The exact contents of citations.db main_text.text for this paper — one flattened LaTeX string, title through conclusion, appendix excluded, unmodified except for removing email addresses. This is what our citation measures are computed over.
30,282 characters
Professional networks and the diffusion of clinical guidelines in opioid prescribing
\maketitle
\begin{abstract}
Large and persistent differences in opioid prescribing across physicians and regions cannot be explained by patient characteristics or physician attributes alone. We developed a behavioral framework in which prescribing evolves through persistence, exposure to peers in professional networks, and heterogeneous responses to a common policy signal that varies with network centrality. Using nationwide Medicare Part D data from 2013 to 2020, covering more than two million physician-year observations, we tested three hypotheses implied by this framework. Physicians exposed to higher peer prescribing subsequently prescribe more; more central physicians reduce prescribing more following the introduction of the 2016 CDC guideline, with no evidence of differential pre-trends; and changes in peer prescribing are closely associated with changes in individual prescribing in the post-guideline period. By 2020, physicians at the 90th percentile of network centrality exhibited prescribing reductions 0.30 percentage points larger than those at the 10th percentile, with the gap widening steadily after the introduction of the CDC guideline. Together, these results indicate that opioid prescribing operates through professional networks, in which policy effects spread through connections and appear to be shaped by network position. This suggests that engaging highly connected physicians may help extend the reach of opioid stewardship programs. It also raises questions about how the burden and benefits of such targeting would be distributed across physicians and patients.
\end{abstract}
\noindent\textbf{Keywords:} opioid prescribing; physician networks; social learning; CDC guideline; network centrality
\section{Introduction}
The opioid epidemic has claimed more than 500,000 lives in the United
States over the past two decades, and physician prescribing behavior has
played a central role in its trajectory. Despite extensive research, large
and persistent differences in prescribing intensity across physicians and
regions remain difficult to explain, even after accounting for patient
characteristics and the clinical context. Between 2007 and 2012, the share
of Medicare Part D recipients receiving prolonged opioid therapy nearly
doubled, and substantial geographic variation persisted after adjusting for
patient factors \citep{kuo2016trends}. Standard explanations emphasize the
patient case mix, regulatory environments, and physician training; however,
these factors do not fully account for the observed heterogeneity. A
central question remains: why do physicians serving similar patients in
similar settings prescribe so differently?
A growing body of literature suggests that professional networks may
explain this variation. Physicians operate in environments where peer
behavior, shared norms, and local information flow shape clinical decisions
\citep{tasselli2014social,cunningham2012health}. Evidence from
pharmaceutical adoption indicates that physicians are more likely to
prescribe new drugs when their peers have already done so, particularly
when they share patients or practice in close proximity
\citep{donohue2018influence,nair2010asymmetric,yang2014there}. More
centrally positioned physicians often act as opinion leaders and influence
the diffusion of clinical practice
\citep{soumerai1998effect,locock2001understanding,hao2018empirical}.
Similar patterns have been documented in antibiotic prescribing
\citep{gong2025peer,wang2023can}. Whether these network mechanisms operate
in opioid prescribing and shape responses to major policy interventions
remains unclear. In particular, two questions remain unresolved. Do peer
effects persist after accounting for physician fixed effects and local
shocks? Does network position influence how physicians respond to national
clinical guidelines?
This study addresses these questions by developing a unified behavioral
framework that links persistence in individual prescribing, exposure to
peer behavior, and heterogeneous responses to a common policy signal. This
framework builds on the DeGroot tradition of network-based social learning
\citep{golub2010naive,chandrasekhar2020testing,choi2023learning}. A key
feature is that the responsiveness to external policy signals varies with
network centrality, reflecting differences in exposure to peers and
information flows \citep{kandhway2016using}. The network structure plays
two roles: determining the extent of peer exposure and shaping how policy
signals are translated into behavioral changes.
The framework yields three testable hypotheses. First, physicians exposed
to higher peer prescribing intensity exhibit higher subsequent prescribing,
reflecting social learning within local professional networks
\citep{gong2025peer,wang2023can,yang2014there}. Second, physicians who are
more centrally positioned in the network respond more strongly to common
policy shocks because their greater exposure to peer behavior amplifies the
transmission of external signals. Third, changes in peer prescribing are
associated with changes in individual prescribing following policy
intervention, consistent with network-mediated propagation of behavioral
adjustment \citep{kreindler2014diffusion,onnela2010influence}.
These hypotheses were evaluated using nationwide Medicare Part D
prescribing data from 2013 to 2020
\citep{donohue2012sources,kuo2016trends,yin2008effect,ang2025expanded}.
The 2016 CDC opioid prescribing guideline provides a well-defined national
policy shock \citep{bohnert2018opioid}. Physician networks are constructed
from geographic co-location within ZIP codes, a tractable proxy for local
professional interaction
\citep{tasselli2015social,cunningham2012health}.
Our empirical analysis proceeded in three steps. First, we documented a
positive association between peer and individual prescribing robust across
specifications with physician and county-by-year fixed effects. Second,
using an event study design around the 2016 CDC guideline, we showed that
physicians with higher network centrality exhibited larger reductions in
prescribing following the policy change, with no evidence of differential
pre-trends. Third, we found that changes in peer prescribing were
positively associated with changes in individual prescribing in the
immediate post-guideline period.
These results are consistent with a process in which peer exposure, network position, and policy responses together shape prescribing behavior. Policy signals do not act on physicians in isolation,
but move through professional connections, generating heterogeneous and
evolving responses across the network. Consequently, policies that account
for professional network structures may produce broader and more sustained
reductions in prescribing than interventions that target physicians
individually
\citep{valente2012network,galeotti2020targeting,cabana1999don,kilaru2014physicians}.
Beyond opioid prescribing, this study speaks to a broader question in the social science of medicine. It asks how professional norms and peer relationships shape the way clinicians respond to external policy. The physician setting lets us observe this social process at large scale and connect it to a major public health intervention.
\section{Methods}
\subsection*{Data and Sample Construction}
Our analysis drew on three publicly available data sources. Our primary
data source was the Medicare Part D Prescriber Public Use Files provided by
the Centers for Medicare and Medicaid Services, covering 2013 to 2020. The
unit of observation was physician-year.
We restricted the sample to seven commonly prescribed opioid analgesics,
which together accounted for approximately 71\% of all opioid claims
(Table~\ref{tab:opioids}). Drugs were selected based on four criteria:
active opioid ingredients classified under ICD-10 code T40.2, mappable
morphine milligram equivalent (MME) conversion factors, consistent
appearance across all sample years in the Medicare Part D data, and common
use as analgesics for routine outpatient pain management. All prescriptions
were converted to MME using conversion factors from the 2022 CDC Clinical
Practice Guideline for Prescribing Opioids for Pain. Physician geographic
information was obtained from prescriber ZIP codes in the CMS data, mapped
to counties using HUD--USPS ZIP Code Crosswalk files.
The primary outcome was prescribing intensity, defined as the total MME
prescribed divided by the number of opioid claims in a given physician-year.
Patient volume was measured as the total number of unique Medicare
beneficiaries associated with each physician in a given year.
\begin{table}[H]
\centering
\caption{Selected opioid drugs and their share of total opioid claims, 2013--2020.}
\label{tab:opioids}
\begin{tabular}{lrr}
\toprule
Drug name & Total claims & Share (\%) \\
\midrule
Hydrocodone/Acetaminophen & 213,122,922 & 39.68 \\
Oxycodone HCl/Acetaminophen & 69,842,984 & 13.00 \\
Oxycodone HCl & 60,481,680 & 11.26 \\
Morphine Sulfate & 27,303,780 & 5.08 \\
Hydromorphone HCl & 6,436,994 & 1.20 \\
Oxymorphone HCl & 1,964,177 & 0.37 \\
Hydrocodone/Ibuprofen & 672,224 & 0.13 \\
\midrule
\textbf{Total (7 drugs)} & \textbf{379,824,761} & \textbf{70.72} \\
\bottomrule
\end{tabular}
\end{table}
\subsection*{Network Construction and Centrality}
Annual physician networks were constructed based on geographic proximity.
Two physicians were connected in a given year if they practiced within the
same ZIP code. The weight between physicians $i$ and $j$ was:
\begin{equation}
w_{ij} = 0.1 \times \min(\text{Tot\_Benes}_i, \text{Tot\_Benes}_j),
\end{equation}
where $\text{Tot\_Benes}_i$ denotes the number of unique Medicare
beneficiaries treated by physician $i$. Network position was measured using
weighted degree centrality, defined as the sum of a physician's edge
weights.
\subsection*{Unified Behavioral Framework}
We modeled physician prescribing as a dynamic updating process following
the DeGroot tradition
\citep{golub2010naive,degroot1974consensus,ding2019consensus}:
\begin{equation}
p_{i,t+1} = \alpha p_{i,t} + \lambda \bar{p}_{N(i),t} + \mu_i \theta G_t + \varepsilon_{i,t+1}.
\end{equation}
Here $p_{i,t}$ denotes prescribing intensity, $\bar{p}_{N(i),t}$ is the
weighted average prescribing intensity of physician $i$'s network peers,
$\alpha$ captures persistence, $\lambda$ captures social learning, and
$G_t$ denotes the external signal corresponding to the 2016 CDC guideline.
We assumed $\alpha + \lambda < 1$. We allowed centrality-dependent
responsiveness:
\begin{equation}
\mu_i = \mu_0 + \mu_1 C_i,
\end{equation}
where $C_i$ denotes physician $i$'s network centrality. We held $\lambda$
constant across physicians and focused inference on $\mu_1$.
\subsection*{Testable Hypotheses and Empirical Specifications}
\textit{Hypothesis~1.} Physicians exposed to higher peer prescribing
subsequently prescribe more:
\begin{equation}
\log(1+p_{i,t+1}) = \beta_1 \log(1+\bar{p}_{N(i),t}) + \beta_2 \log(1+p_{i,t}) + \alpha_i + \lambda_{\mathrm{county}(i),t} + X'_{i,t}\gamma + \varepsilon_{i,t}.
\end{equation}
A positive $\beta_1 > 0$ supports this hypothesis.
\textit{Hypothesis~2.} Physicians with higher network centrality exhibit
greater reductions in prescribing following the policy intervention:
\begin{equation}
\log(1+p_{i,t}) = \sum_{\tau \neq 2015} \delta_\tau \bigl(1[t=\tau]\times C_i\bigr) + \alpha_i + \lambda_t + X'_{i,t}\gamma + \varepsilon_{i,t}.
\end{equation}
$C_i$ is measured using the 2015 network and standardized as a z-score.
Log-transformed centrality violates the parallel trends assumption
($p = 0.0002$) and is not adopted. More negative post-2016 $\delta_\tau$
coefficients are consistent with this hypothesis.
\textit{Hypothesis~3.} Physicians whose peers reduce prescribing more
exhibit larger own reductions:
\begin{equation}
\Delta p_i = \beta\,\Delta \bar{p}_{N(i)} + \Gamma'X_{i,2016} + \varepsilon_i,
\end{equation}
where $\Delta p_i$ and $\Delta \bar{p}_{N(i)}$ denote changes in
individual and peer prescribing between 2016 and 2017. The estimation
sample consists of 246,481 physicians observed in both years with
non-missing peer prescribing changes. Across all specifications, standard
errors are clustered at the county level.
\section{Results}
\subsection*{Network Structure and Aggregate Prescribing Patterns}
Table~\ref{tab:network_summary} reports summary statistics for the same-ZIP
physician network and opioid prescribing behavior from 2013 to 2020. The
network was large and densely connected throughout the sample period, with
more than 300,000 physicians per year and an average degree exceeding 80
even in later years. The network remained highly stable over time, with
node Jaccard similarity around 0.70--0.75 across adjacent years. Total MME
declined steadily after 2016, consistent with the CDC guideline timing.
\begin{table}[H]
\centering
\caption{Summary statistics of same-ZIP physician network structure and opioid prescribing behavior, 2013--2020.
Nodes correspond to physicians; edges indicate co-location within the same ZIP code.
Avg.\ degree is the average number of connections per physician.
Total MME claims measure aggregate opioid prescribing volume.
MME per prescriber is the average prescribing volume per physician.
MME per beneficiary is total MME-weighted claims divided by total beneficiaries.}
\label{tab:network_summary}
\begin{adjustbox}{width=\textwidth}
\begin{tabular}{lrrrrrr}
\toprule
Year & Nodes & Edges & Avg.\ degree & Total MME claims & MME per prescriber & MME per beneficiary \\
\midrule
2013 & 376,021 & 19,411,088 & 103.25 & 63,198,978 & 168.07 & 0.784 \\
2014 & 378,963 & 19,851,482 & 104.77 & 65,115,840 & 171.83 & 0.770 \\
2015 & 362,141 & 18,241,140 & 100.74 & 64,048,801 & 176.86 & 0.758 \\
2016 & 358,240 & 17,987,489 & 100.42 & 63,675,193 & 177.75 & 0.743 \\
2017 & 346,564 & 16,869,380 & 97.35 & 60,439,886 & 174.40 & 0.708 \\
2018 & 324,740 & 14,856,886 & 91.50 & 55,607,602 & 171.24 & 0.668 \\
2019 & 310,098 & 13,758,441 & 88.74 & 51,912,316 & 167.41 & 0.634 \\
2020 & 294,899 & 12,447,808 & 84.42 & 49,981,873 & 169.49 & 0.654 \\
\bottomrule
\end{tabular}
\end{adjustbox}
\end{table}
Figure~\ref{fig:overview} characterizes the network's structural
properties, temporal stability, and geographic heterogeneity. Panel (a)
shows a right-skewed degree distribution that was remarkably stable over
time. Panel (b) shows substantial persistence in network structure. Panel
(c) maps pronounced geographic heterogeneity in prescribing intensity
across states.
\begin{figure}[H]
\centering
\includegraphics[width=0.95\textwidth]{fig1.png}
\caption{Structure and empirical properties of the same-ZIP physician network.
Panel (a) shows the complementary cumulative distribution function of degree, revealing a
heavy-tailed structure with a small number of highly connected prescribers.
Panel (b) shows substantial persistence in network structure over time.
Panel (c) highlights pronounced geographic heterogeneity in opioid prescribing intensity
across counties, indicating meaningful variation in the network environment.}
\label{fig:overview}
\end{figure}
\subsection*{Evidence on Hypothesis 1: Peer Exposure and Prescribing Behavior}
Figure~\ref{fig:peer_effect_regression} shows estimates from a sequence of
specifications that progressively address potential confounding. The pooled
OLS estimate indicated a strong positive association. After including
physician fixed effects and county-by-year fixed effects, the estimated
association remained positive, economically meaningful, and precisely
estimated.
\begin{figure}[H]
\centering
\includegraphics[width=0.85\textwidth]{fig2.png}
\caption{Estimates of the relationship between physicians' prescribing and the average prescribing
of their peers across regression specifications. The figure reports estimates of the coefficient
$\beta_1$ from regressions of physician-level opioid prescribing on the average prescribing of
connected peers. A1: pooled OLS, no fixed effects. A2: adds own lagged prescribing and controls.
A3: adds physician fixed effects. A4: adds county-by-year fixed effects.
Horizontal bars denote 95\% confidence intervals. Standard errors clustered at the county level.}
\label{fig:peer_effect_regression}
\end{figure}
Table~\ref{tab:peer_effect_table} reports the corresponding estimates. In
the preferred specification (A4), the estimated coefficient of peer
prescribing is 0.0631 ($SE = 0.0050$), interpretable as an elasticity: a
1\% increase in peer prescribing is associated with a 0.063\% increase in
individual prescribing in the subsequent period. Moving from the 10th to
the 90th percentile of the peer prescribing distribution yields an increase
of approximately 0.9\% in a physician's own prescribing.
\begin{table}[H]
\centering
\caption{Estimates of the relationship between physicians' prescribing and the average prescribing of
their peers across regression specifications.
Column A1 presents pooled OLS estimates. A2 adds the physician's own lagged prescribing and controls.
A3 further includes physician fixed effects. A4 includes county-by-year fixed effects.
All regressions are estimated at the physician-year level; standard errors clustered at the county level.}
\label{tab:peer_effect_table}
\begin{tabular}{lcccc}
\toprule
& (A1) & (A2) & (A3) & (A4) \\
\midrule
Peer prescribing ($\bar{p}_{N(i),t}$)
& 0.8801 & 0.2005 & 0.0812 & 0.0631 \\
SE
& (0.0061) & (0.0035) & (0.0045) & (0.0050) \\
\midrule
Own prescribing, $p_{i,t}$
& & 0.7732 & 0.2253 & 0.2220 \\
SE
& & (0.0029) & (0.0040) & (0.0039) \\
\midrule
$\log(1+\text{patient volume}_{i,t})$
& & $-$0.0008 & $-$0.0004 & $-$0.0000 \\
SE
& & (0.0001) & (0.0003) & (0.0002) \\
\midrule
Physician FE & No & No & Yes & Yes \\
County $\times$ Year FE & No & No & No & Yes \\
Controls & No & Yes & Yes & Yes \\
\midrule
Observations & \multicolumn{4}{c}{2,005,267} \\
\bottomrule
\end{tabular}
\end{table}
\subsection*{Evidence for Hypothesis 2: Network Position and Differential Responses to Policy}
Figure~\ref{fig:event_study_regression} presents event study estimates of
prescribing responses by physician centrality. Pre-treatment coefficients
for 2013 and 2014 were small and statistically indistinguishable from zero
(F-test, $p = 0.85$), supporting the parallel trends assumption. Following
the introduction of the guideline, physicians with higher centrality
exhibited larger and increasingly pronounced reductions in prescribing
intensity.
\begin{figure}[H]
\centering
\includegraphics[width=0.80\textwidth]{fig3.png}
\caption{Event study estimates of differential prescribing responses by network centrality around
the 2016 CDC opioid prescribing guideline. Coefficients $\delta_\tau$ are from regressions of
physician-level prescribing on event-time indicators interacted with a standardized (z-score)
measure of 2015 network centrality, with 2015 as the omitted baseline year. The vertical dashed
line indicates the timing of the guideline release. Pre-treatment coefficients for 2013 and 2014
are jointly indistinguishable from zero (F-test, $p = 0.85$). Error bars denote 95\% confidence
intervals. All regressions include physician and year fixed effects.}
\label{fig:event_study_regression}
\end{figure}
Table~\ref{tab:event_study_table} reports the corresponding estimates.
Coefficients became negative and grew in magnitude from $-0.0004$ in 2016
to $-0.0014$ in 2020.
\begin{table}[H]
\centering
\caption{Event study estimates of differential prescribing responses by network centrality
around the 2016 CDC opioid prescribing guideline.
Coefficients represent the differential response associated with an increase of one SD in
network centrality. Pre-treatment estimates (2013 and 2014) are jointly indistinguishable
from zero ($p = 0.85$). All regressions are at the physician-year level; standard errors
clustered at the county level.}
\label{tab:event_study_table}
\begin{tabular}{lccc}
\toprule
Year & $\beta$ & SE & 95\% CI \\
\midrule
2013 & $-$0.0001 & (0.0002) & [$-$0.0005,\;0.0003] \\
2014 & $-$0.0001 & (0.0001) & [$-$0.0003,\;0.0002] \\
2016 & $-$0.0004 & (0.0001) & [$-$0.0006,\;$-$0.0001] \\
2017 & $-$0.0005 & (0.0002) & [$-$0.0008,\;$-$0.0002] \\
2018 & $-$0.0010 & (0.0002) & [$-$0.0014,\;$-$0.0006] \\
2019 & $-$0.0013 & (0.0002) & [$-$0.0018,\;$-$0.0008] \\
2020 & $-$0.0014 & (0.0003) & [$-$0.0019,\;$-$0.0009] \\
\midrule
Observations & \multicolumn{3}{c}{2,171,034} \\
\bottomrule
\end{tabular}
\end{table}
Figure~\ref{fig:percentile_comparison} plots event time coefficients
separately for physicians at the 10th, 50th, and 90th percentiles of the
baseline centrality distribution. By 2018, the difference in prescribing
between the 90th and 10th percentile physicians reached 0.2109\% and
continued to increase to 0.2742\% in 2019 and 0.3002\% in 2020 (expressed
as predicted percentage changes, $\exp(\Delta\log)-1$). The gap between
the 50th and 10th percentiles remained substantially smaller at 0.0520\%,
0.0677\%, and 0.0741\% over the same period.
\begin{figure}[H]
\centering
\includegraphics[width=0.80\textwidth]{a1.png}
\caption{Dynamic prescribing responses by baseline network centrality. Event time estimates of
predicted percentage changes in opioid prescribing relative to the 2015 baseline, separately for
physicians at the 10th, 50th, and 90th percentiles of the baseline centrality distribution.
Event time is defined relative to the 2016 CDC guideline. High-centrality physicians (90th
percentile) exhibited substantially larger reductions in prescribing following the guideline.
There is no evidence of differential pre-trends across groups.}
\label{fig:percentile_comparison}
\end{figure}
\subsection*{Evidence for Hypothesis 3: Peer Adjustment and the Propagation of Prescribing Changes}
Figure~\ref{fig:mechanism_test} reports first-difference estimates.
Table~\ref{tab:mechanism_test_table} reports the corresponding estimates.
In specification M1, a one-unit change in network-weighted peer prescribing
is associated with a 0.0225 change in individual prescribing ($p = 0.030$).
The estimate remained stable in specification M2 after controlling for
baseline patient volume ($\beta = 0.0221$, $p = 0.032$), indicating the
relationship is not driven by differences in physician scale.
\begin{figure}[H]
\centering
\includegraphics[width=0.80\textwidth]{fig4.png}
\caption{Estimates of peer adjustment in prescribing behavior following the introduction of the
2016 CDC opioid prescribing guideline. Coefficients $\beta$ from first-difference regressions of
changes in physician-level prescribing intensity between 2016 and 2017 on corresponding changes
in network-weighted peer prescribing. M1 includes only changes in peer prescribing. M2
additionally controls for baseline patient volume. Points denote coefficient estimates; horizontal
bars represent 95\% confidence intervals. Standard errors clustered at the county level.}
\label{fig:mechanism_test}
\end{figure}
\begin{table}[H]
\centering
\caption{Peer adjustment and individual prescribing responses following the 2016 CDC opioid
prescribing guideline. M1 includes only changes in peer prescribing. M2 additionally controls
for baseline patient volume. Standard errors clustered at the county level.}
\label{tab:mechanism_test_table}
\begin{tabular}{lcccccc}
\toprule
& $\beta$ & SE & $p$-value & 95\% CI & $N$ & Controls \\
\midrule
M1 & 0.0225 & (0.0103) & 0.030 & [0.0022,\;0.0427] & 246,481 & No \\
M2 & 0.0221 & (0.0103) & 0.032 & [0.0019,\;0.0424] & 246,481 & Yes \\
\bottomrule
\end{tabular}
\end{table}
\section{Discussion}
This study showed that opioid prescribing is shaped by professional
networks. Policy interventions appear to spread through these networks, producing responses that vary across physicians in ways that individual characteristics alone do not fully explain. All three results point toward a common mechanism:
network position shapes both routine prescribing behavior and the
transmission of policy-relevant behavioral change.
The stronger guideline responses observed among more central physicians are
consistent with a well-established finding that centrally positioned
professionals facilitate information transfer and serve as natural conduits
for behavioral change
\citep{tasselli2015social,cunningham2012health,soumerai1998effect}. Our findings extend aggregate evidence that the 2016 CDC guideline accelerated the decline in high-dose opioid prescribing \citep{bohnert2018opioid}. We show that these changes were concentrated among more central physicians and moved together with the prescribing of their local peers. By 2020, physicians at the 90th percentile of the centrality distribution exhibited prescribing reductions approximately 0.30 percentage
points larger than physicians at the 10th percentile, a gap that widened
steadily over the four years following the guideline's release.
The first-difference evidence is consistent with social learning models in
which agents update by incorporating the behavior of their neighbors
\citep{golub2010naive,chandrasekhar2020testing}. This pattern is also
connected to the clinical literature showing that peer comparison feedback
and local norm setting are among the most effective tools for changing
prescribing practices
\citep{zhang2020systematic,stepan2019development,wetzel2018interventions}.
The absorption of peer variation by county fixed effects in the
first-difference design underscores the general methodological principle
that identification strategies must align with the geographic level at
which peer interactions are defined \citep{marcus2021role,chiu2026causal}.
These findings have direct implications for opioid stewardship programs.
Network-based targeting strategies have been shown to generate substantial
cascading effects in public health contexts
\citep{alexander2022algorithms,kim2015social,valente2012network,valente2020diffusion}.
The theory of optimal network targeting predicts that interventions placed
on high-centrality nodes generate disproportionately large spillovers when
behaviors are strategically complementary \citep{galeotti2020targeting}.
Volume and centrality are related but distinct dimensions of physician
network position, and our results suggest that centrality may be a more
policy-relevant criterion for identifying physicians whose behavior changes
will propagate most broadly through the network
\citep{gouin2022identifying,iyengar2014social,hao2018empirical}. At the same time, lower prescribing is not an unambiguous good. Efforts to reduce prescribing that are too broad or poorly placed can leave patients with legitimate pain undertreated. A network-based strategy would therefore need to be paired with attention to appropriate care, and should not be judged by reductions in prescribing volume alone.
This study had several limitations. The physician network was built from geographic co-location within ZIP codes, which serves as a proxy for local professional interaction. This construction treats all physicians in the same ZIP code as connected. It does not capture actual communication, referral, or shared-patient relationships. Patient-sharing network data
\citep{donohue2018influence} would allow a more precise characterization of
interaction channels. The analysis relies on Medicare Part D data, which
primarily reflect prescribing for older and disabled populations; opioid
prescribing dynamics may differ in other settings \citep{ang2025expanded}.
Two issues limit causal interpretation. The reflection problem makes it hard to separate a physician's influence on peers from peers' influence on the physician \citep{manski1993identification}. Shared local conditions and the sorting of physicians into practice areas can also produce correlated behavior that is not peer influence \citep{goldsmith2013social}. Our physician and county-by-year fixed effects narrow these concerns for the peer-exposure results, and the event study leans on an external policy shock with no detectable pre-trends for the centrality results. Even so, we treat the peer estimates as associations rather than as fully identified causal effects. Our analysis also does not address the distributional consequences of network-based targeting. Concentrating attention on highly central physicians would shift where monitoring, education, and oversight fall. Where those physicians practice, and which patients they serve, carries equity implications that we do not measure here. Future work should ask whether such targeting narrows or widens existing disparities in access to pain care.
More broadly, this study contributes to a growing body of literature
showing that professional behavior is shaped by the structure of the
networks in which individuals are embedded
\citep{lee2025granular,weng2026rewiring,lee2026adaptive}. Future research
could extend this approach using richer network data, additional policy
shocks, or counterfactual simulations of alternative
intervention-targeting strategies.
Physicians do not adjust their prescribing in isolation. They respond to the behavior of peers and to their position within local professional networks. Policies that recognize this social structure, by engaging well connected physicians or using peer feedback, may reach further than those that treat physicians as independent actors. Whether they do so safely and fairly is a question for the next stage of this work.
\section*{Acknowledgments.}
\section*{Author Contributions.}
\section*{Data Availability.}
All data are publicly available from the Centers for Medicare and Medicaid
Services (CMS) Medicare Part D Prescriber Public Use Files
(\url{https://data.cms.gov}). ZIP-to-county crosswalk files are available
from the HUD--USPS ZIP Code Crosswalk. Code for replication will be made
available upon publication.
\section*{Competing Interests.}
The authors declare no competing interests.
\bibliographystyle{unsrt}
\bibliography{ref}
\clearpage