## Pareto.uab.es

Research and the Approval Process
An agent sequentially collects information to obtain a principal's approval, such as a
pharmaceutical company seeking FDA approval to introduce a new drug. To capture suchenvironments, we study strategic versions of the optimal stopping time problem …rst proposedby Wald (1945). Our ‡exible model allows us to consider di¤erent types of rules and commit-ments by the principal as well as strategic withholding of information by the agent. We shedlight on current regulation and proposed reforms of the drug approval process. The modelalso captures situations such as a …rm seeking antitrust approval to merge with a competitor,a manager proposing a project to the …rm's headquarters or an author submitting a paper toan editor.

Keywords: Research, organization, approval, regulation.

JEL Classi…cation: D83 (Search; Learning; Information and Knowledge; Communication;Belief), M38 (Government Policy and Regulation).

Ottaviani acknowledges …nancial support from the European Research Council through ERC Advanced Grant
295835. We thank Umberto Sagliaschi for excellent research assitance.

ySciences Po and CEPRzBocconi and CEPR
Pharmaceutical companies run costly clinical trials on new drugs to demonstrate the safety and
e¤ectiveness necessary to obtain regulatory approval. Similarly, a …rm intending to take over
a competitor searches for evidence of synergies to convince the antitrust authority to approve
the transaction. A manager collects information to push the …rm headquarters to validate his
project. And an author submitting a paper conducts research and robustness checks to convince
the editor to accept. In turn, the regulator, the headquarters or the editor can can also conduct
additional independent research on these issues.

In all these situations an agent searches sequentially for evidence to convince a principal, with
a priori di¤erent preferences, of the desirability of an activity with uncertain private and social
payo¤s; the principal, in turn, also conducts some research. How does the organization of the
approval process a¤ect the information that gets produced and the quality of the …nal decision?
How does the possibility for the agent to withhold information a¤ect the process?
Consider our leading application to the drug approval process. After identifying a promising
compound, pharmaceutical companies conduct an extensive and well de…ned series of clinical
trials to obtain the approval of the regulator in charge of drug safety.1 Pharmaceutical research
is conducted sequentially, so that at each point in time one of three decisions is made: continue
research by acquiring additional information, abandon the project altogether, or ask for approval
for introducing the drug to market.2
This corresponds to the current organization of the drug approval process, in particular in
the US. This process has evolved greatly over time and new issues are currently emerging (as
described in detail in section 2). The historical evolution of legislation on drug approval has
tended to strengthen the powers of the FDA to mandate research and created opportunities for
the agency to commit to standards and to a precise process of approval. In this paper, we will
study the bene…ts from these di¤erent types of commitment. Currently the attention has turned
to both the issue of withholding of information, following some heavily publicized scandals,3 and
1 In the US the Food and Drugs Administration (FDA) regulates the approval of new drugs. In the European
Union, pharmaceutical companies can choose between applying to a EU-wide authority, the European MedicinesAgency (EMA), or to one of the national authorities, such as the Medicines and Healthcare products RegulatoryAgency (MHRA) in the UK or the Agenzia Italiana del Farmaco (AIFA) in Italy.

2 Obtaining additional evidence can mean conducting additional clinical trials or adding patients to a trial. As
reported in Nundy and Gulhati (2005), increasingly Western drug companies conduct trials in India to decreasecosts and bene…t from easier regulatory approval. Moreover, several scandals involve illegal trials not approved bythe Indian authorities.

3 Scandals include the allegation that for several years Merck withheld evidence on adverse e¤ects of its block-
buster drug Vioxx. There has been a recent push to impose stronger disclosure requirements. For example,
the question of post-approval regulation. Our theoretical analysis will shed light on these new
concerns and on the necessary regulatory steps.

Our model captures these types of environments in a simple, tractable and ‡exible way. A
choice needs to be made between rejection and approval. Rejection yields a zero payo¤ while
the bene…ts from approval depend on a binary state of nature; they are positive if the state is
high and negative if it is low. Research can be conducted to obtain additional information about
the state. It is costly on two accounts: there is a direct …nancial cost and, in addition, research
delays decision making with an associated opportunity cost. The arrival of new information is
conveniently modeled in continuous time as a Wiener process with a drift that depends on the
If the same player were in full control of both research and approval, the model would boil
down to a version of the classic single-agent optimal stopping problem that has been widely
analyzed in the statistical decision theory literature on sequential analysis pioneered by Wald
(1945). The well-known solution involves two threshold values (or standards) for the belief, such
that it is optimal to abandon the project when the belief that the state is high is su¢ ciently low
(below the rejection standard) and such that it is optimal to adopt the project when the belief
is su¢ ciently high (above the approval standard). When the belief lies within these thresholds,
it is optimal to continue researching— this is Wald's celebrated sequential probability ratio test.

The payo¤s of agent and principal are typically misaligned. For instance, because the phar-
maceutical …rm does not internalize all externalities, it typically gets a higher payo¤ than the
regulator in the bad state. In most of the paper we thus focus on situations where, for the
stand-alone Wald Problem, the agent searches more than the principal at the lower end, when
the state is bad, and less at the upper end, when the state is good. It is the case most people
have in mind for drug approval: pharmaceutical companies are eager to stop researching, obtain
approval, and adopt the drug earlier than the regulator when information is good and more
reluctant to abandon the project when the state is bad.

In practice, the research and approval processes are not single agent problems; intuitively the
agent controls the research decision, at least initially, while the principal has the hold on the
approval decision. The baseline situation we thus study is one where the agent chooses the lower
benchmark of search and the principal simultaneously chooses the upper benchmark (the Nash
medical journals and regulatory authorities have pushed for early registration of trials and for disclosure of theresults obtained in the trials.

Equilibrium).4 In equilibrium, we show that the principal will compromise at a lower standard
for approval compared to the principal's non-strategic standard. Intuitively, given that the agent
now chooses when to abandon research, the principal's value of information is reduced— thus the
principal becomes more eager to approve.

Two properties of this Nash equilibrium solution stand out. First, the principal can obtain a
negative payo¤ in equilibrium. Indeed he controls the upper benchmark, and can force immediate
approval but cannot force rejection. If the agent has su¢ ciently misaligned preferences and
searches too much at the lower end, the principal at the upper benchmark gets a negative expected
bene…t. The goal of research in such situations is that it will provide su¢ ciently bad news that
the lower benchmark will be reached and the agent can be convinced to abandon. Second, more
research is conducted in the Nash equilibrium solution than in either stand-alone problems.

The principal thus has an interest to commit to a course of action that discourages research
by the agent at the lower end. We …rst consider commitments ex ante to an approval standard,
what we call the Stackelberg outcome. The principal chooses his preferred point on the agent's
best response curve. Depending on the value of the initial belief, di¤erent types of commitments
are optimal. When the belief is low, a high blocking commitment that discourages research by
the agent is preferred: this allows the principal to avoid the negative payo¤ that sometimes
characterizes the nash outcome. For intermediate values of the belief, the principal will commit
to an interior commitment, allowing for some research by the agent. Finally, when the initial
belief is very favorable, the principal will commit to immediate approval.

In practice, this type of commitment is not easy to achieve and furthermore, as suggested by
the above discussion, cannot be uniform and independent of the current state of knowledge. In
fact, as described in section 2, this is not the approach chosen for drug regulation. Rather, there
is a commitment to a well de…ned sequential procedure for approval. We thus study a model
where the interaction occurs in two stages as follows:
1. In the …rst stage the agent conducts research, and then decides when to transfer the decision-
making power to the principal.

2. Once that happen, in the second stage, the principal conducts research and eventually
decides whether to approve.

4 This is in fact the outcome of a game where the principal can mandate research: in each period the agent
chooses between three actions, research, submit or wait and if the agent chooses submit, the principal choosesbetween research, approve and wait.

Clearly, the solution in the second stage corresponds to the non-strategic solution for the principal.

Expecting that outcome, in the …rst stage the agent has less incentive to undertake research at
the bottom because of the extra research conducted at the top in comparison to the baseline
Nash equilibrium. Our analysis shows that this type of commitment is not necessarily optimal
from the point of view of the principal. Whereas in the Nash Equilibrium outcome, the issue was
excessive research by the agent at the bottom, the issue is here one of excessive research at the
top. This suggests that regulation of drug approval should be reoriented towards a commitment
Recent regulation of the drug approval process has focused on the issue of withholding of
information. As recent scandals suggest, the agent can withhold some of the evidence, but this
is nevertheless costly; Merck for instance in the Vioxx case has paid over 4.8 billion dollars for
settling complaints. We enrich the model and suppose that the agent can report any belief, but
is then subject to an expected penalty when the state turns out to be low. The expected penalty
increases in the distance between the actual state of knowledge and the report (the case is harder
to defend in court when the lie is big).

Even though the agent will always lie in equilibrium, by exploiting the knowledge of the
bounds of the agent's research interval, the principal is able to perfectly invert the information
and not be deceived. In fact, we demonstrate that the principal will actually bene…t from the
agent's ability to costly misreport information. The reason is that the cost of lying decreases
the value of information for the agent, leading to a reduction of research at the lower end,
something that is bene…cial for the principal. We show that it is optimal for the principal to
choose a penalty for misrepresentation that is not in…nite so that the agent carries out some
costly misrepresentation in equilibrium.

As initially suggested, our model is rich enough to cover other cases of research and approval
than our main application to the pharmaceutical market. We derive some further results in
this case, varying some assumptions to better …t the applications. We …rst depart from the
assumption that the agent cares about the cost of the principal's research (i.e. has to pay for
research) and revisit the two stage commitment game. We show that even though this introduces
a free riding incentive, the agent still has incentive to do research to move the principal away
from his rejection threshold which is ine¢ ciently high from the agent's point of view.

5 There is some discussion of relevant standards in the literature. For instance, Ocana and Tannock (2011) argue
that, even though the FDA has tended to accept any trial showing statistically signi…cant results, they shouldbecome stricter and request "clinically important di¤erences", i.e., statistical di¤erences large enough to make itworth running the risk of introducing a new drug.

Related Literature.

The original problem of sequential research, examining the tradeo¤ be-
tween the cost of an extra signal and the bene…t of a more informed decision, was introduced by
Wald (1945, 1947) and Wald and Wolfowitz (1948).6 The ensuing applied probability literature
of this non-strategic problem has a large impact on the actual design of clinical trials. Closely
building on Wald's decision-theoretic foundational framework, we focus on the strategic issues
that arise when the decisions to collect information and to make the …nal decision are made by
two di¤erent players.

Our paper thus relates to the literature on strategic experimentation (see Bolton and Harris
1999) and especially to Strulovici (2010), who highlights how the loss of control of decision making
(determined through voting in his model) reduces the incentives to acquire information and thus
induces a status quo bias; see also Fernandez and Rodrik (1991). Our model is closest to Gul and
Pesendorfer (2012), Lizzeri and Yariv (2011), and Chan and Suen (2012) who consider strategic
settings in which public information arrives over time to voters. In Gul and Pesendorfer's (2012)
model information is provided by the party that leads, whereas in Lizzeri and Yariv (2011)
and Chan and Suen (2012) voters decide collectively themselves when to stop acquiring public
information and reach a decision. In their setting information is revealed publicly to all voters,
while we focus on the sequential interaction between an agent who collects private information
and then reports (or possibly misreports) it to a principal who makes the approval decision. We
also analyze the commitment solution in which the principal moves …rst by setting the approval
standard, and then extend the model to analyze approval in multiple stages.

For our baseline analysis we constrain reporting of the belief (corresponding to the …nal re-
sults) to be truthful at the moment of application, for example because misrepresentation is
in…nitely costly as in the disclosure models of Grossman (1981) and Milgrom (1981). We also
consider the possibility of costly misreporting. While Kartik, Ottaviani and Squintani (2007) and
Kartik (2009) characterize the amount of equilibrium costly lying in static models of strategic
communication, in our dynamic model we show that ex post lying costs reduce the ex ante incen-
tives for information collection. See also Shavell (1994), Henry (2009), and Dahma, GonzÃ¡lez,
and Porteiro (2009) for strategic analysis of partial disclosure of research results. In Henry (2009),
pharmaceutical …rms are worse o¤ when their research e¤orts are not observed by the regulator
as they are forced to do additional tests to convince him they are not hiding any evidence. Our
6 Moscarini and Smith (2001) recently advanced this literature on non-strategic sequential analysis by analyzing
a continuous-time model in which the decision maker can the vary number of experiments in each period. Ourformulation is also in continuous time, but we focus on the simpler case with one experiment per period.

setup is di¤erent in the sense that information is not veri…able: in fact the possibility of hiding
information here reduces research because of the cost of lying. This turns out to be bene…cial
for the regulator who wants to limit research at the lower end.

Finally, we do not allow our principal to use monetary transfers, in line with the literature
on mechanism design without transferable utility; see HolmstrÃ{m (1977) and Alonso and Ma-
touschek (2008), Armstrong and Vickers (2010), and Taylor and Yildirim (2011). This approach
delivers a number of important insights on the functioning of approval processes that we observe
in a number of practical settings where, by and large, transfers are actually not used. A com-
plementary literature analyzes the problem of optimal incentive provision for innovation, search,
and experimentation where transfers are allowed; recent papers in this area are Manso (2011),
Lewis and Ottaviani (2008), Lewis (2012), Gerardi and Maestri (2012), HÃ{rner and Samuelson
(2012), and Halac, Kartik, and Liu (2012).

Drug Approval Process
We present a brief overview of the drug approval process, past and present, in the US. This
exercise will guide our theoretical analysis: we will both analyze the possible e¤ects of past
regulations and consider potential consequences of current regulatory e¤orts.

The evolution of the legislation on drug approval was a series of reactions to resounding
scandals. Prior to 1938, the role of the US Food and Drug Administration (FDA) was mostly
limited to preventing misleading statements on drug labeling. In 1937, a drug company developed
a liquid preparation that was not tested prior to marketing and contained a poisonous solvent.

The drug killed over a 100 people. In reaction, the 1938 Food, Drug and Cosmetic Act was
swiftly passed. The main feature of this law is that it required that research results be submitted
to obtain approval for the drugs, although the FDA had little power to mandate further research
if the initial evidence was unsatisfactory.

It was an important step: it introduced the New Drug Application (NDA) necessary to obtain
approval, a procedure that still exists today.7 However, the power of the FDA still remained
limited. For instance it had only 60 days to examine the evidence and there was no speci…cation
of the rules for testing.

A new scandal in 1962 highlighted the need for regulation of the process of testing. A hypnotic
7 The NDA had to include "all clinical investigations, a full list of the drug's components and copies of both the
packaging and labelling of the new drug"
known as thalidiomide was discovered in Europe to lead to birth defects. It was not allowed the
US, but several thousands of samples had been sent to US doctors who gave them to patients
without mentioning it was experimental, leading to a number of cases of a¤ected babies. In
reaction, the 1962 Drug Amendments introduced the process of drug testing as we know it. The
main features of the amendments can be summarized as follows:
It put in place a system of pre-clinical testing noti…cation so that regulators could judge
whether it was safe to start testing on humans
It gave more power to the regulator to mandate research. As explained by Junod (2008),
"FDA was given the authority to set standards for every stage of drug testing from labo-
ratory to clinic".

However the law did not set very strict legal standards for approval. The law required that
there be "substantial evidence"that the drug be e¤ective. As pointed out in Junod (2008),
alternative stronger language such as "preponderance of evidence"or "evidence beyond any
reasonable doubt" could have been used.

The current phase of regulation is another example of a reaction to a scandal, this time
involving misreporting of evidence by …rms, in particular information on side e¤ects of drugs. A
case in particular, the allegation that for several years Merck withheld evidence on adverse e¤ects
of its blockbuster drug Vioxx, has led to a recent push to impose stronger disclosure requirements.

The FDA Modernization Act of 1997 created the clinical trial registry ClinicalTrials.gov. The
FDA Amendments Act of 2007 expanded the types of clinical trials needed to be registered and
the amount of details that should be included. Some legislators are trying to push for further
Model and Best Responses
To capture most of the features of the interactions between an agent and a principal, such as a
pharma …rm and the FDA, we consider the following model. A choice needs to be made between
two alternatives, adoption A or rejection R. The bene…ts derived from these alternatives depend
8 For example, medical journals and regulatory authorities have pushed for early registration of trials and for
disclosure of the results obtained in the trials.

on the state of the world ! that can be either high H or low L. The payo¤ for player i in state
j if the choice is k is given by vi . We assume that the payo¤ from rejection is zero for all
players, regardless of the state. Furthermore we assume that, for any player i, accepting a good
project provides positive payo¤s while accepting a bad one provides a negative one: vi
For convenience, we use the following log-likelihood parametrization of beliefs
so that the probability that the state is high is given by e = (1 + e ). All players share a common
prior 0. Given the restrictions we imposed on the payo¤s, if player i is forced to make a decision
, there exists a threshold value (or standard) of the belief , such that A is chosen if
> and R is chosen if
< . That value solves
Research can be conducted to learn the value of the state. The arrival of new information is
modeled as a Wiener process d . The drift is determined by the state. Speci…cally, the process
has positive drift
and variance 2 if the state is H or drift
and variance 2 if the state is
L. Accumulating information over a period of time dt costs cidt, where the cost of collecting
information can vary across individuals.

Finally, payo¤s are discounted, so that if an alternative is chosen at date t it is discounted at
rate ri. There are therefore two costs associated with searching for more information: the direct
…nancial cost and the opportunity cost associated to the delay in the accrual of the decision
Suppose player i undertakes research until time t. The accumulated information at date t is
t. The log-likelihood ratio of observing
in the two states is given by
where h is the density of a standard normal distribution. According to Bayes' rule, the log
posterior probability ratio is equal to the sum of the log prior probability ratio and the log-
likelihood ratio. Thus, the posterior belief at time t is given by
where d 0 is a Wiener process with drift 2 2= 2 if the state is H and
2 2= 2 if the state is L
and instantaneous variance 4 2= 2.

When the same player i makes both the search and approval decisions, for a belief that is
, there may be a bene…t of searching for more information to make a more informed
decision. This is a standard stopping time problem: there exists two values of , s and S (s
< s the player stops researching and rejects;
< S the player conducts research;
> S the player stops researching and approves.

It is immediate to characterize the utility function of the player when
u( ) = e rdtE[u( + d 0)]
Following Stokey (2009, Chapter 5), starting in the intermediate region, we let T be the …rst
time the belief hits either s or S. The direct monetary cost of searching is given by
c e rT . Once we de…ne, as in Stokey (2009):
2 (s; S) is given by
The …rst line corresponds to the case where the state is high and the upper benchmark S is
reached …rst. The second line is the case where the state is low but the upper benchmark is
reached …rst, and so on.

Best Response Analysis
We start by characterizing the best responses of the research problem. Speci…cally, for a given
value of the lower standard s (resp. upper standard S) we characterize the optimal choice of
the upper standard S = BR(s) (resp. lower benchmark s = br(S)). This best response analysis
allows for a better understanding of our problem and will serve as a building block for the next
sections. We start by characterizing br(S). For the moment we drop the subscript i.

Proposition 1 For a given upper benchmark S:
1. The best response br(S) is independent of the current belief
and is such that br(S) = S
2. The best response br(S) is decreasing in vHA, vLA and increasing in c.

3. The length of the research interval l(S) = S
br(S) is increasing in S and converges to a
…nite value l when S converges to +1, where l is solution to vHA = (eg(l)
The …rst result states that there is dynamic consistency in the sense that the best response
is independent of the current belief. It also states that br(S) = S for values S < . This result
is natural, since when S < , approval at S gives a negative value and the player can guarantee
himself a zero payo¤, when
S, by choosing s = S and imposing rejection.

, the …rst-order condition characterizing the best response (derived in the appendix)
bene…t of gaining more information
…nancial cost of research
1(s; S) > 0 and where:
is the expected bene…t from approval when the belief is
At the lower benchmark s, the tradeo¤ expressed by (3) is clear. There are typically two
costs associated with research: …rst the direct …nancial cost, proportional to c=r and second the
cost of delaying the decision. At the lower benchmark of research rejection yields a zero payo¤.

Thus there is no cost of delay and the only cost is the …nancial one. This expected cost has
to be equal to the expected value of information which is proportional to VA(S), i.e., the value
if the upper standard is reached. Overall this gives condition 3. The comparative statics then
naturally follow. Increasing the cost c naturally decreases research. On the other hand increasing
vHA or vLA has the e¤ect of increasing the value of information without a¤ecting the cost and
thus decreases the lower benchmark.

Interestingly, result 1.3 indicates that the length of the research interval, measured by S
br(S) (an indirect measure of the quantity of research), is increasing in S. The intuition is the
following: for a given length l between s and S, the expected bene…t VA(S) is higher for large
values of S and the expected cost of moving from s to S is lower since there is a higher probability
that the drift will be positive. Thus, if for a certain value of S, the optimal choice is a length
br(S) of the research interval, for higher values of S, the interval will be larger. This
property is visible in Figure 1 where we plot both br(S) and BR(S). As indicated in result 1.3, at
the limit, when S goes to in…nity, the value of l(S) converges to l such that vHA = (eg(l)
This limit value depends only on c=r and vHA: when S goes to in…nity, the player is sure the
state is good and at the lower benchmark br(S) he is indi¤erent between stopping immediately
and getting a zero payo¤ or incurring the cost of research and obtaining vHA when the upper
benchmark is reached.

At the upper benchmark, the tradeo¤ is more intricate since the cost of research now has the
two components mentioned above: the direct cost and the cost of delaying the decision. Overall
Proposition 2 For a given lower benchmark s
1. the best response BR(s) is independent of the current belief
2. the best response BR(s) is decreasing in vHA, vLA and c.

3. the length of the research interval l(s) = BR(s)
s is decreasing in s and converges to a
…nite value l when s converges to
1, where l is solution to
vLA = (e g(l) + 1) cr
The …rst order condition characterizing the best response to a given value of s can be expressed
in the following way (where
2(s; S) > 0 and
bene…t of information
cost / bene…t of delaying decision
…nancial cost of research
For the interpretation of these conditions we distinguish between the case where VA(S)
(i.e S > ) and the case VA(S) < 0. Consider the …rst case, that occurs when s is not too low.

At the upper benchmark S the cost of research is composed of the direct …nancial cost and of
the cost of delaying the decision, which is proportional to VA(S). Information has value since
it can lead to avoiding the negative payo¤ vLA if the state is in fact low. Condition (4) re‡ects
this tradeo¤ between cost and value of information. In the second case, when s is very low, the
tradeo¤ is di¤erent: for these values, it will be too costly in terms of expected cost of research, to
choose a value of S > p. At the upper benchmark S, the player will thus incur a loss VA(S) < 0.

Research then has value to try since it can allow to reach the lower benchmark where a zero
payo¤ can be obtained. Thus, in these cases, at the upper benchmark S the loss VA(S) has to
be equal to the expected cost of research needed to reach the lower benchmark.

Finally, result 2.3 indicates that the length of the research interval decreases with s. As stated
above, when s is small, the purpose of research at the upper benchmark S is to avoid incurring
the loss VA(S) by performing research to reach s and get a zero payo¤. When s is small, the
loss is particularly large and furthermore, the expected time cost to reach s will be smaller, since
there are more chances that the drift is negative. This property is represented in Figure 1. At
the limit, l converges to a value l that depends only on
vLA and c=r. The intuition is similar
to that of 2.3. The player is sure that the state is bad and at BR(s), he is indi¤erent between
getting the sure loss
vLA and searching in the hope of reaching the lower benchmark, with a
cost proportional to c=r.

These best responses are a natural building block for the rest of our argument. If the same
player was making both the research and approval decisions, his optimal choice (s ; S ) would be
characterized by the intersection of the best response curves. However in practice these decisions
are typically made by di¤erent agents and these strategic interactions are the focus of our paper.

To clarify the exposition of the rest of the paper, we add more structure on how the preferences
of agent and principal are misaligned. We will focus on the leading case where the agent does
not bear the entire social cost of a wrongful adoption, so that the payo¤s of adoption in the low
> vp . For instance, in the application to drug approval the key concern is
that the pharmaceutical company does not fully compensate patients who su¤er from taking an
unsafe drug because of the di¢ culty in identifying these individuals and the company's ability to
shelter from liability (the judgement proofness problem). According to the previous comparative
statics, this implies that the agent prefers to stop earlier at the upper end but to conduct more
research at the lower end.

The comparison in the high state is less obvious. It seems reasonable to think that va
i.e., that the submitter has more at stake than society at large. For instance an author that has
a paper accepted gets the full private bene…ts from that decision, but does not take into account
the negative externality he imposes on other authors. In the case of a private …rm conducting
research, this can be less obvious. Indeed, it is often thought that a …rm cannot capture the full
social bene…t from an innovation; see for instance Bloom, Schankerman, and Van Reenen (2012).

Of course the factor mentioned above, that goes in the other direction, is still present: the …rm
that innovates, in the case of non radical innovations, takes some pro…ts away from the current
market leader, an e¤ect typically not internalized. Our results for the leading case hold provided
the externality associated to adoption in the high state is either negative or positive but lower
than the negative externality associated to adoption in the low state: va
Given this type of con‡ict in preference, the bliss point of the principal (sp; Sp) and of the
agent (sa; Sa) are such that the agent conducts more search at the lower end sa < sp and less
search at the upper end Sa < Sp as represented in Figure 2. One property is notable and will
9 There is an additional factor speci…c to the pharmaceutical industry that can push vaHA above vp . Given
drug users typically do not pay directly but are reimbursed, pharmaceutical companies might be able to obtainprivate bene…ts higher than social bene…ts.

prove useful in the rest of the paper: the bliss point corresponds to the maximum of the upper
best response and the minimum of the lower best response.

Approval Regulation
In practice, in all the applications we have in mind, the same player does not control the full
research process and at the same time make the approval decision. As described in section 2,
in the case of drug regulation, separating these roles was precisely the purpose of the 1938 law,
that introduced an approval requirement before marketing.

A natural way of thinking of the e¤ects of such a relatively weak regulation is that the …rm
conducts research and, at some point submits the evidence to the principal who, based on the
evidence, makes a decision to approve or wait. The 1938 law did not give the principal power to
mandate further research. Thus, in a subgame perfect equilibrium, the principal approves any
drug when the evidence is above p (i.e. any evidence that gives the principal a positive expected
bene…t). With small con‡icts of interest (va
), the research interval is then the …rst best
of the agent, (sa; Sa). With intermediate con‡ict of interest s.t. Ba (ba ( p)) < p, the agent
searches in (ba ( p) ; p). While with large con‡ict of interest such that p > S de…ned as the
lowest S above p such that S = ba (S), agent does not do any research.

The 1962 Amendments gave the further power to the principal to mandate research. We
examine in Section 4.1, how this extra power a¤ects the outcome of the game. We then show
in Section 4.2 that committing ex ante to an approval standard can improve the payo¤ of the
designer. This was not the path chosen by the lawmakers, who instead chose to organize the
regulation as a sequential process that we examine in Section 4.3. Throughout this section we
will maintain the assumption that the regulator cannot misreport the information he obtains, an
assumption we relax in Section 5.

The e¤ect of granting the principal the power to mandate research (as in the 1962 Amendments)
in her interaction with an agent who initiated the research process, is naturally captured by the
following baseline model. In each period t, agent and principal move sequentially. First, the agent
chooses between three actions research Ra, submit Sa or wait/withdraw Wa. Second, if the agent
submits Sa, the principal chooses between research Rp, approve Ap or wait Wp. Research is the
period's outcome if either the agents chooses research Ra or the agent chooses submit Sa and the
principal chooses research Rp (in that sense the principal can mandate research). Approval A is
the period's outcome if the agent chooses Sa and the principal Ap. Finally withdrawal W is the
period's outcome if the agent chooses Wa or the agent chooses to submit Sa and the principal
We assume that the cost of research enters symmetrically in the agent's and principal's utilities
regardless of who conducts the research. In the case of the FDA, the principal can mandate
research but integrates this research cost in her welfare function, since these costs re‡ect the
costs to patients. We show in the appendix that the outcome of all Markov Perfect Equilibria of
the game above, with
as the state variable, correspond to what we call the Nash equilibrium
solution and denote (sN ; SN ). This equilibrium is at the intersection of the best response curve
of the agent to the upper benchmark bra(S) and the best response curve of the principal to the
lower one BRp(s) . In other words, in this setting, the principal controls the upper standard S
while the agent controls the lower standard s. We have:
Proposition 3 There exists a unique Nash equilibrium such that:
1. the principal conducts less search at the upper end and the agent less search at the lower
end: Sa < SN < Sp and sp > sN > sa;
2. the length of the research interval is larger than for the stand-alone problems: SN
The strategic interaction between the agent and the principal of course a¤ects the research
decision compared to the non strategic benchmark. These changes can be decomposed in two
e¤ects. First, there is an e¤ect on the extensive margin: the range of values of
research is conducted changes. Second, there is an e¤ect on the intensive margin: the length of
the research interval is a¤ected.

Result 3.1 above refers to the extensive margin. Compared to the principal's …rst best, in the
Nash solution, more research is conducted at the lower end and less at the upper end. The logic
of the result is clear. Since both the agent and the principal now control only one benchmark,
the value of information is decreased: the principal conducts less research at the upper end and
the agent less research at the lower end then in their respective stand-alone problems. These
ideas are illustrated in Figure 1. The solid lines correspond to the best response S to a given
s and the dotted ones to the best response s to a given S. The equilibrium (sN ; SN ) is at the
Figure 1: Best replies, the Nash equilibrium, the commitment solution, and comparison with theprincipal's and the agent's unconstrained solutions.

intersection of the lower best response curve bra(S) of the agent and the upper best response
curve of the principal BRp(s). The …gure illustrates the fact that SN < Sp and sN > sa. Indeed,
as the lower benchmark s moves away from sp, the principal's best response decreases (Sp being
the maximal value), since the value of information is decreasing.

Result 3.2 above refers to the intensive margin. Surprisingly, introducing strategic interactions
increases the intensity of research (the length of the research interval is increased). This runs
contrary to the classical intuition that tends to …nd the opposite e¤ect (for instance Strulovici
2010). To understand this result, consider the agent's problem. The principal conducts more
research at the upper end than the agent would like him to do (SN > Sa). Thus, when the agent
considers the choice of the lower benchmark, expressed in equation (3), his incentives to search
are higher: if he reaches the upper benchmark, he gets a higher value (VA(SN ) > VA(Sa)) and
moreover he reaches it faster in expectation since the belief that the state is H is higher for larger
values of S. These two e¤ects (underlying the result of Proposition 1.3) imply that the search
interval is larger SN
We now discuss the payo¤ of the principal and the agent in the Nash Equilibrium solution.

These values are plotted as a function of
in Figure 1. The key message is that the utility of the
principal can be negative at the Nash equilibrium solution. Consider a belief sN <
1 0 A similar logic leads to the result SN
principal was picking both benchmarks alone, he would not be able to obtain a positive utility
and he would thus choose sp > , in other words immediate rejection, to guarantee himself a
zero payo¤. In the Nash equilibrium solution, this is not an option since the agent controls the
lower payo¤ and actually sets it below
Proposition 4 At the unique Nash equilibrium:
the agent gets a positive payo¤ for all values of
the principal gets a negative payo¤ for
2 [sN ; brp(SN )],
In practice, it seems natural that the principal would try to achieve a higher payo¤ by
committing ex ante to a certain behavior. The most natural form of commitment, that we
consider in the next section, would be to commit ex ante to a certain standard of approval.

As highlighted in section 2, this was not the approach chosen in the 1962 Amendments, who
chose rather weak legal language in terms of standards. The legislator chose rather to commit to
perform the evaluation in a prede…ned number of rounds, something we consider in section 4.3.

Commitment: Stackelberg Solution
We study in this section the case where the principal has the ability to commit to an approval
standard that depends only on the current state of knowledge (and not on the path or time
taken to get there). Clearly, if the principal could commit to an approval rule that could be
conditioned on the entire path, the principal would be able to obtain the unconstrained optimal
solution sp; Sp . Such commitment, however, might be di¢ cult to achieve in practice so we
consider a simpler and more realistic commitment to approval rules that depend only on the
current state of knowledge
with the following cuto¤ form: approve if and only if
is the type of commitment, although weak, that was introduced in the 1962 law with the terms
We now characterize the path of the commitment solution; see Figure 4.2. Dynamic con-
sistency no longer applies: the optimal choice of commitment by the principal depends on the
as described in the following result. We use the notation bra for the upper inverse
function constructed by inverting bra for S > Sa.

Proposition 5 In the Stackelberg equilibrium with commitment by the principal, there exist be-
liefs 2 (sp; bp) such that, if the con‡ict of interest is not too large, i.e SN > bp:
1. If the initial belief
< , the principal chooses a blocking commitment
inducing no research:
sa, any commitment above
is part of an equilibrium: Sc( ) 2 ( ; +1);
< , any commitment above bra ( ) is part of an equilibrium: Sc( ) 2
(bra ( ); +1).

2 ( ; SN ), the principal chooses an interior commitment Sc( ) decreasing in .

There is a discontinuity in commitment at : Sc( ) < bra ( ).

> SN , the principal chooses an approval commitment Sc( )
If the con‡ict of interest is large (SN
bp, the principal chooses a blocking
commitment and if
> bp, he chooses immediate approval.

For low values of the initial belief
, the optimal commitment is what we call a blocking
commitment: the principal commits to an upper benchmark that induces the agent to do no
sa, the initial belief is so low that even the agent would not want to do any
research, regardless of the commitment. For
slightly higher, the commitment to be blocking,
i.e induce the agent to do no research, needs to be above bra ( ) (by de…nition of bra(S)). Note
that this minimum blocking commitment is an increasing function of
When the initial belief
, starts to be su¢ ciently favorable, blocking research by the agent
by committing to rejection becomes too costly and there is a preferable interior commitment.

This happens at belief , which is the belief at which the zero iso-utility curve of the principal
is tangent to the lower best response curve of the agent: if the principal chose his preferred
point on the best response curve of the agent bra ( ), he would get a zero utility and he is thus
indi¤erent. If the belief is above that value, an interior commitment is strictly preferable. At
this point there is a discontinuity in the commitment: there is a discrete downwards jump from
a blocking commitment to the optimal interior commitment.

Proposition 5, indicates that the value is above sp. Intuitively, for beliefs lower than sp,
the principal cannot obtain a positive utility even when in full control; a fortiori, the lowest
at which the principal can only obtain a zero utility when the agent controls the lower standard
must be above sp. [In addition, at
= brp(SN ), the principal obtains a zero payo¤ at the
Nash equilibrium solution and could do strictly better by committing to a di¤erent point on the
agent's lower best response curve. The belief for which the principal is indi¤erent between interior
commitment and the blocking commitment (yielding zero payo¤) thus has to occur earlier.]
The fact that interior commitment is strictly above SN for
2 ( ; SN ) re‡ects the tradeo¤
between two e¤ects:
A second-order negative direct e¤ ect: Holding …xed the agent's strategy s, an excessive
amount of research is induced at the upper end, which induces a loss for the principal. This
loss is clearly second order by the envelope theorem because we start from the principal's
optimal choice of S holding …xed the agent's choice of s.

A …rst-order positive strategic e¤ ect: The agent's strategic response of the increase in S is
to increase s given that the agent's best reply is upward sloping in the relevant range—
strategies are strategic complements in the terminology of Bulow, Geanakoplos, and Klem-
perer (1985). Intuitively, the increased loss of control at the upper end further reduces the
agent's value of information at the lower end. Given that the agent's choice of s at the lower
end was originally lower than the principal would have liked, this increase in s bene…ts the
principal. This is …rst-order e¤ect because the envelope theorem does not apply given that
the agent, not the principal, chooses s.

As we show, an increase in
increases the direct e¤ect and reduces the strategic e¤ect. Thus,
as the optimal interior commitment decreases in
= SN the strategic e¤ect becomes
zero, and commitment has e negative value for
> SN . When the initial belief
the optimal choice for the principal is to chose immediate approval, what we call an approval
commitment. This occurs for beliefs below Sp (the point where the principal would chose
immediate approval if in full control), since the principal cannot control the upper benchmark.

In fact, as indicated in Proposition 5, for beliefs above SN immediate approval is optimal.

In terms of welfare, it is clear that the principal can always do weakly better by committing.

In fact, the following proposition indicates that she does strictly better whenever the initial belief
is in (sa; SN ). This result naturally follows from the previous discussion. First, the blocking com-
mitment is chosen for beliefs where the Nash equilibrium gives negative utility to the principal,
so that commitment is strictly preferable. Second, when an interior commitment is chosen, the
principal's optimal commitment is strictly above SN indicating that a better commitment exists.

Proposition 6 In the Stackelberg equilibrium, the principal's payo¤ is:
weakly higher than the payo¤ in the Nash equilibrium for all values of
strictly higher for
2 (sa; SN ).

The …nal essential consideration is the e¤ect of commitment on both the extensive and in-
tensive margins of research discussed in the previous sections.

Proposition 7 In the Stackelberg equilibrium:
The extensive margin of research is decreased compared to the outcome in the stand-alone
principal solution: ( ; SN )
The intensive margin of research is increased compared to both the Nash Equilibrium and
the stand-alone principal outcomes: for
2 ( ; SN ), Sc bra(Sc) > max(SN sN ); Sp sp).

Proposition 7 suggests that, in the Stackelberg solution, there are less instances where some
research is conducted, but for the values where this is the case, more research will be performed.

There are two instances where no research is performed in the Stackelberg solution: for low values
, the principal chooses a blocking commitment and for high values he induces immediate
approval. The …rst result is then due to the fact that the principal, since he does not control the
lower benchmark, wants to prevent research for more values of
. The second result echoes the
result on the extensive margin for the Nash outcome: for beliefs such that research is conducted,
more research will be performed.

Commitment: Sequential Research
In practice a commitment is not always easy to achieve. In fact, as stated earlier, the 1962 law
explicitly chose not to commit to a strict standard. As suggested in the previous section, one of
the main reason could be that the optimal level of commitment is speci…c to the baseline state
of knowledge, which can vary across types of drugs, and thus there is no uniform standard that
can be applied. Of course, the close interaction between the …rm and the regulator could allow
for individualized commitments, but even those are not easy to credibly make. A di¤erent way
of committing is in the way research is organized. Often the communication between the agent
and the principal is organized in a number of rounds. This was the approach chosen by the 1962
law that organized the interaction between the …rm and the regulator in a well de…ned series of
clinical trials.

We consider in this section a model where the interaction is organized in two rounds of
inde…nite length. First the agent conducts research and at some point decides to transfer this
information to the principal. The principal then decides how much additional research to perform
before making the approval decision.11 We maintain the assumption that the agent cares about
the research cost of the principal. This assumption is sensible in the application to the FDA that
can mandate research. It is less so for other applications and we consider those in section 7. We
denote (sseq; Sseq) the choice of the agent, where seq stands for sequential.

When the agent submits the information to the principal, the principal performs research and
makes the approval decision as in the stand-alone case since there will be no more interaction
with the agent: the principal will then perform research if
2 (sp; Sp). Thus the agent will never
want submit before Sp is reached because the agent bears the full research cost regardless of
who performs the research but would lose from submitting to the principal before Sp is reached
because then principal would carry too little research at the lower end in the eyes of the agent.

Furthermore, at the lower end the agent chooses the best response to Sp: sseq = bra(Sp). These
results are summarized in the following proposition:
Proposition 8 In the sequential problem:
1. In equilibrium the agent conducts research whenever
is in (bra(Sp); Sp), abandons for
lower values and submits the evidence for higher ones
1 1 In most applications, there could be additional round but we will focus on the one round case without loss of
generality of the message.

2 (sN ; ), the principal obtains a higher payo¤ in the sequential than in the Nash
2 ( ; Sp), the principal obtains a higher payo¤ in the Nash equlibrium than in the
4. The length of the research interval is larger in the sequential than in the Nash or in the
stand-alone problems
The Nash equilibrium, the Stackelberg commitment solution considered in section 4.2 and
the sequential research procedure correspond each to a di¤erent point on the best response curve
of the agent, bra(S). Clearly, the Stackelberg point results in highest expected payo¤ for the
principal. Results 8.2 and 8.3, indicate that the comparison between the Nash and sequential
outcomes is potentially ambiguous. This suggests that the type of commitment put in place
through the 1962 law was not necessarily welfare enhancing, in particular in cases where the
initial belief is quite favorable as in 8.3.

Does the principal prefer the Nash equilibrium outcome or the sequential research outcome?
Compared to the Nash equilibrium, the sequential procedure results in more research at the top
and less research at the bottom. The principal is not necessarily better o¤, as illustrated in Figure
2. On the one hand, the principal bene…ts from reduction of s and increase in S along p's BRp(s);
the movement toward North-East along BRp (s) from (sN ; SN ) to (bra (Sp) ; BRp (bra (Sp))) in-
creases the principal's expected payo¤. On the other hand, the principal loses for additional
increase in S; the upward movement from (bra (Sp) ; BRp (bra (Sp))) to (bra (Sp) ; Sp) results in
a reduction in the principal's expected payo¤. The dashed indi¤erence curve corresponds to
a setting in which the principal prefers the sequential solution to the Nash equilibrium; the
opposite ranking holds with the continuous indi¤erence curve. As can be seen graphically, a
su¢ cient condition for the principal to prefer the sequential solution to be Nash equilibrium is
that SC > Sp = Sseq; otherwise the ranking is ambiguous.

In the region where the blocking commitment is the optimal Stackelberg commitment (
(sN ; )), the principal prefers sequential commitment to the Nash outcome because sequential
commitment results in the same outcome as the Stackelberg commitment in some cases and limits
the amount of research performed in others. When instead
2 ( ; Sp), the optimal commitment
is to approve immediately and this is also the outcome with the Nash solution, whereas additional
research is performed under the sequential commitment thus leading to a lower payo¤ for the
Figure 2: Welfare comparison between the Nash solution and sequential research.

Misrepresentation of Information
The new phase of regulation has started focusing on the regulation of the disclosure of clinical trial
results. The alleged withholding of negative results by pharmaceutical companies in the recent
cases of Vioxx (an anti-in‡ammatory drug proven to increase the risk of cardiovascular events) or
Paxil (an anti-depressant that could increase the suicide rates among children) generated major
uproar and large demands for compensation.

Withholding information has a potential cost for the …rm itself: in the case of Vioxx, Merck
paid over 4.85 billion dollars for settling individual complaints from patients. In 2011, it agreed
to plead guilty and pay 950 million to the federal government to settle the criminal and civil
charges …led against it. These costs seem to be an increasing function of the size of the misrep-
resentation. It is because Merck was shown to have withheld evidence that the penalties were of
that magnitude. A larger lie makes it easier for the plainti¤s to win their case in court.

To capture such situations, we enrich the model and assume that the agent who collects the
evidence can misreport it at no cost.12 However, if the state turns out to be low the agent expects
a …ne F . The probability of obtaining this sentence depends on the size of the misreporting. If
the agent has collected evidence showing that the state is
and conveys information
probability of being convicted is given by P (
0) and we denote the overall expected sentence
0) where C (i.e., P ) is increasing and C(0) = C0(0) = 0.

1 2 In our model, evidence comes in in…nitesimal amounts so that the agent can for any state, always hide a
su¢ cient quantity of negative results to be able to present veri…able evidence consistent with that state.

We concentrate on the commitment case as in Section 4.2 where the principal can commit to
an approval standard, that we denote SM (M stands for withholding). The rule is thus: "approve
if and only if the reported evidence is above SM ". In this setting, the agent is in full control
of both extremities of the search interval, but the agent expects a lower value when deciding to
< SM provided that the state turns our to be low.

Two limit cases are of special interest. When F = +1 we are back to the commitment
case of section 4.2 where truthful reporting was assumed, (sC; SC). When F = 0, the agent's
stand-alone solution of Section 3 results, (sa; Sa). However nothing refrains the principal from
choosing intermediate values of F . He now has two instruments available that are indeed used
in practice: the …ne F imposed for misreporting and the standard for acceptance SM .

In the context where F takes intermediate values, the cost of lying has two e¤ects on research:
The cost of lying weakly decreases the payo¤ from approval and thus a¤ects research at
The cost of lying creates an additional incentive to conduct more research at the upper
end, to accumulate further positive evidence in order to decrease the expected …ne.

We show in the following results that this modi…ed research problem has a unique solution
(sM ; SM ), which are the lower and upper benchmark of the search interval chosen by the agent
when the principal sets a standard for acceptance SM . In other words, there is full information
revelation in equilibrium. The agent will lie but the principal will know exactly the state at
which search was stopped. Sequential search is key for this result— if, instead, the agent was
just observing the state and reporting, di¤erent types would pool on the same report given the
binary decision made by the principal.

Proposition 9 With private information and misreporting, there is a unique solution (sM ; SM )
such that the agent conducts research if and only if
2 (sM ; SM ) and such that, if SM > Sa,
lying occurs in equilibrium, i.e the agent chooses an upper benchmark of search strictly lower than
the standard for acceptance: SM < SM .

Proposition 9 indicates that there will always be lying in equilibrium. The intuition is that
the marginal cost of lying when the report is close to the truth is zero while the added value of
getting more information is strictly negative since the standard set by the principal is greater
than the optimal upper benchmark of the agent (SM > Sa). Thus the agent prefers to stop a bit
earlier and incur the expected lying cost.

We now examine the modi…ed best responses in this case. For the lower benchmark, the best
reply to S is characterized by a condition identical to condition (3),
bene…t of gaining more information
…nancial cost of research
except that the value VA upon stopping is given by
Clearly, for given values of S less than SM the agent will search less at the lower end than
he would in the absence of misreporting since his value in case the state turns out to be bad
will be lower due to the expected penalty. It thus decreases the value of information which is
proportional to VA.

For the best response to s, there is an additional e¤ect: searching more provides evidence
that directly decreases the expected cost in case the state is in fact low. The …rst order condition
bene…t of information
value of evidence
cost of delaying decision
…nancial cost of research
At the upper benchmark three factors push the agent to do more research than in the stand-alone
The value of information (proportional to
S)) increases since there is a
higher cost to avoid in case the state is shown to be low.

The cost of delaying the decision, proportional to VA is smaller.

There is a value of collecting evidence (independent of the value of information), propor-
S), this extra evidence making a court case less likely.

We can now use this best response analysis to compare the welfare of the principal in this case
compared to the commitment case studied in Section 4.2. Note that the commitment case with
no misreporting is equivalent to this model with an in…nite penalty. The question we therefore
address is the following: Should penalties be in fact limited?
In Section 4.2 we characterized the optimal commitment of the principal SC in the case of no
misreporting, i.e., the principal's preferred point on the agent's best response bra(S). In what
follows, we focus on the case of interior commitment, i.e according to Proposition 5, the case
where the initial belief
2 ( ; ).13 The …rst thing to note is that in the case of
limited penalties, the principal can always choose a large value of SM > SC such that the agent
chooses in equilibrium the upper benchmark of search to be SC. Indeed, for any value of s the
best response BRa(s) is an increasing function of SM . So SM can be chosen so as to induce the
agent to do the same amount of research at the upper end as in the commitment case analyzed in
Section 4.2. However in this case with misreporting, the agent will also search less at the lower
end for the reasons outlined above. If this extra research is not excessive, this will be bene…cial
for the principal:
Proposition 10 For
2 ( ; ), the principal can always …nd a combination of instruments
(F; SM ) with F < +1 that results in a strictly higher expected payo¤ than in the commitment
case without misreporting.

Proposition 10 indicates that it is optimal in equilibrium to allow some misreporting. Mis-
reporting induces the agent to conduct less excessive research at the lower end. The intuition is
presented in Figure 9.

1 3 In the case where without misreporting the optimal commitment is a blocking commitment, this can also be
obtained with misreporting.

Post-Approval Regulation
After the approval of the drug by the regulator, the drug is prescribed to patients and further
information is thus accumulated on a large scale. If very bad information is accumulated, the
principal can decide to order a withdrawal of the drug. However, this information accumulation
strongly depends on the monitoring e¤ort of the principal.14 In practice, it appears that the
monitoring e¤ort of the FDA in the United States is rather weak: for instance there is no central
recording system of all adverse e¤ects. The FDA intervenes only in the case where litigation
starts for serious injuries due to the drug.

We thus extend our model to this setting and it further demonstrates the ‡exibility and
applicability of our setup. We assume that while the drug is on the market, the regulator obtains
a ‡ow bene…t w and the …rm a ‡ow pro…t . However, if the state is in fact low, at some point in
time very serious adverse e¤ects will occur. We assume that the players then gets a negative payo¤
Pi (where i 2 a; p) whose arrival time is distributed according to an exponential distribution of
parameter . Furthermore we assume that Pi is large enough so that if the regulator is sure the
state is bad, he will immediately withdraw (condition formally stated below).

Initially, before the start of the game, the principal chooses how closely he will monitor the
drug post approval. Speci…cally, we assume he chooses a costless monitoring e¤ort e.15 Given a
choice of e, post approval, there is a probability
(e) that principal obtains further information.

In this case he chooses a benchmark of withdrawal W . With probability 1
information is collected.

The timing of the game is therefore:
1. The principal chooses e at no cost
2. The agent and the principal simultaneously choose s and S (Nash equilibrium as in Section
3. If the decision is to approve, with probability (e) further information is obtained and the
principal chooses the withdrawal benchmark W
Stage 2 is unchanged compared to the Nash Equilibrium resolution of section 4.1. However,
the addition of stage 3 puts structure on the payo¤s in the di¤erent states (vLA and vHA):
1 4 It also depends on the amount of sales, something that could also give rise to a di¤erent extension of our
1 5 In the case of the FDA, this corresponds to setting up a central system to collect information on adverse
If there is no additional information accumulated post approval (probability 1
agent and the principal have no decision to make, they just collect bene…ts and potentially
su¤er from penalties. If the state is high, the principal obtains w and the agent
state is low, they collect these bene…ts until the adverse e¤ect arises: the principal obtains
p) and the agent obtains r
If additional information is accumulated (probability
(e)), the principal has to choose W
the withdrawal benchmark. In the bad state, bene…ts will be obtained up to the point
where either the belief reaches the withdrawal benchmark W or the adverse shock occurs
leading to withdrawal. In the good state, bene…ts will be continuously obtained unless the
belief reaches the withdrawal benchmark W
Given the choice of W we denote fL(tj ) (resp. fH(tj )) the distribution of the …rst time the
belief reaches the withdrawal benchmark W < S for a belief
. The distribution fL(tj ) is an
inverse Gaussian of parameters (
Conditional on the low state, the expected bene…t of the principal for a choice of W at belief
fL(T j ; W )dT dt
e ( +r)tP [T > t
When the state is high, this is
If the state is low, there are two countervailing e¤ects of an increase in W : it decreases the
chance of getting the penalty but increases the expected ‡ow of bene…ts w obtained. If the
penalty is large enough, the …rst e¤ect dominates and vLA increases in W . If the state is high,
an increase in W unambiguously decreases the expected payo¤ vHA. Thus, if the belief becomes
su¢ ciently negative, the regulator has an incentive to withdraw:
Proposition 11 In the three stage game with post approval, there exists a value P such that, if
1 6 Distribution of the …rst time a Brownian motion hits a particular value
the principal imposes a withdrawal benchmark W >
1 and achieves a strictly higher
payo¤ than when no information is obtained.

the agent's lower best response and the principal's upper best responses are decreasing in e.

When the penalty is high, choosing partial monitoring, i.e. picking a low value of e, is a way
for the principal to decrease the value obtained when the state is bad vLA and thus to commit
to perform more research at the upper end. It also forces the agent to do less at the lower end.

This could explain the behavior of the FDA that chooses not to have a systematic review of post
approval evidence.

Other Applications: Research with Private Costs
Our model can be extended to analyze a wide range of situations involving an agent initiating
research and a principal making approval decisions. As discussed in the introduction, most
relevant are the cases of an author trying to convince an editor, …rms wanting to merge trying
to in‡uence the antitrust authority or a manager trying to push his project at the headquarters.

This section extends the model to address these applications.

Throughout the previous sections, we maintained the assumption that both the principal and
the agent cared about the full cost of the research, even if the other party was performing it. It
was natural in the case of drug approval: the …rms conduct the research (since the regulator can
mandate it) and thus care about the cost and the regulator also cares as a social planner. In the
applications discussed above, instead, the agent does not typically bear the cost of the research
undertaken by the principal.

Of course, it is always possible for the principal to charge the agent for the cost of his research.

For instance the competition authority could charge a fee proportional to the number of sta¤
members it puts on a particular case, or the …rm headquarters could ask the manager to pay a
fee for examination of the case. However, it does not seem to be the approach taken in most
applications. We examine in this section, under what conditions it is optimal for the principal
to charge the agent for the additional research.

We place ourselves in the context of the sequential two stage game considered in section 4.3.

A di¤erent interpretation of the setting considered in that section is that it corresponds to the
case where the agent has to pay for the principal's research cost. We are now going to derive the
other polar case.

The …rst thing to note is that, regardless of whether the agent pays or not for his research
e¤ort, once he transfers the information, the behavior of the principal will be identical: after the
agent's report is received, if the report of the agent is in the interval (sp; Sp), the principal will
conduct additional research using these stopping times (sp and Sp).17
Sseq) the agent's search interval when he does not pay for the principal's
research ((sseq; Sseq) is the analogous interval in the case where he does pay, derived in section
4.3), we see that from the principal's point of view, the welfare bene…ts from charging for the
principal's research, will only depend on the di¤erence between b
sseq and sseq since at the upper end
the …nal outcome will be identical (search will be conducted until Sp). We obtain the following
result that shows that it is not necessarily optimal for the agent to charge for his research:
Proposition 12 It is optimal for the principal to charge for his research if and only if
In the case where the agent pays for the principal's research (case of section 4.3), the agent
searches in the interval (br(Sp); Sp). Indeed, he knows that if he stops searching before Sp,
the principal will keep searching and he will run the risk that bad information arrives and the
principal abandons the search ine¢ ciently early since they do not agree on the lower benchmark
of search. Since he pays in any case for the research, he has no incentive to stop before Sp.

As a reminder, in this case, the lower best response to S is characterized by:
bene…t of gaining more information
…nancial cost of search
In the case where he does not pay, a tradeo¤ emerges: the agent might want to stop earlier
to save on research costs even though he runs the risk that the principal stops early. When
he stops at a belief S < Sp, the agent knows that the principal will conduct additional search
and thus uncertainty remains. If the state is high, rather than expecting a payo¤ of vHA, he
p(S; H )vHA that we denote vHS and if the state is low,
p(S; L)vLA replaces vLA that
we denote vHS, where
p(S; I ) < 1 characterizes the expected time for the principal to reach his
upper benchmark when the state is I.

1 7 Given the result of Proposition 2, the boundaries of the principal's search interval will be independent of the
report of the agent.

The …rst-order condition characterizing the best reply is then
bene…t of gaining more information
…nancial cost of search
So it is the same condition as in the case with the fee except for the value of information (VH
rather than VA). The key to result in Proposition 12 is that VH can be lower (resp. higher than)
vLA is small (resp. large) compared to vHA. The intuition is clear: the fact that the
principal conducts more search has a negative impact if the state is high (delay good decision)
and a positive one if the state is low (potentially avoid bad decision). If
vLA is large compared
to vHA, the positive impact outweighs the negative one and the agent would in fact search more.18
For the sake of concreteness, most of the paper focused on illustrating how our model sheds light
on current and considered regulation of the drug approval process. Similar considerations are
relevant for approval regulation in other areas, from competition and consumer policy to …nancial
regulation. These applications are particularly relevant given the recent trend toward increased
consumer protection and the move toward using the ex ante approval approach when regulating
systemically important …nancial institutions.

Given its tractability, the model can be extended to address a number of other applications.

In the case of managers proposing projects to the headquarters, it would be natural to consider
competing agents. Considering an agent facing a sequence of principals would be a natural
extension. For example, in the case of authors submitting papers, upon rejection in one journal,
authors can submit to another outlet. In some applications, the cost of research might di¤er
across players. It is worth revisiting the question of delegation of decision rights by the principal
to the agent within this model.

1 8 Note that reasoning on the lower best response functions is su¢ cient since, given the smooth pasting condition,
the equilibrium in both the case with and without the fee will correspond to the maximum of the best reply.

We can rewrite the utility is the research area as
where, for a given process with mean
2 and optimal stopping times s and S,
standard results as in Stokey (2009) give
. We can derive useful relations among
these values using the fact R1 + R2 = 1. We have
e(S s) e(R2 1)(S s)
= e S+s+R2(S s) e S+s+R1(S s)
= es S eR2(S s) eR1(S s)=e
We now determine the best response to a given upper benchmark S. The …rst order
condition w.r.t. s is
Substituting the expressions for @ =@s
R2) e R1(S s) e R2(S s)
R2eR2(S s) + R1eR1(S s)
We can rewrite the …rst-order condition as
This establishes the …rst part of 1.1: the best response is independent of the current belief
fact, br(S) is implicitly de…ned by
It is useful to introduce the following variables,
Dividing the previous expression by eS s we can rewrite the br(S) as
, we obtain as in the main text
We show below that
1(s; S) > 0. We have eg > 1 (eg is increasing in S
s and is equal to 1 for
s = 0), so the sign is that of egeS s
R1e R2(S s) + e S eR2(S s)
R1eR1(S s) + eR2(S s)
1(s; S) > 0.

Comparative Statics for br (S)
We use to derive the comparative statics, the following expression for br(S)
Comparative Statics with Respect to vHA
Taking derivatives of this …rst order condition, we have
Comparative Statics with Respect to vLA
As above, we have
Using the same arguments as above, we have
Comparative Statics with Respect to c
Follow immediately from the fact that
1(s; S) > 0.

br(S) increasing in S
br(S). Equation (13) can be rewritten as
Taking derivatives with respect to S, we have
Following the same arguments as above we can show @g > 0 and @g0 < 0.

At l = 0 (i.e s = S), eg = 1 so that @l > 0. Equation 16 then implies that @g0 < 0 is always
increasing and asymptotically approaches the value l solution to vHA = (eg(l)
1) c , as indicated
in result 3.

We use the notation
( ; H) and for ( ; H). Using the expression for the utility
derived in the proof above, the …rst order condition with respect to S is given by:
We now derive the expressions for the partial derivatives of
and with respect to S. We
R1e R1(S s) + R2e R2(S s)
From the de…nition of we have
Given that R1 + R2 = 1, collecting eS we obtain
@ = (R2 R1)eS e R2s R1
After adding and subtracting R2 to the …rst exponential in the numerator and, similarly, after
adding and subtracting R1 to the second exponential in the numerator, this is equal to
@ = (R2 R1)eS e R2s R1 +R2 R2
Substituting the de…nition of
( ; L) = eR2(S s) eR1(S s)
( ; H), we conclude that
Exploiting the following relation
eR1(S s) = e(1 R1)(S s)e(1 R2)(S s)
= e(S s) e R1(S s)
rewrite the condition above in a more compact way as
( R2e R2(S s) R1e R1(S s)
( R2e R2(S s) R1e R1(S s)
We can now express the best reply as
We can rewrite the implicit equation de…ning BR(s) as
(s; S)VA + (s; S) c =
(s; S) = ef es S(1 + eS) > 0
(s; S) = ef0 > 0:
(s; S), we use the fact that
Note that ef > 1 and we can show that ef e S
e s + ef0e s is of the same sign as
Comparative Statics
To derive the comparative statics on BR(s), we use the following characterization of the best
Comparative Statics with Respect to vHA
Taking derivatives of (19)
R1e R1(S s) + R2e R2(S s)
if VA(S) > 0, we clearly have
If VA(S) < 0, we can rewrite the expression
Given that @f < 1, we also obtain the result.

Comparative Statics with Respect to vLA
Thus the same arguments allow us to establish that
Comparative Statics with Respect to c
Follow immediately from the fact
(s; S) > 0.

s decreasing in s
We denote l = BR(s)
s. Equation (19) can be rewritten as
Taking derivatives with respect to s, we have
This can be written as
Following the same arguments as for the proof of the comparative statics with respect to vHA,
we have that the denominator is positive. Since ef > 1, the numerator is negative and we obtain
We now establish the result by taking the limit of the above …rst order condition. We obtain
(ef es S + ef0) vLA +
We can show that ef es S + ef0 = eg, so that indeed, l is solution to
vLA = (e g(l) + 1) cr
Foundation for the Nash Equilibrium Solution
At each instant t, agent and principal move sequentially:
First, the agent chooses research Ra, submit Sa or wait/withdraw Wa
Second, if the agent submits Sa, the principal chooses research Rp, approve Ap or wait Wp
R results if Ra or (Sa; Rp); A results if (Sa; Ap), W results if Wa or (Sa; Wp)
We solve for the Markov Perfect Equilibria where the state variable is given by the current in-
formation . We show that the unique MPE outcome is: R for
< sN and A for
This unique outcome can supported by multiple equilibrium strategies.

We prove the result in a number of steps:
bri(S) reaches its minimum at S and BR
i(s) reaches its maximum at si
According to the smooth pasting condition, given that rejection always yields a zero value,
Taking derivatives with respect to S, taking into account that s = br(S) yields
Furthermore, given that @ui (S ) = 0, we have
The best response function br(S) reaches a minimum for S = S .

Taking derivatives with respect to s yields
The best response function BR(s) reaches a maximum for s = s .

If bra and BRp cross it is for values such that BRp(s) is increasing in s and bra(S)
is decreasing in S.

In Proposition 1 and 2, we showed that bra(S)
brp(S) and BRp(s)
BRa(s). According to
step 1 and 2, BRp crosses brp at (sp; Sp) where BRp is maximum. Given that bra(S)
and that brp is single peaked, it has to be the case that bra and BRp cross for values such that
BRp(s) is increasing in s. The same logic applies to show that bra(S) is increasing in S when
bra and BRp cross.

@BR(s) < 1.

We derived in the proof of Proposition 2, that BR(s) is de…ned by
s, and taking derivatives of the implicit equation above, we have
This can be rewritten as
Furthermore we have
Given that ef > 1 the right hand side is positive. Furthermore, we have that along BR(s) that
> 0. Therefore, since we established that @f > 0, we conclude that
@br 1(s) > 1 for values such that br(S) is decreasing in S.

br(S) is de…ned by the following implicit function
We have shown that for s > s , br(S) is an increasing function. On this interval, br 1(s) is a
well de…ned function. Taking derivatives of the implicit equation above, we have
Grouping terms we have
We showed in the proof of Proposition 1 that: e Seg0 + e S
ege s < 0. Given that vLA < 0,
to have the equation above satis…ed, since we are deriving a property of the curve over which
@S > 0, it has to be that
We also showed in the proof of Proposition 1 that @g0 < 0 and @g > 0. So the above inequality
implies that @S > 1.

Step 5: Deriving the results.

Step 3 implies that if a crossing between bra(S) and BRp(s) occurs, it will be for values of
(s; S) such that the properties of step 4 and 5 apply. Furthermore these properties imply both
that the curves do cross and that the crossing is unique: when the curves cross, @br 1(s) < 1 and
@BR(s) < 1, so the curves cannot cross again.

Result 1. then follows from step 1.

Finally from Proposition 1.3, we know S
br(S) is increasing in S, so since SN < Sa and
sN = bra(SN ), we have SN
sa. Similarly, using Proposition 2.3, we have SN
sp. This establishes result 3.

Result 1: We have SN > Sa > a. So, if
SN , the agent gets a strictly positive payo¤
< SN he can guarantee himself a zero payo¤ by choosing s = . Thus the agent's payo¤
is positive for all values of
Result 2: Given a value of the upper benchmark SN , the principal gets a zero payo¤ when
he chooses the lower benchmark
= brp(SN ) since it implies rejection. In the Nash equilibrium
solution, since sN = bra(SN ) < brp(SN ), he gets a strictly negative payo¤. The result is true
2 [sN ; brp(SN )] since even if the principal was choosing alone, he could only
guarantee a zero payo¤. However, when sN > , the principal, by de…nition of s, gets a zero
We …rst examine the case of small con‡ict of interest:
sa, the agent never experiments, since even when he controls both bench-
marks, he gets a negative payo¤ when experimenting. Thus, in this case, as indicated in Propo-
sition 5.1, any commitment Sc( ) >
will be equivalent for the principal and will give the zero
rejection payo¤. Furthermore, choosing a commitment below
leads to immediate approval and
a negative payo¤. We conclude on the …rst result: if
sa then Sc( ) 2 ( ; +1).

> sa, the principal has the choice between an interior commitment that leads the agent
to perform research and a blocking commitment, such that the agent performs no research: even
though the principal cannot force rejection because the lower threshold s is controlled by the
agent, the principal can induce the agent to reject immediately by committing to a su¢ ciently
high adoption threshold S such that bra (S) = s
. The lowest possible such blocking
commitment Sc = bra ( ), where bra is the upper inverse function constructed by inverting
bra for S > Sa. Clearly, any S strictly above bra ( ) is equally blocking and results in exactly
the same outcome and gives the principal a zero payo¤.

Consider the other option of an interior commitment. We …rst show that if, for a starting
, an interior commitment gives a strictly positive value (i.e is strictly preferred to a
blocking commitment), then it will also be the case for
0 > . Consider any point (bra (S); S)
(point on the agent's best response curve). The utility is given by
) + R1eR1(S ) < 0:
On the contrary, we have
We need to show that the second e¤ect dominates to prove that
@u(s; S; ) > 0
and thus if u(s; S; ) > 0, then for
0 > , u(s; S; 0) > 0. Therefore, there exists a value of the
initial belief such that the interior commitment is preferred to the blocking commitment if and
We conclude the proof of result 1 by showing that 2 (sp; p). For
(s; S) gives the principal a negative or zero payo¤, so naturally > sp. For
equilibrium outcome (sN ; SN ) gives the principal a positive payo¤ since he could always choose
S = p and guarantee a zero payo¤. So we have
Result 2. According to result 1, if
is above and close to , an interior commitment is
optimal. We …rst show that the interior commitment is decreasing in
where @up and @up are derived in the proof of Propositions 1 and 2 and ds corresponds to the
lower best response of the agent to a choice of S. The …rst term is the non strategic direct
e¤ect of S on the principal's utility. The second is the strategic e¤ect of S working through
the indirect impact of S on the agent's choice.

We …rst examine how the direct non strategic e¤ect reacts to an increase in
Using the formulas derived in Proposition 2, we have that @2up is proportional to
Both terms are positive so that
For the strategic e¤ect, we use the fact that the agent's best response to S is independent
, so that ds is independent of
. Using the formulas derived in Proposition 1, we have that
@2up is proportional to
The …rst term is positive while the second is negative. Thus we have
Overall we need to show that the strategic e¤ect dominates and we have
In a second step we show that
We showed in the proof of result 2 that @2up < 0 and ds > 0.

We have established in the proof of Proposition 2
In the proof of Proposition 2 when we examine the comparative statics with respect to vHA, we
established that the second term is also decreasing in S. Thus we have
We also have d2s , so that overall, the utility the principal is a concave function of the level of
commitment S.

This establishes that up has a unique interior maximum de…ned by dup = 0. Furthermore, we
established that d2up < 0, so that u
p will reach a maximum for lower values of S as
i.e the optimal commitment SC( ) is decreasing in
Result 3. We now identify the belief where the principal will switch from an interior com-
mitment to an approval commitment where the principal chooses immediate approval. Denote
, the …rst belief such that the principal is indi¤erent between commitment and immediate ap-
proval. Such a belief exists: indeed, the payo¤ from approval and from interior commitments are
continuous and, for
Sp, the optimal commitment is Sc =
since even when the principal has
full control, he would adopt immediately. For belief
, we must have SC( ) = . Furthermore,
since we showed above that the interior commitment SC( ) is decreasing in , we will also have
the interior commitment is below
and the principal therefore chooses immediate
We conclude the proof by showing that
= SN . Note that we necessarily have
< SN , at the Nash equilibrium, the principal could choose immediate approval but does not
do so, in other words, for those values of , a commitment SC = SN does better than immediate
= SN , an interior commitment at S = SN gives the payo¤ of immediate approval.

We conclude the proof by showing that at
= SN , the optimal interior commitment is indeed
SC = SN . We have
At (s; S) = (sN ; SN ), we have @up (s
N ; SN ) = 0. Furthermore, for
) (multiple of ) is equal to zero. Thus, we have for
dup (sN;SN; = SN) = 0:
As we established before, the unique interior commitment is characterized by dup = 0 so that
the optimal commitment is SC = SN . This concludes the proof.

This follows immediately from the results in Proposition 5. For
2 (sa; ), the optimal
commitment is either a blocking commitment or an interior commitment that gives a strictly
higher payo¤ than the Nash. For
< sa, both yield a zero payo¤. For
approval payo¤.

In Proposition 5 we showed that > sp and
< Sp. This establishes the …rst result:
SN . We showed in Proposition 1 that S
bra(S) is increasing
bra(SN ). The fact that Sc
immediately from Proposition 3.

Result 1: As explained in the main text, in the second phase, the principal will search in the
interval (sp; Sp). By backwards induction, the agent will thus search in the interval (bra(Sp); Sp).

Indeed, since the agent pays the cost of research even if the principal searches, stopping for
decreases the utility of the agent since sp > bra(Sp) and thus with some probability, the principal
would stop too early.

2 ( ; Sp), the Nash equilibrium outcome coincides with the optimal com-
mitment and is strictly di¤erent from the outcome of sequential search. The principal will thus
achieve a strictly higher level of utility
Result 4: The length of the research interval is larger in the sequential case than in the
Nash since (bra(Sp); Sp) is a point on the agent's best response curve for a higher value of S.

The result in Proposition 2.3 yields the result. It is also clearly larger than for the standalone
problem since the value of S = Sp is the same but the lower benchmark s = bra(Sp) < brp(Sp).

Given an acceptance standard SM , when the agent chooses a search interval (s; S) with
SM , the utility is
The change compared to the benchmark case is that an additional cost is incurred if the state is
low. The best response to the upper benchmark is given by
Thus the best response is above the best response in the benchmark case.

We now examine the best response to the lower benchmark. The …rst order condition with
respect to S is slightly modi…ed
( R2e R2(S s) R1e R1(S s)
We can then rewrite the …rst order condition in a more compact way as
(s; S)VA + (s; S) =
We conclude the proof by showing that the upper benchmark chosen SM by the agent is such that
SM < SM . Suppose this was not the case and the agent chose SM = SM . Then the best response
above would be the same as the best response for the agent's stand-alone problem. However,
given that SM = SM > Sa and that the agent's stand-alone upper best response reaches its
maximum at a value S = Sa, SM cannot be on the agent's best response curve. We therefore
reach a contradiction. As can be veri…ed grapically in Figure 5, the agent's upper best response
with misreporting is strictly below the horizontal line S = SM .

We are in the case
2 ( ; ), i.e the case of an interior commitment. At S = Sc( ), by
de…nition of the interior commitment, the iso-utility curve of the principal is tangent to the lower
best response curve of the agent for a …ne F = O (bra(S), i.e. the best response of the agent
used in all the previous sections).

For a …ne F > 0 and an approval standard SM = Sc, the best response curve of the agent
bra(S; F ) is modi…ed. It is also tangent to bra(S) at S = Sc( ) since C0(0) = 0. However, locally,
for F large enough it is to the right of the iso-utility curve of the agent: for a given S, s is
strictly increasing in F . As a consequence, the intersection with the agent's upper best response
BRa(S; F ), (sM ; SM ) belongs to a higher iso-utility curve. All these properties are represented
in Figure 9.

If the state is low, the expected bene…t of the principal for a choice of W at belief
by (5), is decreasing in W for su¢ ciently large values of Pp. If the state is high, (6) is increasing
in W . Thus, there will be a benchmark belief W where these two e¤ects cancel each other and
such that withdrawal should be chosen, whenever
Given that the expected payo¤ upon approval are
vHA = p(e)vHA(W ) + (1
vLA = p(e)vLA(W ) + (1
LA is increasing in e and vHA is
increasing in e. For Pp su¢ ciently large, the e¤ect on vLA dominates and according to Proposition
2, the best response of the agent is decreasing in e. A similar argument can be made for the
lower best response of the agent.

The utility of the agent is given by
where the expected value upon stopping research depends on the uncertainty due to the principal's
research process,
We …rst examine the …rst order condition characterizing the lower benchmark of search s. Since
p(S; H ) depends on s, the …rst order condition is identical to the base
case except for the value of VA. We have
VH (S) = 1(s; S) c=r
using the fact that
p(S; H ), we have
We examine how VH compares to VA. We have that for S = Sp, VH(S) = VA(S). Thus, we look
at the comparative statics of VH (S) with respect to Sp, for S
1+eS e R1(Sp sp) e R2(Sp sp)
R2e R2(Sp sp) (vHA + e Sp vLA)
e Sp (e R1(Sp sp)
1+eS e R1(Sp sp) e R2(Sp sp)
R2e R2(Sp sp) vHA + e Sp
R2e R1(Sp sp) + R1e R2(Sp sp)) vLA :
Therefore we have @VH
R1e R1(Sp sp) + R2e R2(Sp sp)
e Sp R2e R1(Sp sp)
R1e R1(Sp sp)+R2e R2(Sp sp)
i < 1. We use the notation
R2e R1(Sp sp) R1e R2(Sp sp)
R1e R1(Sp sp) + R2e R2(Sp sp)
e Sp R2e R1(Sp sp)
Note that this condition is independent of S.

Thus we have that if
1(Sp ; sp), VH (S) > VA(S) for all S
Sp. Thus in this case,
the lower best response is lower in the case where the agent does not pay (he searches more). For
the same reasons as in the proof of Proposition 3, the equilibrium is reached at the point where
the best response is maximized. Thus in equilibrium, we will have b
sseq < sseq. The reverse holds
Alonso, Ricardo, and Niko Matouschek. 2008. "Optimal Delegation." Review of Economic
Studies, 75(1): 259–293.

Armstrong, Mark, and John Vickers. 2010. "A Model of Delegated Project Choice." Econo-
metrica, 78(1): 213–244.

Bloom, Nicholas, Mark Schankerman and John Van Reenen. 2012. "Identifying Technology
Spillovers and Product Market Rivalry." Econometrica, forthcoming.

Bolton, Patrick, and Christopher Harris. 1999. "Strategic Experimentation." Econometrica,
67(2): 349–374.

Bulow, Jeremy I., John D. Geanakoplos and Paul D. Klemperer. 1985. "Multimarket Oligopoly:
Strategic Substitutes and Complements." Journal of Political Economy, 93(3): 488–511
Chan, Jimmy, and Wing Suen. 2011. "Does Majority Rule Produce Hasty Decisions?" Unpub-
lished working paper.

Dahma, Matthias, Paula GonzÃ¡lez, and NicolÃ¡s Porteiro. 2009. "Trials, Tricks and Trans-
parency: How Disclosure Rules A¤ect Clinical Knowledge." Journal of Health Economics,
28(6): 1141–1153
Fernandez, Raquel, and Dani Rodrik. 1991. "Resistance to Reform: Status Quo Bias in the
Presence of Individual-Speci…c Uncertainty." American Economic Review, 81(5): 1146–
Gerardi, Dino and Luca Maestri. 2012. "A Principal-Agent Model of Sequential Testing."
Theoretical Economics, 7(3): 425–463.

Grossman, Sandford J. 1981. "The Informational Role of Warranties and Private Disclosure
about Product Quality." Journal of Law and Economics, 24(3): 461–483.

Gul, Faruk, and Wolfgang Pesendorfer. 2012. "The War of Information." Review of Economic
Studies, 79(2): 707–734.

Halac, Marina, Navin Kartik, and Quingmin Liu. 2012. Optimal Contracts for Experimentation.

Unpublished working paper.

Henry, Emeric 2009. "Strategic Disclosure of Research Results: The Cost of Proving Your
Honesty." Economic Journal, 119(539): 1036–1064.

HolmstrÃ{m, Bengt. 1984. "On the Theory of Delegation."In: Bayesian Models in Economic
Theory, eds. Marcel Boyer and Richard Kihlstrom. Amsterdam: Elsevier.

HÃ{rner, Johannes, and Larry Samuelson. 2012. "Incentives for Experimenting Agents." Un-
published working paper.

Junod, Suzanne White. 2008. "FDA and Clinical Drug Trials: A Short History," in: A Quick
Guide to Clinical Trials, eds. Madhu Davies and Faiz Kerimani. Washington: Bioplan,
Kartik, Navin. 2009. "Strategic Communication with Lying Costs." Review of Economic Stud-
ies, 76(4): 762–794.

Kartik, Navin, Marco Ottaviani, and Francesco Squintani. 2007. "Credulity, Lies, and Costly
Talk." Journal of Economics Theory, 134(1): 93–116.

Lewis, Tracy R. 2012. "A Theory of Delegated Search for the Best Alternative."RAND Journal
of Economics, 3(1): 391–416.

Lewis, Tracy R. and Marco Ottaviani 2008. "Search Agency." Unpublished working paper.

Lizzeri, Alessandro, and Yariv, Leeat. 2011. "Sequential Deliberation." Unpublished working
Manso, Gustavo. 2011. "Motivating Innovation." Journal of Finance, 66(5): 1823–1860.

Milgrom, Paul R. 1981. "Good News and Bad News: Representation Theorems and Applica-
tions. Bell Journal of Economics, 12(2): 380–391.

Moscarini, Giuseppe, and Smith, Lones.

"The Optimal Level of Experimentation."
Econometrica, 69(6): 1629–1644.

Nundy, Samiran, and Chandra Gulhati. 2005. "A New Colonialism? Conducting Clinical Trials
in India." New England Journal of Medicine, 352(16): 1633–1636.

Ocana, Alberto, and Ian Tannock. 2011. "When are ‘Positive'Clinical Trias in Oncology Truly
Positive?" Journal of National Cancer Institute, 103(1): 16–19.

Shavell, Steven. 1994. "Acquisition and Disclosure of Information Prior to Sale."RAND Journal
of Economics, 25(1): 20–36.

Stokey, Nancy. 2009. "The Economics of Inaction. " Princeton University Press.

Strulovici, Bruno. 2010. "Learning while Voting: Determinants of Strategic Experimentation."
Econometrica, 78(3): 933–971.

Taylor, Curtis, and Huseyin Yildirim. 2011. "Subjective Performance and the Value of Blind
Evaluation." Review of Economic Studies, 78(2): 762–794.

Wald, Abraham. 1945. "Sequential Tests of Statistical Hypotheses." Annals of Mathematical
Statistics, 16(2): 117–186.

Wald, Abraham. 1947. Sequential Analysis. New York: John Wiley and Sons.

Wald, Abraham, and Jacob Wolfowitz. 1948. "Optimal Character of the Sequential Probability
Ratio Test." Annals of Mathematical Statistics, 19(3): 326–339.

Source: http://pareto.uab.es/research/sempdf/micro/2014/Ottaviani.pdf

AgentSheets®: an Interactive Simulation Environment with End-User Programmable Agents Alexander Repenning AgentSheets Inc., Gunpark Drive 6560, Boulder, Colorado, 80301, email: [email protected] Center for LifeLong Learning and Design, Department of Computer Science, University of Colorado at Boulder, Campus Box 430, Boulder, Colorado 80309-0430, email: [email protected]

BW Clip Real Time 2 OR 3 YEAR H S, CO, O , SO SINGLE-GAS DETECTOR Part Number 50115903-101 OM-EN-BWCRT-A3 BW Clip Real Time Operator Manual BW Clip Real Time is a portable, single-gas detector manufactured by BW Technologies by Honeywell (BW). BW Clip Real Time continuously displays the concentration of a specific gas in the ambient environment, and activates an alarm when concentrations exceed alarm