: 10.1016/j.athoracsur.2018.12.022CrossrefMedlineGoogle Scholar
  • 17. Buuren SV. MICE: multivariate imputation by chained equations in R.J Stat Softw. 2011; 45:1–67. doi: 10.18637/jss.v045.i03CrossrefGoogle Scholar
  • 18. Buuren SV. Flexible imputation of missing data. Second edition. ed. CRC Press, Taylor & Francis Group; 2018.CrossrefGoogle Scholar
  • 19. Heymans MW, Eekhout I. Applied missing data analysis with SPSS and (R)studio. 2019.Google Scholar
  • 20. Doove LL, Van Buurenca S, Dusseldorp E. Recursive partitioning for missing data imputation in the presence of interaction effects.Comput Stat Data Anal. 2014; 72:92–104. doi: 10.1016/j.csda.2013.10.025CrossrefGoogle Scholar
  • 21. Jaeger BC, Cantor R, Sthanam V, Xie R, Kirklin JK, Rudraraju R. Improving outcome predictions for patients receiving mechanical circulatory support by optimizing imputation of missing values.Circ Cardiovasc Qual Outcomes. 2021; 14:e007071. doi: 10.1161/CIRCOUTCOMES.120.007071LinkGoogle Scholar
  • 22. Sommet N, Morselli D. Keep calm and learn multilevel logistic modeling: a simplified three-step procedure using stata, R, Mplus, and SPSS.Inter Rev Soc Psychol. 2017; 30:203–218. doi: 10.5334/irsp.90CrossrefGoogle Scholar
  • 23. Bates D, Machler M, Bolker B, Walker S. Fitting linear mixed-effects models using lme4.J Stat Softw. 2015; 67:1–48. doi: 10.18637/jss.v067.i01CrossrefGoogle Scholar
  • 24. Fonarow GC, Abraham WT, Albert NM, Stough WG, Gheorghiade M, Greenberg BH, O’Connor CM, Sun JL, Yancy CW, Young JB; OPTIMIZE-HF Investigators and Coordinators. Influence of beta-blocker continuation or withdrawal on outcomes in patients hospitalized with heart failure: findings from the OPTIMIZE-HF program.J Am Coll Cardiol. 2008; 52:190–199. doi: 10.1016/j.jacc.2008.03.048CrossrefMedlineGoogle Scholar
  • 25. Gilstrap LG, Fonarow GC, Desai AS, Liang L, Matsouaka R, DeVore AD, Smith EE, Heidenreich P, Hernandez AF, Yancy CW, et al. Initiation, continuation, or withdrawal of angiotensin-converting enzyme inhibitors/angiotensin receptor blockers and outcomes in patients hospitalized with heart failure with reduced ejection fraction.J Am Heart Assoc. 2017; 6:e004675. doi: 10.1161/JAHA.116.004675LinkGoogle Scholar
  • 26. Hernandez AF, Mi X, Hammill BG, Hammill SC, Heidenreich PA, Masoudi FA, Qualls LG, Peterson ED, Fonarow GC, Curtis LH. Associations between aldosterone antagonist therapy and risks of mortality and readmission among patients with heart failure and reduced ejection fraction.JAMA. 2012; 308:2097–2107. doi: 10.1001/jama.2012.14795CrossrefMedlineGoogle Scholar
  • 27. Bhagat AA, Greene SJ, Vaduganathan M, Fonarow GC, Butler J. Initiation, continuation, switching, and withdrawal of heart failure medical therapies during hospitalization.JACC Heart Fail. 2019; 7:1–12. doi: 10.1016/j.jchf.2018.06.011CrossrefMedlineGoogle Scholar
  • 28. Roderick JAL, Donald BR. Statistical Analysis with Missing Data. 3rd ed. Wiley; 2019.Google Scholar
  • 29. Enders CK. Multiple imputation as a flexible tool for missing data handling in clinical research.Behav Res Ther. 2017; 98:4–18. doi: 10.1016/j.brat.2016.11.008CrossrefMedlineGoogle Scholar
  • 30. Vach W, Blettner M. Biased estimation of the odds ratio in case-control studies due to the use of ad hoc methods of correcting for missing values for confounding variables.Am J Epidemiol. 1991; 134:895–907. doi: 10.1093/oxfordjournals.aje.a116164CrossrefMedlineGoogle Scholar

Source link