Physician Division of Labor and Patient Selection for Outpatient
The Wharton School
University of Pennsylvania
Mark D. Neuman
School of Medicine
University of Pennsylvania
Little is known about the effect of incentives on physicians’ choices of settings for medical procedures. We advance a model
where physicians’ division of labor between ambulatory surgery centers (ASCs) and hospital‐based outpatient departments
(HOPDs) affects the medical complexity of patients treated in low‐acuity settings (i.e. ASCs). Analyses of outpatient surgical
procedure data show that physicians working exclusively in low‐acuity settings (i.e. ASCs) treat patients of significantly
higher medical complexity in these settings than do physicians who also practice in higher‐acuity settings (i.e. HOPDs).This
discrepancy shrinks with increasing procedural risk and with increasing distance between ASCs and acute care hospitals.
JEL Classification: I11; I12; D8
Keywords: Outpatient Care; Patient Selection; Physician Behavior
We thank David Abrams, Charles Branas, Lee Fleisher, Alex Gelber, Amy Hillier, Andrea Millman, Daniel Polsky, John Rizzo,
Sandy Schwartz, and Jeff Silber for their helpful comments. Erin Quinn, Phil Saynisch, and Victoria Perez provided excellent
research assistance. Financial support from the Leonard Davis Institute of Health Economics, University of Pennsylvania is
The motives and ability of physicians to influence the medical services used by their patients has received much attention in
the health economics literature (Arrow, 1963; McGuire, 2000). Seminal work in this area has focused on demand
inducement, where financial incentives may increase the quantity of services recommended and delivered by physicians
beyond the point at which the medical benefits of such services justify their costs (Evans, 1974; McGuire and Pauly, 1991;
Labelle et al., 1994; Gruber and Owings, 1996). Prior work on referrals by physicians for office‐ versus hospital‐based care
has highlighted the effects of financial incentives on decisions regarding the utilization of resources for the care of acute
conditions (Marinoso and Jelovac, 2003; Blomqvist and Léger, 2005; Bain and Morrisey, 2007; David and Helmchen, 2010).
However, little attention has been paid to incentives that may influence decisions regarding the choice of setting for
otherwise identical medical procedures in cases where such choices may influence patient outcomes.
In the context of outpatient care in the United States, decisions regarding care settings are of increasing importance due to
the rapid growth over time in alternatives to traditional hospital‐based outpatient departments (HOPDs) for the provision
of diagnostic and therapeutic procedures. In particular, patient visits to freestanding ambulatory surgery centers (ASCs),
facilities physically separate from acute‐care hospitals, increased by 300 percent between 1996 and 2006 (Cullen, Hall, &
Golosinskiy, 2009) with the number of ASCs in the U.S. growing from 240 in 1983 to 5,174 by 2008 (Medicare Payment
Advisory Commission, 2009). The increased prevalence of ASCs has created distinct responsibilities for physicians related to
the choice of care setting (Lynk and Longley, 2002). Prior approaches to patient selection in procedural care have focused
exclusively on balancing the anticipated benefits of a given procedure with the probability of a complication as determined
by patient and procedural factors (Bryson et al., 2004). As ASCs and HOPDs differ in their access to hospital care, physicians
must also decide on the appropriate care setting based on the probability of a surgical complication and the accessibility to
hospital services that may be critical to the management of such a complication.
The decision of a physician regarding the location of care for a patient of a given risk profile should therefore be sensitive to
incentive structures that vary according to patterns of physician division‐of‐labor. More specifically, we identify two groups
of ASC physicians who differ in their access to HOPDs, distinguishing “splitters,” those who perform outpatient procedures
at both ASCs and HOPDs, from “non‐splitters,” those who work exclusively at ASCs. We advance a simple model of physician
agency, where physicians derive utility from both clinical appropriateness and monetary rewards, to show that the
opportunity cost of providing care in one location versus another depends on the physician’s division of labor. Non‐splitters
face a relatively greater opportunity cost of referring high‐risk patients to HOPD‐based care. As a consequence, splitters
would deliver care to higher‐risk patients overall, but to lower‐risk patients within ASCs, compared to their non‐splitter
counterparts. The first prediction stems from the appropriateness of HOPDs for patients with elevated surgical risk. The
second prediction stems from the higher profitability of a self‐referral compared with an out‐referral. Moreover, as the
consequences of surgical complications at ASCs may depend on their distance from the nearest acute‐care hospital, our
model demonstrates that, while patient‐level risk would decrease with increasing distance from hospital care, differences in
risk selection between physician types would diminish as a function of distance.
As a test of our model’s predictions, we examine the outpatient surgical risk profiles of 1,326,337 ASC and 464,568 HOPD
patient visits for two common outpatient procedures performed in Florida between 2004 and 2007. We use the patient‐
level Charlson Comorbidity Index, (Charlson et al, 1987; Quan et al, 2005), a common measure of medical severity, to
quantify the patient‐level operative risk observable to the physician. We use a rich set of area, facility, patient, physician,
and procedure variables to study the relationship between physician splitter status and patients’ setting selection on risk.
As risk selection behaviors may be endogenous to splitter status, we instrument for splitter status using an indicator
variable that equals one if the physician completed medical training in Florida and zero if the physician completed medical
training in a different state in the U.S. Medical training in Florida is highly correlated with retaining admitting privileges in
hospital outpatient departments, which we hypothesize may be due to preservation of professional networks and
affiliations established at the time of training. Hence, we observe a higher likelihood of splitter status among former in‐
state trainees. At the same time, controlling for foreign medical graduate status, completion of medical education in Florida
is likely to be uncorrelated with other variables affecting patient selection on risk, as we hypothesize the quality of medical
training within Florida will be similar to the average quality of training available in other U.S. states.
As expected, we find that, compared to non‐splitters, splitters treat more medically complex patients overall, with the
most complex patients being concentrated in the HOPD setting; within ASCs, we find risk profiles to be lower among
patients treated by splitters when compared to those treated by non‐splitters. We find case selection by splitters to be
related not only to the site of care (i.e., ASC vs. HOPD), but also to the distance between the ASC and the nearest hospital.
Finally, we find a growing similarity in case‐level risk for non‐splitters and splitters as the distance between the ASC and the
nearest acute‐care hospital increases.
All results are consistent for cross‐sectional regressions and for those using county fixed‐effects, which account for
potential confounding of the relationship between splitter status and risk selection by variation across individual ASCs in
given geographical regions. Our results persist under our instrumental variable strategy, which account for potential
endogeneity of risk selection to splitter or non‐splitter status. Further robustness checks include the use of alternate
definitions of splitting, the use of alternative dependent variables, and alternate instruments.
Our findings are consistent with the argument that patient selection by physicians for care in ASCs is sensitive to differences
in the opportunity cost of sending a patient to the alternate, more resource‐intensive setting of the HOPD. This effect
persists despite adjustment for procedure factors, secular time, physician factors, county fixed effects, and potential
omitted variable bias. Such observations provide a clear illustration of deviation from perfect agency in medical decision
making, which extends beyond the quantity of care.
The paper proceeds as follows: section 2 presents a simple model of physician choice, in which asymmetric information
regarding surgical risk and variation in division of labor for providers dictates the site of care for patients. Section 3
describes the data and estimation, section 4 discusses the results, and section 5 concludes.
2. Conceptual Framework
In this section we model a “downstream” medical decision, in which patients have already been determined to need a given
procedure. Rather than exploring induced utilization, as much past research has, we focus on the setting of care and the
context in which that procedure is provided. While physicians in our model act as agents for patients in that they help
patients make decisions regarding the site of care, they also posses superior medical information regarding each patient’s
risk of surgical complications and hence the most appropriate site for care. Information asymmetries allow physicians in our
model to act as imperfect agents for their patients, deriving utility from both the clinical appropriateness as well as the
monetary rewards associated with each setting. In particular, the opportunity cost of out‐referral in the case of non‐
splitters exceeds the cost of self‐referral in the case of splitters, as non‐splitters face lost income from the patients they
Patients generate a value, V(θ), from receiving an outpatient procedure as a function of severity level, θ (with large values
of θ corresponding to higher surgical risk). Assume that the distribution from which surgical risk is drawn is bounded by the
interval [θL, θH]. For simplicity, we assume that information asymmetry between surgeons and patients are such that
patients cannot observe θ.1 Nevertheless, all patients have θL < θ < θH, such that their corresponding value of surgery
exceeds the value of not getting the procedure at all. Classic demand inducement would suggest the potential for service
provision to patients for whom the benefit from the intervention is extremely low (i.e. θ < θL) and/or patients for whom the
surgical risk associated with the intervention is extremely high (i.e. θ > θH). This model ignores such extensive margins
expansions/deviations, as it focuses on incentives that influence decisions regarding the (intensive margins) choice of
setting for medical care.2 To the extent that the choice of setting is driven, in part, by non‐clinical grounds, it constitutes a
deviation from perfect agency that may influence patient outcomes through inappropriate patient risk selection.
Procedures in the interval [θL, θC], where θC<θH, are clinically appropriate for ASCs while all procedures in the interval [θL,
θH] are appropriate for HOPDs. The cost to a physician of performing a procedure in an ASC is lower than the cost of
performing the same procedure in a HOPD, (Casalino LP, Devers KJ & Brewster LR, 2003) i.e.
, where the
1 The model’s results are robust to the assumption that patients can imperfectly observe θ.
2 Alternatively, we can define θL to be zero surgical risk (i.e. no θ<θL), and θH to be sufficiently high, such that no physician would choose
to operate on a patient with such elevated surgical risk.
cost is monotonically increasing in surgical risk. The payment for the procedure is fixed at p. Hence, physician profit margins
are higher at ASCs.
Physicians maximize utility,
which is influenced by the patient’s surgical risk, θ, and the setting of care (ASC or
HOPD), s. We assume that all physicians have the ability to access an ASC but have a fixed cost, F, levied on them if they
choose to maintain a split practice (i.e. become a splitter). This fixed cost is physician specific, as it stems from the time cost
of operating in two facilities and the administrative costs of maintaining admitting privileges in HOPD.3 Since profit margins
are higher at ASCs, when
θ θ ≤
the surgeon will always decide to operate in an ASC and her utility would equal
. However, when
the non‐splitter surgeon will choose to operate in an ASC only if the profit
margins in the ASC,
, outweigh the disutility from deviating from the clinically appropriate for ASCs, described as
a quadratic loss function,
, where α measures the value (or weight) placed on this deviation. α could be
thought of as the degree of risk attributable to the procedure itself, or, alternatively,α could be thought of as the distance
between the ASC and the nearest emergency department.4 If α=0 the ASC has full access to backup capabilities or provides
a procedure with zero risk to the patient and therefore there is no disutility from treating patients with higher surgical risk.
Equation (1) provides the non‐splitter surgeon’s utility by a patient’s surgical risk and setting:
θ θ >
the splitter surgeon will choose to operate in an ASC only if the relatively higher margins in the ASC
compared with a HOPD,
, outweigh the disutility from deviating from the clinically appropriate for ASCs,
3 While legal statute disallows denial of hospital operating privileges to a physician on the basis of ASC affiliation or ownership, reports
of efforts to deny privileges for this reason suggest that the choice to operate in an ASC may come with varying time and psychic costs
for hospital physicians (for additional discussion see Lynk and Longley, 2002).
4 Quadratic loss functions generally represent the economic cost or regret associated with a deviation from a set target, such as the
level of medically appropriate care (for additional discussion see David and Helmchen, 2010).
described as a quadratic loss function scaled by distance,
. Equation (2) provides the splitter surgeon’s utility
by a patient’s surgical risk and setting:
For splitters, the level of surgical risk, θS, above which a patient will be treated in a HOPD is the solution to the following
. For non‐splitters, the level of surgical risk, θNS, above which a patient will out
referred for surgery in a HOPD is the solution to the following condition:
for any level of surgical risk, the cutoff point of patient risk chosen by a splitter for self‐referral to the HOPD is lower than
that for out‐referral chosen by a non‐splitter. The case of the non‐splitter is similar to Blomqvist and Léger (2005) where
patients perceive their health status within broad intervals, and as a result, primary care physicians have an incentive to
understate severity, in order to prevent having to refer patients out to specialists. Moreover, the cutoff point for surgical
risk chosen by both splitters and non‐splitters exceeds the clinically desired one. That is,
Figure 1 summarizes these results. The solid line maps the level of surgical risk to the utility of a non‐splitter (i.e. a
combination of the first and second expressions in Equations (1) and (2)). The dotted line maps the level of surgical risk to
the utility of from treating a patient in HOPD (i.e. the third expression in Equation (2)). Hence, the splitter’s utility function is
the external envelope (i.e. the solid line up to θC and the dotted line from θC to θH). The area of interest is the grey area
(between θS and θNS) where patients seen by a splitter surgeon would be treated in a HOPD while patients seen by a non‐
splitter surgeon would be treated in an ASC. Based on clinical experience, we assume that patients do not exhibit
preferences for splitters.
Additionally, this figure suggests that patients treated by splitters in HOPD are of markedly higher risk, as they include
patients with surgical risk at the interval [θNS, θH], which is out referred by non‐splitters. Finally, since θC<θS, even surgeons
with split practice choose a risk cutoff above the clinically desired level, θC, as they also respond to incentives and prefer to
treat patients with surgical risk between θC and θS in ASCs. While the first two predictions can be tested in the data, the
third prediction requires an objective measure of θC, which cannot be calculated using administrative data and requires
detailed patient chart data.
Note that if patients seen by surgeon i, faced with fixed splitting costs, Fi, are distributed uniformly along the interval [θS,
θH], splitting will occur if:
As α, the distance between the ASC and the nearest emergency department, increases the weight placed on deviation from
θC which in turn lowers both θS and θNS. Additionally, θNS ‐ θS decreases with α, such that there is growing similarity between
splitters and non‐splitters as the distance between ASCs and hospitals increases.5
3. Data and Sample
We examined records from the Florida Agency for Health Care Administration’s (AHCA) 2005‐2007 Ambulatory Patient
Databases (APD), which includes visit‐level data from all freestanding ambulatory surgical centers and short‐term acute care
hospitals in the state of Florida. The public use data file contains information on 33 variables, including year and quarter of
data collection, facility name, operating physician license number, type of facility, patient age, gender, residential ZIP code,
principal payer, principal and up to nine secondary procedures performed, principal and up to nine secondary diagnoses,
and patient discharge status. Procedure coding uses Current Procedure Terminology (CPT) coding and the International
Classification of Disease‐9 (ICD‐9) coding system. To quantify pre‐procedure patient risk, we calculated the Charlson
Comorbidity Index, a common non‐negative integer measure of medical severity, for each patient based on ICD‐9‐CM
5 A higher α, implies greater benefits from splitting, however, the greater the distance between an ASC and a hospital is for a surgeon,
so is the fixed costs, Fi, associated with splitting. Therefore, this model cannot predict changes in the likelihood of splitting that occur as
a function of ASCs‐to‐hospital distance.
diagnosis codes contained in the APD, using algorithms described by Quan et al. (2005). While the Charlson Index does not
include patient age, it is highly correlated with patient age; we thus report results from models with and without age as a
Data regarding the location of medical school training were collected from the Florida Department of Health’s physician
information database, to which all Florida physicians are required by law to submit data regarding medical training.6 Data
were collected on the state in which medical school was completed, foreign medical graduate status, graduation year, and
completion of subspecialty postgraduate training.
To identify hospitals that may have been available to provide care to an adult procedural outpatient for each calendar
quarter, we used the AHCA inpatient data file, which includes visit‐level data on patient, provider, facility, and encounter
variables, including variables that identify the source of admission (e.g. Emergency Department, physician referral, hospital
transfer). We classified a hospital as “open” in a given quarter if it reported one or more discharges of patients 18 years or
older whose source of admission was listed as “Emergency Room.” We excluded specialty surgical hospitals and facilities
providing exclusively pediatric, psychiatric, or rehabilitation care, as we considered such facilities to be unlikely to offer
emergency services to procedural outpatients. We verified our results using web searches and AHCA directories.
We obtained ASC and hospital addresses from AHCA directories, and calculated the point‐to‐point (Euclidean) distance
between each location where a procedure occurred and the nearest acute‐care hospital in operation at that time. Note
that since HOPDs are located in hospitals, all HOPD‐to‐ED distances are zero. We focused exclusively on common
gastrointestinal endoscopy procedures, namely upper endoscopy (UE) and colonoscopy, for patients 18 years of age or
older.7 We examined gastrointestinal endoscopies for three reasons: first, UE and colonoscopy are two of the three most
6 All searches were conducted between August 3, 2009 and September 28, 2009, using physician license numbers to identify individual
providers. The website used is: http://ww2.doh.state.fl.us/IRM00PRAES/PRASLIST.ASP. Facilities were located by street addresses on
maps made available by the U.S. Census Bureau, and locations were verified by comparison of ZIP codes between facility addresses and
7 Upper endoscopy (UE) and colonoscopy are identified by presence of any of the following principal CPT codes: 43234, 43235, 43239,
43250, 43251, 43255, 43258, 43241, 43243, 43244, 43245, 43246, 43247, 43248, and 43249 (UE); and 45378, 45379, 45380, 45381,
45382, 45383, 45384, and 45385 (colonoscopy).
common ambulatory procedures performed in the U.S. (Cullen, Hall, & Golosinskiy, 2009)8 Second, these two procedures
are performed at high rates in both ASCs and HOPDs, increasing the likelihood that scenarios may exist in which a patient
may be referred either to an ASC or an HOPD based on pre‐procedure risk factors. Third, these procedures are typically
provided by gastroenterologists, a group of specialized physicians with similar amounts of training, allowing for
comparisons of decision‐making among physicians with comparable levels of training. Finally, we duplicated all analyses
using a subset of our data, restricted to patients undergoing screening colonoscopy for colon cancer.9 The American Cancer
Society guidelines recommend initiation of colon cancer screening in adults at age 50. Focusing on a routine, low‐
complexity procedure with a uniform indication mitigates concerns about procedural mix.
Physicians were defined as splitters or non‐splitters based on the fraction of all outpatient procedural cases performed over
the study period at ASCs. While regression analyses considered only UE and colonoscopy or screening colonoscopy alone,
our definitions of division‐of‐labor examined all outpatient procedures to obtain the most comprehensive description of
physician access to ASCs and HOPDs. We defined splitters as those physicians for whom at least 5% and up to 95% of all
outpatient cases were performed at ASCs.10 We define non‐splitters as those for who over 99% of all outpatient cases were
performed at ASCs.11 While we employed a definition of splitting based on all cases performed over the three year period,
we also examined the percentage of cases performed by each physician at an ASC by calendar quarter to rule out
misclassification of physicians switching from an all‐ASC (or HOPD) to an all‐HOPD (or ASC) practice as splitters. We
excluded low volume physicians, defined as providing fewer than 150 procedures over the entire study period.
By these definitions, we classified 47 physicians as non‐splitters and 739 as splitters. The average splitter in our sample
performed 68% of cases in an ASC and 32% in an HOPD. The number of total procedures observed per physician did not
8 The third category of ambulatory procedure is lens and cataract surgery, which involves no surgical risk and is rarely performed in
hospital outpatient departments.
9 Screening colonoscopies were identified by presence of one of three principal ICD‐9‐CM diagnosis codes (V76.41, V76.50, or V76.51)
in any patient aged 50 years or older.
10 Our results are robust to using different window definitions.
11 While our main analysis excludes physicians locating between 95% and 99% of their practice at ASCs, all results are robust to
alternative definitions of non‐splitters as all physicians with over 95% of their practice at an ASC.
differ significantly between groups, with the average splitter performing 67.6 cases per month (ASCs and HOPDs), and the
average non‐splitter performing 59.7 cases per month (ASCs only).
We identified 192 ASCs and 196 HOPDs in our dataset. 57.8 percent of ASCs where less than a mile away from an acute
hospital, with fewer than 1% of ASCs located on that hospital campus. The average distance from acute care hospitals for
the remaining ASCs was 2.8 miles, with the greatest distance between an ASC and the nearest hospital being 12.8 miles.
Excluding cases performed by physicians working only in HOPDs, which were not included in our regression analyses, the
average ASC had more visits per month than did the average HOPD (190.5 vs. 72.3), despite a lower average number of
physicians at ASCs versus HOPDs (6.41 vs. 9.35).
Table 1 reports summary statistics for patients cared for by splitter and non‐splitter physicians for both the full patient
sample and for a sample undergoing screening colonoscopy only. Medical complexity was highest among patients receiving
care in HOPDs. Within ASCs, non‐splitters cared for patients of greater medical severity than did splitters. Significant
differences existed between patient groups for other patient‐level characteristics, including race, insurance, and
performance of secondary procedures within the visit. Notably, in both samples, non‐splitters delivered care to patients of
greater average age; we found this to be fully accounted for by distributional effects, as non‐splitters delivered care to a
greater proportion of patients over age 65 than did non‐splitters.
The distribution of patient‐level risk across care settings by splitters and non‐splitters is further depicted in Figure 2. Among
patients receiving care from splitters, average operative risk, as measured by the Charlson Comorbidity index, is more than
eight times greater for those treated in HOPDs compared with those treated in ASCs. Comparing patients treated within
ASCs by splitters and non‐splitters, average Charlson Score is consistently greater for those patients treated by non‐
splitters. Comparing the full sample of patients (upper panel of Figure 2) with the sample restricted to screening
colonoscopies only (lower panel of Figure 2) we find consistent patterns for the two samples. Differences in Charlson scores
for splitters and non‐splitters is greater for screening colonoscopy compared with the full sample (both for the full sample,
and in samples restricted to patients below 65 years of age and patient aged 65 and over). Similarly, these differences are
greater for patient aged 65 and over compared with patients below 65 years of age (regardless of the set of procedures).
4. Empirical Models
Broadly speaking we use two mutually exclusive strategies: we use variation in physician type (splitters vs. non splitters) to
study risk selection in ASCs, and use variation in settings type (ASCs vs. HOPDs) to study risk selection by physicians who
operate in both settings (i.e. within splitters).12 The latter analysis is important for establishing the notion that a physician
with access to both an ASC and an HOPD would schedule patients with higher medical complexity at the HOPD.
To determine whether patients’ surgical risks vary systematically across splitters and non‐splitters, the following model is
corresponds to the Charlson Comorbidity Index score for patient i, physician j, county k, in period t. The coefficient,
β , on the splitter variable,
kj S,, captures the extent to which splitters select patients with lower (or higher) surgical risk
is a vector of specific encounter characteristics, including insurance status, principal CPT code, and
presence of additional CPT codes to indicate higher levels of procedural complexity.
iP is a vector of patient characteristics;
M is a vector of physician characteristics, including years in practice, presence of fellowship‐level training in
gastroenterology, and graduation from a U.S.‐based medical school.
tφ are time fixed‐effects, corresponding to the twelve
μ are county fixed‐effects. Finally,
is an error term (results are reported in Table 2).
12 Note that a simple difference‐in‐difference approach, looking across physician types and setting type is not feasible since non‐splitters
operate solely in ASCs.
As discussed in the previous section, the Charlson Comorbidity score is a non‐negative integer measure of medical severity.
The bulk of patients in our data have a Charlson score of zero, and therefore this severity measure exhibits overdispersion
(i.e. the conditional variance exceeds the conditional mean). To address this issue we use negative binomial models, zero‐
inflated models, and a matching strategy to balance our sample across high and low surgical risk patients using one‐to‐one
nearest neighbor matching along all observable, patient‐level, dimensions.
Our study will directly address two important sources of bias in assessing differences in risk selection across these two
physician groups (splitters and non‐splitters). First, the effect of variation across geographical areas may confound the
relationship between splitter status and patient risk profiles. For example, if non‐splitters tend to work in ASCs located in
neighborhoods with older or higher‐risk patients, non‐splitters may appear to take on more risks. Similarly, as ASC’s are
located at varying distances from acute‐care hospitals, variation may exist in the time required to transport patients to
emergency care resources in the event of a complication. To address these potential confounding effects, we will use
county fixed‐effects analyses to compare splitters and non‐splitters within a given area.
Second, selection of patients based on risk may be endogenous to splitter or non‐splitter status. For example, if non‐
splitters are less risk‐averse than splitter physicians, or more highly skilled at providing care in ASC, patient risk profiles and
splitting status could be considered to relate to unobserved confounding variables. To account for such potential
unobserved variable bias, the present study will employ instrumental variable analyses treating physician splitting status as
endogenous. We use an indicator variable for the completion of medical training in the state of Florida as an instrument for
splitting. Since medical training is predominantly provided in acute‐care hospital settings, physicians who have completed
periods of medical training in the state of Florida are more likely (compared with physicians trained outside of Florida, yet in
the U.S.) to maintain a clinical practice at an acute care hospital in Florida. Moreover, in‐state training is plausibly unrelated
to other factors influencing patient selection practices. To the extent that in‐state training and risk taking behavior are
uncorrelated, in state training will provide an exogenous source of variation in splitting (results are reported in Table 3).
Even with the use of instrumental variables and county fixed‐effects models, our analysis relies on comparisons across
groups of physicians, and is unable to address patterns of risk selection within individual physicians. To directly examine
variation within individual physicians in the selection of settings on patient risk, we exploit the fact that splitters, by
definition, operate in both an ASC and a HOPD, with ASCs varying in their geographical distance from a functioning
emergency department (ED). Prior research on Emergency Medical Services has identified associations between increasing
time required for transportation to emergency care and worsened patient outcomes in a range of care‐sensitive conditions
(Carr et. al., 2009).13 Increasing distance from an emergency department implies an increasing level of procedural risk due
to additional time required for transport to care, and, consequently, an increased potential for delays in required care in
the event of a complication.
We begin with the choice of setting (ASC vs. HOPD) within individual splitter physicians. Note that the sample includes
procedures performed in both ASCs and HOPDs, but is restricted only to physicians who operate in both settings (i.e.
splitters). The following model is estimated:
β , on the setting variable,
ASC,, captures the extent to which treating a patient in an ASC is associated
with lower (or higher) surgical risk for a given physician. ψj are physician fixed‐effects. All other variables are described
following our initial regression model in (4)(results are reported in the upper panel of Table 4).
In addition, we examine patterns of risk selection within splitters as a function of facility‐to‐ED distance. This analysis sheds
light on the nature of the tradeoff made by splitters, who are in a better position to choose the site of care (ASC or HOPD)
that is the best match to their patients’ risk profile. We use the following model:
13 Rapid access to emergency care represents a core element of safe outpatient procedural care as severe complications of ambulatory
surgery may lead to emergencies such as myocardial infarction, congestive heart failure, acute respiratory failure.
where instead of a dichotomous variable for setting of care,
D,is the distance, in miles, between each ASC and the
nearest functioning emergency department as identified through the American Hospital Association Survey and Florida
ACHA records. The coefficient,
β , on the distance variable, captures the extent to which treating a patient further away
from an emergency department is associated with lower (or higher) surgical risk for a given physician. ψj are physician
fixed‐effects, developed as dummy variables corresponding to each physician’s medical license number, to control for
variation in risk selection within physicians (results are reported in the lower panel of Table 4).
Finally, we are interested in the case‐level risk for non‐splitters and splitters as the distance between the ASC and the
nearest acute‐care hospital increases. Similar to (4) the analysis considers only procedures performed in ASCs by both
splitters and non‐splitters. We use the following model:
where the coefficient, α , on the interaction of splitters and distance captures the extent to which splitters select patients
with lower (or higher) surgical risk at an ASC as the distance of that ASC from acute‐care hospital increases (results are
reported in Table 5).
Tables 2 through 7 report results from estimating equations  through . For robustness, each table includes six sample
definitions and five regression models for each one of these samples in Tables 2 through 7. We use two alternative
definitions of procedures: the right hand side of each table considers outpatient gastrointestinal endoscopy procedures, the
left hand side of each table considers only screening colonoscopies for colon cancer, a set of procedures with uniform
indications and constant risk attributable to non‐patient factors. Beyond the full sample of patients (upper panel) we
stratify the sample to patients below age 65 (middle panel) and patients 65 and above (lower panel), as patients in these