Is there a physician peer effect? Evidence from new drug prescriptions.
Yang, Muzhe ; Lien, Hsien-Ming ; Chou, Shin-Yi 等
I. INTRODUCTION
Is there a peer effect among physicians? The answer may not be so
obvious as people commonly believe. Take the existence of two
conflicting practices as an example. The pharmaceutical industry spends
almost twice as much on promotion as it does on R&D (Gagnon and
Lexchin 2008). In spreading its pharmacological innovations, the
industry has targeted opinion leaders who are research-active
specialists, ostensibly in response to the power of peer influence
(Nair, Manchanda, and Bhatia 2010). Conversely, in promoting
evidence-based medicine, practitioners and researchers generally believe
that the process of information diffusion is very slow. (1) Therefore,
one of the policy recommendations in the Institute of Medicine report is
to "enhance dissemination efforts to communicate evidence and
guidelines to the general public and professional communities"
(Institute of Medicine 2001). Limited evidence on the role of physician
peer networks in medical knowledge diffusion and technology adoption
perhaps could explain these conflicting practices. (2)
In this study, we investigate the existence and the pattern of peer
effects through social learning among physicians. We refer to social
learning as a process in which decision makers collect information by
observing others in their social network. (3) Through this process,
individuals' behaviors may change the behaviors of others in the
network via revealed information, and an information externality can
arise. If social learning occurs, then peer effects can be detected and
revealed by changes in an individual's behavior in response to that
of his or her peers. (4) Our study contributes to the existing
literature in three ways. First, we examine not only whether peer
effects exist, but also where they exist. If peer effects are
heterogeneous, identifying the circumstances to promote physician social
interactions can better nurture the dissemination of medical
innovations. Second, estimates of peer effects are important for policy
consideration. It has long been recognized that health policies play a
role in affecting the diffusion of medical innovation (Weisbrod 1991).
To gauge the full impact of a policy to be implemented, researchers need
to know an aggregate coefficient, including both individual direct
responses and a social multiplier based on the peer effects (Glaeser,
Sacerdote, and Scheinkman 2003). Third, we examine the persistence of
peer effects. Our results will shed important light on the duration and
mechanism that lead to productivity spillovers and the well-documented
geographic variations in health care provisions (Phelps 2000; Chandra
and Staiger 2007).
Specifically, we study whether and how a physician's
prescription of second-generation antipsyehotics (SGA) for schizophrenia
patients could be influenced by the prescription decisions of his or her
colleagues working in the same hospitals We focus on drug prescriptions
not only because new drugs can play important roles in the dissemination
of medical innovations, but also because, in contrast to new technology
or equipment, drug prescriptions usually do not require a significant
amount of fixed inputs. Therefore, its peer effect is less likely to be
confounded by the externalities of large input costs, one example of the
correlated effects discussed in Manski (1993). (6) We focus on
schizophrenia patients treated by antipsychotic drugs because there have
been serious debates about the effectiveness of SGA despite the rapid
shift from the first-generation antipsychotics (FGA) to SGA. (7) Such
controversies will provide strong incentives for physicians to
continuously acquire new knowledge about SGA.
Identifying peer effects is difficult because of self-selection of
a peer group, unobserved heterogeneities of a peer group, and the
simultaneity (or "reflection") problem (Manski 1993, 2000;
Moffitt 2001; Brock and Durlauf 2007). (8) We confront these problems by
using a unique dataset derived from antipsychotic medications
prescribed for schizophrenia patients in Taiwan between January
1997 and December 2010. The dataset has several features allowing us to
address some major challenges impeding the empirical studies on peer
effects. First, we have unique and consistent identifiers for patients,
physicians, and hospitals over the 14-year period. These identifiers
enable us to construct a longitudinal dataset consisting of complete
prescriptions of antipsychotic medications for each observed
hospital-physician--patient pair during the sample period. To identify
the peer effect, we use variations over time within each
hospital-physician-patient pair, as opposed to relying exclusively on
cross-sectional variations that are often confounded by time-invariant
unobserved heterogeneities within each hospital-physician-patient pair
(such as a physician's training background, a patient's
genetic factors, a hospital's location, and the matching between
physicians and patients based on time-invariant unobservables).
Second, a challenge in studying peer effects is how to define peers
properly. In the health care industry, physicians working in the same
hospital may form a natural peer group, which is especially the case in
our empirical setting: most physicians in Taiwan work only in one
hospital, and they may interact with their colleagues in the same
hospital frequently. However, such interactions may attract physicians
with similar characteristics to work in the same hospital and therefore
bias the peer effect estimates. Although the self-selected peer group is
a legitimate concern, it seems unlikely that a physician is hired, or
that his or her peer group is self-selected, based on the propensity to
prescribe SGA, given that the effectiveness of SGA is still in debate.
Third, because physicians can affect each other simultaneously, it
is difficult to identify the causal effect of peers. Although we are
unable to directly address the simultaneity problem given our empirical
setting, we conduct falsification checks on the presence of the
simultaneity problem. We apply the falsification checks to both the
fixed-effect (FE) estimator and the first-difference (FD) estimator used
in our study. Although both estimators control for the time-invariant
unobserved heterogeneity at the hospital-physician-patient pair level,
the consistency of the FD estimator requires a weaker assumption on the
exogeneity of regressors than the strong exogeneity assumption required
for the consistency of the FE estimator (Cameron and Trivedi 2005, 730).
Therefore, compared with the FE estimator, the FD estimator is
relatively less likely to suffer from the simultaneity problem that
invalidates the strong exogeneity assumption. By comparing the FD and
the FE estimates, we may gauge the extent of the simultaneity bias of
the peer effect estimates.
Our empirical findings are consistent with the presence and
persistency of peer effects. Our results also indicate that the FE
estimator can overestimate peer effects when the strong exogeneity
assumption falls (likely due to the simultaneity problem). Overall, we
find that peer effects among physicians are small. Our estimates suggest
that an increase of 10 percentage points in the SGA prescription share
of a physician's peers during a month will induce an increase of
approximately 0.07-0.10 percentage points in the physician's own
SGA prescription share. Our estimates are comparable to the findings of
0.30-0.40 percentage-point increase (in response to a 10
percentage-point increase in opinion leaders' drug prescriptions)
in Nair, Manchanda, and Bhatia (2010). Nevertheless, if the effect is
persistent over a long period of time, the cumulative effects are
nontrivial. For example, we find that about 69% of the increase in SGA
prescriptions in Taiwan over the 14 years (between January 1997 and
December 2010) can be explained by a multiplier effect in the presence
of the peer effect. Although the peer effect can manifest its impact
over time, the long duration for research advancement to reach clinical
practice can be costly and harmful to a society (Lenfant 2003). (9)
Furthermore, we find the peer effects to be heterogeneous. First,
positive effects appear to exist among physicians of similar age, while
in some cases inter-generational peer effects are negative. (10) Second,
peer effects are stronger among the peer group that has existed longer,
or when the group's composition is more stable, or when the group
is larger. Third, peer effects are larger when the drugs are newly
approved; the magnitude of peer effects appears to decline over time.
Lastly, peer effects diminish if the increase in the SGA prescription is
induced by an exogenous shock that adds little information about SGA
effectiveness. Our findings suggest that it is important to take into
account the heterogeneity of peer effects when designing policies to
promote or facilitate social learning among physicians.
The rest of the paper is organized as follows. Section II describes
the empirical setting. Section HI discusses the identification strategy
and lays out the econometric specifications. Section IV presents the
empirical findings. Concluding remarks are in Section V.
II. EMPIRICAL SETTING
For our empirical study we combine several data sources from
Taiwan; all of them come from the National Health Insurance (NHI)
database, which contains the medical claims and eligibility files of all
NHI enrollees. Because enrollment in NHI is mandatory in Taiwan, we
actually have the utilizations of all (more than 20 million) individuals
since the beginning of NHI. Our primary data source is the psychiatric
inpatient medical claims (PIMC) database, which records the inpatient
and also the outpatient care that occurred between January 1997 and
December 2010 of those who ever had inpatient admissions for psychiatric
treatment between 1996 and 2007. (11) For each inpatient admission and
outpatient visit, we have information on the date, diagnosis, payment,
and a list of medical codes that indicate the type and units of services
provided (e.g., drugs or procedures), from which we can identify whether
SGA was prescribed. Moreover, each claim has three identifiers for a
patient, a physician, and a hospital, respectively. Using hospital and
physician identifiers, we link PIMC with several files containing
characteristics of hospitals and physicians, such as hospital type, age,
gender, and specialty of a physician (e.g., psychiatry or neurology). We
also link prescriptions from the PIMC data with the file detailing the
descriptions of all drugs approved by NHI, which includes the
drug's name, formula, price, dosage, and approval date. As
discussed later, the information about the drug allows us to further
examine the peer effect in newly approved SGA by different approval
years.
Although the PIMC database only includes schizophrenia patients who
had at least one inpatient admission for psychiatric treatment between
1996 and 2007, it actually includes about 60% to 70% of schizophrenia
patients in Taiwan during our study period (1997-2010), because
schizophrenia is a chronic and relapsing illness and episodes of
inpatient care are common. Based on other NHI datasets we identify the
schizophrenia patients who never had inpatient admissions for
psychiatric treatment between 1996 and 2007, and we compare them with
our study population from the PIMC. We find that on average the patients
of our study population are about 1 to 6 years younger and are slightly
more likely to be male; they also have about one more outpatient visit
per year, have higher average treatment and drug expenses per visit, and
receive more SGA prescriptions per visit. We also find that the PIMC
database includes the majority of the outpatient visits (70%-80%) of all
schizophrenia patients.
Following early studies (Duggan 2005), we use diagnosis codes to
identify individuals who were treated for schizophrenia. To ensure the
accuracy of the diagnosis codes among outpatient claims, we restrict our
sample to those whose visits were seen by specialists (12) (i.e.,
psychiatrists and/or neurologists) and who were given at least one week
of medications. (13) In the end, our sample used for estimation based on
first differences includes 1,100 physicians (psychiatrists and/or
neurologists), 72,273 patients, and 373 hospitals, among which we
observe 291,821 hospital-physician-patient pairs in the period of
January 1997-December 2010. (14)
We conduct our study at the level of a hospital-physician-patient
pair. The dependent variable of our study is the SGA prescription share
for each hospital-physician-patient pair. To calculate the SGA
prescription share, we first identify whether each prescribed drug is
SGA based on the drug identifier recorded in our PIMC data. Next, we
count the number of SGA prescribed for each hospital-physician-patient
pair on each date of the treatment (or the physician visit). Then, we
divide the SGA count by the total number of drugs prescribed, which
ranges from one to three in our PIMC data, for that
hospital-physician-patient pair and on that treatment (or the physician
visit) date. Using the treatment (or the physician visit) date
information, we calculate the age of the physician and the age of the
patient, as of the drug prescription date (which is the treatment or the
physician visit date). For the physician's age, we divide the
elapsed days between the drug prescription date and the physician's
birth date by 365, and we use the same calculation for the
patient's age based on the patient's birth date. Thus, patient
age and physician age are measured in days. (15)
Our main estimation uses data values averaged monthly (from January
1997 to December 2010) by each hospital-physician-patient pair. The
sample size for our main estimation is 2,772,966. In our main estimation
sample and averaged monthly across hospital-physician-patient pairs, the
distribution of the number of prescribed drugs is the following: one
drug prescribed, accounting for 70.69% of the observations; two drugs
prescribed, accounting for 25.20% of the observations; and three drugs
prescribed, accounting for the rest (i.e., 4.11%) of the observations.
For patients with major psychiatric disorders, it is a common clinical
practice to use multiple antipsychotic drugs (i.e., polytherapy), such
as combinations of two SGAs, or older and newer antipsychotics. However,
the clinical benefits of antipsychotic polytherapy have not been well
studied (Centorrino et al., 2004).
During our sample period and on a monthly basis, the SGA
prescription share increased substantially from 3.5% in January 1997 to
62.4% in December 2010, which is shown in Figure 1. During this period,
the FGA prescription share dropped while several SGAs were subsequently
introduced to Taiwan. (16) Despite the continual entries of new SGAs as
they were approved for use during our study period, the prescription
share for each of the SGAs increased over time, that is, the SGAs are
not cannibalizing share from one another. (17) Our study focuses on drug
prescriptions by a hospital-physician-patient pair at the drug category
level, that is, whether an SGA is prescribed or not. We do not further
investigate any possible substitution pattern among SGAs prescribed by a
physician for his or her patient. In fact, SGAs all have different
pharmacological properties, and schizophrenia patients may also respond
to the same SGA differently. As a result, there is still no consensus
regarding which SGA should be prescribed first for schizophrenia
patients (Johnsen et al. 2010), and therefore, it is not surprising to
see the coexistence of several SGAs with different years of market
entries.
Summary statistics based on our main estimation sample are reported
in column (1) of Table 1. In our study, we define the peer physicians as
those physicians who work with the focal physician, excluding the focal
physician himself or herself, in the same hospital in each month. The
average number of peers based on this definition and in our main
estimation sample is 14.752 (shown in column [1] of Table 1). (18) The
peer physicians' average SGA prescription share is 0.476. Note that
the calculation of peers' average SGA prescription share is based
on the remaining physicians (i.e., excluding the focal physician) who
work in the same hospital with the focal physician in each month. In
comparison, the average SGA prescription share for a
hospital-physician-patient pair in our main estimation sample is 0.477
(not shown in the table).
[FIGURE 1 OMITTED]
A notable feature of the prescription drug market in Taiwan is
that, similar to Japan, health care providers can prescribe and also
dispense drugs. To purchase prescription drugs, hospitals (or
occasionally private-practice physicians) usually bargain with
pharmaceutical companies to set acquisition prices, which are not
regulated. In comparison, drug reimbursements to health care providers
are fixed and predetermined by the Bureau of National Health Insurance.
In Taiwan, most physicians are hospitals' salary-based employees.
Therefore, employed physicians may have the same incentive as the
hospital to choose prescription drugs according to the markup. There has
been anecdotal evidence that drug dispensing has been a profitable venue
for physicians and hospitals in Taiwan. In our data we do not have the
information on acquisition prices, but only the reimbursement price per
prescription. (19) In our regression analysis we include the average
reimbursement price per prescription by a hospital-physician-patient
pair as a proxy variable for the financial incentive that a physician or
a hospital may have. (20)
III. IDENTIFICATION STRATEGY AND SPECIFICATION
To identify a peer effect, we need to deal with three problems
extensively discussed in Brock and Durlauf (2007), Lee (2007), Manski
(1993, 2000), and Moffitt (2001)--the problems of the simultaneity, the
correlated unobservables, and the endogenous group formation. In our
context, the simultaneity problem stems from the fact that physicians
affect each other simultaneously; as a result, it is likely to
overestimate the effect of peers on a focal physician. The associated
reflection problem occurs because it is difficult to break the
collinearity between peers' average SGA prescriptions (generating
the endogenous peer effect) and peers' average characteristics
(generating the exogenous effect). Correlated unobservables arise if
there are commonly shared factors, such as hospital resources,
pharmaceutical marketing, patients' severity, or learning
mechanisms, that are unobservable to a researcher but are correlated
with the SGA prescriptions of both the focal physician and his or her
peers. Finally, the endogenous group formation in our empirical setting
means that physicians choose a particular hospital because of similar
preferences, motivations, and other unobserved characteristics that
influence the SGA prescriptions.
To deal with these problems, we take advantage of our longitudinal
data with information on antipsychotic drug prescriptions by each
hospital-physician-patient pair over time. With the longitudinal data we
could address the problems of correlated unobservables and endogenous
group formation as long as the unobserved heterogeneities are
time-invariant at a hospital level (e.g., pharmaceutical marketing or
other common learning sources), at a physician level (e.g., preference,
motivation, or training background), at a patient level (e.g.,
preference or innate factors), and at a hospital-physician-patient pair
level (e.g., the matching between physicians and patients in a hospital
based on time-invariant unobservables). We could also circumvent the
reflection problem to the extent that peers" average
characteristics (such as gender, educational attainment, and training
background) are time-invariant, whereas their SGA average prescription
shares change over time. (21) In all our econometric specifications, we
also control for time effects, which can arise from advertising that
varies over time.
In our empirical setting the focal physician's SGA
prescription is measured at the hospital-physician-patient pair level,
while the peer physicians' SGA prescriptions are measured at the
level of peer physicians who work in the same hospital with the focal
physician, not at the level of hospital-physician-patient pairs. In the
latter case, the SGA prescriptions of the focal physician and his or her
peers can affect each other if they share the same patient. In the
former case (which is our case), that simultaneous influence might be
mitigated but will not be avoided completely. Thus, we are unable to
completely solve the simultaneity problem with our empirical setting.
Instead, we use falsification checks to examine the presence of the
simultaneity problem.
We start our estimation of the peer effect with the following
regression model:
(1) [MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII],
where we denote a physician by i, a patient by j, a hospital by k,
physician i's peer group by ([G.sub.ik], t), and t indexes the
year-month from January 1997 to December 2010. The dependent variable,
[y.sub.ijk,t], is the SGA prescription share, measured by the proportion
of SGA prescriptions out of total drug prescriptions for the focal
hospital-physician-patient pair indexed by ijk at time t; peer
physicians defined as those who work with the focal physician i
(excluding the focal physician himself or herself) in the same hospital
k in each month are indexed by ([G.sub.ik], t), and [bar.y]
[G.sub.ik,t], is the SGA prescription share averaged across the peer
physicians of the focal physician i, who work with the focal physician i
in hospital k at time t. Other control variables include the number of
physicians in the focal physician's peer group ([x.sub.l]
[G.sub.ik,t]), the focal physician's age ([x.sub.2ik,t]), the
patient's age ([x.sub.3jk,t]), the price of the drug prescribed by
the focal hospital-physician-patient pair averaged across the
reimbursement prices of the drugs prescribed by that
hospital-physician-patient pair ([[bar.p].sub.ijk,t]). In this
regression model, we use the hospital-physician-patient pair fixed
effect ([[alpha].sub.ijk]) and the monthly time effect
([[delta].sub.t]). Note that both the physician's age and the
patient's age are measured by the elapsed days (divided by 365)
between their birth dates and the date of the treatment (i.e., drug
prescriptions) or the physician visit. Thus, the month-to-month
variation in the physician's age and the patient's age can be
different because the day of the treatment (i.e., drug prescription) or
the physician visit can be different from month to month. We also take
into account the within hospital-physician-patient pair clustering in
the conditional variance-covariance matrix of the disturbance term
([u.sub.ijk,t]) for panel-robust statistical inference (Cameron and
Trivedi 2005, 727).
In Equation (1), the peer effect is indicated by the parameter
[gamma]. To examine the presence of the simultaneity problem, which
hinders the identification of [gamma] we conduct the following
falsification check:
(2) [MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII].
Under the null hypothesis (which is known to be true),
[[gamma].sub.0] equals zero because the current peer physicians'
average SGA prescription should not predict the focal physician's
past SGA prescription (based on chronology). Rejecting the null
hypothesis means a nonzero [[gamma].sub.0], which indicates a
correlation between [y.sub.ijk,t-1] and [bar.y] [G.sub.ik,t] And, this
correlation can be driven by the impact of [y.sub.ijk,t-1] on [bar.y]
[G.sub.ik,t], suggesting the presence of a simultaneity bias.
We use the fixed-effect (FE) estimator to estimate Equations (1)
and (2). The consistency of the FE estimator depends on the strong (or
strict) exogeneity assumption, which requires that the disturbance term
in Equation (1) should not be correlated with any leads or lags of the
regressors, because these leads and lags are used for the within-panel
variations in the FE estimator (Cameron and Trivedi 2005, 727). A
nonzero [[gamma].sub.0] in Equation (2) means that [u.sub.0ijk,t-1] is
correlated with [bar.y][G.sub.ik,t], which provides direct evidence
against the strong exogeneity assumption.
Alternatively, we use the following first-difference (FD) model to
estimate [gamma]:
(3) [MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII].
Here, we take into account the within hospital-physician-patient
pair clustering in the conditional variance--covariance matrix of the
disturbance term ([DELTA][u.sub.ijk,t]) for panel-robust statistical
inference (Cameron and Trivedi 2005, 730). The FD estimator uses
variations only in adjacent periods and thus imposes a weaker assumption
on the exogeneity of regressors than the strong exogeneity assumption
that is required for the consistency of the FE estimator (Cameron and
Trivedi 2005, 730). Next, we conduct the following falsification check
using the FD estimator:
(4) [MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII].
Under the null hypothesis (which is known to be true based on
chronology), [[gamma].sub.0] equals zero because the change in peer
physicians' average SGA prescription between t and (t+1) should not
affect the change in the focal physician's SGA prescription between
(t - 2) and (t - 1). A nonzero [[gamma].sub.0] in Equation (4) would
suggest the association between the focal physician's past SGA
prescription and peer physicians' future SGA prescription, which
indicates the presence of a simultaneity bias.
IV. ESTIMATES OF LEARNING-BASED PEER EFFECTS
In this section, we examine the presence and the heterogeneity of
peer effects. (22) For example, between senior and junior physicians,
physicians may be more willing to learn from the same generation. Thus,
intra-generational (as opposed to inter-generational) social learning is
likely to generate positive (instead of negative) information
externality and lead to positive (as opposed to negative) peer effects.
Furthermore, we also examine whether the peer effects vary with the
changes in the environment of social learning. We first investigate the
peer effects by the stability and the size of the peer group. If peer
effects are indeed driven by social learning, then the effects would
become more salient when the group is more stable or when the group size
is larger. Next, we examine peer effects by drugs' approval years.
We expect social learning to be more relevant when drugs are more
recently approved, and thus the peer effects could be stronger. Peer
effects may decline over time when more information is revealed and
physicians have more knowledge about the new drug's effectiveness.
A. Presence and Heterogeneity of Peer Effects
Table 1 presents our first set of peer effect estimates.
Considering the colleagues working with the focal physician in the same
hospital in the same month, we find significant peer effects. Notably,
in columns (2) and (4), the FE peer effect estimate is larger than the
FD estimate. The source of this discrepancy could be indicated by the
falsification checks based on Equations (2) and (4). The results of the
falsification checks are reported in columns (3) and (5).
The FE estimator uses all within-panel variations over time,
requiring that all of the leads and the lags in the time-varying
regressors should be uncorrelated with the contemporaneous disturbance
term. This identifying assumption rules out any effect of the unobserved
heterogeneities in the current period on future observed
heterogeneities. In column (3) of Table 1, we find a significant nonzero
estimate of [[gamma].sub.0] (0.172), of which the true value is known to
be zero. This provides evidence against the strong exogeneity
assumption, which invalidates the consistency of the FE estimator. This
also suggests that the peer effect estimate (0.201, shown in column [2])
by the FE estimator is biased when in fact the focal physician's
SGA prescription affects the peer physicians' SGA prescriptions.
Thus, the FE estimate (0.201) will be an overestimate, because it
fails to take into account the effect from the focal physicians on his
or her own peers.
In contrast, the FD estimator uses variations in time-varying
regressors from the adjacent periods only, which requires that the
response from the unobserved heterogeneities in the current period
should not affect the observed heterogeneities in the immediate next
period--a weaker assumption than the strong exogeneity assumption needed
for the consistency of the FE estimator (Cameron and Trivedi 2005, 730).
In column (4), we see that the FD estimate is much smaller than the FE
estimate; this reduction in magnitude could be explained by the
elimination of the possible effects of focal physicians' SGA
prescriptions in the current period on their peer physicians' SGA
prescriptions in the future periods. Based on Equation (4), we find the
FD estimate of [[gamma].sub.0] to be statistically insignificant, which
supports the null hypothesis that the peer effect described in Equation
(4) is known to be zero. Since the FD estimator is less likely to suffer
from the simultaneity problem than the FE estimator, we focus on FD
estimations in the following analyses.
Our results suggest that on a monthly basis there is an increase of
approximately 0.07 percentage points in the SGA prescription share of
the focal hospital-physician-patient pair in response to an increase of
10 percentage points in the peer physicians' SGA prescription share
(shown in column [4]). This is our baseline peer effect estimate. It has
been suggested that a patient's response to an antipsychotic
treatment during the first 1-2 weeks is highly predictive of the
long-term effectiveness (Stauffer et al., 2011). If there is no response
to the treatment or no symptom improvement during the first 1-2 weeks,
then it should be considered to switch to other antipsychotic drugs.
Thus, it is plausible for us to examine the peer effect during a month
as physicians who consider the switching may need to learn about certain
SGA effectiveness from peers' SGA prescriptions.
If the peer effects stem from social learning, then as long as the
uncertainty of SGA effectiveness exists, we would expect that such
effects will persist over time. To test for this persistence, we include
four additional terms for peers' lagged SGA prescription shares to
Equation (3). The augmented regression model is specified as follows:
(5) [MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII],
where the parameters [[gamma].sub.1], [[gamma].sub.2],
[[gamma].sub.3], and [[gamma].sub.4] indicate the influence from the SGA
prescription shares of peers in the past 1-4 months on the focal
hospital-physician-patient pair. The parameter [[gamma].sub.0] captures
the contemporaneous peer effect, which is indicated by [gamma] in
Equation (3). The estimates in column (6) of Table 1, based on Equation
(5), confirm this persistency feature in the learning-based peer
effects. The contemporaneous peer effect estimate (0.009) is similar to
the baseline estimate (0.007). In addition to this contemporaneous
learning-based peer effect, we find that peers' past SGA
prescriptions over the last 1-4 months have continued to exert
significant influence on the focal hospital-physician-patient pair, with
the magnitude of the effect ranging from 0.003 to 0.010. If the amount
of knowledge transferred from peers to the focal physician increases
with the length of the learning period, then we would expect the
peers' lagged SGA prescription decisions to have a greater impact
than their contemporaneous ones. Our estimates in column (6) of Table 1
have confirmed this. The influence from peers' prescription
decisions in the past month increases by roughly 11% when compared with
the peer influence in the current month. We also notice that the peer
influence begins to diminish from the past month and becomes marginally
significant after 3 months.
The finding of a significant peer effect implies a social
multiplier (Glaeser, Sacerdote, and Scheinkman 2003) approximately equal
to 1.007. (23) This social multiplier derived from the monthly peer
effect appears small. However, its cumulative effect over time is not
trivial. Based on our empirical finding that the peer effect is
persistent, we could uncover a total multiplier effect over the
168-month period (from January 1997 to December 2010) approximately
equal to 3.228. (24) Our data show that the monthly SGA prescription
share increased from 0.035 in January 1997 to 0.624 in December 2010, a
nearly 17-fold increase over the 168-month period. It is important to
recognize that approximately 69% of this observed nearly 17-fold
increase in the SGA prescription share could potentially be explained by
the underlying social multiplier in the presence of the peer effect.
(25)
Next, we examine the heterogeneity of peer effects. We examine the
influences of peer effects among junior, medium-aged, and senior
physicians. We modify Equation (3) by using three interaction terms
based on three binary indicators: one for peers aged under 35 (junior
physicians), one for peers aged between 35 and 55 (medium-aged
physicians), and another for peers aged above 55 (senior physicians).
The regression model is specified as follows:
(6) [MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII],
where [[gamma].sub.0], [[gamma].sub.1], and [[gamma].sub.2],
capture the peer influence among junior, medium-aged, and senior
physicians, respectively. We estimate Equation (6) for three subgroups
of focal physicians: junior physicians under 35, senior physicians above
55, and physicians aged between 35 and 55. Separate results are reported
in Table 2.
Given that insufficient knowledge about the effectiveness of SGA is
likely and thus learning among physicians is possible, we find that such
learning appears to be most salient for physicians aged 35-55 (shown in
column [2] of Table 2): an increase of 10 percentage points in the SGA
prescription share of physicians who work in the same hospital and in
the same month with the focal physician aged 35-55 is associated with an
increase of approximately 0,07-0.08 percentage points in that focal
hospital-physician-patient pair's own SGA prescription share. The
results in columns (1) and (2) suggest an intra-generational peer
effect, which is probably due to similar backgrounds and experiences
among physicians of similar age: physicians could regard the
prescription decisions of their peers of similar age as a relevant
source of information because they might share similar backgrounds or
experiences, The results in columns (1) and (2) also suggest a pattern
for inter-generational peer effects. In column (1) we find that junior
physicians' SGA prescription decisions respond to their medium-aged
peer physicians. Similarly, in column (2) we find that medium-aged
physicians' SGA prescription decisions respond to their senior peer
physicians. In contrast, in column (3) we find that all peer effects for
senior physicians are insignificant. These findings suggest that in the
presence of insufficient knowledge about SGA effectiveness, the
differences in backgrounds or experiences between senior physicians and
junior (or medium-aged) physicians may countervail the perceived
information or knowledge transferrable from peer physicians'
prescription decisions. Thus, "wait-and-see" could arise,
especially for senior physicians. In contrast, for junior and
medium-aged physicians, similar backgrounds or experiences shared by
them may facilitate knowledge transfer from peers and thus increase the
likelihood of "following the crowd."
Note that in our main estimation sample (used for column [4] of
Table 1) there are 291,821 hospital-physician-patient pairs and 132,064
hospital-patient pairs. Thus, within a hospital, which is the basis of
our peer group definition, a patient could be treated by 2.21 physicians
on average, and there are patients who switched physicians within a
hospital. (26) It is possible that a patient switches to another
physician because the patient wants the new physician to prescribe SGA
for him or her. If this is true and since the peer group of a focal
physician consists of his or her colleagues who work in the same
hospital in the same month and who may have prescribing habits similar
to the focal physician, it would appear that the physicians learn from
one another when in fact it is the patients who switch physicians who
are driving the observed correlation of prescribing behaviors among
physicians.
Comparing patients who switched physicians within a hospital with
those who stayed with the same physician, we find that the average peer
physicians' SGA prescription share among those switchers is
slightly higher, by 0.0005, than the average peer physicians' SGA
prescription share among those non-switchers. This is consistent with
the possibility that patients switch to different physicians for more
SGA prescriptions. But, given the tiny difference in SGA prescription
shares between the two groups, we would expect the influence of the
switchers on physicians' SGA prescriptions to be small. In fact,
when we regress the SGA prescription share of a focal
hospital-physician-patient pair on the dummy variable indicating whether
the patient switches to other physicians in the same hospital (equal to
1) or not (equal to 0), we find the coefficient estimate to be -0.0002,
which is not statistically significant.
Next, we repeat the same analyses conducted in Tables 1 and 2, but
we restrict the sample to patients who never switch physicians in the
same hospital. The results of this subsample analysis are reported in
Table 3. Here, we find the peer effect estimate to be 0.008 (column
[1]), which is very similar to the estimate based on the full sample
(0.007, shown in column 4 of Table 1); the effects are also persistent
(column [2]) and heterogeneous among junior, medium-aged, and senior
physicians (columns [3]-[5]). Here, we also confirm the pattern that
junior and medium-aged physicians' SGA prescription behaviors are
likely to be affected by peers of similar age, but senior
physicians' SGA prescription behaviors are likely to be unaffected
by peers.
B. Monotonicity of Learning-Based Peer Effects
Next, we examine the monotonicity of the learning-based peer
effect--the effect should be larger (or smaller) when an empirical
setting promotes more (or less) social learning.
Group Stability and Size. If the peer effects among
physicians' prescription decisions are driven by learning about SGA
effectiveness, then we would also expect such peer influence to be
strengthened in the group that stays stable. (27) In that group, more
interactions and exchanges of knowledge are likely, and therefore more
social learning would occur. Here, we consider another peer group for
the FD estimator, in which there is no change in group size (but not
necessarily no change in group composition) in adjacent periods.
Comparing columns (1) and (2) in Table 4, we find that the peer effect
estimate doubled (from 0.004 to 0.008) in the peer group that remains
constant in size between two adjacent months. Next, we conduct the
estimation by group size and report the results in columns (3) and (4).
We choose 10 as the cutoff number of physicians in a focal
physician's peer group, which is close to the median of the size of
the peer group when it is defined on a monthly, weekly, or quarterly
basis. Here, we find that the peer effect estimate is greater (0.010 vs.
0.007) in a larger peer group. This suggests that peer influence
increases with the size of the peer group in which the opportunity of
social learning may increase. (28)
Peer Effects by SGA Approval Years. In the presence of social
learning, peer effects could be strongest when drugs are just approved
for use. Over time the peer effect for one specific drug may decline as
physicians can learn about the drug through their own studies or their
experiences of treating patients. In our dataset we have information on
drug approval years, which allows us to further examine the peer effects
by SGA approval years. Here, we consider the following two sets of SGA
approved for use by the Bureau of National Health Insurance in Taiwan:
Zyprexa in 1999, and Geodon and Solian in 2003. (29) In Table 5 we
report the peer effect estimates for SGA approved for use in 1999 (in
Panel A), and for SGA approved for use in 2003 (in Panel B). In Panel A
(or B), the SGA prescription share, for the focal
hospital-physician-patient pair and for the peer physicians, is measured
by the proportion of prescribed SGA approved for use in 1999 (or 2003)
out of total drug prescriptions. Peer effects are estimated in each year
after the approval year, through 2010, based on Equation (3).
Overall, we find that peer effects are strongest within the first
year of approval. In Panel A, within the first year of approval
(1999-2000), the peer effect estimate is 0.179. The magnitude of the
peer effects decreases by about 53% (from 0.179 to 0.084) within the
second year (2000-2001) after the approval year, and decreases by about
56% (from 0.084 to 0.037) within the third year (2001-2002), and
decreases by about 51% (from 0.037 to 0.018) in the fourth year
(2002-2003). In Panel B, we also find that the peer effect appears to be
the strongest within the first year after the SGA approval year, and
there is a decline in the peer effect between the first year and the
second year after the SGA approval year. These results imply that the
peer effect could be stronger soon after a drug is approved for use,
when there is more uncertainty about the drug's effectiveness so
that social learning is more relevant or important. These results also
imply that for a specific SGA, an individual physician, through his or
her own experience of treating patients with that SGA, may rely
decreasingly on peers.
One factor that may contribute to the decreasing pattern is that
physicians may substitute a new SGA for an old SGA when the new one is
introduced. For example, physicians may substitute Geodon or Solian
(approved in 2003) for Zyprexa (approved in 1999) after 2003. If most
physicians have this substitution, then the prescriptions of Zyprexa
will be likely to go down for both the focal and peer physicians, which
will lead to an overestimate of peer effect. For SGA approved in 1999
(Panel A), the estimate for 2003-2004 (column [5]) is slightly larger
than the one for 2002-2003 (column [4]), but the difference is almost
negligible. It suggests that the SGA substitutions are not likely to
bias our estimates for the peer effects.
C. Alternative Explanation
In this section, we investigate whether the learning-based peer
effect can be falsified by an exogenous shock in which there is possibly
no knowledge about SGA effectiveness learned from peers'
prescription decisions. If significant peer effects are found in this
situation, we would suspect that the earlier peer effect estimates could
have an alternative interpretation--for example, being the result of the
social norm under a common shock unobserved to researchers.
Here we use the global budgeting (GB) policy, which imposed a cap
on total expenditures in hospital care, to conduct the falsification
test. Because drug expenditures are subtracted from the cap first, the
reimbursement for drug expenditures is not subject to any uncertainty.
Consequently, hospitals and thus their employed physicians have stronger
incentives to increase drug expenditures by prescribing more drugs, such
as SGA (Chou et al. 2010). In this situation, the increase in SGA
prescriptions from peers could contain little new information about SGA
effectiveness.
The GB policy took effect on July 1, 2002, which we treat as a
break point in our estimation sample. To investigate this policy-induced
peer effect, we use a before-after (relative to the break point) design
for a series of subsamples of physicians working within 1 to 30 days
before and after this break point. The peer physicians are defined on a
weekly basis, because our focus is the short-term impact (under this
policy change) of peers who have worked lately with the focal physician.
Next we use GB (a binary dummy variable equal to 1 indicating the
periods after July 1, 2002) together with the other two related
variables--the number of days (within 1 to 30 days) that the focal
physician works before or after the break point and its interaction with
GB--to construct orthogonality conditions for the disturbance term in
Equation (3). Then we estimate Equation (3) using a two-step efficient
generalized method of moments (EGMM) based on the three orthogonal
conditions derived from the GB, the number of days, and the interaction
of the two. We also obtain the p values associated with the
Hansen's J-statistic for testing the validity of those three
orthogonality conditions, based on the GMM criterion function evaluated
at the EGMM estimates.
Our peer effect estimates reported in columns (1)-(30) of Table 6
passed this falsification test. In a range of within 1 to 30 days before
and after the break point, we do not find significant peer effects
induced by GB. Our falsification test conducted on a series of
subsamples provides additional support for the learning-based peer
effects on physicians' SGA prescription decisions on the basis of
confirming no peer effect when there is possibly no knowledge about SGA
effectiveness learned from peers' prescription decisions.
V. CONCLUSION
We provided empirical evidence consistent with the presence of peer
effects among physicians and further examined whether social learning
could be an important driving force behind the peer effects.
Specifically, we examined how a physician's SGA prescription
decision could be influenced by his or her peers. We found that positive
peer effects are more likely to exist when peers are of similar age,
presumably having similar (and comparable) experience and background,
Peer effects are stronger when the peer group is more stable, when the
peer group is larger, or when the period of social learning (through
observation) is longer. Peer effects also are stronger when drugs are
newly approved.
Our findings have several implications. First, although the peer
effect among physicians is small in general, it is persistent,
heterogeneous, and could manifest its impact over time. One possible
implication of our findings, and something worth further research to
understand, is the extent to which more interaction, facilitated through
changes in physical infrastructure, could increase peer effects through
a multiplier effect. As Berwick (2003, 1974) points out, "the
crucial interface between the early adopter and the early majority
cannot be effectively supported by memoranda or publications. Spread
requires social interaction." Peer effects will be stronger when
physicians have direct interactions with their colleagues through
observations or conversations.
Second, peer effects are the strongest when the innovation is newly
introduced. Based on our estimates, the multiplier effect in the first
year of introducing new drugs could be eight times as large as the
baseline multiplier effect. (30) It implies that promoting medical
innovations or new scientific findings that are beneficial to a society
can be most effective when the new drug, technology, device, or practice
is first introduced.
Third, our results contribute to the literature on the productivity
spillovers and geographic variations in health care provisions (Chandra
and Staiger 2007). Phelps (2000) and his colleagues propose a Bayesian
learning process to explain the persistency of local treatment styles
once they emerge. The initial beliefs usually were formed during
physicians' medical schooling and residency training. Through the
Bayesian learning process by observing others, treatment styles are
expected to eventually converge and persist within the same geographic
area. Thus, different areas represent different clusters of treatment
styles. It is learning among physicians that could lead to treatment
style clustering.
There are several caveats to our study. First, our peer effect
estimated at a hospital-physician-patient pair level may capture
patients' herd behavior too, because some patients may share the
same physician. In this case, a patient may request that his or her
physician prescribe a particular SGA when more and more patients of the
same physician request this particular SGA prescription. However, we
suspect that the effects attributable to physicians' learning about
SGA effectiveness still dominate the effects due to patients' herd
behavior. If the opposite is true, then we would not be able to detect
negative peer effects from the intergenerational social learning.
Second, the peer effect in physicians' SGA prescription decisions
is detected based on a generic definition of peer groups. This can be an
underestimate because the actual social learning network is unknown to
us. And, there will be measurement errors in our defined peer groups,
causing the attenuation bias in the peer effect estimate. Third, we do
not have the data on acquisition prices paid by health care providers
for drug purchases. We have the data only on prices reimbursed by the
payer, which is the Bureau of National Health Insurance in Taiwan. It is
possible that physicians choose certain drugs based on the markup, or
the payer steers physicians away from certain drugs for cost reasons,
which can induce similar drug prescription behaviors among physicians
and thus bias our peer effect estimates upward.
ABBREVIATIONS
EGMM: Efficient Generalized Method of Moments
FD: First-Difference
FE: Fixed-Effect
FGA: First-Generation Antipsychotics
GB: Global Budgeting
NHI: National Health Insurance
PIMC: Psychiatric Inpatient Medical Claims
SGA: Second-Generation Antipsychotics
doi: 10.1111/ecin.12022
Online Early publication June 13, 2013
TABLE A1
Peer Effect Estimates Based on Peers Formed on a Weekly Basis
(1) (2) (3)
FD Falsification FD Estimates
Estimates Check of Lagged
Effects
[DELTA] Peers' average SGA 0.001 0.004
prescription share (t) (0.003) (0.004)
[DELTA] Peers' average SGA 0.004
prescription share (t + 1) (0.003)
[DELTA] Peers' average SGA 0.006
Prescription share (t - 1) (0.005)
[DELTA] Peers' average SGA 0.010*
prescription share (t - 2) (0.005)
[DELTA] Peers' average SGA 0.001
prescription share (t - 3) (0.005)
[DELTA] Peers' average SGA 0.001
prescription share (t - 4) (0.004)
Number of observations 296,595 97,729 59,739
(4) (5) (6) (7)
Group Group Group Size Group
Size Size [less than Size
Not Fixed or equal >10
Fixed to]10
[DELTA] Peers' average SGA -0.002 0.002 0.001 0.001
prescription share (t) (0.005) (0.003) (0.003) (0.012)
[DELTA] Peers' average SGA
prescription share (t + 1)
[DELTA] Peers' average SGA
Prescription share (t - 1)
[DELTA] Peers' average SGA
prescription share (t - 2)
[DELTA] Peers' average SGA
prescription share (t - 3)
[DELTA] Peers' average SGA
prescription share (t - 4)
Number of observations 123,464 173,131 213,650 82,945
Notes: Estimation results in columns (1)-(3) are based on
Equations (3)-(5), respectively. Estimation results in columns
(4)-(7) are based on Equation (3). Group size refers to the
number of physicians in a focal physician's peer group. [DELTA]X
represents the change in the value of variable X from period t - 1
to period t. Other control variables include the change in the
number of physicians in a focal physician's peer group (except in
columns [4]-[7]), the change in the focal physician's age, the
change in the patient's age, the change in the average price of
drugs prescribed by the focal hospital--physician--patient pair.
and the dummy variables for weekly fixed effects. Standard errors
(reported in parentheses) are robust to the
hospital--physician-patient level clustering in the conditional
variance--covariance matrix of the regression disturbance term.
*** Significant at the 1% level; ** significant at the 5% level;
* significant at the 10% level.
TABLE A2
Peer Effect Estimates Based on Peers Formed on a Quarterly Basis
(1) (2) (3)
FD Falsification FD
Estimates Check Estimates
of Lagged
Effects
[DELTA] Peers' average SGA 0.041 *** 0.049 ***
prescription share (t) (0.002) (0.003)
[DELTA] Peers' average SGA -0.003
prescription share (t + 1) (0.003)
[DELTA] Peers' average SGA 0.040 ***
prescription share (t - 1) (0.004)
[DELTA] Peers' average SGA 0.032 ***
prescription share (t - 2) (0.004)
[DELTA] Peers' average SGA 0.017 ***
prescription share (t - 3) (0.003)
[DELTA] Peers' average SGA 0.008 **
prescription share (t - 4) (0.003)
Number of observations 1,411,475 835,608 574,566
(4) (5)
Group Group
Size Size
Not Fixed
Fixed
[DELTA] Peers' average SGA 0.037 *** 0.042 ***
prescription share (t) (0.003) (0.003)
[DELTA] Peers' average SGA
prescription share (t + 1)
[DELTA] Peers' average SGA
prescription share (t - 1)
[DELTA] Peers' average SGA
prescription share (t - 2)
[DELTA] Peers' average SGA
prescription share (t - 3)
[DELTA] Peers' average SGA
prescription share (t - 4)
Number of observations 679,216 732.259
(6) (7)
Group Group
Size Size
[less >10
than
or equal
to] 10
[DELTA] Peers' average SGA 0.035 *** 0.076 ***
prescription share (t) (0.003) (0.006)
[DELTA] Peers' average SGA
prescription share (t + 1)
[DELTA] Peers' average SGA
prescription share (t - 1)
[DELTA] Peers' average SGA
prescription share (t - 2)
[DELTA] Peers' average SGA
prescription share (t - 3)
[DELTA] Peers' average SGA
prescription share (t - 4)
Number of observations 789,987 621,488
Notes: Estimation results in columns (1)-(3) are based on
Equations (3)-(5), respectively. Estimation results in columns
(4)-(7) are based on Equation (3). Group size refers to the
number of physicians in a focal physician's peer group. [DELTA]X
represents the change in the value of variable X from period t -
1 to period t. Other control variables include the change in the
number of physicians in a focal physician's peer group (except in
columns [4]-[7]), the change in the focal physician's age, the
change in the patient's age, the change in the average price of
drugs prescribed by the focal hospital-physician-patient pair,
and the dummy variables for quarterly fixed effects. Standard
errors (reported in parentheses) are robust to the hospital-
physician-patient level clustering in the conditional variance-
covariance matrix of the regression disturbance term.
*** Significant at the 1% level; ** significant at the 5% level;
* significant at the 10% level.
TABLE A3
Peer Effect Estimates Based on Peers Formed on a Monthly Basis
(1) (2)
[DELTA] Peers' average SGA prescription 0.007 ***
share (0.001)
[DELTA] Peers' average SGA prescription 0.003 *
share due to those who conducted the (0.001)
first month of work in their hospitals
[DELTA] Peers' average SGA prescription
share due to those who conducted the
last month of work in their hospitals
[DELTA] Number of physicians in a peer 0.000 ** 0.000 *
group (0.000) (0.000)
[DELTA] Physician's age 0.029 *** 0.029 ***
(0.005) (0.005)
[DELTA] Patient's age -0.000 -0.000
(0.000) (0.000)
[DELTA] Average price (NT$ 1,000) of 0.084 *** 0.084 ***
prescribed drugs (0.002) (0.002)
Number of observations 2,772,966 2,772,966
(3) (4)
[DELTA] Peers' average SGA prescription
share
[DELTA] Peers' average SGA prescription 0.003 *
share due to those who conducted the (0.001)
first month of work in their hospitals
[DELTA] Peers' average SGA prescription -0.003 * -0.003 *
share due to those who conducted the (0.001) (0.001)
last month of work in their hospitals
[DELTA] Number of physicians in a peer 0.000 ** 0.000 *
group (0.000) (0.000)
[DELTA] Physician's age 0.029 *** 0.029 ***
(0.005) (0.005)
[DELTA] Patient's age 0.000 -0.000
(0.000) (0.000)
[DELTA] Average price (NT$ 1,000) of 0.084 *** 0.084
prescribed drugs (0.002) (0.002)
Number of observations 2,772,966 2,772,966
Notes: Estimation results in columns (1)-(4) are based on
Equation (3). [DELTA]X represents the change in the value of
variable X from period t - 1 to period t. Other control variables
include the dummy variables for monthly fixed effects. Standard
errors (reported in parentheses) are robust to the
hospital--physician--patient level clustering in the conditional
variance--covariance matrix of the regression disturbance term.
*** Significant at the 1% level; ** significant at the 5% level;
* significant at the 10% level.
REFERENCES
Balas, E. A., and S. A. Boren. "Managing Clinical Knowledge
for Health Care Improvement,, in Yearbook of Medical Informatics 2000:
Patient-Centered Systems, edited by J. Bemmel and A. T. McCray.
Stuttgart, Germany: Schattauer Verlagsgesellschaft mbH, 2000, 65-70.
Bandiera, O., and I. Rasul. "Social Networks and Technology
Adoption in Northern Mozambique." Economic Journal, 116(514), 2006,
869-902.
Bertrand, M., E. F. P. Luttmer, and S. Mullainathan. "Network
Effects and Welfare Cultures." Quarterly Journal of Economics,
115(3), 2000, 1019-55.
Berwick, D. M. "Disseminating Innovations in Health
Care." Journal of the American Medical Association, 289(15), 2003,
1969-75.
Brock, W. A., and S. N. Durlauf. "Identification of Binary
Choice Models with Social Interactions." Journal of Econometrics,
140(1), 2007, 52-75.
Burke, M. A., G. M. Fournier, and K. Prasad. "Physician Social
Networks and Geographical Variation in Medical Care." Working
Paper, Florida State University, 2003.
Cameron, A. C., and P. K. Trivedi. Microeconometrics: Methods and
Applications. New York: Cambridge University Press, 2005.
Caplin, A., and J. Leahy. "Miracle on Sixth Avenue:
Information Externalities and Search." Economic Journal, 108(446),
1998, 60-74.
Centorrino, F., J. L. Goren, J. Hennen, P. Salvatore, J. P.
Kelleher, and R. J. Baldessarini. "Multiple versus Single
Antipsychotic Agents for Hospitalized Psychiatric Patients: Case-Control
Study of Risks versus Benefits." American Journal of Psychiatry,
161(4), 2004, 700-06.
Chandra, A., and D. O. Staiger. "Productivity Spillovers in
Health Care: Evidence from the Treatment of Heart Attacks." Journal
of Political Economy, 115(1), 2007, 103-40.
Chou, S.-Y., M. E. Deily, H.-M. Lien, and J.-H. Zhang. "Global
Budgets and Provider Incentives: Hospitals' Drug Expenditures in
Taiwan," in Advances in Health Economics and Health Services
Research, Vol. 22, Pharmaceutical Markets and Insurance Worldwide,
edited by A. Dor. Bingley, UK: Emerald Group Publishing, 2010.
Coleman, L, E. Katz, and H. Menzel. "The Diffusion of an
Innovation among Physicians." Sociometry, 20, 1957, 253-70.
--. Medical Innovation: A Diffusion Study, New York: Bobbs-Merrill,
1966.
Conley, T. G., and C. R. Udry. "Learning About a New
Technology: Pineapple in Ghana." American Economic Review, 100(1),
2010, 35-69.
Devillanova, C. "Social Networks, Information and Health Care
Utilization: Evidence from Undocumented Immigrants in Milan."
Journal of Health Economics, 27(2), 2008, 265- 86.
Duflo, E,, and E. Saez. "The Role of Information and Social
Interactions in Retirement Plan Decisions: Evidence from a Randomized
Experiment." Quarterly Journal of Economics, 118(3), 2003, 815-42.
Duggan, M. "Do New Prescription Drugs Pay for Themselves? The
Case of Second-Generation Antipsychotics." Journal of Health
Economics. 24(1), 2005, 1-31.
Epstein. A. J., and S. Nicholson. "The Formation and Evolution
of Physician Treatment Styles: An Application to Cesarean
Sections." Journal of Health Economics, 28(6), 2009, 1126-40.
Escarce, J. J. "Externalities in Hospitals and Physician
Adoption of a New Surgical Technology: An Exploratory Analysis."
Journal of Health Economics, 15(6), 1996, 715-34.
Foster, A. D., and M. R. Rosenzweig. "Learning by Doing and
Learning from Others Human-Capital and Technical Change in
Agriculture." Journal of Political Economy, 103(6), 1995, 1176-209.
Gagnon, M.-A., and J. Lexchin. "The Cost of Pushing Pills: A
New Estimate of Pharmaceutical Promotion Expenditures in the United
States." PLoS Medicine, 5(1), 2008, el.
Glaeser, E. L., B. I. Sacerdote, and J. A. Scheinkman. "The
Social Multiplier." Journal of the European Economic Association,
1(2-3), 2003, 345-53.
Institute of Medicine. Crossing the Quality Chasm: A New Health
System for the 21st Century, Washington, D.C.: National Academy Press,
2001.
Johnsen, E., R. A. Kroken. T. Wentzel-Larsen, and H. A. Jorgensen.
"Effectiveness of Second-Generation Antipsychotics: A Naturalistic.
Randomized Comparison of Olanzapine, Quetiapine, Risperidone, and
Ziprasidone." BMC Psychiatry, 10, 2010, 26.
Lee, L-f. "Identification and Estimation of Econometric Models
with Group Interactions, Contextual Factors and Fixed Effects."
Journal of Econometrics, 140(2), 2007, 333-74.
Lenfant. C. "Shattuek Lecture: Clinical Research to Clinical
Practice Lost in TranslationT' The New England Journal of Medicine,
349(9), 2003, 868-74.
Lieberman, J. A., T. S. Stroup, J. P. McEvoy, M. S. Swartz, R. A.
Rosenheck, D. O. Perkins, R. S. E. Keefe, S. M. Davis, C. E. Davis, B.
D. Lebowitz, J. Severe, and J. K. Hsiao. "Effectiveness of
Antipsychotic Drugs in Patients with Chronic Schizophrenia." The
New England Journal of Medicine, 353(12), 2005, 1209-23.
Manski, C. F. "Identification of Endogenous Social Effects:
The Reflection Problem." Review of Economic Studies, 60(3), 1993,
531-42.
--. "Economic Analysis of Social Interactions." Journal
of Economic Perspectives, 14(3), 2000, 115-36.
Miguel, E., and M. Kremer. "Networks, Social Learning, and
Technology Adoption: The Case of Deworming Drugs in Kenya." Working
Paper, UC Berkeley, 2003.
Moffitt. R. A. "Policy Interventions, Low-Level Equilibria,
and Social Interactions," in Social Dynamics, edited by S. N.
Durlauf and H. P. Young. Washington, D.C.: Brookings Institution Press,
2001, 45-82.
Moretti, E. "Social Learning and Peer Effects in Consumption:
Evidence from Movie Sales." Review of Economic Studies, 78(1),
2011, 356-93.
Munshi, K. "Social Learning in a Heterogeneous Population:
Technology Diffusion in the Indian Green Revolution." Journal of
Development Economics, 73, 2004, 185-213.
Nail H. S., P. Manchanda, and T. Bhatia. "Asymmetric Social
Interactions in Physician Prescription Behavior: The Role of Opinion
Leaders." Journal of Marketing Research, 47(5), 2010, 883-95.
Phelps, C. "Information Diffusion and Best Practice
Adoption," in Handbook of Health Economics, edited by A. J. Cuyler,
and J. P. Newhouse. Amsterdam: Elsevier Science, 2000.
Rosenheck, R. A., D. L. Leslie, J. Sindelar, E. A. Miller, H. Lin,
T. S. Stroup, J. McEvoy, S. M. Davis, R. S. E. Keefe, M. Swartz, D. O.
Perkins, J. K. Hsiao, and J. Lieberman. "Cost-Effectiveness of
Second-Generation Antipsychotics and Perphenazine in a Randomized Trial
of Treatment for Chronic Schizophrenia." The American Journal of
Psychiatry, 163(12), 2006, 2080-89.
Sorensen, A. T. "Social Learning and Health Plan Choice."
Rand Journal of Economics, 37(4), 2006, 929-45.
Stauffer, V. L., M. Case, B. J. Kinon, R. Conley, H. Ascher-Svanum,
S. KoUack-Walker, J. Kane, J. McEvoy, and J. Lieberman. "Early
Response to Antipsychotic Therapy as a Clinical Marker of Subsequent
Response in the Treatment of Patients with First-Episode
Psychosis." Psychiatry Research, 187(1-2), 2011, 42-48.
Weisbrod, B. A. "The Health-Care Quadrilemma--An Essay on
Technological Change, Insurance, Quality of Care, and Cost
Containment." Journal of Economic Literature, 29(2), 1991, 523-52.
MUZHE YANG, HSIEN-MING LIEN and SHIN-YI CHOU *
* We are thankful to Marianne Bitler (Co-Editor), an anonymous
reviewer, James Dearden, Thomas McGuire, J. Niels Rosenquist, and Judy
Shinogle (deceased) for their valuable and insightful comments and
suggestions. We also thank the seminar participants at Lehigh
University, National Taiwan University, Academia Sinica, Southeastern
Health Economics Study Group, and the Third Biennial Conference of the
American Society of Health Economists. All errors are our own.
Yang: Assistant Professor, Department of Economics, Lehigh
University, Bethlehem, PA 18015. Phone 1-610-758-4962, Fax
1-610-758-4677, E-mail muzheyang@lehigh.edu
Lien: Professor, Department of Public Finance, National Chengchi
University, Wenshan, Taipei 11623, Taiwan. Phone 886-2-29378870, Fax
886-2-29380074, E-mail: hmlien@nccu.edu.tw
Chou: Professor, Department of Economics, Lehigh University,
Bethlehem, PA 18015. Phone 1-610-758-3444, Fax 1-610-758-4677, E-mail
syc2@lehigh.edu
(1.) For example, in Balas and Boren (2000), it states that
"relying on the passive diffusion of information to keep health
professionals' knowledge up to date is doomed to failure in a
global environment...."
(2.) Early studies by Coleman, Katz, and Menzel (1957 and 1966)
have shown the impact of interpersonal networks on a physician's
adoption of new drugs. Escarce (1996) finds that early adoption of
laparoscopic cholecystectomy by some surgeons in a hospital would lead
other surgeons in the same hospital to adopt it nearly 1 year earlier
than they otherwise would have done. Burke, Fournier, and Prasad (2003)
find that a patient will be more likely to receive angiography or
surgical interventions (such as bypass surgery or angioplasty) if the
attending physician is in a group that performs more of those
procedures. Epstein and Nicholson (2009) find that an increase in the
overall c-section rate of a physician's local peer group leads to
an increase in his or her own rate.
(3.) Social learning has been studied widely in a variety of
contexts including, but not limited to, employer-sponsored health plan
choices (Sorensen 2006), retirement plan choices (Duflo and Saez 2003),
welfare program participation (Bertrand, Luttmer, and Mullainathan
2000), health care utilization in Milan (Devillanova 2008), consumption
of movies (Moretti 2011), and other examples, such as crime and labor
market outcomes, which are cited in those studies.
(4.) The literature generally suggests that such social
interactions will affect individuals' behaviors through two
mechanisms (Manski 2000). One is social learning, and the other is
social norms. Our study focuses on documenting social learning to
describe the mechanism of social interaction. Our research goal is
analogous to the studies in agriculture (Foster and Rosenzweig 1995;
Munshi 2004; Bandiera and Rasul 2006; Coniey and Udry 2010), which
examine the importance of social learning among farmers on the adoption
of new crops or new technologies.
(5.) Antipsychotic drugs are often used for controlling symptoms of
schizophrenia, which is one of the most serious, relapsing, and
disabling mental illnesses. Schizophrenia patients often have difficulty
in working or even conducting basic social functions. Most common
symptoms include hallucinations, delusions, disordered thinking, and
cognitive deficits.
(6.) That is, followers of a treatment adoption may incur fewer
fixed costs than the first adopter ff the treatment involves getting new
technology or equipment in a hospital. Once the technology is present,
all physicians in the hospital can have access to it, and the peer
effects can be driven purely by the open access to the common resource
in the hospital.
(7.) The adherence of FGA (such as haloperidol) is low due to very
uncomfortable side effects, such as tardive dyskinesia or parkinsonism.
Since the approval of SGA (such as clozaril) by the Food and Drug
Administration in the United States in 1990, there has been a rapid
shift from FGA to SGA, possibly because SGA produces more tolerable side
effects. However, whether SGA performs better than FGA in terms of fewer
side effects and lower total health care expenditure is still the
subject of debate (Duggan 2005). According to a recent large-scale
experiment (CATIE), SGA, which is more cosily, is found to be no more
effective than FGA (Lieberman et al. 2005; Rosenheck et al. 2006).
(8.) The reflection problem arises when "data on equilibrium
outcomes cannot distinguish endogenous interactions from contextual
interactions" (Manski 2000, 128). We henceforth refer to the
reflection problem as the collinearity problem between the average
behavioral outcome in a peer group and the within-group average
characteristics of peers. More discussions on the reflection problem are
provided by Brock and Durlauf (2007) and Lee (2007). "Manski (1993)
has considered a group effect model where social interaction is modeled
with expected outcomes and the expected outcomes are solutions from
social equilibrium. Manski has pointed out some difficult identification
issues on his social effect model as the expected outcome from social
equilibrium might be linearly dependent on observed exogenous variables
of a group in the model--the 'reflection' problem. The
reflection problem refers to the difficulty to distinguish between
behavioral and contextual factors" (Lee 2007, 334).
(9.) Studies have suggested that it takes about 17 years on average
for research advancement to reach clinical practice (Balas and Boren
2000).
(10.) It has been shown that the information externality can lead
individuals to strategically delay an action and wait for more
information revealed from their peers (Caplin and Leahy 1998; Bandiera
and Rasul 2006). Miguel and Kremer (2003) offer another explanation for
their finding that individuals who are randomly exposed to more
information about deworming drugs through their social network are
significantly less likely to take the drugs and more likely to believe
that the drugs are not effective. That is because they have overly
optimistic prior beliefs about private drug benefits.
(11.) PIMC selects patients who ever had inpatient admissions
between 1996 and 2007, and who had the ICD-9 diagnostic codes between
290 and 319 under the supervision of the department of psychiatry; among
those patients, PIMC obtains their complete claims (including both
inpatient and outpatient use) between January 1997 and December 2010.
(12.) Thus, in our sample all SGA prescriptions are to specialists.
This restriction keeps about 90% of the visits in the original sample.
(13.) More than 87% of the outpatient visits with the diagnosis
with schizophrenia involved at least one week of medications.
(14.) Among the 291,821 observed hospital-physician-patient pairs,
there are 276,746 observed physician-patient pairs, which implies that
nearly all physician-patient pairs did not change their hospitals
(292,821/276,746=1.06).
(15.) We conduct a robustness check, using age measured in years,
and confirm that our estimates are not affected by this change.
(16.) The most important new pharmaceutical products (on which SGA
is defined) approved and included in the NHI formulary are the following
(with the associated drug names in parentheses): Zyprexa (Olanzapine) in
1999; Seroquel (Quetiapine) and Lodopin (Zotepin) in 2000; and Geodon
(Ziprasidone) and Solian (Amisulpride) in 2003.
(17.) For each SGA and each year, we calculated the percentage of
SGA prescriptions out of all drug prescriptions using our full sample.
Our calculations, for example, show that the prescription shares of
Clozapine and Rispefido, both of which were introduced to Taiwan prior
to our study period, increased from about 4% in 1997 to about 17% in
2010, and from about 1% in 1997 to about 16% in 2010, respectively.
(18.) In our main estimation sample (used for column [4] of Table
1), there are 3,418 observed hospital-physician pairs and 373 hospitals,
which implies that on average there are 9.16 physicians per hospital.
Note that the average peer group size is 14.752, which is greater than
9.16. This can be explained by the fact that the number of physicians
working in large hospitals is greater than the number of physicians
working in small hospitals, and the number of large hospitals outweighs
the number of small hospitals.
(19.) If drugs prescribed to a patient exceed a certain amount,
then the patient will pay a modest copayment for the drugs. In our
sample 87% of the cases pay no copayment; 12% pay less than NT$200; and
1% pay the maximum copayment of NT$200.
(20.) Our regression model uses hospital-physician-patient pair
fixed effects and also controls for the average reimbursement price per
prescription measured in NT$1,000 unit (New Taiwan dollar). We compared
the estimation results from the regression models with and without
including the average reimbursement price of prescribed drugs by a
hospital-physician-patient pair. The results are very similar, which
suggests that the financial incentive of an individual physician could
be uncorrelated with his or her peers' SGA prescription decisions,
after controlling for the hospital-physician-patient fixed effects.
Those results are available upon request.
(21.) We herein rely on the time-invariant feature of a peer
group's characteristics, such as gender, educational attainment,
and training background, to break the collinearity between
"exogenous effects" and "endogenous effects" (Manski
1993). This reflection problem discussed by Manski (1993) is pervasive
in empirical studies based on Manski's linear-in-expectation
models, unless models proposed by Brock and Durlauf (2007) for studying
discrete choices with social interactions are employed.
(22.) If social learning drives peer effects, then we could detect
heterogeneous peer effects resulting from different information
externalities. For brevity, we herein skip a theoretical model
(available upon request) that explains why we would expect peer effects
to exist among physicians' prescriptions of a new drug when there
is insufficient knowledge about the drug's effectiveness and
information externalities are likely to occur.
(23.) That is, 1/(1- 0.007) [approximately equal to] 1.007. This is
calculated based on the estimated peer effect equal to 0.007 on a
monthly basis for repeated cross sections. If an individual's
outcome rises by [alpha] (where 0 < [alpha] < 1) as his or her
peers' average outcome rises by 1, then the social multiplier
equals roughly 1/(1- [alpha]) for large enough groups (Glaeser,
Sacerdote, and Scheinkman 2003). In our empirical setting, for not quite
large groups, if we can reasonably assume that the interaction between
the focal physician and his or her peer physicians occurs many times
within a month and each time during that month with a peer effect equal
to a, then the calculation above is still valid.
(24.) That is, [1.007.sup.168] [approximately equal to] 3.228. This
calculation is based on the result that peer effect is persistent
between period t and period t - 1 so that the monthly-based social
multiplier carries over month to month.
(25.) Let [DELTA] y be the outcome change between the first and the
last period (i.e., [DELTA]y = 0.624 - 0.035 = 0.589). Let
[DELTA]x'[beta] be the change (in the outcome) explained by
exogenous factors ([DELTA]x) between the first and the last period, such
as drug advertising and payment structures. Removing the multiplier
effect which inflates the outcome change in each month, we uncover the
change in the outcome due to the changes in exogenous factors in the
absence of the multiplier effect, which is [DELTA]x'[beta] =
0.589/3.228 [approximately equal to] 0.182. In this sense, those
exogenous factors in the absence of the multiplier effect can account
for approximately 31% (0.182/0.589 [approximately equal to] 0.309) of
the total increase in the SGA prescription share over the 168-month
period.
(26.) In our estimation sample, 42.39% of patients stayed with the
same physician in the same hospital; 17.78% of patients had two
physicians in the same hospital; 10.29% of patients had three physicians
in the same hospital; and 29.54% of patients had more than three
physicians in the same hospital.
(27.) If peer influence in physicians' SGA prescription
decisions is induced by learning through observing peers'
decisions, then we would expect such influence to be strengthened (or
weakened) once the learning period is longer (or shorter), because
learning takes time. In the Appendix Tables A1 and A2, we consider two
alternative definitions of the focal physician's peer group, one on
a weekly basis (with a shorter learning period) and the other on a
quarterly basis (with a longer learning period). Thus, the peer group of
a focal physician comprises those working with him or her in the same
hospital in the same week, or in the same hospital in the same quarter.
If the learning period becomes longer, then we would expect the peer
effect estimates in Table 1 (columns [4] and [6]) to increase, which
will be consistent with the presence of social learning. Our estimates
in Appendix Tables A1 and A2 are consistent with this prediction: the
peer effects become larger when the peer group is formed over a longer
period. For the peer group formed on a quarterly basis, we also confirm
that the peer effect is larger when the group is more stable or larger
(results reported in Appendix Table A2).
(28.) In Appendix Table A3, we consider SGA prescriptions of two
particular types of physicians in a peer group, who may generate more
plausibly exogenous source of variation in a focal physician's
information source. The first type of physicians includes those entering
a hospital in a given month. We identify those entering physicians if
their first month of work is that given month and they work in the same
hospital with the focal physician (but we do not count January 1997 as
the first month of work since it is the start month of our sample
period). The second type of peers includes those exiting a hospital in a
given month. Similarly, we identify those exiting physicians if their
last month of work is that given month and they work in the same
hospital with the focal physician (but we do not count December 2010 as
the last month of work since it is the end month of our sample period).
Since we cannot identify the actual turnover of physicians in a
hospital, we view those results as robustness checks. For comparison
purpose, we report column (4) of Table 1 again in column (1) Columns (2)
and (3) are based on the regressions controlling for the change in peer
physicians' average SGA prescription share due to the entering
physicians and due to the exiting physicians, respectively. Here we find
an asymmetric effect from those entering and exiting peers: having an
entering peer who prescribes more SGA increases the SGA prescription of
the focal physician, while having an exiting peer who prescribes more
SGA reduces the SGA prescription of the focal physician. This asymmetric
pattern could be driven by the addition or reduction of information
source about SGA for the focal physician, In the joint estimation with
results reported in column (4), we find that the peer effect estimates
(0.003 and -0.003) are almost the same as the ones based on the two
separate estimations (columns [2] and [3]). This suggests that, in our
empirical setting, a common "shock" to all physicians in a
hospital might not be a major concern because if that common shock
exists, then the entering physicians' SGA prescription shares would
be correlated with the exiting physicians' SGA prescription shares.
In this case, results of the joint estimation (column [4]) would differ
from the results from the separate estimations (columns [2] and [3]).
(29.) To be more precise, SGA is defined on the basis of
pharmaceutical products that were approved and included in the NHI
formulary. In Taiwan, the following pharmaceutical products were
approved by the Department of Health: Olanzapine, Ziprasidone, and
Amisulpride. These pharmaceutical products are sold under the following
drug names Zyprexa, Geodon, and Solian, respectively.
(30.) Taking the coefficient estimate of peer effect in SGA
approved for use in 2003 (column [5] in Table 5) as an example, the
multiplier effect over a 1-year period is equal to
[(1/(1-0.167)).sup.12] [approximately equal to] 8.96. The baseline
multiplier effect over a 1-year period is [(1/(1-0.007)).sup.12]
[approximately equal to] 1.09.
TABLE 1
Peer Effect Estimates Based on Peers Formed on a Monthly Basis
(1) (2) (3)
Summary FE Falsification
Statistics Estimates Check
Based on
Column (4)
Peers' average SGA 0.476 0.201 *** 0.172 ***
prescription share (t) [0.228] (0.004) (0.005)
Number of physicians in a 14.752 0.001 *** 0.001 ***
peer group [16.753] (0.000) (0.000)
Physician's age 41.411 0.035 *** -0.010
[7.459] (0.007) (0.009)
Patient's age 40.537 -0.002 ** -0.003 ***
[11.851] (0.001) (0.001)
Average price (NT$1.000) of 0.088 0.110 *** 0.090 ***
prescribed drugs [0.348] (0.002) (0.003)
[DELTA] Peers' average SGA
prescription share (t)
[DELTA] Peers' average SGA
prescription share (t + 1)
[DELTA] Peers' average SGA
prescription share (t - 1)
[DELTA] Peers' average SGA
prescription share (t - 2)
[DELTA] Peers' average SGA
prescription share (t- 3)
[DELTA] Peers' average SGA
prescription share (t - 4)
[DELTA] Number of physicians
in a peer group
[DELTA] Physician's age
[DELTA] Patient's age
[DELTA] Average price
(NT$1,000) of prescribed
drugs
Number of hospital-- 291,821
physician--patient pairs
Number of observations 2,772,966 4,190,722 2,703.993
(4) (5) (6)
FD Falsification FD Estimates
Estimates Check of Lagged
Effects
Peers' average SGA
prescription share (t)
Number of physicians in a
peer group
Physician's age
Patient's age
Average price (NT$1.000) of
prescribed drugs
[DELTA] Peers' average SGA 0.007 *** 0.009 ***
prescription share (t) (0.001) (0.002)
[DELTA] Peers' average SGA -0.000
prescription share (t + 1) (0.001)
[DELTA] Peers' average SGA 0.010 ***
prescription share (t - 1) (0.002)
[DELTA] Peers' average SGA 0.007 ***
prescription share (t - 2) (0.002)
[DELTA] Peers' average SGA 0.006 ***
prescription share (t - 3) (0.002)
[DELTA] Peers' average SGA 0.003 *
prescription share (t - 4) (0.001)
[DELTA] Number of physicians 0.000 ** -0.000 0.000
in a peer group (0.000) (0.000) (0.000)
[DELTA] Physician's age 0.029 *** -0.041 *** 0.027 ***
(0.005) (0.006) (0.007)
[DELTA] Patient's age -0.000 0.001 * -0.000
(0.000) (0.000) (0.000)
[DELTA] Average price 0.084 *** -0.006 *** 0.083 ***
(NT$1,000) of prescribed (0.002) (0.002) (0.004)
drugs
Number of hospital--
physician--patient pairs
Number of observations 2,772.966 1,817.871 1,338.784
Notes: Estimation results in columns (2)-(6) are based on
Equations (1)-(5), respectively. [DELTA]X represents the change
in the value of variable X from period t - 1 to period t. Other
control variables include the dummy variables for monthly fixed
effects. Sample means of the regressors used in column (4) are
reported in column (1). Standard deviations of the regressors
used in column (4) are reported in brackets in column (1).
Standard errors (reported in parentheses) are robust to the
hospital--physician-patient level clustering in the conditional
variance--covariance matrix of the regression disturbance term.
*** Significant at the 190 level; ** significant at the 59(
level; * significant at the 1090 level.
TABLE 2
Intra-Generational and Inter-Generational Peer Effects
Based on Peers Formed on a Monthly Basis
(1) (2) (3)
Junior Physicians Senior
Physicians (age 35-55) Physicians
(age <35) (age >55)
[DELTA] Junior (age <35) peers' 0.010 *** 0.007 *** 0.005
average SGA prescription share (0.003) (0.001) (0.008)
[DELTA] Medium-aged (35-55) (1.009 *** 0.007 *** 0.004
peers' average SGA (0.003) (0.001) (0.007)
prescription share
[DELTA] Senior (age >55) peers' -0.001 0.008 *** -0.001
average SGA prescription share (0.009) (0.003) (0.008)
[DELTA] Number of physicians in 0.000 0.000 0.001
a peer group (0.000) (0.000) (0.000)
[DELTA] Physician's age 0.026 ** 0.030 *** 0.046 **
(0.011) (0.006) (0.023)
[DELTA] Patient's age 0.000 0.000 0.000
(0.001) (0.000) (0.000)
[DELTA] Average price (NT$ 0.077 *** 0.085 *** 0.102 ***
1,000) of prescribed drugs (0.005) (0.003) (0.010)
Number of observations 562,872 2,064,536 145,558
Notes: Estimation results in columns (1)-(3) are based on
Equation (6) and are obtained from the subsamples of physicians
aged under 35, between 35 and 55, and above 55, respectively.
[DELTA]X represents the change in the value of variable X from
period t - 1 to period t. Other control variables include the
dummy variables for monthly fixed effects. Standard errors
(reported in parentheses) are robust to the hospital--physician-
-patient level clustering in the conditional variance--
covariance matrix of the regression disturbance term.
*** Significant at the 1% level; ** significant at the 5% level;
* significant at the 10% level.
TABLE 3
Peer Effect Estimates Based on Peers Formed on a Monthly Basis and
Patients Who Never Switch Physicians Within Hospitals
(1) (2) (3)
Main Lagged Junior
Results Effects Physicians
(age <35)
[DELTA] Peers' average SGA 0.008 *** 0.012
prescription share (t) (0.003) (0.004)
[DELTA] Peers' average SGA 0.008 *
prescription share (t - 1) (0.005)
[DELTA] Peers' average SGA 0.004
prescription share (t - 2) (0.005)
[DELTA] Peers' average SGA 0.006
prescription share (t - 3) (0.004)
[DELTA] Peers' average SGA -0.001
prescription share (t - 4) (0.004)
[DELTA] Junior (age <35) peers' 0.053 ***
average SGA prescription share (0.013)
[DELTA] Medium-aged (35-55) peers' 0.024 **
average SGA prescription share (0.011)
[DELTA] Senior (age >55) peers' 0.038
average SGA prescription share (0.030)
[DELTA] Number of physicians in 0.000 0.001 ** 0.000
a peer group (0.000) (0.000) (0.001)
[DELTA] Physician's age 0.010 -0.020 -0.026
(0.019) (0.026) (0.055)
[DELTA] Patient's age 0.000 0.000 0.000
(0.000) (0.000) (0.000)
[DELTA] Average price (NT$1,000) 0.103 *** 0.106 *** 0.091 ***
of prescribed drugs (0.008) (0.013) (0.018)
Number of observations 225,941 121,854 26,875
(4) (5)
Physicians Senior
(age 35-55) Physicians
(age >55)
[DELTA] Peers' average SGA
prescription share (t)
[DELTA] Peers' average SGA
prescription share (t - 1)
[DELTA] Peers' average SGA
prescription share (t - 2)
[DELTA] Peers' average SGA
prescription share (t - 3)
[DELTA] Peers' average SGA
prescription share (t - 4)
[DELTA] Junior (age <35) peers' 0.002 0.006
average SGA prescription share (0.003) (0.038
[DELTA] Medium-aged (35-55) peers' 0.007 ** -0.012
average SGA prescription share (0.003) (0.019)
[DELTA] Senior (age >55) peers' -0.001 0.017
average SGA prescription share (0.008) (0.021)
[DELTA] Number of physicians in 0.000 0.004 *
a peer group (0.000) (0.002)
[DELTA] Physician's age 0.008 0.128 *
(0.021) (0.078)
[DELTA] Patient's age 0.000 0.000
(0.000) (0.000)
[DELTA] Average price (NT$1,000) 0.100 *** 0.134 ***
of prescribed drugs (0.009) (0.027)
Number of observations 183,343 15.723
Notes: Estimation results in columns (1) and (2) are based on
Equations (3) and (5), respectively. Estimation results in
columns (3)-(5) are based on Equation (6) and are obtained from
the subsamples of physicians aged under 35, between 35 and 55,
and above 55, respectively. [DELTA]X represents the change in the
value of variable X from period t - 1 to period t. Other control
variables include the dummy variables for monthly fixed effects.
Standard errors (reported in parentheses) are robust to the
hospital--physician--patient level clustering in the conditional
variance-covariance matrix of the regression disturbance term.
*** Significant at the 1% level; ** significant at the 5% level;
* significant at the 10% level.
TABLE 4
Peer Effects FD Estimates by the Stability and
Size of Peer Groups Formed on a Monthly Basis
(1) (2)
Group Size Group Size
Not Fixed Fixed
[DELTA] Peers' average SGA 0.004 ** 0.008 ***
prescription share (0.002) (0.001)
[DELTA] Physician's age 0.028 *** 0.030 ***
(0.008) (0.006)
[DELTA] Patient's age 0.000 0.000
(0.000) (0.000)
[DELTA] Average price (NT$ 1.000) 0.086 *** 0.083 ***
of prescribed drugs (0.003) (0.003)
Falsification check Passed Passed
Number of observations 1,016,439 1,756,527
(3) (4)
Group Size Group Size
[less than >10
or equal
to]10
[DELTA] Peers' average SGA 0.007 *** 0.010 ***
prescription share (0.001) (0.003)
[DELTA] Physician's age 0.029 *** 0.034 ***
(0.005) (0.007)
[DELTA] Patient's age 0.000 0.001
(0.000) (0.001)
[DELTA] Average price (NT$ 1.000) 0.084 *** 0.084 ***
of prescribed drugs (0.002) (0.003)
Falsification check Passed Passed
Number of observations 2,772,002 1,195,758
Notes: Estimation results in columns (1)-(4) are based on
Equation (3). The falsification check is based on Equation (4).
Group size refers to the number of physicians in a focal
physician's peer group. [DELTA]X represents the change in the
value of variable X from period t - 1 to period t. Other control
variables include the dummy variables for monthly fixed effects.
Standard errors (reported in parentheses) are robust to the
hospital--physician--patient level clustering in the conditional
variance-covariance matrix of the regression disturbance term.
*** Significant at the 1% level: ** significant at the 5% level:
* significant at the 10% level.
TABLE 5
Peer Effects by SGA Approval Years
(1) (2) (3) (4)
1999- 2000- 2001- 2002-
2000 2001 2002 2003
Panel A
Peer effect in SGA 0.179 *** 0.084 *** 0.037 *** 0.018 ***
approved for use in (0.018) (0.008) (0.005) (0.004)
1999
Number of observations 286.735 330,837 381,759 407,034
Panel B
Peer effect in SGA
approved for use in
2003
Number of observations
(5) (6) (7) (8)
2003- 2004- 2005- 2006-
2004 2005 2006 2007
Panel A
Peer effect in SGA 0.021 *** 0.019 *** 0.005 0.007 **
approved for use in (0.003) (0.003) (0.003) (0.003)
1999
Number of observations 429,780 458.513 486.448 496,968
Panel B
Peer effect in SGA 0.167 *** 0.139 *** 0.089 *** 0.084 ***
approved for use in (0.015) (0.010) (0.006) (0.005)
2003
Number of observations 429,780 458,513 486,448 496.968
(9) (10) (11) (12)
2007- 2008- 2009- Approval
2008 2009 2010 year 2010
Panel A
Peer effect in SGA 0.018 *** 0.029 *** 0.037 *** 0.026 ***
approved for use in (0.003) (0.003) (0.003) (0.002)
1999
Number of observations 492,079 495.889 488,855 2,565,656
Panel B
Peer effect in SGA 0.051 *** 0.013 *** 0.008 *** 0.052 ***
approved for use in (0.004) (0.003) (0.003) (0.003)
2003
Number of observations 492,079 495,889 488,855 1,897,162
Notes: Estimation results in columns (1)-(12) are based on
Equation (3) for peers formed on a monthly basis and by each
subsample. The subsamples used in columns (1)-(11) are selected
by 1 up to 11 years after the year when the SGA, which was
prescribed by both the focal physician and the peer physicians,
was approved for use. Other control variables include the change
in the number of physicians in a focal physicians peer group, the
change in the focal physician's age, the change in the patient's
age, the change in the average price of drugs prescribed by the
focal hospital-physician-patient pair, and the dummy variables
for monthly fixed effects. Standard errors (reported in
parentheses) are robust to the hospital--physician--patient level
clustering in the conditional variance-covariance matrix of the
regression disturbance term.
*** Significant at the 1% level; ** significant at the 5% level;
* significant at the 10%n level.
TABLE 6
Information Externality from Weekly Peers under a Policy Intervention
(1) (2)
[+ or -]1 Day [+ or -]2 Days
Change in peers' average SGA 0.298 -0.009
prescription share (0.193) (0.274)
Change in average price of 3.265 *** 3.137 ***
prescribed drugs (0.851) (0.691)
Number of observations 191 268
p value of Hansen's J-statistic .727 .590
(3) (4)
[+ or -]3 Days [+ or -]4 Days
Change in peers' average SGA -1.044 -1.208
prescription share (0.950) (0.736)
Change in average price of 2.341 *** 1.111
prescribed drugs (0.682) (0.705)
Number of observations 383 512
p value of Hansen's J-statistic .592 .327
(5) (6)
[+ or -]5 Days [+ or -]6 Days
Change in peers' average SGA -0.795 -(1.811
prescription share ((1.670) (0.597)
Change in average price of 1.372 ** 1.248 ***
prescribed drugs (0.593) (0.478)
Number of observations 604 693
p value of Hansen's J-statistic .531 .413
(7) (8)
[+ or -]7Days [+ or -]8 Days
Change in peers' average SGA -0.760 -0.355
prescription share (0.553) (0.552)
Change in average price of 1.143 ** 0.520
prescribed drugs (0.455) (0.345)
Number of observations 905 991
p value of Hansen's J-statistic .568 .408
(9) (10)
[+ or -]9 Days [+ or -]10 Days
Change in peers' average SGA -0.395 -0.161
prescription share (0.456) (0.455)
Change in average price of 0.465 0.610 *
prescribed drugs ((1.291) (0.334)
Number of observations 1,070 1,209
p value of Hansen's J-statistic .666 .585
(11) (12)
[+ or -]11 Days [+ or -]12 Days
Change in peers' average SGA
prescription share -0.228 0.141
Change in average price of (0.381) (0.501)
prescribed drugs 0.608 * 0.536 *
Number of observations (0.328) (0.292)
p value of Hansen's J-statistic 1,344 1,435
.579 .538
(13) (14)
[+ or -]13 Days [+ or -]14 Days
Change in peers' average SGA
prescription share 0.109 0.247
Change in average price of (0.503) (0.429)
prescribed drugs 0.538 ** 0.664 **
Number of observations (0.270) (1).281)
p value of Hansen's J-statistic 1,516 1,721
.743 .893
(15) (16)
[+ or -]15 Days [+ or -]16 Days
Change in peers' average SGA
prescription share 0.384 0.972
Change in average price of (0.609) (1.124)
prescribed drugs 0.621 ** 0.490'
Number of observations (0.269) (0.260)
p value of Hansen's J-statistic 1,795 1,880
.920 .791
(17) (18)
[+ or -]17 Days [+ or -]18 Days
Change in peers' average SGA
prescription share 0.095 -0.956
Change in average price of (0.939) (1.575)
prescribed drugs 0.509 ** 0.669 **
Number of observations (0.249) (0.269)
p value of Hansen's J-statistic 2,008 2,141
.426 .700
(19) (20)
[+ or -]19 Days [+ or -]20 Days
Change in peers' average SGA
prescription share -2.747 -1.692
Change in average price of (3.966) (2.109)
prescribed drugs 0.779 * 0.755 **
Number of observations (0.399) (0.3(13)
p value of Hansen's J-statistic 2,236 2,316
.976 .948
(21) (22)
[+ or -]21 Days [+ or -]22 Days
Change in peers' average SGA -1.595 -1.575
prescription share (1.589) (1.479)
Change in average price of 0_688 ** 0.689 **
prescribed drugs (0.287) (0.281)
Number of observations 2,527 2,603
p value of Hansen's J-statistic .926 .945
(23) (24)
[+ or -]23 Days [+ or -]24 Days
Change in peers' average SGA -1.682 -1.586
prescription share (1.684) (1.415)
Change in average price of 0.710 ** 0.660 **
prescribed drugs (0.292) (0.265)
Number of observations 2,695 2,831
p value of Hansen's J-statistic .977 .996
(25) (26)
[+ or -]25 Days [+ or -]26 Days
Change in peers' average SGA -0.947 -1.010
prescription share (0.871) (0.897)
Change in average price of 0.730 *** 0.747 ***
prescribed drugs (0.215) (0.208)
Number of observations 2,944 3,047
p value of Hansen's J-statistic .984 .961
(27) (28)
[+ or -]27 Days [+ or -]28 Days
Change in peers' average SGA -1.022 -1.073
prescription share (0.907) (0.818)
Change in average price of 0.783 *** 0.816 ***
prescribed drugs (0.204 (0.213)
Number of observations 3,123 3,325
p value of Hansen's J-statistic .991 .964
(29) (30)
[+ or -]29 Days [+ or -]30 Days
Change in peers' average SGA -1.486 -1.721
prescription share (1.180) (1.292)
Change in average price of 0.924 *** 0.961 ***
prescribed drugs (0.251) (0.255)
Number of observations 3,400 3,485
p value of Hansen's J-statistic .979 .979
Notes: Estimation results in columns (1)-(30) are based on
Equation (3) for peers formed on a weekly basis. The column X
uses a Subsarnple of physicians working X days before or after
July 1, 2002, where X = 1, 2, ..., 30. Three instrumental
variables are used for the change in peers' average SGA
prescription share in each subsamPle regression based on Equation
(3). These three instrumental variables are: (a) the binary
indicator of whether the physician worked after July 1, 2002; (b)
the number of days that the physician worked before or after July
I, 2002; and (c) the interaction between (a) and (b). The two-
step efficient GMM is used, allowing for the hospital-level
clustering in the conditional variance--covariance matrix of the
regression disturbance term. Other control variables include the
change in the number of physicians in a focal physician's peer
group. Standard errors are reported in parentheses. Also
presented are the p values of the Hansen's J-statistics for
testing the overidentifying restrictions based on the GMM
criterion function.
*** Significant at the 1% level; ** significant at the 5% level:
* significant at the 10% level.
