Physician Network Connections Associated with Faster De-Adoption of Dronedarone for Permanent Atrial Fibrillation

Abstract

Background:

Physicians’ professional networks are an important source of new medical information and have been shown to influence the adoption of new treatments, but it is unknown how physician networks impact the de-adoption of harmful practices.

Methods:

We analyzed changes in physicians’ use of dronedarone after the PALLAS trial (November 2011) showed that dronedarone increased the risk of death from cardiovascular events among patients with permanent atrial fibrillation (AF). De-identified administrative claims from the OptumLabs^® Data Warehouse were combined with physicians’ demographic information from the Doximity database and publicly available data on physicians’ patient-sharing relationships compiled by the Centers for Medicare & Medicaid Services. We used a linear probability model with an interrupted linear time trend specification to model the impact of the PALLAS trial on physicians’ dronedarone usage between 2009 and 2014.

Results:

Prior to the PALLAS trial, the use of dronedarone was increasing by 0.22 percentage points per quarter [95% CI, 0.19 – 0.25] in our Medicare Advantage sample (N=343,429 patient-quarter observations) and 0.63 percentage points per quarter [95% CI, 0.52 – 0.75] in our commercially insured sample (N=44,402 patient-quarter observations). After the PALLAS trial and subsequent FDA black box warning, physicians in the Medicare Advantage sample with an above median number of network connections to other physicians decreased their quarterly usage of dronedarone by 0.12 percentage points more per quarter [95% CI, −0.20 – −0.04; p=0.031] than physicians with equal to or below the median number of network connections. Similar patterns existed in the commercially insured sample (p=0.0318).

Conclusions:

After controlling for a wide range of patient, physician, and geographical characteristics, physicians with a greater number of network connections were faster de-adopters of dronedarone for patients with permanent AF after the PALLAS trial and subsequent FDA black box warning detailed the harmfulness of dronedarone for these patients. Policies for improving physicians’ responsiveness to new medical information should consider utilizing the influence of these important professional network relationships.

Treatment practices and the use of evidence-based cardiology care vary significantly across facilties^1,2 and physicians^3–5 in the United States. These differences have predominantly been measured by physicians’ rate of adoption of new treatment methods, but an equally important and understudied clinical decision is the de-adoption of treatments found to be unsafe, or less safe than previously understood.⁶ Physicians’ professional networks are a known source of new medical information and treatment advice^7,8 and have been shown to significantly influence the adoption of new treatments,^9,10 but it is unknown how physician networks impact the de-adoption of harmful practices.¹¹

Physicians’ professional networks are commonly estimated using administrative health data, such as health insurance claims, by connecting two physicians who are observed treating the same patient within a calendar year.¹² Defining professional relationships between two physicians who are observed sharing patients in this way has been validated as a strong measure of physicians’ professional connections in several clinical settings,^13,14 and a physician’s number of network connections has been significantly associated with their use of specific treatments,^15,16 patient health outcomes,^17,18 and health care costs.^7,19 However, the association between network connections and physicians’ rate of de-adoption of harmful treatments has yet to be investigated. Additionally, many past studies on physician network connections did not consider physicians’ clinical specialization or other physician characteristics that might simultaneously impact their professional relationships and treatment behaviors.

It has been repeatedly shown that physicians’ clinical experience and specialization are associated with the adoption of new evidence-based treatments^20,21 and better patient outcomes.^22–24 Accordingly, patients may be referred to more experienced physicians or specialists in pursuit of higher quality care, and thus physicians with a large number of network connections may also be these more experienced or specialized physicians. This means that a physician’s clinical experience and specialization may be simultaneously related to both their number of network connections and their treatment choice. To properly account for these simultaneous relationships, we estimated the association between physicians’ number of network connections and their rate of de-adoption of a harmful treatment practice while controlling for physicians’ experience and specialization, and we contrasted the relative association of these important physician characteristics on the de-adoption process.

Specifically, this study examined changes in physicians’ use of dronedarone in response to new evidence describing the harmful effects of dronedarone for patients with permanent atrial fibrillation (AF). Dronedarone is an antiarrhythmic drug used to treat intermittent AF that was approved by the FDA in July 2009. The 2011 PALLAS trial was conducted to assess whether patients with permanent AF would also benefit from dronedarone, and in November 2011 the trial was stopped early due to clear safety concerns: the researchers reported that dronedarone doubled the rate of cardiovascular-related death and stroke compared to the control group.²⁵ The FDA quickly responded by issuing a Risk Evaluation and Mitigation Strategy (REMS) and a black box warning in December 2011, which advised physicians about the increased risk associated with dronedarone use for patients with permanent AF.²⁶ In our analyses, we examined dronedarone use among a cohort of patients with permanent AF before and after the publication of the PALLAS trial and the FDA REMS and black box warning, and estimated the association between physicians’ de-adoption of dronedarone (measured as a decrease in their quarterly trend of dronedarone use) and their clinical experience, specialization, and number of professional network connections.

Studying these associations between physicians’ characteristics and their de-adoption of a harmful medication can reveal important physician-level determinants of the use of evidenced-based care. All physicians are expected to adhere to the latest clinical care guidelines and medical evidence, which requires that physicians stay informed of the latest medical research and clinical treatment guidelines. In our setting, providing evidence-based care for patients with permanent AF required physicians to de-adopt the use of dronedarone after the PALLAS trail and subsequent FDA black box warning was issued in December 2011. Since physicians’ professional peers are a common source of new medical information and often consulted on treatment practices,^7,8,27 we hypothesized that the physicians with a greater number of network connections would be the faster de-adopters of dronedarone for their patients with permanent AF. Having more network connections means that a physician had more professional peers who could have shared the PALLAS research findings or communicated about the FDA black box warning, which would increase the physician’s likelihood of de-adopting dronedarone. Other mechanisms may also be contributing to an observed association between physicians’ network connections and their rate of dronedarone de-adoption though, and we examine these alternative explanations in the Discussion section.

Methods

Data Sources

This study used de-identified administrative claims from the OptumLabs^® Data Warehouse (OLDW) between 2009 and 2014, which included medical and pharmacy claims, laboratory results, and enrollment records for commercially insured and Medicare Advantage enrollees. The data contained longitudinal health information on enrollees, representing a diverse mixture of ages, ethnicities and geographical regions across the United States.²⁸ OptumLabs combined the OLDW claims data with physicians’ demographic information from the Doximity database using National Provider Identifiers. Doximity is a data resource that contains information on physicians’ gender, age, clinical specialty, medical school, and time since residency; Doximity has been described in detail and validated in several prior studies.^29,30 We also used publicly available data on physicians’ patient-sharing relationships observed from the treatment of Medicare and Medicaid beneficiaries, which was compiled by the Center for Medicare & Medicaid Services (CMS) in response to a Freedom of Information Act request. Specifically, we used the annual data sets (2009 – 2014) that list all pairs of physicians who filed claims for 11 or more of the same Medicare or Medicaid beneficiaries over a 180-day interval during the same calendar year. The analytic codes are available in the supplementary materials, but due to the sensitive nature of the data and our Data Use Agreements with OptumLabs, CMS, and Doximity we are unable to share the data or the proprietary code used to construct the analytic patient cohorts.

Patient Population

Our analytic sample contained quarterly observations of dronedarone use for each quarter that a patient was identified as having permanent AF. To identify patients with permanent AF, we first required that patients had at least 12 months of continuous plan enrollment. We then considered patients as potentially having permanent AF if they had either two or more outpatient claims or one or more inpatient claim(s) with an AF diagnosis (ICD-9-CM 427.31 or ICD-10-CM 148.91) in the preceding year, or if they had five or more total AF claims of any type between 2009 and 2014. Among these patients, we identified a patient as having permanent AF if their number of AF claims was equal to or greater than the 90^th percentile among all patients with AF claims in the past 12 months.³¹ Then, because permanent AF patients should no longer receive active management for their condition, we excluded all patients with a cardioversion procedure (CPT 92960–61 and ICD-9 procedure 99.61 and 99.62) or a catheter ablation procedure (CPT 93650–56 and ICD-9 procedure 37.34, 37.26, and 37.27) in any setting in the past 12 months. We tested whether our results were sensitive to the requirement that patients not have a cardioversion or ablation in the past twelve months by including patients with evidence of cardioversion or ablation in our analytic sample in analyses presented in the supplementary materials.

Physician Attribution Algorithm

We attributed each patient-quarter observation to the physician most likely to have been responsible for their AF care using a multi-step attribution algorithm based on patients’ evaluation and monitoring (E&M) office visits and prescription drug claims, as done in prior studies.^31–35 In the first step, we ranked physicians based on their number of E&M office visits and pharmacy claims for each patient, and attributed patients to the unique physician who had the lowest rank (most claims) on these two dimensions of care. For the patients not attributed to a single physician in this first step, our second step considered only physicians’ E&M office visits rank or physicians’ pharmacy claims rank if the patient did not have any E&M visits in the preceding year. If a patient was still unattributed based on E&M and pharmacy claims, our third step attributed patients to their provider from the previous year. If we were unable to attribute a physician to a given patient, that patient was excluded from our analytic sample. Additional details on the specifics of this attribution algorithm are provided in the supplementary materials, and Table 1 displays the proportion of patient-quarter observations attributed to a physician at each of these three steps. This algorithm was informed both by prior validation studies^31–35 and by the research team’s experience in cardiology care; it prioritized attributing patients to the physicians on E&M office visits because these physicians are more likely to be responsible for writing new prescriptions, whereas physicians on pharmacy claims may not be involved in the prescription decision and are more likely to be filling prescription refills.

Table 1:

Cohort Construction

Step:	Medicare Advantage (patient-quarter obs.)	Commercial (patient-quarter obs.)
1.) Claim with AF in last 12 months between 2009 Q3 and 2014 Q3 (excluding washout period)	5969746	2657656
2.) 12 months of continuous insurance enrollment	5927323	1903535
3.) ≥1 inpatient claim for AF in last 12 months, ≥2 outpatient claims for AF in last 12 months, or ≥5 claims for AF since the beginning of study period	2814944	723924
4.) Observation in 90th percentile of total AF claims	819232	116465
5.) No evidence of ablation in last 12 months	800432	107311
6.) No evidence of cardioversion in last 12 months	733507	78581
7.) No missing patient information (age, sex)	733507	78499
8.) Observation attributed to single physician for AF care	707406	77589
9.) No missing provider information from Doximity	559662	62530
10.) No missing PCSA information	559032	62178
11.) Observation linked to CMS network characteristics	343429	44402

Note: AF, atrial fibrillation. PCSA, primary care service area; CMS, Centers for Medicare and Medicaid Services

Physician Networks: Construction and Measurement

In order to calculate physicians’ number of network connections, we first needed to estimate physicians’ professional networks. For each physician attributed to the cardiology care of a patient(s) with permanent AF in our data,^31–35 we estimated their professional network from publicly available data on their patient-sharing relationships with other physicians for all Medicare and Medicaid beneficiaries.³⁶ Specifically, the national CMS data defined a network connection³⁷ between two physicians who treated the same patient within 180 days of each other in the same calendar year, which is a validated technique for constructing professional networks in this workplace setting.^13,14 We restricted our analysis to the network connections between physicians with clinical specialties most likely to provide care for permanent AF (e.g. internal medicine, cardiology, family medicine), and constructed annual physician networks from these individual connections for each year in our study period (2009 – 2014).

The extent to which physicians are connected with peers in their professional network is measured using physicians’ annual number of network connections (often called “degree centrality” in the social networks literature).^7,17,18 While the number of network connections defined from Medicare and Medicaid beneficiaries is not likely to capture all of a physician’s clinical collaborations and professional peers, it is expected that this measure is proportional to physicians’ total number of network connections. This follows from recent research that found physicians’ number of network connections (degree centrality) was highly correlated when estimated using administrative data from different payers (e.g. Medicare vs. private insurance).³⁸ Additionally, this measure of network connections has been repeatedly associated with treatment quality and health outcomes.^15,17

Additional Physician, Patient, and Geographical Covariates

A physician’s number of network connections is likely a function of many physician attributes, such as their experience, specialization, practice location, and preferences. In order to properly estimate the relationship between physicians’ network connections and their rate of dronedarone de-adoption, we included additional physician-level covariates to control for these important confounding, simultaneously occurring relationships.

We derived physician-level variables from the Doximity data. First, we constructed an identifier for physicians who went to a top 20 medical school in the U.S. based on whether the school was in the top 20 of either the U.S. News and World Report’s annual primary care medical school ranking or research medical school ranking in over half of the years between 2007 and 2014. We also identified whether a physician went to a non-U.S. medical school based on a search of U.S. medical school accreditation records. These identifiers were combined with information on physicians’ gender, age, clinical specialty, and years since residency. We excluded all patient-quarters without an attributed physician with these observable characteristics from Doximity.

Physicians’ use of dronedarone may also be correlated to patient- and geographic-level attributes, which is why we also included model covariates at these two levels. Patient-level measures were obtained from the OLDW claims data, including patient’s gender and age. We additionally constructed a time-varying measure of annual health status based on the number of Elixhauser comorbidities³⁹ recorded over the past year from each patient-quarter observation in our analytic sample.

Geographical information was linked to patients based on their zip code of residence from the OLDW data, which we used to assign patients to a primary care service area (PCSA) as defined by The Dartmouth Atlas;⁴⁰ PCSAs are aggregations of census tracts meant to reflect patterns in how people utilize primary care. We then used PCSA-level data from The Dartmouth Atlas on the per capita number of internists and the percent of PCSA residents who are female, Black, White, Hispanic, and older than 65. We also calculated the Agency for Healthcare Research and Quality Socioeconomic Status (SES) index for each PCSA, which is derived from a principal component analysis that considers household crowding, property values, poverty, household income, education, and unemployment; higher values correspond to higher levels of SES.⁴¹ We also identified patients as living in a metropolitan PCSA if the population-weighted proportion of census tracts that are classified by the United States Department of Agriculture rural-urban commuting area codes of 3 or lower⁴² is greater than or equal to 50%. Finally, we used measures of the PCSAs’ malpractice environment, defined by national rankings of both legal risk (the extent to which malpractice laws are unfavorable for physicians) and insurance risk (the cost of malpractice insurance premiums).⁴³

Statistical Analyses

Summary statistics are presented as frequency and percentage (for categorical variables), or as median and standard deviation (for continuous variables). A χ² test was used to compare categorical variables, and the Wilcoxon rank sum test was used to compare continuous variables. All analyses were conducted for complete observations only.

To model the impact of the PALLAS trial and FDA black box warning on dronedarone usage, we used a linear probability model (LPM) with an interrupted linear time trend specification.^44,45 Our outcome variable was an indicator that equaled one if a patient received a dronedarone prescription in a given quarter-year, and zero otherwise. We used LPMs for ease of interpretation,^46,47 and because non-linear logistic or probit models were computationally intractable for these large datasets. All LPM specifications included the patient, geographical, and physician covariates described previously, including physicians’ clinical specialty and annual permanent AF patient caseload, which was measured by the number of permanent AF patients in the analytic dataset that were treated by a given physician in each calendar year. Standard error estimates were double-clustered at the physician and patient levels. We used a quarterly linear time trend, an indicator variable for the quarters after the PALLAS trial results were published and FDA black box warning was issued, and the interaction of the linear time trend and this post-PALLAS indicator variable to estimate the change in dronedarone use over this period. A negative coefficient on the interaction term indicated that the quarterly trend in dronedarone use declined after the PALLAS trial and FDA black box warning.

To model the associations between physician-level characteristics and the decision to de-adopt dronedarone, we included additional interaction terms between the interrupted time trend terms and each physician characteristic. For example, being a cardiologist was included as its own binary covariate as well as in interaction terms with the quarterly linear time trend, the post-PALLAS indicator, and with the interaction of the quarterly time trend and post-PALLAS indicator. In this way, a negative coefficient on the three-way interaction of the cardiologist indicator, the quarterly time trend, and the post-PALLAS indicator demonstrated that cardiologists de-adopted dronedarone at a faster rate after the PALLAS trial than non-cardiologists.

To more easily compare the association between a physician’s number of annual network connections and their dronedarone de-adoption with the other dichotomous physician characteristics, the measure of annual network connections was dichotomized to equal one if it was greater than the median value observed over the study period, and zero otherwise. Similarly, physicians’ annual permanent AF patient caseload was dichotomized to equal one if it was greater than the median value observed over the study period, and zero otherwise. A second regression model was estimated using the logarithm of the number of annual network connections to quantify the association of a one percent increase in network connections on physicians’ rate of de-adoption. Additional robustness analyses tested variations in the sample construction and definitions used to dichotomize physicians’ number of network connections, and these results are presented in the supplementary materials.

Results

Descriptive Statistics

Overall changes in dronedarone use along with patient and physician characteristics for the 343,429 patient-quarters in the Medicare Advantage sample and 44,402 patient-quarters in the commercially insured sample are presented in Table 2, and the quarterly trends over the sample period are plotted in Figure 1. After dronedarone was approved by the FDA in 2009, the use of dronedarone for patients with permanent AF rose from 0.69% in the Medicare Advantage and 3.69% in commercially insured samples in 2010 to 1.78% of the Medicare Advantage and 5.235% of the commercially insurance patient-quarter observations in 2014. On average, the Medicare Advantage enrollees were older than the commercially insured, and were more likely to have 6 or more comorbidities and be female. In terms of the geographic covariates, the Medicare Advantage populations were more likely to be in areas with fewer primary care physicians (PCPs) but in higher socioeconomic status areas than the commercially insured patients. Finally, the physicians treating the Medicare Advantage sample were less likely to be under 50 years old, and more likely to be female, to have attended a non-US medical school, and to be a primary care physician than the physicians treating the commercially insured. Among these physicians, the median number of network connections (i.e. degree centrality) was 206 and 290 for the physicians treating Medicare Advantage enrollees and the commercially insured, respectively. We also found that physicians’ network connections were strongly correlated with being a cardiologist and having a large caseload of patients with permanent AF (see supplementary materials), which supports the need for a multivariable regression model.

Table 2:

Patient, PCSA, and Physician Characteristics across Analytic Samples

	Medicare Advantage		Commercially Insured		Pvalue
Dronedarone use 2010, %	0.69		3.69		<0.0001
Dronedarone use 2014, %	1.78		5.23		<0.0001
Sample size (patient-quarter obs.), n	343429		44402
Patient Characteristics (n = patient-quarters)
Age, median (IQR)	80	(10)	63	(8)	<0.0001
Female, n (%)	177573	(51.7)	14102	(31.8)	<0.0001
3 or Fewer Comorbidities, n (%)	62668	(18.2)	15048	(33.9)	<0.0001
4 or 5 Comorbidities, n (%)	88903	(25.9)	13208	(29.7)	<0.0001
6 or 7 Comorbidities), n (%)	76598	(22.3)	7847	(17.7)	<0.0001
8 or More Comorbidities, n (%)	115260	(33.6)	8299	(18.7)	<0.0001
Multiple CHF annual visits, n (%)	86341	(25.1)	8475	(19.1)	<0.0001
Physician Attributed in Step 1, n (%)	187892	(54.7)	29143	(65.6)	<0.0001
Physician Attributed in Step 2, n (%)	148209	(43.2)	14377	(32.4)	<0.0001
Physician Attributed in Step 3, n (%)	7328	(2.1)	882	(2.0)	0.0423
Number of unique patients, n	81593		13649
PCSA Characteristics (n = patient-quarters)
Number of PCPs, median (IQR)	21	(43)	24	(48)	<0.0001
Female, median (IQR) %p	51	(15)	51	(14)	<0.0001
Black, median (IQR) %p	5	(12.5)	7	(15.3)	<0.0001
Hispanic, median (IQR) %p	6	(9.0)	7	(118)	<0.0001
White, median (IQR) %-	84	(20.7)	80	(21.8)	<0.0001
Older than 64, median (IQR) %p	14	(4.8)	13	(4.8)	<0.0001
Metro PCSAs, n (%)	297411	(86.6)	39214	(88.3)	<0.0001
AHRQ SES Index median (IQR)	53.6	(4.7)	54	(5.5)	<0.0001
Malpractice ins. Rank, median (IQR)	31	(35)	31	(20)	<0.0001
Malpractice legal rank, median (IQR)	32	(24)	32	(25)	<0.0001
Number of unique PCSAs, n	4286		3063
Physician Characteristics (n = patient-quarters)
Under 50 years old, n (%)	110413	(32.2)	15485	(34.9)	<0.0001
Annual AF caseload, median (IQR) n	12	(18)	8	(13)	<0.0001
Female, n (%)	44254	(12.9)	4368	(9.8)	<0.0001
Top 25 US Med. School, n (%)	53561	(15.6)	7186	(16.2)	0.0013
Non-Top 25 US Med. School, n (%)	193464	(56.3)	25511	(57.5)	<0.0001
Non-US Med. School, n (%)	96404	(28.1)	11705	(26.4)	<0.0001
Cardiologist, n (%)	196643	(57.3)	33961	(76.5)	<0.0001
PCP, n (%)	141569	(41.2)	9914	(22.3)	<0.0001
Other specialty, n (%)	5217	(15)	527	(12)	<0.0001
Less than 20 yrs. post-residency, n (%)	137181	(39.9)	19414	(43.7)	<0.0001
Number of unique physicians, n	24779		8870
Physician Networks
Network connections, median (IQR)	206	(307)	290	(310)	<0.0001

Note: Patient, PCSA, and physician characteristics are presented for the Medicare Advantage and commercially insured analytic samples describing dronedarone use by patient-quarter observations from 2009q3 – 2014q4. P-values correspond to Wilcoxon rank sum tests or χ² tests when appropriate. Physicians were attributed to a patient according to an algorithm described in the methods section and detailed in the supplementary materials, and the proportion of patient-quarter observations attributed to a physician in each of the three steps of the algorithm as displayed in the first panel. Annual AF caseload is defined as the number of permanent AF patients treated in each calendar year. Physician network connections were calculated as a physician’s annual number of patient-sharing relationships, or equivalently, the number of clinical care collaborations with other physicians. CHF; congestive heart failure; PCSA, primary care service area; PCP, primary care physician; pop. total population; IQR, interquartile range; AHRQ, Agency for Health care Quality and Research; SES, socioeconomic status; AF, atrial fibrillation.

Figure 1:

These plots display the unadjusted use of dronedarone over the sample period (2009q3 – 2014q4). The plot on the left was estimated for the Medicare Advantage sample and the plot on the right was estimated for the commercially insured sample.

Network Connections and Dronedarone De-adoption

Table 3 displays the adjusted change in dronedarone use between 2009 and 2014 that was estimated by the interrupted time series LPM model that included all of the available patient, geographical, and physician characteristics, including physicians’ clinical specialty, annual permanent AF patient caseload, and physicians’ network connections. The two columns display the different trends in dronedarone de-adoption estimated within the Medicare Advantage and commercially insured patient samples, and the results show that dronedarone use was increasing by 0.22 percentage points per quarter [95% CI, 0.19 – 0.25] in the Medicare Advantage sample and 0.63 percentage points per quarter [95% CI, 0.52 – 0.75] in the commercially insured sample prior to the PALLAS trial. After the PALLAS trial and FDA black box warning, the quarterly change in dronedarone use stopped increasing in the Medicare Advantage sample and began decreasing by 0.23 percentage points per quarter [95% CI, −0.27 – −0.19] in the commercially insured sample. This resulted in statistically significant decreases in the quarterly time trends of dronedarone usage in both samples.

Table 3:

Changes in Dronedarone Use by Physicians’ Network Connections

	(1) Medicare Advantage	(2) P Value	(3) Commercially Insured	(4) P Value
Observed Time Trend
Pre-PALLAS Qtly Trend	0.0022 [0.0019,0.0025]	<0.001	0.0063 [0.0052,0.0075]	<0.001
Post-PALLAS Qtly Trend	-0.0001 [−0.0004,0.0001]	0.3495	-0.0006 [−0.0018,0.0007]	0.3989
Change in Trend	-0.0023 [−0.0027,−0.0019]	<0.001	-0.0069 [−0.0087,−0.0051]	<0.001
Observations	343429		44402

Note: This table displays the adjusted quarterly change in dronedarone use separately among the Medicare Advantage and commercially insured patient samples. The Pre-PALLAS Trial Qtly Trend is a quarterly time trend in dronedarone usage before 2011:Q3, the Post-PALLAS Qtly Trend is a quarterly time trend after 2012:Q1, and Change in Trend is the difference in these two quarterly rates of change, where dronedarone usage was also adjusted for all available patient, PCSA, and physician characteristics (displayed in Table 2). 95% confidence intervals are displayed in brackets.

PALLAS, PALbociclib CoLlaborative Adjuvant Study (2011q3); Qtly, quarterly.

The results displayed in Table 4 show the estimated interaction between quarterly time trends and the indicated dichotomous physician characteristic from separate LPM models that also estimated the overall quarterly time trends (as in Table 3), where the coefficients show how the quarterly rate of change in dronedarone usage differs both before and after the PALLAS trial according to each physician attribute. The results in the first panel, titled Difference in Trend for Physicians with Above (vs. Below) Median Network Connections, indicate that the physicians with above median number of network connections were increasing their use of dronedarone prior to the PALLAS trial by 0.13 percentage points more per quarter [95% CI, 0.08 – 0.19] than physicians with an equal to or below the median number of network connections in the Medicare Advantage sample. After the PALLAS trial and FDA black box warning, these physicians decreased their quarterly trend in dronedarone use by 0.12 percentage points more per quarter [95% CI, -<0.20 – −0.04] than physicians with a low number of network connections. Similar patterns existed for the commercially insured sample, where physicians with an above median number of network connections were faster adopters of dronedarone prior to the PALLAS trial, and then were faster de-adopters of dronedarone after the PALLAS trail and FDA black box warning. These observed differences in dronedarone use between physicians with above and below the median number of network connections are graphically displayed in Figure 2. Additionally, these differences in dronedarone de-adoption by physicians’ network connections are robust to changes in the cohort construction methods and when using the 25^th percentile (instead of the median) to define the dichotomous measure of physicians’ network connections (see supplementary materials).

Table 4:

Changes in Dronedarone Use According to Physician Characteristics

	(1) Medicare Advantage	(2) P Value	(3) Commercially Insured	(4) P Value
Model 1: Difference in Trend for Physicians with Above (vs. Below) Median Network Connections
Pre-PALLAS Qtly Trend	0.0013 [0.0008,0.0019]	<0.001	0.0011 [−0.0012,0.0034]	0.3304
Post-PALLAS Qtly Trend	0.0001 [−0.0004,0.0006]	0.5775	-0.0029 [−0.0055,−0.0002]	0.0349
Change in Trend	-0.0012 [−0.0020,−0.0004]	0.0029	-0.0040 [−0.0077,−0.0003]	0.0324
Model 2: Difference in Trend for Physicians with Above (vs. Below) Median AF Patient Caseloads
Pre-PALLAS Qtly Trend	-0.0001 [−0.0006,0.0005]	0.8455	-0.0015 [−0.0039,0.0009]	0.2137
Post-PALLAS Qtly Trend	0.0001 [−0.0004,0.0006]	0.6112	0.0010 [−0.0018,0.0037]	0.4947
Change in Trend	0.0002 [−0.0006,0.0010]	0.6396	0.0025 [−0.0013,0.0062]	0.1959
Model 3: Difference in Trend for Female Physicians
Pre-PALLAS Qtly Trend	-0.0005 [−0.0012,0.0003]	0.2193	0.0015 [−0.0027,0.0057]	0.4896
Post-PALLAS Qtly Trend	-0.0004 [−0.0010,0.0003]	0.2618	0.0039 [−0.0004,0.0083]	0.0771
Change in Trend	0.0001 [−0.0009,0.0011]	0.8566	0.0025 [−0.0036,0.0085]	0.4231
Model 4: Difference in Trend for Cardiologists
Pre-PALLAS Qtly Trend	0.0017 [0.0012,0.0023]	<0.001	0.0026 [0.0003,0.0049]	0.0270
Post-PALLAS Qtly Trend	-0.0002 [−0.0007,0.0003]	0.3771	-0.0021 [−0.0049,0.0007]	0.1498
Change in Trend	-0.0020 [−0.0027,−0.0012]	<0.001	-0.0046 [−0.0084,−0.0009]	0.0144
Model 5: Difference in Trend for Physicians with <20 yrs post-residency
Pre-PALLAS Qtly Trend	0.0006 [<0.001,0.0012]	0.0439	0.0038 [0.0013,0.0062]	0.0024
Post-PALLAS Qtly Trend	-0.0002 [−0.0007,0.0003]	0.5009	-0.0005 [−0.0031,0.0021]	0.7189
Change in Trend	-0.0008 [−0.0017,0.0001]	0.0718	-0.0043 [−0.0080,−0.0006]	0.0244
Model 6: Difference in Trend for Physicians Who Attended a Top 25 US Med. Sch.
Pre-PALLAS Qtly Trend	-0.0007 [−0.0014,- [<0.001]	0.0426	-0.0016 [−0.0046,0.0015]	0.3139
Post-PALLAS Qtly Trend	0.0001 [−0.0006,0.0008]	0.8488	0.0013 [−0.0023,0.0049]	0.4738
Change in Trend	0.0008 [−0.0002,0.0017]	0.1304	0.0029 [−0.0021,0.0079]	0.2593

Note: The Pre-PALLAS Trial Qtly Trend is a quarterly time trend in Dronedarone usage before 2011:Q3, the Post-PALLAS Qtly Trend is a quarterly time trend after 2012:Q1, and Change in Trend is the difference in these two quarterly rates of change, where dronedarone usage was also adjusted for all available patient, PCSA, and physician characteristics (displayed in Table 2). 95% confidence intervals are displayed in brackets. Each panel represents a separate regression model that includes additional interaction terms between the quarterly time trends and the indicated physician characteristic. For example, Model 1: Difference in Trend for Physicians with Above (vs. Below) Median Network Connections includes additional interaction terms between the quarterly time trends and the dichotomous measure of physicians’ number of network connections that compares physicians with above the median number of network connections to physicians with equal to or below the median number of network connections. Network connections were calculated as the number of patient-sharing relationships, or equivalently, the number of clinical care collaborations with other physicians, and the results show that physicians with above median network connections significantly reduced their quarterly trend in dronedarone usage after the PALLAS results. Model 2: Difference in Trend for Physicians with Above (vs. Below) Median AF Patient Caseloads shows that physicians with annual permanent AF patient caseloads above the median did not have a significantly different rate of dronedarone adoption pre-PALLAS, nor did they have a faster rate of de-adoption post-PALLAS than physicians with equal to or below the median annual permanent AF patient caseload. Similarly, Model 4: Difference in Trend for Cardiologists shows that cardiologists had a significantly faster rate of dronedarone adoption pre-PALLAS, which was then significantly reduced post-PALLAS relative to other physician types. PALLAS, PALbociclib CoLlaborative Adjuvant Study (2011q3); AF; atrial fibrillation; Qtly, quarterly.

Figure 2:

These plots display the adjusted use of dronedarone over the sample period (2009q3 – 2014q4) after controlling for patient, physician, and regional characteristics, and with separate linear time trends estimated for physicians with equal to or below the median number of network connections and for physicians above the median number of network connections over the study period. The plot on the left was estimated for the Medicare Advantage sample and the plot on the right was estimated for the commercially insured sample. We used our model estimates from the full sample to estimate trends for physicians with high vs. low connections at the mean values of all other independent variables.

Other Physician Characteristics and Dronedarone De-adoption

The association between other physician characteristics and changes in physicians’ use of dronedarone before and after the PALLAS trial are presented in subsequent panels of Table 4. The results from Models 2, 3, and 6 show that a physician’s annual permanent AF patient caseload (measured as a dichotomous variable equal to one if a physician’s annual permanent AF patient caseload was above the median, and zero otherwise), gender, and rank of medical school had no significant association with their dronedarone usage over this period. Instead, Model 4 shows that cardiologists were significantly faster adopters of dronedarone prior to the PALLAS trial, and then cardiologists reduced their use of dronedarone significantly more after the PALLAS trial and FDA black box warning than non-cardiologists. This significant difference in dronedarone de-adoption was observed in both the Medicare Advantage and commercially insured populations. Additionally, Model 5 shows that physicians with fewer than 20 years since their residency were slightly more likely to both increase their use of dronedarone prior to the PALLAS trial and more quickly de-adopt dronedarone after the PALLAS trial relative to physicians with 20 or more years since their residency; however, these differences are only significant at the 5% and 10% level.

Additionally, given that physicians with many network connections were more likely to be cardiologists, in Table 5 we present the association between physicians’ network connections and de-adoption rates while simultaneously controlling for differences in de-adoption rates according to physicians’ clinical specialization. The results show that both having above median network connections and being a cardiologist are simultaneously associated with faster de-adoption of dronedarone after the PALLAS trial and FDA black box warning. The association between a greater number of network connections and a faster rate of dronedarone de-adoption is significant at the 1% level for the Medicare Advantage sample and significant at the 5% level for the commercial sample.

Table 5:

Association Between Physicians’ Network Connections and Dronedarone De-Adoption Controlling for Difference in Trend by Clinical Specialty

	(1) Medicare Advantage	(2) P Value	(3) Commercially Insured	(4) P Value
Difference in Trend for Physicians with Above (vs. Below) Median Network Connections
Pre-PALLAS Qtly Trend	0.0012 [0.0002,0.0022]	0.0434	0.0011 [−0.0006,0.0029]	0.1299
Post-PALLAS Qtly Trend	0.0006 [−0.0007,0.0019]	0.1066	-0.0027 [−0.0058,0.0004]	0.1076
Change in Trend	-0.0006 [−0.0001,−0.0011]	0.0068	-0.0026 [−0.0048,−0.0004]	0.0276
Difference in Trend for Cardiologists
Pre-PALLAS Qtly Trend	0.0017 [0.0008,0.0026]	0.0002	0.0027 [0.0001,0.0067]	0.0606
Post-PALLAS Qtly Trend	-0.0007 [−0.0014,<.0001]	0.0671	-0.0003 [−0.0038,0.0032]	0.8694
Change in Trend	-0.0024 [−0.0037,−0.0012]	<0.001	-0.0030 [−0.0051,−0.0009]	0.0293
Observations	343429		44402

Note: The regression results are from a single model of dronedarone use on the Pre-PALLAS Trial Qtly Trend (a quarterly time trend in dronedarone usage before 2011:Q3), the Post-PALLAS Qtly Trend (a quarterly time trend after 2012:Q1), the Change in Trend (the difference in these two quarterly rates of change), as well as these three interrupted time series terms separately interacted with both the dichotomous measure of physicians’ network connections and an identifier for whether physicians’ were cardiologists. These two sets of interaction terms were included simultaneously in the same model, which also adjusted for all available patient, PCSA, and physician characteristics (displayed in Table 2). 95% confidence intervals are displayed in brackets. PALLAS, PALbociclib CoLlaborative Adjuvant Study (2011q3); Qtly, quarterly.

Discussion

In our study of Medicare Advantage and commercially insured patients with permanent AF from 2009 to 2014, we found that the physicians with more network connections were significantly more likely to exhibit a faster rate of adoption of dronedarone prior to the PALLAS trial as well as faster de-adoption of dronedarone in response to the PALLAS trial and subsequent FDA black box warning. Importantly, these associations were statistically significant when simultaneously estimating time trends in dronedarone use according to both physicians’ number of network connections and clinical specialization. These results expand on existing research that had demonstrated a faster rate of dronedarone de-adoption after the PALLAS trial and subsequent FDA black box warning among cardiologists³¹ by identifying the simultaneously significant relationship of physicians’ network connections on the de-adoption decision. Prior studies on the association between physicians’ network connections and their treatment behavior had not explicitly controlled for physicians’ clinical specialization , but we found evidence that the physicians with a large number of network connections were also the most specialized, thus necessitating a multivariate estimation strategy.

These finding support our hypothesis that a great number of network connections are likely to provide a physician with more opportunities to learn about the latest medical research and changes to federal treatment guidelines, and thus increase the likelihood that the physician is providing the latest evidence-based care. The estimated association between faster de-adoption of dronedarone and physicians’ network connections expands on existing research that had similarly proposed that more network connections can improve a physician’s medical knowledge and treatment quality.^8,14,48 Additionally, prior studies have found that physicians with a larger number of network connections were faster to adopt new treatments than less well-connected physicians,^9,10,49 and our results extend these findings to the de-adoption of harmful treatments, further suggesting that physicians with more network connections are more responsive to changes in medical evidence and treatment guidelines. Our models testing alternative cutoffs for high vs. low network connections (presented in the supplementary materials) also provide suggestive evidence that physicians’ de-adoption decision is no longer positively associated with their number of network connections above the 75th percentile (roughly 370 in the Medicare Advantage sample and 430 in the commercially insured sample).

The results also show that dronedarone de-adoption was significantly faster for the commercially insured than the Medicare Advantage patient sample. However, there are many potential reasons for this difference that we were unable to properly disentangle in this research. First, the patients significantly differ between these two samples, and this may have resulted in physicians using different risk assessment methods. Additionally, a higher rate of dronedarone use existed among the commercially insured sample prior to the PALLAS trail, so the faster rate of de-adoption in this sample could reflect that the de-adoption process occurs more rapidly when physicians use the medication at a higher rate. We did not investigate this potential heterogenous de-adoption rate according to physicians’ pre-PALLAS rate of dronedarone use nor did we estimate three-way interactions between physicians’ number of network connections and other physician or patient attributes, which are all avenues for future research.

These results document a significant association between physicians’ number of network connections and dronedarone de-adoption. Even though our results cannot be interpreted as causal, the relationship we document has implications for health care policy that warrant further research. Specifically, our results show that a higher level of evidence-based cardiology care is likely to be found among more well-connected physicians, and conversely, that the physicians most likely to be practicing outdated methods and using recalled medications are those with few network connections to other physicians. Given the effect of physicians’ professional networks on their treatment behaviors that has been identified in other settings,^10,17,49 these results suggest that increasing collaborations between physicians may increase the de-adoption of harmful medications among rural or otherwise disconnected physicians, where increasing physician connections could be facilitated through telemedicine or other electronic communication methods that are becoming more commonplace in clinical settings across the US. While existing research has shown that an urban-rural disparity did not exist in the de-adoption of dronedarone,³¹ urban-rural disparities have been repeatedly documented for other treatment practices and quality of care measures across the US.^50,51 Rural physicians’ number of network connections to other physicians may be one component of these disparities, and future research should investigate the role of physicians’ network connections in these settings. Additionally, future research should test whether telemedicine or other electronic communication methods are a feasible and cost-effective method of connecting physicians and facilitating increased information transfer across physician networks.

Limitations

While our analyses controlled for a wide range of observable patient, physician, and geographical characteristics, our sample construction methods and the information available in our data resulted in several limitations to this study. First, the analyses were conducted on a Medicare Advantage and commercially-insured subpopulation, so our findings may not generalize to other privately insured or the publicly insured patient populations. Second, since there is not an existing methodology for identifying patients with permanent AF using ICD-9 diagnosis codes, we developed an identification algorithm based on our clinical expertise, but we still may have incorrectly defined the analytical patient sample (although our results are robust to variations on the inclusion/exclusion criteria in our analytic sample). Third, physician-level characteristics from the Doximity database and the CMS patient-sharing relationships data were not available for all of the physicians attributed to patients with permanent AF in our data. While there were no observable differences between the physicians who did and did not have data in these additional data sources, there could be unobservable differences that also make these results less generalizable or potentially bias our findings. Fourth, our results are not causal, and future research should attempt to experimentally manipulate physicians’ number of network connections or use other empirical identification strategies to generate a causal estimate for the effect of physicians’ networks on their de-adoption decision. Fifth, we model changes in dronedarone linearly. This approach fits the observed trends well, but may overlook non-linear patterns in dronedarone use over this period. Finally, these analyses did not control for other external influences of physicians’ de-adoption of dronedarone, such as changes in physicians’ practice- or hospital-level policies, public health campaigns, or pharmaceutical companies’ promotional activities.

Conclusions

Physicians with a greater number of network connections exhibited faster de-adoption of dronedarone for patients with permanent AF after the PALLAS trial and FDA black box warning detailed the harmfulness of dronedarone for these patients. These trends were significant after controlling for a wide range of patient, physician, and geographical characteristics, including other physician characteristics that have previously been associated with the use of evidence-based care. Most notably, while we observe that cardiologists also exhibited faster adoption and de-adoption of dronedarone before and after the PALLAS trial, we find that physicians’ network connections were still associated with faster de-adoption of dronedarone even after adjusting for clinical specialization. Future studies should seek to confirm the relationship between physicians’ number of network connections and the de-adoption of harmful practices in order to design new policies for improving physicians’ responsiveness to new medical information and federal treatment guidelines.

Non-standard Abbreviations and Acronyms:

AF

atrial fibrillation

PALLAS

PALbociclib CoLlaborative Adjuvant Study

REMS

Risk Evaluation and Mitigation Strategy

OLDW

OptumLabs^® Data Warehouse

CMS

Center for Medicare & Medicaid Services

E&M

evaluation and monitoring

PCSA

primary care service area

SES

Socioeconomic Status

LPM

linear probability model

PCPs

primary care physicians