Journal of Health Economics: Pierre Dubois, Tuba Tunc El

Journal of Health Economics 78 (2021) 102461
Contents lists available at ScienceDirect
Journal of Health Economics

journal homepage: www.elsevier.com/locate/econbase
Identifying the effects of scientific information and

recommendations on physicians’ prescribing behavior夽
Pierre Dubois a,∗ , Tuba Tunçel b
a
Toulouse School of Economics, Université de Toulouse Capitole, Toulouse, France
b
Department of Applied Economics, HEC Montreal, Canada
a r t i c l e i n f o a b s t r a c t
Article history: We investigate how the prescribing behavior of physicians reacts to scientific information
Received 13 June 2020 and recommendations released by public authorities. Taking the example of antidepres-
Received in revised form 29 March 2021 sant drugs, we use French panel data on exhaustive prescriptions made by a representative
Accepted 31 March 2021
sample of general practitioners to more than 110,000 depressed patients between 2000 and
Available online 17 April 2021
2008. New results revealing an increase in suicidal thinking among children taking selec-
tive serotonin reuptake inhibitors (SSRIs) were reported in 2004 and prompted the release
JEL classification:
of new guidelines by public health authorities. We identify the effect of this unexpected
I10
warning on physicians’ drug choices while addressing the possibility that patients hetero-
D12
C25 geneity may be correlated with unobserved physician characteristics. While the warning
decreased the average probability of prescribing SSRIs, we find that physicians’ responses
Keywords: to the warning were very heterogeneous and larger if the physician had a higher preference
Physician behavior for prescribing SSRIs before the warning.
Prescription
© 2021 Elsevier B.V. All rights reserved.
Antidepressants
Mixed logit
1. Introduction
Understanding physicians’ prescribing behavior is

夽 We would like to thank the Editor, Luigi Siciliani, and anonymous ref- important for public health and public finance. Physician
erees, James Banks, David Cutler, Brigitte Dormont, Elise Coudin, Bernard prescription activity depends on physicians’ judgment and
Salanié, Jim Hammitt and seminar participants at the NBER IFS work-
continuous updating of their medical knowledge through
shop on the Value of Medical Research, the Risk Analysis Conference at
Harvard University, the European Meeting of the Econometric Society
scientific information and the public recommendations of
in Toulouse, and the Chaire Santé Workshop in Dauphine. This working health authorities. Moreover, prescription of treatments
paper has received the support of Health Chair – a joint initiative by PSL, to patients is a difficult and partially subjective choice
Université Paris-Dauphine, ENSAE, MGEN and ISTYA under the aegis of that implies cost–benefit trade-offs depending on drug
the Fondation du Risque (FDR). This research was also supported by co-
efficacy, patient condition and the evaluation of both by
funding from the joint National Institute on Ageing and Economic and
Social Research Council Valuing Health Research Network (NIA Award the physician.
Number R24 AG048059 to the National Bureau of Economic Research, and Using an important medical information change dis-
ESRC award number ES/M008673/1 to the Institute for Fiscal Studies). TSE seminated through a public warning by health authorities,
also acknowledges funding from the French National Research Agency
we study whether and how recommendations affect
(ANR) under the Investments for the Future program (Investissements
d’Avenir, grant ANR-17-EURE-0010). All errors are ours.
physicians’ decision-making, using the example of antide-
∗ Corresponding author. pressant drugs in France. We use panel data covering 2000
E-mail addresses: pierre.dubois@tse-fr.eu (P. Dubois), to 2008 and containing exhaustive prescriptions made by a
tuba.tuncel@gmail.com (T. Tunçel).
https://doi.org/10.1016/j.jhealeco.2021.102461
0167-6296/© 2021 Elsevier B.V. All rights reserved.
P. Dubois and T. Tunçel Journal of Health Economics 78 (2021) 102461
representative sample of 386 general practitioners to more Finally, thanks to a counterfactual simulation of the
than 110,000 depressed patients. As medical journals pub- decision model, we evaluate the substitution of SSRIs
lish new evidence and public health authorities adjust their towards other drug categories that would result from a ban
recommendations, doctors may update their prescribing on prescriptions of SSRI drugs to kids and adolescents and
behavior. During the study period, important new evidence compare the substitution patterns with those in the case of
on antidepressants’ efficacy and side effects were published the warning. Banning SSRIs for kids and adolescents leads
and transmitted through new official recommendations to a substitution to other alternatives just like the warn-
to physicians. There were new results in 2004 showing ing does, but of course all SSRI prescriptions disappear and
that using selective serotonin reuptake inhibitors (SSRIs) are replaced by alternatives. However, the ban leads to a
for depression treatment increases suicidal thinking in lower level of substitution to other antidepressants and
kids and adolescents. After such events and medical warn- a higher level of substitution towards drugs other than
ings, physicians must update their beliefs on different drug antidepressants, relative to the warning. In other words,
treatments and may react differently to these warnings. a physician stopping to prescribe SSRIs to kids and adoles-
We first show some difference-in-difference estima- cents after a warning is more likely to substitute with any
tion of the change in the prescription probability of SSRI other antidepressant than he or she would do in the case of
drugs before and after the warning for kids and ado- a ban. This happens because the warning seems to less neg-
lescents relative to other antidepressants and to older atively affect the preferences for non-SSRI antidepressants
patients. The evidence shows a decline in the prescriptions compared to drugs other than antidepressants.
of SSRIs and placebo regressions show that it coincides with Our work adds some empirical evidence on the role
the time of the warning. As such difference-in-difference of information in physicians’ prescribing behavior. Pre-
estimation cannot disentangle the effect of the warning vious literature on prescribing behavior has addressed
on the preferences for the different types of drugs, we issues related to physician-induced demand (Mcguire,
use a discrete choice model to estimate physician prefer- 2000; Dickstein, 2016) and its relationship to drug prices,
ences. patient co-payments and the availability of generic drugs,
We develop a model of prescribing behavior with as well as physician learning (Ching, 2010; Coscielli and
physician and patient heterogeneity and show how we Shum, 2004; Crawford and Shum, 2005; Dickstein, 2018;
can identify the effect of a warning on individual physi- Janakiraman et al., 2009; Ching et al., 2013; Ching and
cians’ specific preferences when unobserved heterogeneity Lim, 2020). Our work relates to the evidence on the role
in patients’ health state may be correlated with physi- of physicians’ heterogeneity of skills, beliefs and prefer-
cians’ heterogeneity. Such a correlation could be the result ences, which has been documented recently (Berndt et al.,
of endogenous matching on unobservable characteristics 2015; Currie and Macleod, 2017, 2020; Cutler et al., 2019).
between physicians and patients. Assuming stable pref- Cutler et al. (2019) shows how much regional variation
erences of physicians during the periods before and after in health-care expenditures in the US comes from patient
the warning, we can assess whether the heterogeneity in demand-side factors as opposed to physician supply-side
treatments is due to unobservable differences in patient factors. The results show that the most important factor
or physician preferences (on drug efficacy or side effects, is physician beliefs about treatment. Berndt et al. (2015)
for example). We are able to test not only whether chang- shows that many psychiatrists have significantly hetero-
ing scientific information affects physicians’ prescriptions geneous prescription patterns and concentrate on distinct
but also whether it affects physicians differently. Our deci- drugs. The authors find some evidence of a relationship
sion model allows us to conduct counterfactual analysis between prescription volumes and prescribing behavior
and investigate the impact of different policies on physi- that is consistent with a learning-by-doing model among
cians’ prescribing behavior which would not be possible physicians. Stern and Trajtenberg (1998) show that the
with reduced form analysis. exercise of physician authority is likely to be related to
Our empirical results show that physicians’ behav- skills. Finally, Currie and Macleod (2020) investigate how
ior is very heterogeneous in terms of propensity to physician diagnostic skills, tastes, and beliefs impact physi-
prescribe different kinds of antidepressants and that gov- cian decision-making. The authors use a model in which
ernment warnings also have very heterogeneous effects physician experimentation allows for learning about the
on physicians’ prescribing behavior. We find that physi- match quality between a particular drug and an individual
cians prescribe antidepressants to kids and adolescents less in the case of antidepressant medication.
often after the warning, but many still do not adhere to While there is extensive literature on physicians’ learn-
the recommendation. SSRIs are still prescribed to this age ing and experimentation, papers studying the role of new
group by 62% of physicians, despite the warning advising scientific evidence and public recommendations on physi-
against this. We observe that prescription of SSRIs to kids cians’ prescriptions are sparse. Some have evaluated how
and adolescents decreases in favor of either serotonin and prescriptions change after drug withdrawal. Collins et al.
norepinephrine reuptake inhibitors (SNRIs) or drugs other (2013) show that the Vioxx withdrawal had both positive
than antidepressants. We also find that after the warn- and negative effects for specific substitute drugs and led
ing, the probability of prescribing an SSRI to young adults, to an overall increase in the usage of competing products.
adults and elderly people changes very heterogeneously Berez et al. (2018) study the physicians’ response to infor-
across physicians. It seems that some physicians interpret mation on ineffectiveness of pulmonary artery catheters
the warning as “good” or “bad” news for age groups other (PACs). They find that the use of PACs declined after the
than kids and adolescents. release of evidence on the lack of benefit from their use
2
and that older physicians’ use of PACs was influenced by subclasses according to their effect on the concentra-
the practice patterns of their junior colleagues. Howard tion of serotonin and norepinephrine in the brain. These
et al. (2017) use a negative informational shock (namely the subclasses are selective serotonin reuptake inhibitors
fact that a common knee operation does not improve out- (SSRIs), with the active ingredients citalopram, escitalo-
comes for patients with osteoarthritis) to study the impact pram, fluoxetine, paroxetine, and sertraline; serotonin-
of physicians’ agency relationships on treatment decisions. norepinephrine reuptake inhibitors (SNRIs), with the
They show that providers at physician-owned surgery molecules milnacipran and venlafaxine; and “other antide-
centers abandoned common knee surgeries at a lesser pressants”, which include medicines with the molecules
rate than physicians practicing in hospitals. Howard and mianserine, mirtazapine and tianeptine.
Hockenberry (2019) show that episiotomy rate declined
from 54% in 1994 to 13% in 2010 but that older physi- 2.2. Health care system
cians were more likely to perform it and were slower to
adjust their practices in response to evidence showing Health insurance is mandatory in France, and all resi-
that routine episiotomy is unnecessary. Physician beliefs dents are automatically enrolled in the insurance system
are crucial to explaining their heterogeneous prescribing depending on their occupational status under the French
behavior (Berndt et al., 2015) and are also directly affected national health insurance system. A total of 90% of the
by both scientific knowledge and personal experience with population has supplementary health insurance to cover
their patients. Our new approach and results shed light on benefits not covered under mandatory health insurance.
how to evaluate the impact of medical warnings on physi- Even though health insurance plans differ across occu-
cians and on their wide heterogeneity of responses. pational groups, they are all regulated under the same
The paper proceeds as follows: In Section 2, we statutory framework (Rodwin, 2003). As in the case of the
first present some background descriptive information on Italian market, discussed by Crawford and Shum (2005),
antidepressants, public health warnings and recommen- plans cannot compete by lowering insurance premiums,
dations, the data and some stylized descriptive statistics. and physicians have uniform per-visit payments that atten-
Section 3 shows reduced form evidence of the effect of the uate the agency problem, which may come into play in
warning on prescriptions. Section 4 presents our model and the case of a market with heterogeneous third-party pay-
identification strategy. Section 5 shows the results of the ers. The heterogeneous constraints on physicians’ choices
empirical estimation on antidepressants and depression induced by drug formularies in the US market do not come
treatment in France, and Section 6 concludes. into play in the French market.
2. Institutional background, data and stylized facts 2.3. Scientific information release
2.1. Depression and antidepressants Authorities such as the Food and Drug Administration
(FDA) in the US or other health authorities in European
Depression affects 20% of French residents during their countries monitor the use of drugs and outcomes in terms
lifetimes. According to the World Health Organization, it of public health to check and evaluate the efficacy and scru-
is the leading cause of ill health and disability worldwide tinize the side effects or unintended effects of drugs, even
(James et al., 2018). It is also costly because patients suf- after drugs are authorized and marketed. When new scien-
fer from a decrease in their productivity. More than 60% of tific evidence appears after drug introductions, it is usually
depressed people have symptoms severe enough to keep diffused through scientific publications and then taken into
them from performing daily tasks (Kessler et al., 2003). account by health authorities in their recommendations to
Depression also increases suicide attempts and hence mor- prescribers. In France, the public health authority, named
tality: the risk of suicide is 13–30 times higher among ANSM (Agence Nationale de Sécurité du Médicament), is
depressed people than among nondepressed people, and in charge of authorizing drugs and of regulating the use of
suicide is among the top leading causes of death in high- prescription drugs by giving usage conditions and recom-
income countries (and is the second leading cause of mendations to physicians.
death among 15-to-29-year-old1 ). Finally, depression also We collected all the information on the recommenda-
increases health-care expenditures. Depressed people visit tions of the French authority on antidepressant usage. We
their generalists for somatic complaints three times more also examined the US FDA recommendations and warn-
often than nondepressed people (Kessler et al., 2003). ings as well as the medical literature to verify whether the
The most commonly used modern antidepressants French health authority was giving all relevant informa-
are those from the second generation, which gener- tion that could influence physicians. These data show that
ally dominate those from the first generation. The only recommendations and warnings between 2000 and 2008
first-generation antidepressants still used are those in usually occur in France around the same time as they do in
the category of tricyclic antidepressants (TCAs), with the the US and closely follow the medical literature. All impor-
active ingredients amitriptyline, clomipramine, dosulepin, tant scientific news is monitored by these agencies and
imipramine, maprotiline, and trimipramine. Molecules of processed into official warnings and recommendations.
the second generation are classified into three distinct During the period examined in this study, three important
warnings were released. The first recommended not pre-
scribing SSRI-type antidepressants to kids and adolescents
1
https://www.who.int/en/news-room/fact-sheets/detail/depression. and was issued in December 2004 in France (a few weeks
3
Table 2.1
Share of drugs prescribed for depression diagnoses.
Group All ages Kids and ado. (2–18) Young adults (18–25) Adults (26–65) Elderly people (65+)
SSRIs 0.50 0.50 0.58 0.52 0.43

SNRIs 0.09 0.05 0.10 0.10 0.06
TCAs 0.07 0.05 0.02 0.01 0.11
Other antidepressants 0.11 0.07 0.08 0.09 0.15
Other drugs 0.23 0.33 0.23 0.22 0.25
No. of visits 517,241 2,564 16,795 372,406 125,441
after the US FDA warning). The second one, released in June health (chronic diseases, height, weight). The data include
2006, partially contradicted the 2004 warning by recom- all information recorded at physician visits, including diag-
mending Prozac (the fluoxetine molecule of the SSRI group) noses, prescriptions, and exam results transmitted to the
for use by adolescents and kids above 8 years of age with physician. Thus, we observe the diagnosis and all drugs and
moderate to severe depression. Finally, another warning treatments (drug, dosage, renewal) that were prescribed
was released in February 2008 for three different molecules by the physician on each visit. The unique patient- and
that were deemed not effective enough to be prescribed physician-anonymized identification numbers allow us to
except in the case of severe depression. These varying follow physicians and patients during the nine years that
warnings also reflect the scientific debate about the role of the data cover, unless patients changed their general prac-
SSRI drugs in depression treatment and their relationship titioner.
with suicide, as shown in Gibbons et al. (2006), Gibbons Table 2.1 shows the shares of each drug prescription for
et al. (2007) and Ludwig et al. (2009). Thus, although the depression diagnoses. SSRIs are the most commonly pre-
health authorities’ warnings and recommendations may scribed antidepressants. Across all age groups, more than
clearly recommend not prescribing SSRIs to kids and ado- 50% of the patients receive an SSRI-type antidepressant
lescents, this debate and the posterior evidence show that it prescription upon depression diagnosis. The prescription
is conceivable that physicians had knowledge that may not rate of “other drugs” that are not antidepressants ranges
align with recommendations, leading them not to follow from 22% for adults to 33% for kids and adolescents.
recommendations. These other drug treatments prescribed by physicians are
In the context of these warnings released by the French antipsychotics, anxiolytics, hypnotics, or antiepileptics. For
health authority from the beginning of 2000 to the end of any visit to these physicians, as soon as the physician diag-
2008, we are particularly interested in the impact of the noses a depression (even mild), he or she almost always
warning on December 2004, which informed physicians prescribes some medicine, which can be an antidepres-
that they should not prescribe SSRIs to kids and adoles- sant or another drug from these other classes (only 0.4%
cents under 18 due to the association of such drugs with of depression diagnostic end up with no drug prescrip-
an increase in suicidal thinking at this age. We focus on the tion).
period from January 2000 to June 2006 to avoid contami- Next, Table 2.2 shows the shares of drug prescriptions
nation from the June 2006 warning. for depression diagnoses for the periods before and after
the warning about SSRIs in 2004. For all age groups, the
2.4. Data and descriptive statistics share of SSRI-type antidepressant prescriptions decreases
after the warning, with the largest decrease being in pre-
We use a large panel data set on the exhaustive prescriptions for kids and adolescents, from 51% to 46%. It is
scriptions made by 386 general practitioners to all of striking to see that this decrease is far from an exact com-
their patients in France between 2000 and 2008. This pro- pliance with the warning and that the warning also leads to
prietary data set was provided by CEGEDIM (called the decreases in other age categories. While prescribing fewer
“Cegedim Strategic Data Longitudinal Patient Database”), SSRI drugs, physicians switch to other antidepressants and
a global technology and services company specializing to drugs other than antidepressants. For kids and adoles-
in health care. The data contain information on physi- cents, the share of “other drugs” increases by 10 percentage
cians, patients and patient visits. At the physician level, points after the warning, whereas for other age groups,
the data set includes age, gender and region of opera- the share of SNRI-type antidepressants and “other drugs”
tion. At the patient level, it includes sociodemographic both increase by 2–4 percentage points. However, these
information (age, gender, employment) and information on averages mask large heterogeneity across physicians.
Table 2.2
Drug prescription average probabilities – before and after the warning.
Group All ages Kids and ado. (2–18) Young adults (19–25) Adults (26–65) Elderly people (65+)
Before After Before After Before After Before After Before After
SSRIs 0.51 0.48 0.51 0.46 0.59 0.54 0.53 0.50 0.44 0.42
SNRIs 0.09 0.11 0.05 0.03 0.09 0.12 0.10 0.13 0.05 0.08
TCAs 0.08 0.06 0.05 0.02 0.02 0.02 0.07 0.05 0.12 0.10
Oth. antidep. 0.11 0.10 0.07 0.07 0.08 0.07 0.10 0.09 0.15 0.14
Oth. drugs 0.22 0.24 0.31 0.41 0.22 0.26 0.21 0.23 0.24 0.26
4
Table 2.3
Quantiles of average prescription probabilities across physicians before and after the warning.
Group All ages Kids and adolescents Young adults (18–25) Adults (26–65) Elderly people (65+)
Quantiles Quantiles Quantiles Quantiles Quantiles
25% 50% 75% 25% 50% 75% 25% 50% 75% 25% 50% 75% 25% 50% 75%
SSRIs Before 0.43 0.50 0.57 0.20 0.50 0.73 0.45 0.58 0.70 0.45 0.52 0.59 0.33 0.44 0.55
After 0.40 0.47 0.54 0.00 0.44 0.67 0.36 0.52 0.66 0.42 0.49 0.57 0.32 0.42 0.54
SNRIs Before 0.04 0.07 0.12 0.00 0.00 0.00 0.00 0.04 0.12 0.04 0.08 0.13 0.01 0.03 0.08
After 0.08 0.11 0.17 0.00 0.00 0.00 0.00 0.07 0.17 0.08 0.13 0.18 0.02 0.07 0.14
TCAs Before 0.04 0.07 0.11 0.00 0.00 0.00 0.00 0.00 0.03 0.03 0.06 0.10 0.03 0.09 0.17
After 0.03 0.05 0.08 0.00 0.00 0.00 0.00 0.00 0.00 0.02 0.04 0.07 0.02 0.06 0.13
Oth. Before 0.07 0.10 0.14 0.00 0.00 0.10 0.00 0.04 0.12 0.05 0.09 0.13 0.07 0.12 0.21
antidep. After 0.05 0.09 0.13 0.00 0.00 0.00 0.00 0.02 0.10 0.04 0.07 0.12 0.06 0.10 0.19
Oth. Before 0.16 0.21 0.29 0.10 0.25 0.50 0.12 0.20 0.33 0.15 0.20 0.28 0.13 0.23 0.32
drugs After 0.18 0.23 0.32 0.15 0.45 0.71 0.14 0.25 0.38 0.17 0.22 0.30 0.15 0.24 0.35
Table 2.3 shows the 25th percentile, the median and the in quantiles in the population can come from a few physi-
75th percentile across physicians of the prescription prob- cians dramatically changing their prescription behavior or
ability of each drug class for the periods before and after many physicians changing their behavior but by very lit-
the SSRI warning in 2004. We observe a substantial level tle. Our modeling of treatment decisions by physicians will
of heterogeneity across physicians. For instance, before the thus try to disentangle the effect of physician preferences
warning for kids and adolescents, 25% of the physicians pre- from that of patient heterogeneity, and identify individual
scribe an SSRI less than 20% of the time when they diagnose changes themselves.
depression, whereas 25% prescribe an SSRI more than 73%
of the time when they diagnose depression. We observed 3. Reduced form evidence
heterogeneity in physicians’ prescribing behavior for other
age groups as well. Before developing our discrete choice model to iden-
The comparison before and after the warning shows tify the effect of the warning on physicians preferences, we
that the probability of prescribing SSRIs decreases at each document some reduced form evidence on the effect of the
quantile for every age group. However, there is still a sub- warning.
stantial level of heterogeneity across physicians even after
the warning. For instance, for kids and adolescents, the 3.1. Changes in aggregate prescriptions
value for the first quartile for SSRI prescription probabil-
ity is 20% before the warning and 0% after the warning. Fig. 3.1 shows the shares of each drug category among all
This shows that at least 25% of physicians never pre- prescriptions for the kids and adolescents which is the tar-
scribe SSRIs to kids and adolescents after the warning, get age group of the warning and for all the other ages. The
thus following the recommendation perfectly. Similarly, SSRI warning seems to decrease the propensity to prescribe
the value for the third quartile is 73% before the warn- SSRI while not substantially increasing the other categories
ing and decreases to 67% after the warning. Moreover, the of antidepressants, meaning that the outside option of not
average prescription probabilities for SNRIs and TCAs also prescribing any antidepressant becomes more important
decrease for a large part of the distribution, as many physi- for kids and adolescents after the warning. For other age
cians stop prescribing SNRIs and TCAs. They increase their categories, we can see a decrease in the market share of
prescriptions of drugs other than antidepressants, which SSRI prescriptions while that of SNRI goes up and compen-
are mainly drugs approved for other mental disorders and sate the decrease such that the share of not prescribing any
that are used off-label for depression treatment. Off-label antidepressant is stable.
use of a drug, that is prescription of a drug for an indi- Appendix A.1 compares characteristics of patients
cation other than the indications it is approved for, was before and after the warning and performs a balancing
allowed in France during our sample period (for details see test of these observable characteristics. Results show sta-
(Tunçel, 2021)). It thus seems that the warning on SSRIs tistically significant difference only in mean age for adults
does not simply reduce prescriptions of SSRIs that would between periods before and after the warning and no
be substituted by other drugs in equal proportion to the difference in the mean age of kids and adolescents. The
prescription probability before the warning. In contrast, gender composition of patients by physician does not
the reduction of SSRI prescriptions is accompanied by a change with the warning, either. Even though the one
reduction of SNRI and TCA prescriptions for many physi- year increase in mean age of adults is unlikely to be an
cians, with an increase in other drug prescriptions. Such important confounding factor of the before-after estima-
a pattern may come from the fact that patients are het- tion of prescription choice preferences by physicians, our
erogeneous and physicians have different preferences on difference-in-difference estimation controls for observable
how different depressed patients should be treated in the characteristics such as the age category and gender of the
absence of treatment with SSRI drugs. Moreover changes patient.
5
Fig. 3.1. Market shares over time. Notes: Horizontal axis denotes time in semester of each year. Market shares do not sum to one because the alternative
of not prescribing any antidepressant is not reported.
3.2. Difference-in-difference estimation the warning with the kids and adolescents age category
dummy can be interpreted as the impact of the warning
We estimate a difference-in-difference model of the rel- on prescriptions for this age group. However, if the warn-
ative impact of the warning on prescriptions of SSRI to kids ing also has some effect on adults, then this coefficient can
and adolescents, controlling for observable characteristics be interpreted as the differential effect on kids vs. adults.
of patients and physicians. We regress the binary decision We can then assume that these coefficients allow identify-
of prescribing SSRI or not on patient and physician char- ing the full policy effect under the additional assumption
acteristics such as age and gender, a linear annual trend that there is no time effect concomitant with the warning.
and a dummy variable for the period after the warning Column (1) shows the estimation without physician fixed
(December 2004) as well as the interaction of this dummy effects while column (2) adds physician fixed effects. Under
variable with the indicator that the patient is in the kids the assumptions above, the difference in difference iden-
and adolescents age category. The coefficient of this inter- tification conditions are valid, the results show that SSRI
action should identify the relative impact of the warning on prescriptions decrease with the warning and decrease even
the choice of prescription of SSRI to kids and adolescents more for kids and adolescents. Results of column (1) show
that the warning is supposed to deter. Table 3.1 reports that the warning decreases the probability to prescribe SSRI
the results of a difference-in-difference estimation of the across all patients by 3.2% and the decrease is of an addi-
effect of the warning on the probability to prescribe an SSRI tional 9.8% for kids and adolescents. Column (2) reports
drug using a linear regression model. The difference-in- the results when we control for physician fixed effects in
difference is across kids-adolescents and other age groups which case the effect is a 2.8% decrease for all patients and
and the before and after warning periods. Then, under the an additional 9.9% decrease for kids and adolescents.
assumption that there is no spillover effect of the recom- We then perform an additional set of robustness regres-
mendation on adults, the coefficient of the interaction of sions because one may wonder if this warning effect could
Table 3.1
Difference-in-difference estimation of SSRI prescription probability.
Variables (1) (2) (3) (4)

Warning Warning Placebo Placebo
Kids–adolescents −0.062 −0.089* −0.058 −0.084

(0.053) (0.049) (0.057) (0.055)
Female patient 0.040*** 0.040*** 0.040*** 0.040***
(0.006) (0.006) (0.006) (0.006)
Female physician −0.017** −0.019***
(0.007) (0.007)
Physician’s age −0.001** −0.001***
(0.000) (0.000)
Warning −0.032*** −0.028*** 0.002 0.005
(0.004) (0.004) (0.004) (0.004)
Kids-adolescents ∗ Warning −0.098* −0.099* 0.040 0.040
(0.052) (0.051) (0.056) (0.053)
Annual time trend Yes Yes Yes Yes
Physician fixed effect No Yes No Yes
Observations 517,241 517,241 379,687 379,687
Notes: Columns (1) and (2) use data from July 2000 to July 2006 and the warning happening in December 2004. Columns (3) and (4) use data from July
2000 to December 2004 and a placebo warning in December 2002. Standard errors are clustered by patient as some patient may have several physician
visits with depression treatment. Significance levels: *** p < 0.01, ** p < 0.05, * p < 0.1.
6
with depression. We assume that each physician i receives

patient j with some depression state that the physician is
able to observe. In a given sample period, a physician has
J patients diagnosed with depression (J does not need to
be the same across physicians). We examine the physi-
cian prescription choices at time t(j) when depression is
diagnosed for patient j.
We use all prescriptions over the sample period if a
patient has multiple depression spells. We can abstract
from within-patient learning by focusing on the pre-
scription at the first time the patient is diagnosed with
depression, as at the first visit physicians still have not
yet learned any patient-specific responses to treatments.
For follow-up depression treatments, learning about the
patient’s response to each drug may play a role in a physi-
cian’s choice of treatment. However, when we select only
Fig. 3.2. Coefficient of year dummies interacted with kids and adoles-
the first depression-related visit for each patient, we obtain
cents indicator on SSRI prescription probability. Notes: Graph shows the
coefficients and their 95% confidence interval of the kids ∗ year interac-
estimates of physician preferences that are highly corre-
tion when the year dummy is defined as the date on horizontal axis. The lated with those obtained using all patient visits. Learning
omitted year is 2003 which is the year before the warning happening at may still occur, but our approach only allows for the mod-
the end of 2004. eling and identification of the effect of the warning with
heterogeneity across physicians, not accounting for possi-
be confounded by some time effect that would present a ble within-patient learning.
structural break at the time of the warning. We thus regress Each physician i (she) can choose to prescribe some
the binary decision of prescribing SSRI or not on patient and antidepressant d to patient j (he) depending on his char-
physician characteristics, an indicator on kids and adoles- acteristics j and her own taste or preference parameter,
cents age category, dummy variables for each year as well denoted ˇid t for drug d at period t. The patient char-
as the interaction of these dummy variables with the indi- acteristics j may include observable characteristics and
cator that the patient is in the kids and adolescents age unobservable ones, such as his depression state. As all
category. The estimation leaves out year 2003, the year patients are almost fully reimbursed, we do not consider
before the warning, as the reference group. Fig. 3.2 plots the price of drugs as a determinant of this choice, although
these estimated interaction coefficients, for the years in this is a testable assumption.
the horizontal axis. The figure shows that SSRI prescrip- We assume that the utility of prescription decisions
tion probability for kids and adolescents decreases with depends on physician preferences and patient character-
the SSRI warning of December 2004 until the Fluoxetine istics, including his depression state. Each physician can
recommendation happening in June 2006. choose among D + 1 treatments, indexed by d = 0, 1, . . ., D,
To provide further evidence that the warning effect in and we assume that the decision by physician i to pre-
December 2004 is unique, we also implement the same scribe treatment d for patient j visiting at t(j) is based on
estimation by difference-in-difference using only the data maximizing
before the warning and test if some placebo warning
t(j)
would be identified at some point during the middle of the v(ˇid , j ) − εijd (4.1)
2000–2004 period. Columns (3) and (4) of Table 3.1 present
the results of this estimation, and show no significant effect where v(., .) is non-decreasing in both arguments and εijd
of such placebo warning during that time period. is an individual idiosyncratic deviation for treatment d per-
Given this reduced form evidence, we develop a pre- ceived by physician i and specific to patient j. The random
scription choice model that will allow us to understand term εijd allows decisions to be non-deterministic functions
of the patient depression state j assessed by the physician.
which changes in preferences occurred andgenerated the
t ,

estimated reduced form effects, and under which condi- We normalize v ˇi0 j = 0, where by convention treat-
tions we can interpret those changes as changes in the ment 0 corresponds to drugs other than antidepressants.
preferences of physicians or changes in patients unob- We denote the treatment chosen by physician i for
served states. Moreover, our model will also allow us to patient j as yij ∈ 0, 1, . . ., D . Using (4.1) and the assump-
perform counterfactuals. tion that εijd are i.i.d. type-I extreme value, the probability
that physician i prescribes treatment d to patient j is
4. Discrete choice model and identification of t(j)
exp(v(ˇid , j ))
preference change t(j)
P(yij = d|ˇi , j ) = D t(j)
1+ d̃=1
exp(v(ˇ , j ))
id̃
4.1. Discrete choice model with unobserved
heterogeneity We observe many patients and physicians, which
implies that we can identify averages of these probabilities
We develop a discrete choice model of antidepres- in the population. Such identification needs to account for
sant prescriptions of physicians for each patient diagnosed the fact that patients who visit a given physician may have
7
heterogeneous distributions of health that is possibly cor-

0 + ω1 (j))
exp(ωdij dij
related with physicians’ preferences for treatments. This P(yij = d|i, (j)) = D
could come from common unobserved correlated effects 1+ d̃=1
exp(ω0 + ω1 (j))
d̃ij d̃ij
or from patients having some information on physicians’ 0 1 0 1
abilities and preferences. Thus, the cumulative distribution dF(ω1ij , ω1ij , . . ., ωDij , ωDij |i, (j))
of patients’ j may not be identical across physicians and
Although the warning concerns only one of the drugs d,
depend on the physician preferences ˇit = (ˇi1 t , . . ., ˇt )
iD we allow all utilities for each drug to be affected by the
and on the period t. warning, as it is possible that the new information also
Denoting as F j |ˇit , t the cumulative distribution
affects the physician’s beliefs about other drugs. Moreover,
function of j conditional on the physician preferences ˇit
we allow the full distribution of preferences of each physi-
during t, the probability that a physician i prescribes drug
cian across all patients’ status to vary with the warning, and
d to a patient is an average of the conditional probability to
for example not only the mean preferences. With the nor-
each patient type, with
mality assumption made later, we will allow the mean and

t , )
variance of physician specific preferences to change with
t(j)
exp(v ˇid j t(j) the warning.
P(yij = d|ˇi , t) = D dF(j |ˇi , t)
1+ d̃=1
exp(v(ˇt , j )) Now, we also assume that the distribution of patients’
id̃
unobservable characteristics, such as their health state, are
identical before and after the warning, that is,
which is a function of physician preferences ˇit . Matching
between patients and physicians generates some possi-
F j |i, t ≤ t1 = F j |i, t > t1
ble dependence in the cumulative distribution function
F(j |ˇit , t). However, even if patients are randomized to or equivalently
physicians such that F(j |ˇit , t) = F(j ), it remains the case
that the prescription probability depends on the distribu- 0
F(ω1ij 1
, ω1ij 0
, . . ., ωDij 1
, ωDij |i, (j) = 0)
tion function F(j ) and on the preferences of physician i.
0 1 0 1
Disentangling the distribution of unobserved heterogene- = F(ω1ij , ω1ij , . . ., ωDij , ωDij |i, (j) = 1) (4.3)
ity j from this mixture model is a difficult problem of
deconvolution. This implies that the differences in treatment before and
However, we can separately identify the change in pref- after the warning for a patient with characteristics j comes
erences and patient heterogeneity by assuming that the only from the change in preferences of the physician and
distribution of patient characteristics is stable over time there will be no change in prescription probability at the
and that physician preferences ˇid t may change only at the physician level before and after the warning if prefer-
0 = ˇ1 for ∀d, then ω1 =
ences do not change. Indeed, if ˇid
warning at time t1 , hence they may be different in the peri- id dij
ods before and after the warning. While a time trend in 1 , ) − v(ˇ0 , ) = 0 and
v(ˇid j j
id
preferences could be added, it would lead to prohibitively
hard numerical estimation problems in our empirical spec- 0 )
exp(ωdij
ification. We thus assume that: P(yij = d|i, (j) = 0) = D
1+ d̃=1
exp(ω0 )
d̃ij
t
ˇid = 0
ˇid if t ≤ t1 (before warning)
0 1 0 1
(4.2) dF(ω1ij , ω1ij , . . ., ωDij , ωDij |i) = P(yij = d|i, (j) = 1)
1 if t > t
= ˇid (after warning)
1
Denoting as (j) = 1{t(j)>t1 } the dummy variable for The stability assumption that physician preferences
whether patient j visits physician i before or after the warn- may only change at the time of the warning allows us
ing, we define the physician preference for drug d before to identify whether preferences change even if there is
0 ≡ v(ˇ0 , ) and the change in preferences
the warning ωdij id j endogenous matching between patients and physicians.
1 ≡ v(ˇ1 , ) − v(ˇ0 , )
for drug d due to the warning ωdij j j
The assumption of stability of preferences is usual in many
id id
such that: studies of discrete choice prescribing behavior when con-
tinuous learning is not the focus (for example, Dickstein,
t(j)
v(ˇid 0 , )(1 − (j)) + v(ˇ1 , )(j)
, j ) = v(ˇid 2016). Allowing physicians preferences to vary over time
j id j
0 + ω1 (j)
≡ ωdij more freely than just allowing a difference between before
dij
and after the warning is conceptually possible with our
approach but numerically too difficult. Moreover, our
This implies that the probability that physician i prescribes
reduced form evidence of Section 3 shows that the time
d to patient j at time period (j) is:
of the warning seems to be the time where we can observe
0 + ω1 (j)) a structural break in prescription decisions of SSRIs.
exp(ωdij dij
P(yij = d|i, j, (j)) = D
1+ d̃=1
exp(ω0 + ω1 (j)) 4.2. Econometric specification
d̃ij d̃ij
and the average probability for physician i to prescribe d is To estimate the model, we need to specify a paramet-
then: ric distribution for unobservables. We assume that the
8
0 , ω1 , . . ., ω0 , ω1 are independent across alternatives

ω1ij the average marginal effect on the prescription of drug d
1ij Dij Dij
d, that is for physician i is

D
P yij = d|(˛0di , ˛1di , di0
2 2
, di1 di )d=1,...,D
0 1 0 1 0 1
F ω1ij , ω1ij , . . ., ωDij , ωDij |i = F ωdij , ωdij |i

d=1 0
exp(ωdij + ωdij
1
) 0
exp ωdij
≡ D − D (4.5)
and F(.) is jointly normal with 1+ exp(ω0 + ω1 ) 1+ exp ω0

d̃ =1 d̃ ij d̃ ij d̃ =1

d̃ ij
0 1
iid ˛0di 2
di0 di dF (ω0 , ω1 ) |(˛0 , ˛1 , 2 , 2 , d̃ i )
ωdij , ωdij ∼N , 2
d̃ ij d̃ ij d̃ =1,...,D d̃ i d̃ i d̃ i0 d̃ i1 d̃ =1,...,D
˛1di di di1
The heterogeneity across physicians of parameters ˛0di ,
0
where we allow some nonzero correlation between ωdij
˛1di , di0
2 , 2 , combines the potential heterogeneity of
di1 di
1 , implying that we allow the change in physician
and ωdij behavior of physicians and the potential heterogeneity of
i’s preference for drug d due to the warning, ωdij 1 , to be cor- patients since endogenous matching is possible, and can-
related with the physician’s preference before the warning, not be disentangled without additional assumptions. For
0 . example, before the warning, the distribution of ωdij0 for
ωdij
We obtain a discrete choice model that corresponds to a physician i can come from the fact that the physician has
random coefficient discrete choice logit for each physician very varying preferences of what the value of a treatment
i. While we add functional form restrictions for prac- d is across mildly varying health states of patients or from
tical estimation, McFadden and Train (2000) show that the fact that her patients have very heterogenous health
mixed logit (random coefficient logit) models are flex- states. However, assuming stability of the distribution of
ible enough to approximate any discrete choice model. patients for a given physician before and after the warning
0
allows us to interpret differences in the distributions of ωdij
The conditional choice probability that physician i chooses
yi1 = d1 , yi2 = d2 , .., yiJ = dJ for her J patients is 1 as changes in preferences for a given physician. The
and ωdij
maintained intuitive argument behind our assumptions is

J
that the warning can change physician preferences but
P yij = d|i, (j) would not change the depression likelihood and intensity
j=1 of patients.
where
5. Empirical results

0 + ω1
exp(ωdij dij
(j))
P yij = d|i, (j) = D 5.1. Model estimates
1+ d̃=1
exp(ω0 + ω1 (j))
d̃ij d̃ij
We thus implement the estimation of this random

D
coefficient logit model for each physician. We consider
dF ω0 , ω1 |˛0 , ˛1 , di0
2
, 2 , d̃i (4.4)
d̃ij d̃ij d̃i d̃i d̃i1 the alternative choices of antidepressant classes as SNRIs,
d̃=1 SSRIs, TCAs, and “other antidepressants” (mianserine, mir-
tazapine, tianeptine) while the category “other drugs” is the
With a large number of patients J per physician, we can normalized outside option and gathers drugs that are not
identify the parameters ˛0di , ˛1di , di0
2 , 2 , for all physi-
di1 di antidepressants. The latter are mostly drugs not approved
1
cians i = 1, . . ., I. Thus, if ˛di = 2 =
/ 0 or di1 / 0, it will mean for depression treatment but used off-label in depression
that physician preferences have changed with the warning, treatment by physicians. These drugs are mostly antipsy-
and we identify changes in the full distribution of prefer- chotics (i.e., olanzapine) or anxiolytics (i.e., alprazolam,
ences of each physician and not only a single preference bromazepam, prazepam). The discrete choice model thus
parameter in case we would not allow heterogeneity across has 5 alternatives that almost all physicians prescribe,2 and
patients. we ignore coprescriptions, which represent less than 3% in
The marginal effect of the warning on physician i’s pre- depression treatment.
scription probability to patient j is We allow the patient’s observable characteristics, such
as gender (gj ) and age (aj ) to affect the mean util-
0 + ω1 )
exp(ωdij dij

P yij = d|i, j ≡ D ity of the discrete choice model such that ˛0di gj , aj ≡
1+ d̃=1
exp(ω0 + ω1 ) g
˛0di + ˛di gj + ˛adi aj . The estimation of the random coeffi-
d̃ij d̃ij
cient logit model thus has 8 random effects ωdij 0 , ω1
0
exp ωdij dij
− D for d = 1, 2, 3, 4 at the patient level j and 28 parame-
g
1+ d̃=1
exp ω0 ters ˛0di , ˛1di , ˛adi , ˛di , di0 , di1 , di for d = 1, 2, 3, 4 for each
d̃ij
and after identifying the parameters ˛0di , ˛1di , di0

2 , 2 ,
di1 di
for all physicians, we can obtain any moment or quantile of 2
Among all the physicians, only 3 never prescribe an SNRI, 8 never
the distribution of the marginal effect on the prescription prescribe a TCA and only one never uses other antidepressants. All of the
of drug d both within and across physicians. For example, physicians prescribe SSRIs.
9
physician i = 1, . . ., 386. For 48 physicians, the model is less than 0.41, whereas for 25%, it is more than 0.64.
parameters cannot be estimated even with added restric- We also observe heterogeneity in terms of their response
tions because of the existence of too few patients with to the warning. At every quantile, the probability of pre-
depression diagnoses. For 91 physicians, the correlation scribing SSRIs decreases for every age group. However, for
di is very imprecisely estimated, with a very large stan- kids, adolescents and young adults, the decrease grows
dard error, in which case we estimate the same model with larger as the quantile grows larger, in terms of both
the additional restriction of no-correlation (di = 0). For an percentage and percentage points, suggesting that the
additional subset of 32 physicians, the variance coefficients physicians prescribing SSRIs more often before the warn-
di0 or di1 are too imprecisely estimated, in which case ing decrease their prescriptions more after the warning. We
we also impose that di0 = 0 and di1 = 0. As a result, there can also see that most of the substitution is towards SNRI
are no restrictions on parameters for the remaining 215 drugs.
physicians. We thus obtain all parameter estimates for 338 Fig. 5.1 shows the decrease in percentage of the baseline
physicians (we have imposed di = 0 for 91 of them and prescription probability for each physician. The decrease
di = di0 = di1 = 0 for 32). is on average slightly higher and less variable across the
Table 5.1 reports the results of this random coefficient physicians who have a large prescription probability of SSRI
model for one of the physicians. The results show that the before the warning.
warning makes this physician’s preference towards SSRIs We also compute the change after the warning in the
decrease, as ˛1di is significantly negative for SSRIs, and that prescription probabilities for each drug category and for
the warning increases his preference towards SNRIs, as ˛1di each physician. Table 5.3 reports the quantiles across physi-
is positive for SNRIs albeit significant only at the 10% level. cians for the change in prescription probabilities. For all
The parameter di0 is positive and significant, showing that the age groups but the elderly, 25% of physicians decrease
there is large heterogeneity in treatments before the warn- their probability of prescribing SSRIs by at least 12 percent-
ing. This heterogeneity is not surprising and is due to the age points. For elderly patients, the 25% of physicians who
heterogeneity of patients for this physician. The parameter decrease SSRI prescriptions the most show a decrease of
di1 is also positive, showing, for example, that this physi- a maximum of 9 percentage points. In contrast, across all
cian’s preferences are affected by the warning such that her age groups, 25% of the physicians either do not respond
decision utility for SSRIs has an even larger variance after to the warning at all or increase their average probability
the warning. The parameter di being positive for SSRIs of prescribing SSRIs (the 75% quantile is +0.01). Accord-
shows that the larger the variance before the warning, the ing to Tables 5.3 and 5.4, physicians who decrease their
larger it is after. As a result, this physician decreases SSRI prescriptions of SSRIs substitute towards SNRIs and “other
prescriptions after the warning and substitutes towards drugs”.
SNRIs and the reference alternative, “other drugs”. The dis- Table 5.4 shows, for each drug class, the mean
tribution of estimated parameters across all physicians is and standard deviation of the averages across patients
provided in Table A.2 in the appendix. of the physician-level prescription probability for the
As all parameters change with the warning, it is easier periods before and after the warning. It also reports
to look at changes in prescription probabilities, as we do in the within-physician changes in variance of prescrip-
the following section. tion probabilities. It shows that the average probability
of prescribing SSRIs decreases with the warning by
5.3 percentage points. Physicians substitute away from
5.2. Effects of the warning on choice probabilities SSRIs towards SNRIs. It also shows that the heterogene-
ity across physicians increases after the warning and
Using the model estimates, we can now predict the increases more for other drug classes than SSRIs. It is as
choice probabilities before and after the warning for each if the warning is interpreted differently by heterogeneous
physician for any patient of any age and gender group. physicians.
Table 5.2 reports the quantiles of prescription probability 19 months after the 2004 warning, in 2006, the health
for each choice alternative with before and after warn- authorities released some new information, partially con-
ing preferences. These predicted probabilities should be tradicting the 2004 warning by recommending one SSRI
equal to those in Table 2.3 if there is no estimation error molecule (Fluoxetine, sold under trade name Prozac) for
and if the model specification is correct. We can see that kids and adolescents. One may wonder if some physicians’
the results are similar, although our model imposes the response to the 2004 warning is limited because of pri-
restrictions that age and gender can affect only the mean vate information that Fluoxetine should indeed be used.
utilities and not the variance. This shows that the choice To address this concern, for each physician, we compute
modeling allows us to replicate moments of the physician- the share of Fluoxetine prescriptions in the period before
level choice probability distribution. As we can see below, the SSRI warning and in the period after the SSRI warning
the model also allows us to identify the physician-level but before the Fluoxetine warning. Among the physicians
heterogeneity of prescriptions within her set of patients. who prescribe SSRIs after the warning, only 19% of them
We observe a substantial level of heterogeneity across prescribe Fluoxetine more than half of the time in the
physicians, not only in terms of initial prescription prob- after-warning period, whereas 27% of them prescribe Flu-
abilities but also in their responses to the warning. For oxetine more than half of the time before the warning.
instance, for kids and adolescents, for 25% of the physi- Moreover, only 16% of the physicians prescribing SSRI after
cians, the before-warning probability of prescribing SSRIs the warning increase their share of Fluoxetine prescrip-
10
Table 5.1
Random coefficient logit estimation for a single physician i.
Drugs Patient’s Baseline Warning
Age Gender Mean SD Mean SD Correlation

g
˛di ˛adi ˛0di di0 ˛1di di1 di
SSRIs −0.07 −0.47 4.04 3.21 −1.86 8.26 0.93

(.01) (.39) (.66) (.36) (.52) (1.35) (.99)
SNRIs −0.06 0.19 1.85 3.24 0.78 5.30 −1.58
(.01) (.45) (.74) (.33) (.47) (1.08) (.96)
TCAs −0.01 −0.29 −2.64 2.77 0.83 4.49 0.56
(.02) (.61) (1.18) (.41) (.99) (1.58) (.71)
Other antidep. −0.18 −0.36 6.02 3.45 −2.69 7.55 6.49
(.04) (.60) (1.59) (.42) (1.38) (2.46) (1.98)
No. of visits 1397
Notes: A negative gender coefficient means that the physician has a lower preference for this drug for female patients (dummy is 1 for female and 0 for
male). Standard errors in parentheses.
Table 5.2
Heterogeneity across physicians of average prescription probabilities before/after warning.
Quantiles (%) Quantiles (%) Quantiles (%) Quantiles (%) Quantiles (%)
25 50 75 25 50 75 25 50 75 25 50 75 25 50 75
SSRIs Bef. 0.40 0.48 0.54 0.41 0.52 0.64 0.41 0.52 0.62 0.40 0.48 0.56 0.33 0.41 0.50
Aft. 0.34 0.42 0.48 0.36 0.45 0.55 0.36 0.45 0.53 0.34 0.42 0.49 0.3 0.38 0.46
SNRIs Bef. 0.06 0.10 0.14 0.05 0.10 0.18 0.06 0.10 0.16 0.06 0.10 0.15 0.05 0.08 0.14
Aft. 0.08 0.14 0.22 0.07 0.15 0.23 0.07 0.15 0.23 0.08 0.14 0.22 0.07 0.13 0.20
TCAs Bef. 0.04 0.07 0.10 0.01 0.03 0.06 0.01 0.03 0.06 0.03 0.06 0.09 0.05 0.10 0.17
Aft. 0.01 0.06 0.11 0.00 0.03 0.07 0.01 0.03 0.08 0.01 0.05 0.10 0.02 0.07 0.15
Oth. Bef. 0.07 0.11 0.16 0.03 0.07 0.13 0.03 0.08 0.14 0.06 0.10 0.15 0.07 0.13 0.20
antidep. Aft. 0.06 0.11 0.18 0.03 0.09 0.16 0.03 0.09 0.17 0.05 0.11 0.18 0.06 0.13 0.21
Oth. Bef. 0.16 0.22 0.28 0.12 0.19 0.28 0.13 0.19 0.28 0.16 0.21 0.27 0.15 0.22 0.30
drugs Aft. 0.18 0.22 0.29 0.16 0.22 0.29 0.16 0.22 0.29 0.18 0.22 0.29 0.17 0.22 0.29
Fig. 5.1. Effect of warning on average prescription probability by physician and drug as percentage of baseline prescription probability. Notes: Scatter
plot of the change in predicted physician-level prescription probability after the warning as a percentage of the before-warning probability. Each point
represents a physician-level probability. The left graph shows the mean choice probabilities and the changes for any patient, and the right graph shows
the mean choice probabilities and the changes for kids and adolescents.
tions with the warning. Among physicians who strictly share of Fluoxetine is 12% before the warning and 10% after
follow the recommendation to stop SSRI prescriptions to the warning, showing no significant change. These results
kids and adolescents, the average share of Fluoxetine in show that among those who deviate from the 2004 warn-
SSRI prescriptions before the warning is 17% (and none ing, only 19% do so in a way that is consistent with the
after since they stop prescribing SSRIs). However, among 2006 recommendation. Therefore, it is not necessarily the
those who do not follow the recommendation, the average case that the physicians that do not strictly follow the 2004
11
Table 5.3
Heterogeneity across physicians in change in prescription probabilities due to the warning.
Quantiles (%) Quantiles (%) Quantiles (%) Quantiles (%) Quantiles (%)
25 50 75 25 50 75 25 50 75 25 50 75 25 50 75
SSRIs −0.12 −0.05 0.01 −0.13 −0.06 0.01 −0.13 −0.06 0.00 −0.13 −0.05 0.00 −0.09 −0.03 0.04
SNRIs −0.02 0.04 0.10 −0.02 0.04 0.09 −0.02 0.04 0.10 −0.02 0.04 0.10 −0.02 0.04 0.10
TCAs −0.05 −0.01 0.03 −0.02 0.00 0.03 −0.02 0.00 0.03 −0.04 −0.01 0.03 −0.07 −0.02 0.02
Oth. antidep. −0.05 0.00 0.05 −0.03 0.00 0.05 −0.03 0.00 0.05 −0.04 0.00 0.05 −0.06 −0.01 0.05
Oth. drugs −0.02 0.02 0.05 −0.01 0.03 0.06 −0.02 0.02 0.06 −0.03 0.02 0.06 −0.03 0.01 0.06
Table 5.4
Distribution of physician prescription probabilities before/after warning.
Across physicians Within-physician
Mean Standard deviation Standard deviation
Drug Before After Change Before After Change Before After Change
SSRIs 0.470 0.417 −0.053 0.111 0.117 0.006 0.354 0.388 0.034
SNRIs 0.106 0.152 0.046 0.065 0.101 0.036 0.209 0.268 0.059
TCAs 0.077 0.073 −0.004 0.052 0.074 0.022 0.185 0.179 −0.006
Oth. antidep. 0.119 0.125 0.006 0.068 0.091 0.023 0.221 0.242 0.021
Oth. drugs 0.227 0.241 0.014 0.093 0.103 0.010 0.248 0.305 0.057
Note: Mean and standard deviation across physicians in the first six columns and within-physician standard deviation of prescription probabilities in the
next three columns.
warning are “better” physicians who could anticipate the ing that they are the ones substituting away from SSRIs
2006 Fluoxetine recommendation. towards SNRIs and/or “other drugs”. For kids and ado-
Our model also allows us to identify the effect of the lescents, an even higher number of physicians are in the
warning on the within physician variation of preferences. upper-left corner of the graph.
The results show an increase in the within-physician vari- As we have seen earlier, the warning affects not only the
ance of the probability of prescribing SSRIs and SNRIs, mean physician preference towards each drug but also its
meaning that after the warning, physicians make less variance, meaning that it affects the way physicians pre-
homogeneous decisions across patients than before. The scribe heterogeneously across patients. Fig. 5.4 plots the
figures below help clarify the changes across the different within-physician variance of the prescription probability
drug categories. with before-warning preferences on the horizontal axis
Fig. 5.2 plots the average physician prescription prob- and after-warning preferences on the vertical axis for all
ability for all drugs with before-warning preferences on patients. The figure shows that the physician-level vari-
the horizontal axis and after-warning preferences on the ance in the probability of prescribing SSRIs increases for
vertical axis. The first row reports the average prescription almost all physicians except for those with a lower vari-
probability by physician for all patients and the second row ance before the warning, who do not seem to be affected.
for only kids and adolescents. We see that with the warning, This shows that the warning does not lead physicians to
a majority of physicians decrease their SSRI prescriptions prescribe uniformly across patients after the warning, and
and increase their SNRI prescriptions. We do not observe the second row of graphs in Fig. 5.4 shows that this is also
a clear trend for other choice alternatives. For instance, true within the age category of kids and adolescents. For
for TCAs and “other antidepressants”, half of the physi- a majority of the physicians, the within-physician vari-
cians increase their prescriptions of these drugs after the ance in the probability of prescribing SNRIs and “other
warning, whereas the the other half decrease their pre- drugs” also increases after the warning. We do not see
scriptions of these drugs. The figure shows substitution such a clear effect for other alternatives. For TCAs and
from SSRIs towards SNRIs. For kids and adolescents, we “other antidepressants”, the within-physician variance of
observe similar responses to the warning, except that many the prescription probability slightly increases for approxi-
more physicians substitute SSRIs with “other drugs”, not mately half of the physicians and slightly decreases for the
only with SNRIs. other half. We observe very similar patterns for kids and
Fig. 5.3 shows the substitution patterns between SSRIs adolescents even though the warning concerns only the
and other drugs using estimates of the marginal effect of patients in this age group. Contrary to what may have been
the warning on each probability as in Eq. (4.5). The left expected, the warning does not lead to more uniform treat-
(right) panel plots, for each physician, the change in the ment choices across physicians because the effect of the
probability of prescribing SSRIs on the horizontal axis and warning on their preferences proves to be very heteroge-
the change in the probability of prescribing SNRIs (“other- neous.
drugs”) on the vertical axis. The majority of the physicians Another way to examine the heterogeneity of the effects
are located in the upper-left corner of the graph, mean- of the warning consists in looking at the changes in the
12
Fig. 5.2. Effect of warning on average prescription probability by physician and drug as function of baseline prescription probability. Notes: Scatter plot
of predicted physician-level prescription probability after the warning as a function of the before-warning probability. Each point represents a physician-
level probability. The first row shows the mean choice probabilities for any patient, and the second row shows the mean choice probabilities for kids and
adolescents.
Fig. 5.3. Effect of warning on average prescription probability by physician. Notes: Plots of changes in the physician-level mean probability of prescribing
SSRIs vs. SNRIs and “other drugs”.
distribution of prescription probabilities across physicians When looking at the correlation of the physician-level
depending on their before-warning choice probability. probabilities of prescribing any of these drug categories
Fig. 5.5 and Fig. 5.6 plot these densities of the average with observable physician characteristics, we find no sig-
change in physician prescription probability by quartile nificant correlation with physician age and only some
of the ex ante prescription probability for all patients positive between SNRI preferences and gender before the
and for kids and adolescents, respectively. For SSRIs, the warning showing that female physicians like SNRI drugs
largest decrease in prescription probability after the warn- more than males before the warning. We do not find signifi-
ing is among physicians in the highest quartile (quartile cant correlation with age unlike (Howard and Hockenberry,
4) in terms of ex ante probability of prescribing SSRIs. 2019) who show that older physicians react more slowly
The smallest decrease is among physicians in the low- to new information. However, we find some correlation
est quartile (quartile 1). Similarly, the largest increase in between the number of depressed patients per year seen
the probability of prescribing SNRIs and “other drugs” by the physician and the physician-level prescription prob-
is among those who were prescribing those categories abilities. The more patients seen by a physician, the higher
least often before the warning (quartiles 1 and 2 in the is her probability of prescribing SSRIs both before and after
figures for SNRIs and “other drugs”). The patterns are the warning (without correlation with the change) and
similar across all patients and for kids and adolescents the lower her probability of prescribing SNRIs and ‘Other
only. Drugs’ (for details see Appendix A.3).
13
Fig. 5.4. Effect of warning on within-physician variance of prescription probability. Notes: Scatter plot of physician-level variance of prescription probability
after the warning as a function of the before-warning probability. Each point represents one physician-level variance observation. The first row shows the
variance of choice probabilities for any patient, and the second row shows the variance of choice probabilities for kids and adolescents.
Fig. 5.5. Effect on prescription probability by quartile – all ages. Notes: Kernel density estimates of physician-level changes in prescription probability by
quartile of ex ante choice probability.
14
Fig. 5.6. Effect on prescription probability by quartile – kids and adolescents. Notes: Kernel density estimates of physician-level changes in prescription
probability by quartile of ex ante choice probability.
5.3. Comparing the effects of the warning with a ban of patient types (j in our model) will substitute differently
SSRI by other drugs. As we allow this heterogeneity to be
In the previous section, we show that the warning on physician specific because of possibly endogenous match-
SSRIs on average reduces physician prescriptions of SSRIs ing of patients with physicians, this has consequences on
but also has very heterogeneous effects. Given that the the aggregate substitution rate of SSRI towards other drugs.
warning was clear on the fact that SSRIs should not be pre- Banning SSRI drugs for use by kids and adolescents
scribed (or should only be prescribed as a last resort) to kids could, however, not only change the ability to prescribe
and adolescents, we may consider the possible effect of a SSRIs but also affect the preferences of physicians towards
complete ban like those sometimes imposed on drugs that other drugs, just as the warning has done. As we do not
are uniformly considered too unsafe. A ban can be inter- observe such a ban, we compare the effects of both the ban
preted as strict guidelines adherence by physicians. This is and the warning using the ex ante and ex post physician
what happened, for example, when the anti-inflammatory preferences (before and after the warning).
Vioxx was pulled from the market. We thus look at the Our model allows us to simulate the prescription prob-
counterfactual effects of a ban of SSRI drugs for kids and abilities in the absence of SSRIs as follows. With the same
adolescents to compare physicians’ substitution of drug notation as in Eq. (4.4), the choice probability of any drug d
prescriptions. Of course, in the case of a ban, SSRI pre- that is not an SSRI based on prewarning ((j) = 0) or post-
scriptions to kids and adolescents would disappear, while warning ((j) = 1) preferences is:
the warning is far from yielding such an effect. That said,
the model allows us to compare the substitutions to other
P yij = d|i, (j), noSSRI
drugs in the case of a ban with the case of the warning,
thanks to counterfactual prescription simulations. Substi-

0
exp(ωdij + ωdij
1
(j))
tution of SSRI by other drugs depends not only on the mean = dF(ω0 , ω1 )
d̃ ij d̃ ij
(5.1)
1+ exp(ω0 + ω1 (j))
{d̃ =
/ SSRI}
preferences of physicians that are heterogenous but also d̃ ij d̃ ij
{d̃ =
/ SSRI}
on the variance term that depends on the distribution of

patient types within a physician, which is what our model Table 5.5 shows the mean choice probability of each
also allows us to obtain. For example, two physicians hav- drug category with or without the ban using pre- or post-
ing the same preference but having different distributions warning preferences. Given that the decrease in SSRIs
15
Table 5.5
Effects of an SSRI ban vs. the warning on physician prescription probabilities (kids and adolescents (2–18)).
With prewarning preferences With postwarning preferences Warning only
No ban With ban Change No ban With ban Change Change
Drug Pd0 0
Pd,ban 0
Pd,ban − Pd0 Pd1 1
Pd,ban 1
Pd,ban − Pd1 Pd1 − Pd0
SSRIs 0.517 0.000 −0.517 0.452 0.000 −0.452 −0.065

SNRIs 0.122 0.207 +0.085 0.158 0.231 +0.073 +0.036
TCAs 0.043 0.078 +0.035 0.048 0.075 +0.027 +0.005
Oth. antidep. 0.101 0.176 +0.075 0.112 0.169 +0.057 +0.011
Oth. drugs 0.215 0.537 +0.322 0.234 0.532 +0.298 +0.019
Notes: Column titles denote the mean prescription probability for any kid or adolescent patient across all physicians. Pd0 is the mean prescription probability
0
of drug d under prewarning preferences, and Pd,ban is the mean prescription probability of drug d under prewarning preferences when SSRIs are banned.
Pd1 and Pd,ban
1
denote the same mean probabilities using postwarning preferences.
is obviously much larger under a ban, the ban mostly type antidepressant prescriptions decreased after 2004 for
leads to substitution to non-antidepressant drugs rather kids and adolescents, but the physicians responded to the
than to SNRIs or other antidepressants. The SSRI warn- new information very heterogeneously. The drug warning
ing leads to a modest decrease in SSRI prescriptions, half increased the variance of physician prescribing behavior
of which is directed towards SNRI drugs (see the last col- both across physicians and within individual physicians.
umn of Table 5.5); however, while the ban on SSRIs leads One important result is that the warning reduced the prob-
to a much larger effect, more than half of the decrease ability of prescribing SSRIs to all patients in addition to
in SSRI prescriptions goes to drugs other than antide- kids and adolescents and that this reduction was larger
pressants. This means that the ban on SSRI drugs has a but also more heterogeneous for physicians with a higher
very different effect from that of the SSRI warning. We mean probability of prescribing SSRIs before the warning.
can see that the effect of the ban on SSRIs using post- Using counterfactual simulations that our discrete choice
warning preferences proportionately benefits other drugs model estimation allows us to perform, we compare the
more (0.298/0.452 = 0.66 is larger than 0.322/0.517 = 0.62). effect of the SSRI warning with a strict compliance with
Of course, the ban on SSRIs also has a quite different effect the recommendation (a ban) to not prescribe SSRIs by
on the within-physician variance of the prescription prob- kids and adolescents. The results show that in the case
ability, as it lowers the variance in prescribing SSRIs (since of a warning, physicians who follow the recommendation
the probability of prescribing SSRIs becomes zero for any by not prescribing SSRIs to kids and adolescents substi-
patient of any physician), while the warning has the effect tute more towards other antidepressant classes (SNRIs in
of increasing the variance. particular) than what they would do if there was a strict
ban in which case they substitute more towards drugs
6. Conclusion that are not antidepressants. This shows that the warn-
ing has, perhaps unexpectedly, positive spillover effects
In this paper, we study how scientific information on preferences for other antidepressants compared to a
released by public authorities, such as a drug warning, strict removal of SSRIs. These results call into question
affects the prescribing behavior of physicians. As physi- the interpretation of drug warnings and recommendations
cian prescribing behavior may depend on both physician by physicians and show how heterogeneous reactions can
preferences and on unobserved, possibly correlated, char- occur in relationship to physicians’ ex ante preference for
acteristics of patients, we estimate a model that allows the different possible treatments.
us to infer the effect of the warning on the full distribu-
tion of each physician’s preferences over her patients. We Appendix A
use the long time dimension of panel data on physician
prescriptions to a large set of patients before and after a A.1 Balancing tests before and after the warning
warning that may have affected physicians’ preferences.
By assuming that the distribution of patient heterogene- This section provides descriptive statistics at physician
ity is stable over time before and after the warning, we level on patient observables (age, gender) before and after
can identify the change in preferences by allowing for the warning. For each physician we compute the average
physician-specific random effects in prescribing behav- age of patients and share of female patients, before and
ior. after the warning in order to compare those characteristics
In the case of antidepressant drugs, new evidence on and see if observable differences seem to be important. The
the increase in suicidal thinking in kids were reported in mean across physicians of the average patients’ age went
2004 for selective serotonin reuptake inhibitors (SSRIs). slightly up from 52.4 years before the warning to 54.1 years
We use French panel data on exhaustive prescriptions of after. However, the mean age of kids and adolescents did
a representative sample of general practitioners to more not significantly change, which was 14.9 years before the
than 110,000 depressed patients between 2000 and 2008 to warning and 14.8 years after. The average of gender com-
estimate the effect of an official warning. We find that SSRI- position by physician stayed stable for all age groups. It was
16
Table A.1
Testing for change in patient observables before and after the warning.
Patients variables All Kids/adolescents Adults
Age Gender Age Gender Age Gender

(1) (2) (3) (4) (5) (6)
Warning 1.699*** −0.001 −0.096 0.000 1.639*** −0.001

(0.391) (0.010) (0.308) (0.046) (0.385) (0.010)
Constant 52.403*** 0.707*** 14.904*** 0.689*** 52.597*** 0.707***
(0.276) (0.007) (0.218) (0.033) (0.272) (0.007)
Observations 772 772 248 248 772 772
Notes: OLS regressions of mean age of patients and share of female patients by physician before and after the warning on a warning dummy variable.
Significance level: *** p<0.01.
71% females both before and after the warning. For kids and kids and adolescents, and columns (5) and (6) are across
adolescents, it was 69% females both before and after the the adult patients only. The results show no significant
warning. change in those characteristics for kids and adolescents
In Table A.1, we present the regression results of the but a significant increase in age for adults by a bit more
mean age of patients and share of female patients at than 1.5 years.
physician level before and after the warning on a warning
dummy variable. We thus have two observations of these A.2 Full model estimates table
statistics by physician, one before and one after the
warning. In columns (1) and (2) the mean age and gender Table A.2 reports the distribution of all coefficient esti-
at physician level are calculated including all the patients. mates of the model across the 386 physicians.
Columns (3) and (4) are for the mean age and gender of
Table A.2
Distribution of coefficient estimates across physicians.
Drugs Baseline parameters Quantiles
25 % 50 % 75 %
g
SSRIs Age ˛di −0.04 −0.01 0.01
Gender ˛adi −0.62 0.07 0.81
Constant ˛0di 0.30 1.55 3.06
Std deviation di0 2.41 3.12 4.10
g
SNRIs Age ˛di −0.05 −0.01 0.02
Gender ˛adi −1.31 0.01 1.19
Constant ˛0di −5.60 −2.75 −0.66
g
TCAs Age ˛di 0.01 0.04 0.09
Gender ˛adi −1.54 0.03 1.77
Constant ˛0di −12.8 −7.73 −4.13
g
Other antidep. Age ˛di −0.02 0.02 0.06
Gender ˛adi −1.48 −0.35 0.58
Constant ˛0di −6.81 −3.32 −0.95
Warning effects
SSRIs Mean ˛1di −0.91 −0.28 0.47

Correlation di −0.46 0.00 0.56
SNRIs Mean ˛1di −4.47 −1.07 0.48
TCAs Mean ˛1di −9.41 −3.04 −0.30
Other antidep. Mean ˛1di −4.82 −1.78 −0.17
Notes: Coefficients of random coefficient logits with 338 physician-specific coefficients. Correlation coefficients di are not identified and thus restricted
to zero for 91 physicians, and all random coefficients are not identified and thus are restricted to zero for 32 physicians. From the original sample, 48
physicians do not have enough visits with depression to be included in the model.
17
Table A.3
Correlation of Physician Observables with Before Warning Prescription Probabilities
Variables (1) (2) (3) (4) (5)

SSRI SNRI TCA Other antidep. Other drugs
Female physician −0.0262 0.0227** −0.0003 −0.0092 0.0155

(0.0164) (0.0096) (0.0077) (0.0102) (0.0137)
Physician’s age −0.0010 −0.0001 0.0005 −0.0002 0.0008
(0.0009) (0.0005) (0.0004) (0.0005) (0.0007)
Number of patients 0.0002 −0.0001 0.0001* 0.0000 −0.0003**
(0.0001) (0.0001) (0.0001) (0.0001) (0.0001)
Constant 0.5056*** 0.1144*** 0.0432** 0.1303*** 0.2065***
(0.0428) (0.0249) (0.0201) (0.0265) (0.0357)
Observations 338 338 338 338 338

R-squared 0.0139 0.0226 0.0151 0.0026 0.0220
Notes: Significance levels: *** p < 0.01, ** p < 0.05, * p < 0.1.
Table A.4
Correlation of physician observables with after warning prescription probabilities.
Variables (1) (2) (3) (4) (5)

Female physician −0.0203 0.0079 0.0054 −0.0065 0.0143

(0.0174) (0.0150) (0.0112) (0.0136) (0.0155)
Physician’s age −0.0007 −0.0002 0.0003 0.0002 0.0007
(0.0009) (0.0008) (0.0006) (0.0007) (0.0008)
Number of patients 0.0003* −0.0003* 0.0001 −0.0001 −0.0003*
(0.0002) (0.0001) (0.0001) (0.0001) (0.0001)
Constant 0.4316*** 0.1770*** 0.0523* 0.1206*** 0.2286***
(0.0454) (0.0393) (0.0289) (0.0354) (0.0404)
Observations 338 337 332 337 330

R-squared 0.0137 0.0119 0.0036 0.0024 0.0148
Notes: Significance levels: *** p < 0.01, * p < 0.1.
Table A.5
Correlation of physician observables with the change in prescription probabilities.
Variables (1) (2) (3) (4) (5)

Female physician 0.0061 −0.0129 0.0209 0.0106 −0.0124

(0.0142) (0.0157) (0.0157) (0.0124) (0.0126)
Physician’s age 0.0003 0.0002 −0.0008 0.0010 −0.0005
(0.0007) (0.0008) (0.0008) (0.0006) (0.0007)
Number of patients 0.0001 −0.0002 −0.0000 0.0000 0.0000
(0.0001) (0.0001) (0.0001) (0.0001) (0.0001)
Constant −0.0759** 0.0477 0.0294 −0.0485 0.0386
(0.0370) (0.0409) (0.0406) (0.0324) (0.0328)
Observations 338 337 332 337 330

R-squared 0.0027 0.0069 0.0124 0.0075 0.0038
Notes: Significance levels: ** p < 0.05.
A.3 Correlation between physician observables and Berndt, E.R., Gibbons, R.S., Kolotilin, A., Taub, A.L., 2015. The
prescription probabilities heterogeneity of concentrated prescribing behavior: theory and
evidence from antipsychotics. J. Health Econ. 40,
26–39.
In this section, we report the correlations between Ching, A.T., 2010. A dynamic oligopoly structural model for the
physician observables and physician-level prescription prescription drug market after patent expiration. Int. Econ. Rev. 51,
1175–1207.
probabilities. We regress physician-level prescription Ching, A.T., Erdem, T., Keane, M.P., 2013. Learning models: an assessment
probabilities for each dug on physician observables. of progress, challenges, and new developments. Market. Sci. 32,
Tables A.3–A.5 show the correlations between physician 913–938.
Ching, A.T., Lim, H., 2020. A structural model of correlated learning and
observables and prescription probabilities before the warn- late-mover advantages: the case of statins. Manag. Sci. 66,
ing, after the warning and the change in probabilities, 1095–1123.
respectively. Collins, J.M., Simon, K.I., Tennyson, S., 2013. Drug withdrawals and the
utilization of therapeutic substitutes: the case of Vioxx. J. Econ.
Behav. Organ. 86, 148–168.
References Coscielli, A., Shum, M., 2004. An empirical model of learning and patient
spillovers in new drug entry. J. Econometr. 122,
213–246.
Berez, J., David, G., Howard, D.H., Neuman, M.D., 2018. Does bad news Crawford, G.S., Shum, M., 2005. Uncertainty and learning in
travel faster? On the determinants of medical technology pharmaceutical demand. Econometrica 73, 1137–1173.
abandonment. J. Human Capital 12, 569–603.
18
Currie, J., Macleod, W.B., 2017. Diagnosing expertise: human capital, injuries for 195 countries and territories, 1990–2017: a systematic
decision making, and performance among physicians. J. Labor Econ. analysis for the Global Burden of Disease Study 2017. Lancet 392,
35, 1–43. 1789–1858.
Currie, J.M., Macleod, W.B., 2020. Understandin doctor decision making: Janakiraman, R., Sismeiro, C., Dutta, S., 2009. Perception spillovers across
the case of depression treatment. Econometrica 88, 847–878. competing brands: a disaggregate model of how and when. J.
Cutler, D., Skinner, J., Stern, A., Wennberg, D., 2019. Physician beliefs and Market. Res. 46, 467–481.
patient preferences: a new look at regional variation in health care Kessler, R.C., Berglund, P., Demler, O., Jin, R., Koretz, D., Merikangas, K.R.,
spending. Am. Econ. J.: Econ. Policy 11, 192–221. Rush, A.J., Walters, E.E., Wang, P.S., 2003. The epidemiology of major
Dickstein, M.J., 2016. Physician vs. Patient Incentives in Prescription depressive disorder: results from the National Comorbidity Survey
Drug Choice, Working Paper NYU. Replication (NCS-R). J. Am. Med. Assoc. 289, 3095–3105.
Dickstein, M.J., 2018. Efficient Provision of Experience Goods., pp. 1–58, Ludwig, J., Marcotte, D.E., Norberg, K., 2009. Anti-depressants and
Working Paper NYU. suicide. J. Health Econ. 28, 659–676.
Gibbons, R.D., Brown, C.H., Hur, K., Marcus, S.M., Bhaumik, D.K., Erkens, McFadden, D., Train, K., 2000. Mixed MNL models for discrete response. J.
J.A., Herings, R.M., Mann, J.J., 2007. Early evidence on the effects of Appl. Econometr. 15, 447–470.
regulators’ suicidality warnings on SSRI prescriptions and suicide in Mcguire, T.G., 2000. Physician agency. In: Culyer, A.J., Newhouse, J.P.
children and adolescents. Am. J. Psychiatry 164, 1356–1363. (Eds.), Handbook of Health Economics. Vol. 1, Chapter 9 of Handbook
Gibbons, R.D., Hur, K., Bhaumik, D.K., Mann, J.J., 2006. The relationship of Health Economics. Elsevier, pp. 461–536.
between antidepressant prescription rates and rate of early Rodwin, V.G., 2003. The health care system under French national health
adolescent suicide. Am. J. Psychiatry 163, 1898–1904. insurance: lessons for health reform in the United States. Am. J.
Howard, D.H., David, G., Hockenberry, J., 2017. Selective hearing: Public Health 93, 31–37.
physician-ownership and physicians’ response to new evidence. J. Stern, S., Trajtenberg, M., 1998. Empirical implications of physician
Econ. Manag. Strategy 26, 152–168. authority in pharmaceutical decisionmaking, NBER Working Paper
Howard, D.H., Hockenberry, J., 2019. Physician age and the abandonment 6851.
of episiotomy. Health Serv. Res. 54, 650–657. Tunçel, T., 2021. Should off-label drug prescriptions be prevented?
James, S.L., Abate, D., Abate, K.H., Abay, S.M., Abbafati, C., Abbasi, N., Empirical evidence from France, Working Paper, HEC Montreal.
Abbastabar, 2018. Global, regional, and national incidence,
prevalence, and years lived with disability for 354 diseases and
19

Journal of Health Economics: Pierre Dubois, Tuba Tunc El

Uploaded by

Copyright:

Available Formats

Journal of Health Economics: Pierre Dubois, Tuba Tunc El

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Journal of Health Economics: Pierre Dubois, Tuba Tunc El

Uploaded by

Copyright:

Available Formats

Journal of Health Economics 78 (2021) 102461

Contents lists available at ScienceDirect

Journal of Health Economics

Identifying the effects of scientiﬁc information and

Understanding physicians’ prescribing behavior is

SSRIs 0.50 0.50 0.58 0.52 0.43

Quantiles Quantiles Quantiles Quantiles Quantiles

Variables (1) (2) (3) (4)

Kids–adolescents −0.062 −0.089* −0.058 −0.084

Observations 517,241 517,241 379,687 379,687

with depression. We assume that each physician i receives

heterogeneous distributions of health that is possibly cor-

0 , ω1 , . . ., ω0 , ω1 are independent across alternatives

and after identifying the parameters ˛0di , ˛1di , di0

Drugs Patient’s Baseline Warning

Age Gender Mean SD Mean SD Correlation

SSRIs −0.07 −0.47 4.04 3.21 −1.86 8.26 0.93

Across physicians Within-physician

Mean Standard deviation Standard deviation

on the variance term that depends on the distribution of

With prewarning preferences With postwarning preferences Warning only

No ban With ban Change No ban With ban Change Change

SSRIs 0.517 0.000 −0.517 0.452 0.000 −0.452 −0.065

Patients variables All Kids/adolescents Adults

Age Gender Age Gender Age Gender

Warning 1.699*** −0.001 −0.096 0.000 1.639*** −0.001

Drugs Baseline parameters Quantiles

SSRIs Mean ˛1di −0.91 −0.28 0.47

Variables (1) (2) (3) (4) (5)

Female physician −0.0262 0.0227** −0.0003 −0.0092 0.0155

Observations 338 338 338 338 338

Variables (1) (2) (3) (4) (5)

Female physician −0.0203 0.0079 0.0054 −0.0065 0.0143

Observations 338 337 332 337 330

Notes: Signiﬁcance levels: *** p < 0.01, * p < 0.1.

Variables (1) (2) (3) (4) (5)

Female physician 0.0061 −0.0129 0.0209 0.0106 −0.0124

Observations 338 337 332 337 330

Notes: Signiﬁcance levels: ** p < 0.05.

You might also like

and after identifying the parameters ˛0di , ˛1di , di0

Warning 1.699* −0.001 −0.096 0.000 1.639* −0.001