Econ664 Empirical Studies in Industrial Organization by cuiliqing


									     Empirical Studies
in Industrial Organization
       by Ginger Z. Jin
         Spring 2009
     At Peking University

    Notes as of 5/26/2009
             Self Introduction

   My name: Ginger Zhe Jin
   B.S. Of Economic Management in the University
    of Science and Technology of China (USTC)
   M.A. Of International Finance in the Graduate
    School of the People's Bank of China
   Ph.D. Of Economics in UCLA
   Now Associate professor of economics at the
    University of Maryland
                This course
 My goal
  Introduce the recent empirical studies in industrial
   organization (IO) --- what is IO?
  Give you some sense of how to conduct empirical
   research in empirical IO
  Offer an opportunity to design your own research
  Learn about Chinese economy and explore research
   topics in China
 Rationale
  learn from each other
  free to criticize or give feedback
  encourage class discussion
      What makes a good Empirical IO
A good empirical IO study should be a good
  match of IO theory, real data, and empirical
 •   start with an interesting and important question
 •   be a good consumer of theoretical literature
 •   be a good consumer of econometrics
 •   be an entrepreneur in data collection
 •   emphasize empirical identification
 •   clarify main story and alternative explanations
    New Syllabus (changes in red)
   5 themes, each theme has:
       – 2-4 papers for lecture,
       – 1 case discussion or 1 student presentation
       – students are expected to read papers before class
       – engage in class discussion (20%)
   2 referee reports (40%)
       – Due dates: April 4, April 24
   Research Proposal due June 15 (40%)
       – Your own idea (original)
       – Review of related literature, clarify your
       – Design data requirements and estimation strategy
       – Present in front of the class (30-min, optional)
            Common short hand
   AEJ: American Economic Journal
   AER: American Economic Review
   EJ: Economic Journal
   EI: Economic Inquiry
   JPE: Journal of Political Economy
   JEMS: Journal of Economics & Management Strategy
   JF: Journal of Finance
   JIE: Journal of Industrial Economics
   JHE: Journal of Health Economics
   NBER-WP: National Bureau of Economic Research (NBER) working
   QJE: Quarterly Journal of Economics
   RAND: RAND Journal of Economics
   ReStat: Review of Economic Statistics
   ReStud: Review of Economic Studies
   RQ: research question
   SSRN: Social Science Research Network
     Theme #1: Pricing Strategy
•   Definition of Price Discrimination
•   Price discrimination in automobiles
     • Ayres and Siegelman AER 95
     • Goldberg JPE 96
•   Why discriminate?
     • List QJE 2004
•   An example of structural estimation on secondary ticket
•   Leslie and Sorensen, Stanford working paper 2008
     Define Price Discrimination

• Similar products are sold to different consumers at
  different prices, and the difference in price cannot
  be explained by any difference in cost.

• Types of price discrimination:
  • First-degree
  • Second-degree
  • Third-degree
         Necessary Conditions
        for Price Discrimination

• There exists some consumer heterogeneity

• Firm(s) have some degree of market power (i.e. downward
   sloping demand)

• No arbitrage
    Why are researchers interested in
         price discrimination?
   The Robinson-Patman Act of 1936 makes it unlawful to
    discriminate by race and gender

    welfare implication on price discrimination is ambiguous
     Profit-maximizing sellers always prefer price
       discrimination to uniform price
     Compared to uniform pricing, consumers that are
       discriminated against will be worse off but the other
       consumers will be better off
             Empirically detect
            price discrimination

   need data on transaction price
   cost difference vs. mark-up difference
   consider the possibility of arbitrage
   define market power: monopoly vs. competitive markets
    Ayres and Siegelman 1995 AER

   RQ: Is there third degree price discrimination in new car retail
   Why interesting?
    • It is unlawful to discriminate by race and gender (Robinson-
      Patman Act of 1936)
    • Rent distribution between car manufacturers and car
    • Shed light on the pricing strategy of dealers/manufacturers
      – why negotiate rather than charge list price?
    Ayres and Siegelman 1995 AER
           Research Design
•   Chicago area, 9 car models
•   Matched pairs of individual consumers, only differ by race
    and gender
•   Follow exactly the same bargaining process (two
•   Record every negotiation price, including the final offers even
    if there is a disagreement
•   Why use a controlled experiment?
   Ayres and Siegelman 1995 AER

               Initial   Final

White male     0         0

White female   200+      130+

Black female   450+      440+

Black male     1000+     1100+
        Ayres and Siegelman 1995 AER
           Possible Interpretations
•   There are significant price differences by race and gender
•   This does not disappear for black dealers
•   Not sure whether it is driven by animus discrimination or by
    statistical discrimination
    • Animus: distaste of doing business with minorities
    • Statistical: use observable characteristics to make statistical
      inference about the reservation value of consumers
•   Little difference between initial and final offers, probably because
    the “consumers” follow exactly the same bargaining process. For
    the same reason, it does not detect any discrimination based on
    bargaining ability
•   Specific to Chicago and the nine tested car models?
         Goldberg 1996 JPE
Comparison with Ayres and Siegelman 95
 • Same question but different research design:
               A & S 95            Goldberg 96

               Audit study         CES observational

   Area        Chicago             Whole US

   Consumers   Ind. testers        Fam./HH

   Strategy    Same bargaining     Potentially different

   Model       9 car models        A large # of models

   Price       Every negotiation
               price               Only the final price, with
                                   measurement error
              Goldberg 1996 JPE
             Findings comparison
                           A & S 95       Goldberg 96

               Initial P        Final P   Final P

White male     0                0         0

White female 200+               130+      120-130+

Black female 450+               440+      426+

Black male     1000+            1100+     274+

• The race and gender difference in final price is
  not as large and not statistical significant
                     Why different?
•   Sample difference
    • Goldberg subjects to sample selection (on only finished
      sales and self-selection into dealerships)
    • Goldberg has measurement errors on both left and right
      hand side variables
    • Goldberg covers larger areas and more car models
•   Goldberg captures discriminations that are assumed away in A
    & S?
    • Discrimination by bargaining
    • Initial offers reflect statistical discrimination by race and
      gender, but discrimination by bargaining cancels out the
      initial discrimination?
               More detection

• First offer does not necessarily reflect the mean transaction
• First offer could differ substantially if the mean transaction
  price is the same across race and gender but the
  distribution of reserve value is more dispersed for
  minorities and first offer reflects the upper tail
• Check variance of price (Table 4), quantile regression
  (Table 3)
                   Follow-up studies
•   Harless and Hoffer (AER March 2002) “Do Women Pay More for New
    Vehicles? Evidence from Transaction Price Data”
     • Use transaction price from J.D. Power
     • Find discrimination by car type and age group, but not by gender
•   Scott Morton et al. (2002) “Consumer Information and Price Discrimination:
    Does the Internet Affect the Pricing of New Cars to Women and Minorities?”
    SSRN #288527
     • The same J.D. Power transaction price data
     • African-American and Hispanics pay 2% ($500) more in offline
       transaction, but 65% of this difference can be explained by other
       observables. Find no evidence for statistical race discrimination
     • Small price premium for women (0.43%, $100)
     • No significant difference in online price
           •    Internet provides less cue for racial discrimination -- no
           •    Internet erases search cost – yes
     • Conclusion: discrimination on non-racial differences across individuals, not
       statistical discrimination based on race -- may be the individual
       characteristics are revealed in the bargaining process?
                    List QJE 2004
• Source of observed discrimination:
    • Animus
    • Statistical
    • Bargaining ability
• Field experiment (different from an audit study like A & S 95)
    • Traders endogenously select into the market
    • More likely have previous experience buying, selling and
    • Active control on confounding factors
             List QJE 2004
             Experimental I

• Treatment B: Nondealer buyers negotiate with dealer
• Treatment S: Nondealer sellers negotiate with dealer
• Four groups: white male aged 20-30
        white female aged 20-30
        non-white male aged 20-30
        white male aged 60+
                List QJE 2004
            results on experiment I
• Results of Treatment B and Treatment S:
   • Initial offers to minorities are inferior.
   • Final offers are inferior to minorities too, but the
     difference is not as great as the initial offers.
   • Previous experience of non-dealers help eliminate the
     minority bias, but it takes longer in negotiation.
   • Experienced dealers exercise more discrimination than
     non-experienced dealers.
• These results prove the existence of price discrimination, but
  do not distinguish between the three possible sources of
              List QJE 2004
        Where is the smoking gun?
• Dictator game of a dealer splitting $5 to a person that belongs to
  a specific demographic group (double blind design with opaque
  envelope and blank paper)
• no evidence of animus discrimination
              List QJE 2004
        Where is the smoking gun?
• Chamberlain Experiment: Buyer reservation values are
  randomly drawn, but sellers may or may not know this
   • Treatment Random
   • Treatment Unclear
   • Why these two treatments can identify statistical
     discrimination from discrimination by bargaining ability?
               List QJE 2004
         Where is the smoking gun?
• Results --- Treatment random vs. Treatment clear
  • Majority buyers do better than minority buyers, but only in
    Treatment Unclear
  • Experienced dealers do better than inexperienced dealers
    (more obvious in Treatment Random)
  • Experienced dealers do worse in Treatment Unclear than
    experienced dealers in Treatment Random
• Conclusion: Statistical discrimination is more likely than
  discrimination by bargaining ability
              List QJE 2004
        Where is the smoking gun?
• The actual distribution of consumer reservation value
   • Solicited by Vickey’s second price auction
   • Ask dealers to match two consumer-value distributions
     with specific demographic groups
          List QJE 2004
Conclusion and remaining questions

• Statistical discrimination is the most likely explanation
• Experienced dealers use their knowledge to exercise more
  statistical discrimination (why does it take long to learn
• Experimenter effect on the dictator game?
• What happened during the negotiation (in Experiment I and
  the Chamberlain market)?
• Will this only apply to negotiation markets?
• Why List (2004) and Scott Morton et al. (2002) reach
  different conclusion?
                 Related studies
• Rust and Hall JPE 2003 “Middlemen versus Market Makers: A
  Theory of Competitive Exchange”
    •   Market makers vs. secret price
    •   Why posting price vs. bilateral bargaining?
    •   Price search is costly
• Gong (2006) dissertation extends Rust and Hall
    •   Endogenous entry of market makers
    •   Collusion of market makers?
    •   Lab experiments
• List and Price (2005 RAND): field experiments in collusion and
  secret price cuts – depends on information and experience
• Matthew White (Upenn Wharton, working in progress): the impact
  of centralized trading on the price of wholesale electricity
               Case Discussion
• The Children's Hospital of Tianjin has started to
  sell three types of consumer cards. They allow
  card buyers a higher priority to make doctor
  appointment at the hospital. (News reported in
  Beijing Qing Nian Bao 2/18/2009)
• Class discussion – five groups of students, each
  represents a different player in this problem
  • Hospital
  • Broker (who issues the cards for the hospital)
  • Patients that have bought or will buy the cards
  • Patients that do not want to buy the cards
  • Regulator
    Leslie & Sorensen (2008) “The Welfare
           Effects of Ticket Resale”
       Widespread hostility against ticket resale, recent
        trends in favor of ticket resale (e.g. MLB, artists)
       Economic intuition: voluntary trade should enhance
        efficiency and result in an allocation that improve
        social welfare
       Economic players:
        Primary producers
        Brokers (scalpers)
        Trading intermediaries (eBay and StubHub)
        Consumers
        Regulators
                        Why resale?

   Consumer demand changes between the primary and
    secondary markets
   Primary market is mis-priced
       sellers take pride in “sold out”, sometimes for
        marketing reasons
       sellers have difficulty anticipating the demand
       seat quality varies to a great deal but only a limited
        number of pricing points are allowed in the primary
                Research Question

   Measure the welfare effect of ticket resale:
         – who is better off or worse off by the
           introduction (or show-down) of ticket resale?
         – Measure it for each player first and then sum
           up the total welfare effects
   primary producers
   Brokers (scalpers)
   Trading intermediaries
   Consumers
   Effect of resale on primary sellers
          – Difference-in-Difference by variations in state anti-scalping
            laws (Williams 1994, Depken 2007)
          – DepVar = primary price, Mixed results
          – Theories: Swofford 99, Courty 03, Karp and Kerloff 05,
            Geng, Wu and Winston 04)
          – 0 or + depending on assumptions about demand
            uncertainty, broker risk attitude/cost/revenue/information
   Effect of resale on consumers
          – Pro: easier access to the event
          – Con: higher price
          – Differ by type of consumers
          – Elfenbein 2005: stricter anti-scalping laws reduces
            competition among resellers, leading to fewer tickets sold
            and higher prices
            How does resale work?

   Stage 0: primary sellers set primary prices
   Stage 1: Consumers and brokers participate in the
    primary market conditional on their expectation
    about the secondary market
   Stage 2: Ticket supply and demand in the secondary
    market depend on the tickets and price sold in the
    primary market
         To model resale empirically
 Can primary sellers change ticket price over time?
 Can primary sellers endorse resale and share
  revenue from the secondary market?
 How would primary sellers adjust primary price given
  different settings of the secondary market?
 Shall we allow individuals to endogenously choose
  whether to participate primary or secondary markets
  due to transaction cost?
 Shall we distinguish brokers from real consumers?
 Shall we allow buyer uncertainty? The authors allow
  schedule conflict, random participation
            Pro and Con of this paper
   Cons:
        – Taking primary producers and their behavior
          as given. Cannot answer how the opening of
          ticket resale market affects producers
   Pros:
        – allow buyers in the primary market to resell in
          the secondary market activities. Allow their
          primary market buy decision depends on their
          expectation of the secondary market
        – distinguish real consumers and brokers
        – Estimate welfare effects
   Primary market: TicketMaster, 103 concerts of 18 artists in the
    summer of 2004
        Transaction price, day of transaction, section/row/seat/lawn,
         seat quality (relative, not absolute)
   Secondary market: all tickets sold on eBay and StubHub in the
    summer of 2004
        Transaction price, day of listing and transaction,
        Does not capture the full secondary market
        2-5% tickets resold, 39% of price mark up from ticket's face
         value, total resale revenue comparable to 6% of ticket
        Some tickets are over-sold, some are under-sold relative to
         face value
        Finer variation by seat quality
             Model – primary market
   Buyers randomly arrive at the primary market, choose
    from the pool of unsold ticket, and cannot buy more than
    1 ticket
   Consumer utility for event k seat j:

   Broker utility from the event is zero
   Consumers and brokers are forward-looking
           Model – secondary market
   All the tickets sold in the primary market is on sale in the
    secondary market via a sequence of private-value,
    second-price auction with limited buyer participation
     • Order of resale goes from the highest quality ticket to
       the lowest quality ticket
     • Buyer participation is random and limited to L bidders
       in each resale auction
             Sources of uncertainty
   Random arrival of buyers           A simplifying assumption
   Buyer's wtp is their own private info, and there is
    uncertainty about the distribution of wtp
     • ==> Buyers are not sure about the allocation of
       tickets in the primary market
     • Otherwise brokers cannot lose money
   Potential schedule conflict, may be related to wtp
   Random participation in resale auction

          Mainly to accommodate the fact that price is not a
         monotonic function of seat quality everywhere, though
            the increasing relationship is true on average
               Objective functions
   Resale price:
   Broker's problem
   Consumer's problem
               Estimation roadmap
   Key parameters are transaction costs and all the
    uncertainty parameters
   For any set of key parameters, we need to model
    forward-looking based on the assumption of rational
   Estimation has two loops:
     • Outer loop: search for key parameters to match
       the empirical data – simulated GMM
     • Inner loop: find rational expectation for forward-
       looking agents – fixed point
          Estimation – inner loop #1
   Given parameter set α, a buyer's expected utility at
    the primary market is represented by function

   The inner loop needs to find the value of α to reach
    the fixed point of the V function
          Estimation – inner loop #2
   How to find a fixed point of V?
     • Given α0, simulate the optimal behavior → V1
     • Regress V1 on variables in V function → α1
     • Repeat until convergence on V
   Convergence on V:

   Where M = # of buyers, S = # of draws
               Estimation – outer loop
   In the outer loop search for key parameters on transaction costs,
    uncertainty parameters, and event fixed effects
   Simulated generalized method of moments (GMM)
   Total 111 moments for 110 parameters:
     • Fraction of available tickets sold in primary mkt for each event (103
     • Average fraction resold by consumers (1)
     • Average fraction resold by brokers (1)
     • Average resale price (1)
     • Average quality of resold tickets (1)
     • 25th and 75th percentile of the resale price distribution (2)
     • 25th and 75th percentile of the resale seat quality distribution (2)
110 parameters:
     Event dummies (103): tickets sold per event
     Consumer premium for high quality tickets (Φ): price-quality curve
     Mean of average consumer wtp (λ-bar): price-quality curve, resale price
      of high-quality tickets
     Stdev of average consumer wtp (σλ): fraction of tickets resold at a loss
     Prevelence of broker (β): fraction of brokers in the resale market
     Consumer transaction cost (τc): fraction of tickets resold by consumers
      at high versus low expected mark-up
     Broker transaction cost (τb): fraction of tickets resold by brokers at high
      versus low expected mark-up
     Probability of schedule conflict (ψ): the relative rate at which consumers
      versus brokers resell below face value
   In the context of rock concerts, allowing resale does not result in
    dramatic welfare improvement, but lower transaction costs will.
   Resale makes some better off (reselling brokers and consumers), some
    worse off (consumers who attend the event, slightly lower revenue for
    primary producers)
   Limitations:
     • No equilibrium analysis on producers, this may apply to this data,
       but it also restricts the potential of applying this paper's conclusion
       to other resale markets
     • A lot of robustness checks are needed before we believe the results,
       especially structural assumptions
     • Can we use more moments to estimate the model or test the fit of
       the estimation?
     • No account for the timing of resale (Andrew Sweeting 2008)
                Theme 2: Collusion
    •   Competition patterns
        •   Perfect competition (firm as price-taker, homogenous
            goods, perfect and complete information, free entry)
        •   Monopolistic competition (firms are price-setters,
            heterogeneous goods, zero-profit in the long run)
        •   Bertrand (firms set price, homogenous goods, fixed # of
        •   Cournot (firms set quantity, homogenous goods, fixed # of
        •   Monopoly (collusion)
    •   Collusion makes more profits because it takes into account the
        cross-elasticity between competing firms
         Wiki Definition of Collusion
    •   Collusion is an agreement, usually secretive, which occurs
        between two or more persons to deceive, mislead, or defraud
        others of legal rights, or to obtain an objective forbidden by
       law typically involving fraud or gaining an unfair advantage
        and can involve "wage fixing, kickbacks, or misrepresenting
        the independence of the relationship between the colluding
    •   A cartel is a formal (explicit) agreement among firms. Cartels
        usually occur in an oligopolistic industry, where there are a
        small number of sellers and usually involve homogeneous
        products. Cartel members may agree on such matters as price
        fixing, total industry output, market shares, allocation of
        customers, allocation of territories, bid rigging, establishment
        of common sales agencies, and the division of profits or
        combination of these. The aim of such collusion is to increase
        individual member's profits by reducing competition.
                            Sherman Act
    •   Collusion is per se illegal
        •    Price fixing
        •    Bid rigging
        •    Territorial or customer allocation among competitors
    •   Identical price does not necessarily entail collusion. Similarly, the fact that all
        prices are not identical does not indicate the absence of collusion
    •   Antitrust evidence
        •    Explicit agreement among competitors
        •    Records of changes or prices
        •    Competitors’ meetings, phone conversation
        •    Testimony from a cartel member
                 Collusion Stability
    •   Collusion is inherently unstable because each member has
        incentive to deviate and secretly undercut the price or expand
       production.
    •   Detection of violators may be tough if there are unpredictable
        random factors in demand and supply.
    •   Even worse if the factors are not perfectly observable to other
        cartel members.
    •   Cartel stability is also subject to potential entries because
        cartel makes more than normal profits.
                 Empirical difficulty
               in detecting collusion
    •   Multiple equilibria by folk theorem: many payoffs between
        perfect competition and monopoly can be supported by cartel
    •   Demand and supply shifters may be unobservable to
        econometricians, but observable to cartel
    •   Need to accommodate entries and potential strategic elements in
        entry deterrence
    •   If the theory works, we will never observe deviation from the
        collusive equilibrium.
    •   Collusive equilibrium depends on beliefs of how competitors
        behave in off-equilibrium paths, which econometricians cannot
    •   Lack of benchmark price distribution that reflects non-collusion
        Literature on Conduct Parameter

    •    Tim Bresnahan ―Empirical studies of industries with market
         power‖, in Schmalansee and Willig edited The Handbook of
         Industrial Organization, vol II, 1989
    •    Ken Corts ―Conduct parameters and the measurement of
         market power‖ Journal of Econometrics 1999
        Literature on Conduct Parameter
    •   Conjectural variation model extends the classical Bertrand and
       Cournot:
         •   firms may hold non-zero conjectures about their rivals’
             responses to changes in their strategies
         •   Firm i expects rivals’ aggregate output is Ri(qi).
         •   Define ri=R’i(qi)
         •   Firm i’s first order condition is: (defines supply func.)

                c i  )'Q
                q 1i P q
               Pi()( r ()i

                                 Conduct parameter
        Literature on Conduct Parameter
    •   Conduct parameter:
        •   0 corresponds to perfect competition
         •   1 corresponds to classical Cournot
         •   N corresponds to collusion

                   ~ Pc
                   i     N

 Conduct                Lerner Index                Industry
     Criticism on Conduct Parameter
 •   No coherent theory for behaviors beyond the three extremes
•   Even not good at testing the three extremes because theta=1
     could be consistent with any level of market power when
     demand is iid
 •   In theory, conduct parameter measures how the competitors
     respond to a change of firm i’s strategy. In reality, it measures
     how equilibrium output varies with shifts in exogenous
     variables. These two aren’t the same, not even asymptotically.
 •   Corts (1999) shows that the mismeasurement of the conduct
     parameter could be quite severe.
     Criticism on Conduct Parameter
 •   No coherent theory for behaviors beyond the three extremes
•   Even not good at testing the three extremes because theta=1
     could be consistent with any level of market power when
     demand is iid
 •   In theory, conduct parameter measures how the competitors
     respond to a change of firm i’s strategy. In reality, it measures
     how equilibrium output varies with shifts in exogenous
     variables. These two aren’t the same, not even asymptotically.
 •   Corts (1999) shows that the mismeasurement of the conduct
     parameter could be quite severe.
                    Porter BELL 1983
    •   RQ: Is the grain price fluctuation during 1880-1886
       attributable to collusive behavior of the Joint Executive
        Committee, or something else?
    •   Why interesting?
        •   A lot of theories, little empirical evidence
        •   Provide more insights for antitrust efforts?
    •   Data: p, q over time
         Porter BELL 1983 -- theory
  •   Collusion with imperfect detection: price may decline (thus q
      may increase) due to demand uncertainty or secret price cut
 •   Collusion is supported by switching to a more competitive
      market supply if the price is too low or the total quantity
      supplied is too high
  •   Given uncertain demand, the optimal choice of detecting
      mechanism does not support the monopoly price. In fact, the
      collusive equ. has larger quantity and lower price than perfect
      collusion (trade off short-run profits with long-run stability)
  •   Green and Porter 1984 shows how to use Cournot as a
      punishment to support perfect collusion. This is not the exact
      model this paper tests, but it motivates the econometrics
  •   Detect fluctuation within the collusive equ, not on/off equ.
            Porter BELL 1983 -- Model
    •   Demand:        
                      g 0 l tt U
                     lQ g 2 
                     o  1 p Li
                       t 
                          o     t

    •   Supply:        
                    o l 
                    lp  Q
                     g  g
                        o S IU
                          t   0 1        t    2t     3t     t
    •   (supply side based on conduct parameter)
                        l ( 
                       g  / 1
                      3  o1    )

ζ=0 for Bertrand; Herfindahl for Cournot, 1 for collusion
    •   Identification:
        •   use entry and regime shifts as supply shifter to identify
        •   use lake seasons as demand shifter to identify supply
        •   regime shifts are unobservable! Assuming Iit conforms iid
            Bernoulli (0/1), use E-M algorithm to identify Iit
        Porter BELL 1983 -- assumptions
    •    Q: Lake season is predictable, but the theory emphasizes
         unpredictable factors. Contradiction?
        •   Lake dummy helps identify the supply curve, but what
             triggers the regime shift is the unobservable and
             unpredictable part of the error term U_1t.
    •    Key assumptions:
         •   Assume away the active role of entry and exit, and
             therefore entry only shifts the supply curve by a constant
         •   Assume independent draws of regime shifting rather than
             path dependent. Claims that this would not generate too
             much bias if the number of reversionary episodes is small
             relative to the sample size
         •   Error terms are normal
Porter BELL 1983 – key identification
    •   The actual error term (U2t) with no account for regime shifts
        looks like a bi-normal distribution, implying that the error is
       generated by two different regimes, of which one is a parallel
        shift of the other
                  Porter BELL 1983
               – remaining questions
    •   What is the null? Perfect competition, bertrand, cournot, or
       something else?
    •   What if we reject the null? Can we say the cartel switch from
        Cournot to Bertrand, or something else?
    •   No, the evidence suggests regime shifting during the sample
        period, but we do not know what triggers the switch and what
        specific competition pattern each regime represents.
    •   Not completely consistent with the theory: no trigger, no
        dependent across regimes, no test of unpredictable demand
    •   Price elasticity is less than one – non-optimal?
    •   Some of these questions are addressed in Ellison 1994 RAND.
            Knittel and Stango AER 2003
    •   RQ: Is clustering at price ceiling attributable to tacit
    •   Why interesting?
        •   Price ceiling is a common form of market regulation,
            other examples include minimum wages and real estate
            agent fees. The paper challenges the common wisdom that
            price ceiling cannot do any harm unless it is binding.
        •   Collusion is illegal per se, but tacit collusion is harder to
            detect than explicit collusion
Price clustering and stickiness
Knittel and Stango AER 2003 -- theory
    •   Price ceiling serves as a focal point to facilitate tacit collusion
        •    Price cluster at the price ceiling, more than it should be
        •    Collusion is more likely with lower price ceiling, higher cost, bigger
             firms, greater market concentration, and lower market demand
        •    Tacit collusion invites entry
    •   Firms charge ceiling price because the ceiling is binding
        •    If the ceiling is binding, price conforms to a distribution censored at
             the ceiling
        •    The ceiling is more likely to bind if the ceiling is low, or the cost if
        •    A binding price ceiling discourages entry
    •   Is it possible that there are secret price cuts, but unobserved in the data?
         Knittel and Stango AER 2003 –
          identified by functional form
    •   Assume error terms conform to the same normal distribution in states with
        and without price ceiling
    •   Take issuers with lower-than-ceiling rate or in states without ceiling as the
        control group to recover the non-collusive price distribution
    •   If no collusion, price ceiling censors the distribution at the right tail
    •   Deviation from such censoring may indicate collusion
    •   Note that ―ceiling‖ observations may reflect both collusion and non-
                                                                By a constant
                    _ cld
                        i o e
                    1f_ lu
              i 
            w                                                   probability,
                    oe ie
                      _ hw
                    0t rs                                       or vary by Zit

                  t i n ig
               p X j
                * i
                 t t   t ol
                        _n
                        _c  i
                            e      
                                    
             t  * _ n i _ o l e
              ih oii_n _ cd
             i  t s  n e r _
                    nb gl o t l
             p pa n d c g n o u
               o i
                tre
α = prob of collusion =
        Knittel and Stango AER 2003 –
    •   Diff-in-diff with structure
    •   Depends on normality assumption for the error term (they
        have empirically tested this)
    •   It predicts the probability of collusion
    •   It assumes away other forms of collusion (i.e. credit card
        issuers do not collude in states without price ceiling)
    •   Harm from price ceiling?
    •   Cost explanations? Default risk or price adjustment cost?
    •   The effect of price ceiling on exit? Maybe the ceiling reduces
        the number of firms, thus the equilibrium price distribution?
            Other studies about collusion
    •   Genesove and Mullin AER 2001
    •   Reality check: compare minutes of a cartel with many theories
        of cartel
    •   Inconsistencies
        •    No fix on p or q, but fix on rules
        •    Contracts are incomplete, reinforced by regular meetings
             instead of renegotiation
        •    Violations did occur
        •    Emphasize detection and the reason for deviation
        •    Mild and limited punishment except for massive violation
        •    No discussion on the length of punishment?
    Other studies about collusion

•   Bid rigging:
     –   Porter and Zona 1993 JPE, bid rotation in
         New York HW construction
     –   Baldwin, Marshall and Richard 1997 JPE,
         timber auction
     –   Porter and Zona 1999 RAND, Ohio milk
     –   Bajari and Ye 2003 ReStat, procurement
             Case Discussion

•   Price wars among Chinese airlines
•   Five players:
     –   China Eastern
     –   The other big 3 airlines
     –   Travel agents
     –   Consumers
     –   Regulators
            Theme #3: Demand Estimation
                  – some history
    •   IO emphasizes supply side decisions, but these decisions are made in
        response to demand. The equilibrium concept stresses the importance of
       demand estimation.
    •   Structure—Conduct—Performance Paradigm (SCPP)  comparison
        across industries
        •    Dep var: accounting profits, tobin’s Q, lerner index (proxy MC by
        •    Indep var: market structure (Herfindahl), R& D, min efficient scale of
             production, buyer concentration, advertising/sales ratio, etc.
    •   Chicago school: efficiency difference between firms could lead to the firm-
        size distribution observed by Herb Simon.
    •   SCPP attributed outcomes to ―market power‖, while Chicago school
        attributed outcomes to endogenous, firm-level differences.
        •    Error includes unobserved industry/firm characteristics, which in turn
             determines market structure.
         Theme #3: Demand Estimation
             – Hedonic approach
    •   Traditional demand and supply equations only work for homogenous
    •   How to estimate the demand of differentiated goods? Dennis Epple JPE
       1987 ―Hedonic Prices and Implicit Markets: Estimating Demand and
        Supply Functions of Differentiated Products‖

            Pfz 
                                                      P: price (e.g. house price)
            ( 
              )                                       Z: prod. Attributes (e.g. # of
            ' z' z A 1 1 v
                  ) 1X 1
            P) f(  ZH 
             (                                        bedrooms)
                                                      X1: demand shifters
            ' z' z A
                  ) ZH 
            P) f( 2 2 2 v
             (         X 2                            X2: supply shifters
    •   Rewrite the demand system, we have z as a function of X1 and v1, so Z is
        not exogenous in the demand
    •   With demand and supply, we can rewrite v1 as a function of v2, X1, X2. So
        X2 cannot be a valid instrument for Z in the demand equation
    •   Overall, the price equation imposes constraints on demand and supply, so
        the whole system is misidentified. It is ok to run price equation alone for
        the purpose of data description
         Goldberg 1995 Econometrica
    •   RQ: impact of volunteer export restraint.
    •   Logic: if the restraint is binding, it amounts to an extra term in
       the first-order-condition. To examine whether this extra term
        exists, we need demand estimation.
    •   Producers facing no constraint maximize:

    •   Producers facing constraint maximize:
         Goldberg 1995 Econometrica
    •   First order conditions:


    •   Demand estimates will tell us elasticity.
            Goldberg 1995 Econometrica
    •   Demand estimation:
        •    Household i chooses car j to maximize her utility:
                 h a {  Xj
                  cx j X j i
                   i g i j i 
                   e a
                   c m
                  or U t
                       j   X t}
                                            
       Goldberg 1995 Econometrica
  •   McFadden: discrete choice model (nested logit) + maximum likelihood
  •   Prob(i choose j) = prob(i choose nest g) * prob(choice j | nest g)
 •   Across nests, parameter ζ (within (0,1)) describes correlation between
      nests. Define the inclusive value of nest g as:
                       D (
                          e
                             x    )

  •   Within a nest,
                              ep       )
                     poi(jg 
                      rb | )      1
                              p1 
                                e (     )
                             jk     

  •   Across nest:                         1
                         r i( )
                         p bg
          Goldberg 1995 Econometrica
  •   IIA assumptions: branches within the same nest are independent and
      irrelevant alternatives

 •   If ζ=0, nests are independent of each other, nested logit boils down to
      simple logit. If ζ is close to 1, branches within the same nest are correlated.
      In other words, branches within the same nest are much closer substitutes
      than branches across nests.
  •   Test of IIA assumption:
      •    Try alternative specifications
      •    reject the nest if estimated ζ is out of [0,1) range, then test if ζ is 0.
      •    Hausman test for IIA property (subsample of choice set)
      •    Chi-2 test for model fitting ((predicted mktshare-actual mktshare)^2)
      •    Out of sample predictions
  •   Assume price and all product attributes being exogenous.
       Goldberg 1995 Econometrica
  •   Once we get the demand estimated, we can calculate demand
      elasticity for each car

 •   Invert the first order condition for each supplier, we can back
      out marginal cost
  •   Regress estimated marginal cost on a function of car
      attributes, and whether the car model is subject to VER
      (japanese*year dummies)
  •   If VER is binding, the coefficients of japanese*year dummies
      should be positive
  •   These coeffs allow us to say whether a specific model faced
      binding VER in a specific year binding 83,84,87,
      nonbinding 85 (why not reporting standard errors?)
  •   Demand estimation allows for policy simulation
Goldberg 1995 Econometrica Summary
    •   Plus:
        •   micro data with household demographics,
        •   detailed demand estimation,
        •   nice framework to back out supply side parameters and
            link them to policy
    •   Minus:
        •   Nests may still be arbitrary and non-exhaustive
        •   Price may be endogenous
        •   Impose first-order-condition as a constraint to identify the
            supply side: assume these FOCs being right …
                  Berry RAND 1994
    •   RQ: how to estimate the demand of
        •   differentiated products
        •   oligopoly
        •   using aggregate data only (i.e. mkt share by prod*time)
    •   Importance:
        •   Methodology contribution
        •   Aggregate data is much more common than individual
            micro data
            Berry RAND 1994: Model
    •   Cost side:
       •   Producer j maximizes her (static) profits
        •   By setting pj (Bertrand competition)

        •   First-order-condition:

        j c j )
               1 s
         , , j ( ]
          q     s
        p ( w [/ ) j           
                                                       j

        •   With p and estimated demand elasticity, we can invert FOC to estimate
            c function
         Why random coefficients?
    •   Without random coefficient, the additive separability of δi and
        εij implies IIA, which means that a consumer who substitutes
       away from a given choice (say Mercedes) will tend to
        substitute toward other popular products (popularity measured
        by market shares), not to other similar products.
    •   Random coefficients allow individual i to have a specific
        substitution pattern across cars that appeal to her
    •   More flexible and less arbitrary than nested logit. Doesn’t
        need the IIA assumption. Doesn’t need a common nesting
        structure applicable to every consumer.
    •   We can estimate the distribution of the taste (later on you will
        see that this estimation could be enriched with consumer
        distribution data or individual micro data)
    How to… (without rand. coeff.)?
    •   Assume logit errors in vij
    •   Price (pj) may be correlated with unobserved product attributes (for
        example reliability, ξj). Not accounting for this endogeneity may
       underestimate price elasticity (for example, people buy reliable cars even if
        they are of higher price)
    •   By logit formula we have
                                                               
                                ej                             e0          good
                     j 
                    s ( )                          0
                                                   s ( )
                                                           
                                 J                             J 
                               e    k
                                                             e 
                                 k0                            k0
    •   Normalize the outside good (not buy) as zero utility

                  n )  
                  ls ls j x j j
                   ( n j
                    ) (
                    j 0     p                  
    •   Run OLS+IV on this equation
    •   IV for price = the other products’ x
        •    Validity based on supply side FOC and the functional form of U
      How to… (with rand. coeff.)?
  •   Assume logit errors in vij
  •   Still need instruments for price
 •   All the previous logic goes through, but we cannot estimate it by simple
      OLS+IV. Now the market share inversion needs to integrate over the
      distribution of taste. Usually requires a numerical simulation to solve for
                  
      GMM framework, where moment conditions are:
             ( j j j  )
             x E 
           E) (( x  )0 x                 pj 

               xp
             E ) (   0
              j          j

             (  ( x  )
             x  Ej
                 j                
                                   j     j   j       j

  •   Minimize these moments’ empirical analogy  parameters
  •   See more details in BLP 1995, BLP 1999 and Nevo 1998
        Berry RAND 1994 summary
    •   Estimate differentiated demand from aggregated market share data
    •   Target two problems:

       •   The curse of dimensions
        •   model reasonable substitutions across choices  random coeff.
        •   Address the endogeneity of price  IV for price
    •   Strong assumptions
        •   utility function (linear, x-j does not appear in uj)
        •   Logit errors (convenient but not necessary)  inversion of market
        •   X-j is a valid instrument for pj,. In other words, x-j is orthogonal to
            ξj. Is Toyota’s choice of engine power really independent of Ford’s
            reliability index?
        •   This may work for short-run decision of price setting, doubtful if
            we consider a dynamic horizon where the producers choose x
                    Alternatively …
    •   Problem #1: substitution pattern
        •   Impose a nest tree (that applies to every one)
        •   Allow individual attributes (say income) interacts with
            prod. attributes. Must have individual data
    •   Inversion of market shares depends on logit error assumption
        •   if individual micro data is rich enough, we can estimate
            product-specific dummies (i.e. δj) as fixed effects. This
            will skip the inversion of market shares, but still face the
            prob. of endogenous price.
    •   Problem #2: endogenous price
        •   Classical IV problem: cost shifters?
BLP 1995 Econometrica and 1999 AER

 •   Apply Berry 1994 to Automobile industry, voluntary
      export restraints
  • BLP 1995 Econometrica emphasizes methodology
      • Aviv Nevo 1998 NBER Tech. Paper #221 ―A
        Research Assistant’s Guide to Random
        Coefficients Discrete Choice Models of Demand‖
  • BLP 1999 AER emphasizes policy implications
                Methodology: data
    • In each market, market shares, price, product
      attributes, and distribution of demographics (the last
      one is not necessary but will be a plus)
    • Better to have multiple markets, where market could
      be defined by geography, time, or product
    • Need to define the overall market size (e.g. all
      households in BLP) and in many cases, the outside
    • Products with extremely small market shares are hard
      to identify in the estimation (i.e. zero market share
      implies δj=-infinity)
                Methodology: steps
 •   Prepare a set of random draws for individual taste (δijk) from an assumed
     distribution (say normal) or from an empirical individual demographic
     distribution (say income from CPS)
•   Start form a parameter value ζ0
 •   For a given set of {δj }, we can use {δijk} and ζ0 to calculate simulated
     market shares.
 •   Equating simulated market shares and the actual market shares, we solve
     for {δj} numerically.
 •   With {δi}, we have
                         x  p
                       j 

 •                       1 
     Moment condition:    Z
                        ()   
 •   Choose ζ to minimize

              o( (
                 ' 
              b )1 ) 
                      
               s _a _
               ai e e EZ
               _s t _ Z
                  e i o '
              A ntst f ( ')
                   Optimal weighting matrix
                      Methodology: IV
 •   BLP style IV: the other products’ attributes x-j (produced by the same firm
     or rival firms)
     •    Invalid if x-j is correlated with ξj because of oligopoly decision in the
          long run
•   Nevo: If we have multiple markets (m) and introduce a brand-specific
     dummy to capture unobserved product attributes for that brand, price that
     is specific to market and brand (pjm) is still endogenous to ξjm but xj and ξj
     will be absorbed in the brand dummies. Nevo proposes pj,-m as instruments
     for pjm under the assumptions that the prices of brand j are linked across
     markets because of common cost shifters but pj,-m is independent of ξjm.
     •    Invalid if there is a national shock of ξj for every m, e.g. consumers
          now value Mercedes’ service more than before.
     •    Invalid if prices of brand j are coordinated across m
 •   Direct cost shifters
 •   A stronger assumption of mean independence (i.e. E(ω|Z)=0 instead of
     E(ωZ)=0) validates any function of Z as IV as well and therefore allow
     more room to use optimal instruments to improve efficiency.
              More on brand dummy
    •   Must have data on more than one markets
    •   Brand dummies improves fit
    •   Brand dummies absorb product attributes (observed or unobserved) that
       are common to all markets, and therefore removes some endogeneity
        problem in price
        •   To recover β, we can regress:

              nx _a 
             ai f t s 
             b f es tx j
             r_ fe e j j 
              d_ im                                     
        •   If price does not vary by market, p will appear on the right hand side
            as well and the BLP arguments apply.
    •   Draw individuals for once and all
    •   Use smooth simulators instead of the naïve frequency counts
       to calculate simulated market share
        •   Reduce simulation errors
        •   Guarantee smooth objective function so the gradient
            methods can be used for the ultimate estimation
    •   The inversion of market share to find out δj can be done by
             uus,)
              t
             (alm p
             jl t ) s t (
              n _ ia , )
                jl n e _                                          t
               Nitty-Gritty continued
    •   Find optimal weights in GMM
        •   Use A=Z’Z to calculate an estimate of ζ, and then use the estimated ζ
            to construct a new A=E(Z’ω(ζ)ω(ζ)’Z)
        •   may iterate until converge
    •   Find optimal ζ. To speed up:
        •   Separate ζ into two parts: ζ1 (linear part in δ) and ζ2 (others such as ζ
            of random coefficients).
        •   Because δ is a linear function of ζ1 and x once we know ζ2, we can
            express ζ1 as a linear function of ζ2 and focus on the optimal search of
        •   Search algorithm: gradient search, Nelder-Mead simplex.
        •   Gradient search can be sped up if we tell the computer the analytical
            form of the gradients
        •   How to ensure global optimal? (never 100% sure, but we can try ..)
                      Check the fit
    •   Do you have reasonable own price elasticity (-) and cross price
        elasticity (+)?
    •   Similar products have large cross elasticities?
    •   Fit of market shares?
    •   How do results differ with and without instruments?
    •   Overidentification test when you have more instruments than
                          Last help
Matlab codes for a random coefficient BLP model is available at
Nevo’s website
                  BLP AER 1999
 •   Research Question: Quantify the impact of the Voluntary
     Export Restraints on Automobiles from 1981 to 1990
•   Potential impact on:
     •   Consumers: -
     •   Producers: Japanese -, Domestic +, European ?
     •   Government: 0 (- if compared to tariff)
                        BLP AER 1999
    •   Methodology: follow BLP 1995 Econometrica
    •   Data: P (listing), Q and attributes by model-year

                         BLP AER 1999
    •   Findings:
        •   VER binding in 84,86,87,88,89,90, not binding in 81,82,83 and 85.
       •   Producers – bindings years only
              •     Domestic producers do not raise price, but sell more cars.
                    Profits go up by 1.5-3 billions per year. Total gain 10.1 billion.
              •     Japanese producers raise price, but sell fewer cars. Profits go
                    down by 0.1 billion per year. Total loss <1 billion.
              •     European producers lower price slightly, and earn 0.1 billion
                    more profits per year. Total gain <1 billion.
        •   Consumers are worse off ($-41 per HH in 1987), especially those who
            purchase Japanese cars. Total loss $13.1 billion.
        •   An equivalent tariff will transfer 11.2 billion from Japanese producers
            to US government.
                        BLP vs. Goldberg
    •   Why same question, but different answers?
         •    Both assume consumer tastes among different car models do not
              change over time
        •    Changes in observable car models are exogenous and independent to
              unobserved car attributes (e.g. Japanese do not introduce ―better‖ cars
              after VER)

        BLP                                   Goldberg
        1971-1990 aggregate data              1983-1987 Ind. Micro data
        Listing price                         Transaction price

        Random coefficients allows for        Impose one nested structure on all
        individual specific substitution      consumers, but allow tastes to
                                              differ by demographics

        With trend                            Without trend
        IV for price                          Assume exog price
                   Petrin 2002 JPE
 •   RQ: Who gains and who loses from the introduction of minivans?

•   Importance:
     •   New products occur frequently in markets. How to value a new
           •   Ex post: we can evaluate anything by assuming revealed
           •   Ex ante: we cannot value something that is completely new, but
               we can consider a new combination of existing attributes
     •   First mover advantage?
     •   Methodology: combine micro and macro data
                       Petrin 2002 JPE
    •   Background:

       •   Ford initiated the first idea of minivan in early 1970s, but the idea did
            not go anywhere due to the concern of cannibalizing their strong
            station wagon sales.
        •   Chrysler introduced Dodege Caravan in 1984 with immediate success
        •   GM and Ford quickly responded with their own minivan in 1985 and
        •   Chrysler dominated the minivan market over time
        •   Wagon sales declined
    Petrin 2002 JPE

                     Petrin 2002 JPE
    •   GMM
        •   Aggregate moments


                            X = observable car attributes
                                                            Instruments for P
                            W = observable cost shifters

        •   Micro moments (predicted purchase by demo = actual avg. in CEX)
A glimpse of micro moments about family vehicles
           More price sensitive 
                                         Why higher
 Demand                                  income HH
                                         are more
estimates                                price


Every consumer benefits from the introduction of minivan, if she buys a
new vehicle
    Predict producer change

             Petrin 2002 JPE summary

 •   Improve demand estimates by matching in micro moments

•   Decompose the welfare gain/loss from a new product
 •   Same criticisms as in BLP
 •   Two inconsistencies:
     •   Dynamic supply but assume static demand
     •   With X changing over time, is it still a good IV for p?
        More materials on demand estimation
                Debate about BLP
    •    BLP 2004 JPE ―Differentiated Products Demand Systems from a
         Combination of Micro and Macro Data: The New Car Market‖
         •   Micro data specify each consumer’s first choice and second choice
               •   Identify δj as product fixed effects, no need for IV
               •   Use the estimated δj to identify price elasticity, need IV
    •    Knittel and Metaxoglou ―Estimation of Random Coeffcient Demand
         Models: Challenges, Difficulties and Warnings‖ available at
    •    Dube, Fox and Su ―Improving the Numerical Performance of BLP Static
         and Dynamic Discrete Choice Random Coeffcients Demand Estimation‖
         available at
        More materials on demand estimation
                  online tutorials

    •    William Greene and David Hensher ―The Mixted Logit Model, State of
         •   recent advance
         •   Caveats to avoid
    •    Kenneth Train: An internet course on ―Discrete Choice Methods with
         Theme #4: Information Issues
•   Imperfect and/or incomplete information characterizes a large amount of
    IO theory
     •   Avinash Dixit’s talk at lunch honoring the 2001 Economics Nobel
         Laureates George Akerlof, Michael Spence and Joseph Stiglitz
•   Information rents (in auctions    )
•   Adverse Selection
•   Moral Hazard
•   Reputation and Learning
•   Explicit Information disclosure
     •   Advertising
     •   Third-party verification
     •   Government regulation
        Hendricks and Porter 1988 AER
    •   RQ: How does asymmetric information explain the bidding
        behavior in federal oil drainage auctions?
         •   asymmetric / symmetric information
        •   collusive / competitive bidding
    •   Why important?
         •   test the theory on asymmetric bidders
         •   antitrust policy – competitive or not?
         •   social planner’s point of view = efficiency:
                  •   is the oil field allocated to the party that values it
                      most or has the lowest cost draining the oil?
         •   government’s point of view
                  •   maximize government revenue from the auction?
                  •   redistribution between government and bidders
               Hendricks and Porter 1988 AER

•   Setting:
     •   each field represents a
        first-price sealed common
         value auction
     •   adjacent fields may be
         positively correlated in
•   Two key data elements:
     •   Researchers observe the
         actual profitability ex post
     •   Drainage vs. wildcat
                Information Asymmetry
    •   Informed: neighbor firms
         •   assume neighbor firms are equally informed
        •   automatically holds if neighbor firms collude
    •   Uninformed: non-neighbor firms
    •   Neighbor firms are ―better informed‖
         •   Non-neighbor firms observe public signals about the field’s
         •   Neighbor firms observe public signals + some private signal
         •   Neighbor firms’ information set dominates non-neighbor
             firms’ information set.
•   Basic facts
     •   Neighbor firms’ participation is a positive signal of the
         potential profitability
     •   Neighbor firms, if participate, are more likely to win
     •   Neighbor firms earn positive profit, but non-neighbor firms
         earn zero profit on average
•   Two thoughts
     •   Non-neighbor firms do not exit the market because their
         entry enforces a contestable market so that neighbor firms
         can only earn rents from information or cost advantage. They
         cannot earn abnormal profits in addition.
     •   Information advantage may disappear if neighbor firms bid
         against each other
  •   Neighbor firms have better information

  •   Neighbor firms has cost advantage
      • Both are more evident for drainage sales than
        for wildcat sales
  •   Neighbor firms are colluding
      • difficult to identify potential ring members for
        wildcat sales
      • why do firms collude more in drainage fields?
        Case #1: information advantage only

    •    Informed bid:                    public signal

             I H V Z
             B f( (| , )
                I  EX )
                                        private signal
    •    Uninformed bid:
                   B  fU(Z
                    U      )

    •    Then:
                 EBU ( | )
                  ( ) B H
                 H I

          • In words, E(BI) over H is close to Bu conditional
            on E(H|Z)
    •    So the ex ante distribution of BI should be equal to
         the distribution of BU if we only condition on Z
                   Tests of Case #1

    •   First,
                    Z e
                  B  I  I
                                      test: ζI = ζU,
                                     ζe(I)= ζe(U)
                  U Z
                  B   UU

    •   Second,
                       I
                    H e
                 I 
                 B ZI I             test: γI > 0,
                 U    U
                    U e
                 B ZU H               γU = 0
          Case #2: cost advantage only

    •   Informed bid:             public signal
                 B f(  )
                  I   Z c
    •   Uninformed bid:
                  B  f (Z)

    •   In this case, BI and BU should be systematically
        different ex ante when we only condition on Z.
    •   Also, the ex-post-revealed H should be insignificant
        in both BI and BU.
                   Tests of Case #2

    •   First,
                    Z e
                  B  I  I
                                      test: ζI NE ζU
                  U Z
                  B   UU

    •   Second,
                       I
                    H e
                 I 
                 B ZI I             test: γI = 0,
                 U    U
                    U e
                 B ZU H               γU = 0
    •   Maximum likelihood, truncated above R, endogenous
            l (t/ t i
             o i RY i
              gB ) i W i
                      t t t         
             iI U

    •   Likelihood:
                   Case #3: collusion
    •   If competitive, π’(# of neighbor firms) <0
    •   If competitive, b’(# of bidders) >?< 0
    •   Sometimes one neighbor firm participates, but other
        neighbor firms don’t.
         • cannot explain this by some non-participating neighbor
           firms receiving bad (negative) signals.
    •   If more than one neighbor firms participate and they collude,
        then all of them should bid lower than what they do non-
    •   If neighboring firms collude, how do they divide the extra
        profits? Or do they simply rotate participation?
    •   Asymmetric information – yes!
    •   Cost advantage – maybe but unlikely
    •   Collusion – likely
    •   More in Porter 1995 Econometrica ―The role of information
        in US offshore oil and gas lease auction‖
                    Other auction papers:
•   How to estimate cost or value from bids of first price auctions
     • Laffont, Ossard and Vuong (1995 Econometrica), must know the
       number of potential bidders (N)
     • Guerre, Perrigne and Vuong (2000 Econometrica), non-parametric
       estimation, must know N
•   Auction with future resale: Phil Haile (2001 AER) Haile (2003 JET)
•   Test of common vs. private values: Haile, Hong and Shum (2007 wp)
•   Spectrum Auctions (Bajari and Fox 2007 wp)
•   Treasury bill auctions (Ali Hortacsu)
•   Structural estimation reasonable? (Bajari and Hortacsu 2005 JPE)
•   eBay auctions (Bajari and Hortacsu 2003 RAND, Greg Lewis 2007)
•   Experimental evidence (winner’s curse, etc. John Kagel, Ohio State)
        Adverse Selection and Moral Hazard
   Information asymmetry Arise from insurance: for example, life
   insurance sold to a population that includes smokers and non-
   Adverse selection:
Akerlof (1970) ―market for lemons‖
asymmetric information before contract
   Model hazard
asymmetric information after contract
   Research topics
          does adverse selection or moral hazard exist?
          the extent of adverse selection and moral hazard?
          the consequence of adverse selection and moral hazard?
          the market/regulatory mechanisms that solve or alleviate the
          information asymmetry?
    How to Identify Info Asymmetry?
•   Use price data to infer the quality difference between
    markets A and B, where A is subject to more information
    asymmetry. The approach depends on:
    • both markets are ―competitive‖
    • the markets are comparable in other dimensions, so we
      can attribute the difference to different info
    • whether we can control for all the other factors that
      affect price
•   Use ex post data to directly measure info asymmetry
•   Both approaches assume that info asymmetry has not been
    completely solved by observables
     Chiappori and Salanie 2000 JPE
•   RQ: is there information asymmetry in French auto
•   Main predictions from Rothschild and Stiglitz 1976 QJE
    ―Equilibrium in competitive insurance market‖
    • observationally equivalent agents will face a menu of
    • within the menu, contracts with more comprehensive
      coverage are sold at a higher (unitary) premium
    • within the menu, contracts with more comprehensive
      coverage are chosen by agents with higher accident
           Alternative Explanations
•   unobservables other than agent risk : e. g. risk aversion
    (sometimes even imply advantageous selection)
•   difference in ex ante risk (adverse selection) vs. difference
    in ex post risk-taking actions (moral hazard)
•   differ not only the probability of accident but also the
    severity of accidents
•   Assumptions:
     • observationally equivalent from the insurer's point of
       view (e.g. insurers observe driving history but driving
       history could be endogenous)
     • market is competitive (i.e. free entry and zero profit)
     • pricing may be dynamic and nonlinear, including other
       tools such as deductibles, experience rating
             Background and Data
•   French car insurance, mandatory
•   RC (cover responsibility) and TR (cover responsibility and
    self damage), with various deductibles
•   all insurers apply a uniform experience rating system
•   insurers can overprice young drivers (<=3yr), up to some
•   data: a 5% survey of insurance contracts conducted by the
    French federation of insurers (FFSA)
•   1989, cover 1.12 million contracts, 120K accidents
                Sampling strategy
•   Focus on 20,716 young drivers to avoid the potential
    omitted variable bias (e.g. no need to worry about
    endogenous driving history).
•   control for an extensive set of observables
•   only observe claimed accidents so there could be an ex post
    moral hazard.
•   Solution: discard all accidents in which only one automobile
    was involved. Assuming claim will be made for every
    accident that involves two parties.
•   Probit:
        y=TR or not
        z=accident at fault or not
        cov(ε,ε)>0 if adverse selection
                (result: -0.029 with std err 0.049)
•   conditional on a cell with ―homogenous‖ members, compute

                                                    ~ χ2(1)

•   summarize across cells:
    K-S test:                                     (0.63<cril value 1.35)
    Count # of cells that reject H0 by T~B(# of cells, 0.05) (5/64)
    Sum all T ~ χ2(# of cells) (p value = 0.15)
            Why no info asymmetry?
•   Young drivers do not know their own risk
        slightly older drivers who do not have accidents in the past
        same conclusion
•   Moral hazard or adverse selection?
•   Include a dummy of premium discount on the RHS of probit, cost
    of accident is lower if this dummy is equal to one
•   If this discount is assigned randomly and there is moral hazard,
    this new variable should have a positive coefficient on y (buying
    TR) and z (more accident). → result is the opposite
•   Either adverse selection does not exist or some form of
    advantageous selection cancels off adverse selection
    what does this imply for the market competition?
    Advertising [Bagwell “The Economic
         Analysis of Advertising”]
•   Consumer response to advertising
     •   Persuasive: consumer taste changes as a result of advertising,
         enhance brand loyalty, create barrier to entry, wasteful and anti-
     •   Informative: direct (price, location) / indirect (quality), reduce
         consumer search cost, pro-competitive
     •   Complementary to the advertised product: social prestige
•   Firm’s optimal strategy in advertising and related areas (pricing, entry,
    R&D, etc.)
•   Regulation? Examples:
     •   Lift the ban of price advertising of liquor stores in RI
     •   Lift the ban of price advertising in eye-glasses
     •   Lift the ban of health-claim advertising by cereal producers
     •   FDA clarification on direct-to-consumer advertising of pres. drugs
    Ackerberg RAND 1997 and IER 2003
•    RQ: Does advertising of a non-durable product have an
     informative effect and/or a persuasive effect on consumers?
•    Informative: existence of the product, signaling quality, etc.
      • Experience good: consumers learn the quality for sure
        after some purchase history.
        informative only for non-experienced consumers
•    Persuasive:
      • Gain utility from advertising via taste change, social
        prestige, peer effects, etc.
        apply to both experienced and non-experienced
    Ackerberg RAND 1997 and IER 2003
•   If both advertising and experience are randomly assigned:

    u o ip  
    b n Ae
    y t y  x d
     r u
    __     d  x
            n A p          
•   Informative if β>0 and γ=0
•   Persuasive if β>0 and γ>0
    Ackerberg RAND 1997 and IER 2003
•   But advertising exposure and experience may both reflect
    some selection. Here:
    • Discrete choice model.
    • Advertising : exogenous
    • Experience driven by omitted HH characteristics
            •   HH random effects
            •   Dep var = the whole purchase sequence per HH
    • Alternatively,
            •   Chamberlain’s conditional logit (which requires
                long panel)
            •   linear probability model: IV?
           Ackerberg RAND 1997
•   RAND model:


         HH random
                              The whole purchase
                              history of HH i

                     Account for the endogeneity of
                     purchase history due to αi
    Ackerberg RAND 1997 and IER 2003
•   RAND results: Table 5

              Ackerberg RAND 1997
•   Alternative explanations?
     • Competitors’ advertising cancel out each other’s
      persuasive effects
     • Experienced buyers get less prestige effects
•   If advertising is primarily informative, why keep on
     • Advertising reinforces consumers’ memory
     • Quality change over time
     • New products within the brand
     • Consumer taste change / new consumers flow into the
     • Other brands introduced in the market
               Ackerberg IER 2003
•   Why structural:
    • More details of learning: exp. consumers may like or
      dislike the brand  greater dispersion ex post
    • Calculate the value of advertising to consumers
    • Simulate firm strategy in terms of optimal advertising
      and optimal price
    • Welfare implications
    Ackerberg IER 2003 – basic model
•   Model: HH i maximizes:

       i  t t
       ( 
      cI) 
                 t | ]
                    i I
                    c t

    i  1
    U it  
    1t  pi  m  
                
              t                    t
                                   i1 3t

                                              Prestige Effects
    HH i’s taste on
                                              of Ads
    estimated as HH
                       Ex post belief of quality
    random effects
                       (ads affect this term via
                       informative effects)
        Ackerberg IER 2003 -- learning
•   Uncertain about the true quality (δi) and the mean
    advertising intensity (a)
•   Observe two signals:
     • Experience: δit+1= δi + vit+1 where vit+1 ~ i.i.d. N(0,ζv2)
     • Advertising: ait = a + ξit where ξit ~ i.i.d. N(0,ζξ2)
•   Initial prior:
       
          Signaling effect of advertising
      a0 1
       N    
        
         
                                         
                                               2  2         2

            m               
                     
                  
         i                                     i  j        1j
       ~m
           ,                             
                     0      a          0         2         22
      a                    0     0         
                                               1j         1j

•   Baynesian updating:
                                   i  
                                 ~0  a , i
                               N  t t
                                   m     
                                 i  
                              a      m t
     Ackerberg IER 2003 -- learning

•   Forward looking:

V 
i i i i i

 [      p 
m ,  ,,, ,
( t tt
  t     )
EU     (  ]
xp ,, ) |
a ( , m V mc
           )      a
            i i i i i
            t t t t t1   i i i i i
                         i1t t t t t
t ,

           today         Tomorrow and
                         thereafter conditional
                         on today’s info
       Ackerberg IER 2003 -- learning

•   Intuition:
     • No learning: purchase decision never changes
     • Learn in one period: purchase decision changes after
       one-consumption, but not thereafter
     • Baynesian learning: purchase decision converges over
       time, some HHs may converge to always buy, and some
       HHs may converge to never buy, but this is gradual
     • Myopic vs. forward looking: more likely to take a
       myopically suboptimal choice  experimentation
                  Ackerberg IER 2003
•   Oveall, the IER model adds:
     •   Informative effects = signaling effect of underlying quality
     •   Baynesian learning structure
              •   Learn about the mean experience utility δi and the
                  mean advertising intensity a
              •   Learn from each ad exposure and each purchase
                  experience. If ads are informative, ads and purchase
                  experience are substitutes in learning about δi
              •   Learning may be gradual
              •   The posterior δi(t+1) enters U as informative effects,
                  and the posterior mita enters U as persuasive effects
     •   Forward looking: max Vi=E(Uit+βVit+1)
                 Ackerberg IER 2003
•   Estimation: simulated likelihood of entire purchase history:
     •   Given parameters:
            •   Given today’s choice : solve the dynamic
                 programming problem for optimal future choices
             •   Compute likelihood on the entire purchase history
                 (Note: logit form simplifies the likelihood
     •   choose parameters to maximize the total simulated likelihood
                         Ackerberg IER 2003
     •      Results: Table 2

                       Myopic w/ trend   Dynamic w/ trend   Dynamic / No
                                                            Prestige Ad

Θ1(price)              -5.54***          -5.49***           -5.49***
Θ3(Prestige)           0.00              -0.02              0
ζi (var around δ)      1.78***           1.84***            1.82***
ζj (prior var of δ)    1.66              0.65**             0.59**

ρ (corr coef in prior, 0.67**            0.13***            0.12***

δ (mean quality)       -0.72             0.90***            0.90***
LogL                   -3942.3           -3943.7            -3944.0
                  Ackerberg IER 2003
•   Welfare analysis
     •   conditional on the assumption how ads signal quality
     •   assume firm optimally sets a single mean price and a single
         advertising intensity in the introduction periods  marginal
         cost of production and cost per unit of advertising
     •   assume advertising bans  reoptimize price and ads
     •   compare consumer welfare and firm profits with and without
              •   ad provides valuable information
              •   ad increases cost and price
              •   welfare: ambiguous in theory, here worse with ads
              •   the observed advertising is non-optimal?
                Chintagunta, Jiang and Jin
             “Information, Learning, and Drug
          Diffusion: the case of Cox-2 Inhibitors”

      •     Two types of learning
      •     Separate (a) learning from patient experience, and (b) information
            from other sources
Information about a RX

Doctors are uncertain about drug

  – General effects applicable to all patients

  – Idiosyncratic match between drug and
Where to learn information?

• FDA decisions

• Patient feedbacks

• Other healthcare providers

• Academic publications

• Newspapers

• Manufacturer advertising
                 FDA regulation
•   Before FDA approval
     – Require clinical trials
     – Short-term efficacy and side effects (general)
     – Voluntary human subjects
     – Compared with placebo

•   Post-marketing
     – Long-term clinical trials
         • 33% mandated, 24% ever completed (TCSDD 2004)
         • Report subject to the will of manufacturers
     – FDA MedWatch: voluntary, ad hoc, passive
     – GAO 2006: “lack of criteria for determining what safety action to take and
       when to take them”
     – IOM 2007: dysfunctional
Proposals for FDA post-marketing

    • Mandate (the reporting of) post-marketing clinical trials
    • A separate agency focusing on post-marketing surveillance
    • Pooling information from daily practice nationwide
    • Intensify academic detailing
    • Report emerging information in a timely manner
    • Ban direct-to-consumer advertising for the first x years after new
      drug approval

      Key Question: how do doctors process all the available
      information and apply it to prescription choice?
     Research Questions

 How do different sources of information
  affect doctor’s prescription choice?

 What happen if
   – No FDA updates
   – Intensify academic publications
   – Nationwide sharing of patient feedbacks?
Contribution to the literature

• Combine across- and within-patient learning
   – Literature on across-patient learning:
       • Ching (2005), Coselli and Shum (2003), Narayanan et al. (2005)
   – Literature on within-patient learning:
       • Crawford and Shum (2005)

• Unique data
   – Patient satisfaction
   – Manufacturer advertising (all 4 types)
   – News coverage and medical articles
   – FDA updates
                IPSOS Data

• Marketing research company, IPSOS, tracks the drug use of
  a national representative sample of patients
   – total 16,000 households
   – six years 1999-2005

• Reports every RX received by the sampled patients
• Include self-reported patient satisfaction since 2001
   – Effectiveness
   – Works quickly
   – How long does it last?
   – Side effects
   – Easy to take
             Regulatory History

Celebrex   Jun. 7, 2002     Labeling change for no advantage in GI risk

           Dec. 23, 2004    Public Health Advisory on increased cardiovascular risk in
                            association with Cox-2s and traditional NSAIDS

           Apr. 7, 2005     New labeling that highlights cardiovascular risk

                            New warnings concerning reduced GI risk and increased
Vioxx      Apr. 11, 2002    cardiovascular risk based on the Vioxx Gastrointestinal
                            Outcomes Research (VIGOR)

           Sept. 30, 2004   Withdrawal (voluntary by Merck)

Bextra     Nov. 15, 2002    New warnings on life-threatening skin irritations

           Dec. 9, 2004     More warnings on skin irritations and cardiovascular risk

           Dec. 23, 2004    Public Health Advisory on increased cardiovascular risk in
                            association with Cox-2s and traditional NSAIDS

           Apr. 7, 2005     Withdrawal (by Pfizer)
               Preview of Results
• RX choice depends on many sources of information
   – Doctors learn from patient feedback
      • Within-patient learning more important than across-learning
   – Academic articles have negative impact on RX, News articles have
     positive impacts
   – FDA updates trail behind the medical literature
   – Patient feedback and academic articles provide different information
     about Cox-2s, and hence do not substitute for each other

• Counterfactuals:
   – No FDA updates have little impact on drug diffusion
   – Nationwide sharing of patient feedbacks increase Cox-2 market shares
     by ~20%
   – Doubling academic articles reduces Cox-2 market shares by 25-30%

• Data description
• Basic evidence of learning
• Structural model set up
• Structural estimations
• Counterfactual simulations
• Conclusion
                                  Cox-2s Advertising

                                  Figure 5: Trend of Total Promotion Expenditure
                                          (DTL+DTCA+Journal, 1999-2003)



US Dollars

             40,000,000                                                                                                                            Vioxx
             30,000,000                                                                                                                            Bextra
                                                                                                                                                   All three


           Medline Articles

                               o eea
                                s +tl
     iu : d a e ih y
     g e M et
     F r7 enr lse t b
               cw e
                  gd           ete
                               na i
      m a ( o J Ai e
       a c 1nA a
        c o
     i p f tr =e Mt l)
                  rc           a l st
                                m a
                                  e s
                               s p tr
                                m d
                               al e







19    00
      20   21
           00   00
                22   23
                     00   00
                          24   25
                               00   00
                        News Articles

          Figure 8: Lexis-Nexis articles weighted          negative
     by circulation (1=one Wall Street Journal article)    Sample starts
                                                           Sample ends







 1999        2000     2001      2002      2003      2004   2005       2006
                 Patient satisfaction
                 (1=extremely satisfied
                5=extremely dissatisfied)

                           allothRx   Celebrex   Vioxx   Bextra

Effectiveness                1.90       1.81     1.83     1.94

Side effects                 1.98       1.81     1.89     1.82

Works quickly                2.03       1.94     1.99     2.00

How long does it last?       2.04       1.93     1.96     1.98

Easy to take                 1.51       1.38     1.38     1.40

Average effectiveness        1.99       1.89     1.93     1.97

Average across five          1.90       1.77     1.81     1.83
    Basic evidence of learning
• Patients switch, and become less likely to switch over the
  course of treatment
• Average switching rate

   – Celebrex (16%)
   – Vioxx (19%)
   – Bextra (23%)
   – Other RX (9%)
• Switch less when more satisfied
• More likely to choose a brand that has greater satisfaction
  across previous patients
              Structural model

• Assume doctor is a perfect agent for the patient, because we have no
  doctor id.
• Doctors share patient experience within a geographic area (nationwide,
  region, division, DMA)
• Focus on prescription choice within NSAIDS (traditional, Celebrex,
  Vioxx, Bextra)
• Doctor considers all the drug information available up to t, but no
    – no strong evidence indicating forward-looking
    – 56% of patients disappear after one RX
    – Potential risk of mal-practice is likely to prevent doctors from experimenting
            Model setup (1)

• Patient p’s CARA utility from a prescription of drug j

• True effect of drug j on patient p is

       Qpj = Qj + qpj

• Doctors are uncertain about :

Qj =Overall quality of drug j that applies to every patient

qpj =Match value between drug j and patient p
           Model setup (2)

• Doctors have priors about Qj and qpj (i.i.d.)
• Each prescription generates a signal

            Q qj p
       R  0R( j p) vjt
        t     .
       : S l fc r
        0 R     a
              ceat so
             0   2
       vjt ~ ( , v)

• Based on patient experiences, doctors form
  posteriors on Qj and qpj
Posterior update

                   Patient feedback

        Choice probabilities

U p p Qj x  p
p p j Q
j U j
tt t
    X
      ~ p jpz
          Zt t
              j t
                j  
                    
                 Absolute risk
                   aversion           Distinguish
                                     Cox2s versus
            x( j
            ep pt )                                  Include FDA
Pjt  J
rp                                                     updates,
             x k
             e ( p
                                                    news reports,
                            RE                      medical articles
        1 2
U  pt   Qjt  xX  zZ
 pt Qj    p  p  jt
                  j t
Qbar0j:   prior of each Cox-2 relative to non-Cox2s, identified from initial market shares
α0, αR:   scale factors of R, αR identified from the correlation
          between drug diffusion paths and patient feedbacks
σv:       error in signal, identified from observed heterogeneity in R
σQo:      dispersion of the prior Q, identified by the speed of diffusion
σqo:      dispersion of the prior q, identified from the speed of within-
          patient updates
βxj:      effects of demographics, identified from RX pattern across different patients
βz:       effects of non-patient info flows, identified from co-movements
          of drug market shares and the info
λpj: patient-drug match observed by doctors only, identified from time-invariant patient-drug
γ:        risk aversion, distaste of the posterior variance, identified from
          functional form in Bayesian updating
            Estimation Sample
• New patients starting on or after 01/01/2001, till 12/31/2003
   – 6,577 patients
   – 17,329 Rxs
• Cover 9 census divisions, assume info pooling by division
• Control for age, sex, income, edu, and 3 insurance variables
• Ignore drug copay (data dirty, 80%+ insured)
• No formulary info
• Include (inversed) advertising, news, med articles, FDA updates
           Structural Estimation
• Step 1: We regress Rpjt on a full set of patient-drug (pj) dummies,
  and compute the residuals’ standard deviation.
   – gives us an unbiased estimate of συ.
   – R-square 0.697, we get συ = 0.496

• Step 2: Use this value to estimate the remaining parameters
               Basic Model
          (both learning, no RE)

α0                     -16.735        ***
αR                        2.569       ***
γ                        set at 0
ζv                     set at 0.496
Qbar0 (celeb’x)           2.697       ***
Qbar0 (vioxx)             2.324       **
Qbar0 (bextra)            2.309       **
ζQ0 (celebrex)            0.018       ***
ζQ0 (vioxx)               0.020       ***
ζQ0 (bextra)              0.029       ***
ζq0                       0.307       ***

*** p<0.01.
              Across- and within- learning

• Compare:
    1) Basic structural model (logL=-11376)
    2) Across-patient learning only (-17259)
    3) Within-patient learning only (-11565)
    4) Conditional logit with no learning structure (-17226)

• Both learning are statistically significant, but across-only fits the data worse than the
  models of within-learning

• This suggests that most learning of patient satisfaction is within-patient (qpj) rather
  than across-patients (Qj).
           Why do non-negative Medline articles
           have negative impact on RX choice?

• Composition of articles
    – positive (28%) neutral (58%) negative (14%)
    – efficacy (57%) side effects (38%)
    – Author affiliation (drug company 8%, university 73%)
    – Company affiliated articles are more likely to be positive
    – Positive articles are more likely to focus on efficacy

• Basic model with more detailed Medline variables
    – Negative response to non-neg articles are driven by negative response to
         positive articles
    – Negative response to company affiliated articles
    – Collinearity if we include many Medline variables
    – Other results robust
                 Unobserved heterogeneity

                                Basic Model      Basic Model + Basic Model +
                                                 2 patient-type 3 patient-type
Probability of type 2             0              0.415     ***  0.324     ***
λ (type 2 on Celebrex)            -              4.178      ***    3.349    ***
λ (type 2 on Vioxx)               -              3.979      ***    3.086    ***
λ (type 2 on Bextra)              -              4.356      ***    3.305    ***
Probability of type 3             0                                0.403    ***
λ (type 3 on Celebrex)            -                                -2.767   ***
λ (type 3 on Vioxx)               -                                -2.885   ***
λ (type 3 on Bextra)             NA                                -2.082   ***
Log L                          -11376            -10181            -10086

All the learning coefficients are similar across specifications.
         Other robustness checks

• Sampling weights
   – Results are qualitatively similar
   – Within-patient learning is even more conspicuous
• Functional form of advertising
   – We tried:
       • Total or log(total) instead of inv(total)
       • Detailing and DTCA separately
       • Flow instead of cumulative sum
       • Estimate depreciation rates
       • Lag advertising by 3,6,9,12 months
   – Advertising coefficients are sensitive, but other learning variables
   – Take advertising as pure control
                              Fit of Model

• Preferred model: Basic + 3 patient-types
• For each month-drug, our prediction of each Cox-2’s market
  share deviates from its actual share by 26.5%.
• Predict the RX choice as observed in data:

Logit without                 Basic, within- Basic, across-     Basic + 3
  learning      Basic model   learning only learning only     patient-types
  structure                                                    (preferred)

  61.17%          78.96%         78.35%         60.58%          85.49%

                                % Change in the market share of

                      Celebrex       Vioxx       Bextra      All others

#1: No FDA update      1.07%         0.78%       1.58%        -0.63%

#2: Nationwide
sharing of patient    15.01%        20.82%      21.05%       -10.74%

#3: double academic   -30.44%       -25.36%     -27.11%       17.03%
                            In summary ...

• Prescription choice depends on many sources of information

• On patient satisfaction

    – Doctors learn both across- and within-patient, but within-patient
      learning is much more important

    – Doctors held a strong prior on the average drug quality (defined by FDA
      approval?), so learning is gradual

• FDA updates add no new info upon the medical literature

• Patient feedback and Medline articles provide different information about
  Cox-2s, and hence do not substitute for each other

• Value of observing actual signals
   – Identifies the effect of actual signals from other confounding factors
   – Dramatically simplifies computation
       • no need to simulate signals
       • no need to estimate true quality
       • observe the noise of signals, no need for normalization

• Why no forward-looking?
   – Malpractice concerns, 56% of patients stop with one RX
   – No strong evidence in favor of forward looking
   – Difficult to identify discount factor from risk aversion
              Limitation and future work

• Limitations
   – We focus on a dramatic class, results may not applicable to
     other drugs
   – We focus on Bayesian learning
   – Patient satisfaction may not reflect all the dimensions of
     drug quality
   – We don’t address the potential endogeneity of advertising

• Future work
   – Learning during and after the withdrawals of Vioxx and
                    Quality Disclosure
• Motivation:
      –    sellers often have better information about product quality.
      –    Akerlof (1970): could lead to market shutdown
      –    sellers lack incentive to maintain high quality
      –    mismatch of consumers and products
• Quality assurance mechanisms
       –    brand / advertising
       –    word-of-mouth / reputation
       –    manufacturer warranties
       –    licensing / minimum quality standard
       –    quality disclosure
• “We define quality disclosure as an effort by a certification agency
  to systematically measure and report product quality for a
  nontrivial percentage of products in a market.“
                  Quality Disclosure
• Three types:
      –    Industry initiated voluntary disclosure
              • the Joint Commission on the Accreditation of
                 Healthcare Organization
              • the Movie Picture Association of America
      –    Government mandated disclosure
              • 1933 Securities Act, 1934 Securities and Exchange
              • 1990 Nutrition Labeling and Education Act
      –    Market-provided disclosure
              • bond ratings (since 1909))
              • college rankings (US News since 1983)
              • Consumer Reports (since 1936)
              • may or may not need seller consent
              • certifiers may or may not be profit driven
          Key Players and Key Questions
• Key players
      –    sellers
      –    buyers
      –    certifiers
      –    gov could be sellers, buyers or certifiers
• Key questions
      –   How do consumers respond to disclosure? Does the response
          depend on the source, content and format of information?
      –   How do sellers respond to disclosure? Why do some sellers
          disclose but not others? Do sellers improve quality after a
          disclosure system is in place? Does disclosure drive out low
          quality sellers?
      –   Do we need mandatory disclosure, or will the market provide
          sufficient quality assurance in the form of voluntary or third-
          party disclosure?
      –   What is the economics of certifiers? Do they have incentives
          to be truthful and thorough?
• Unraveling results
     – Grossman 1981, Milgrom 1981, Jovanovic 1982
• Why unraveling fails in reality?
     – disclosure cost
     – oligopolistic competition
     – sellers do not know their own product quality
     – consumers do not pay attention, do not understand, or
         do not infer correct information from disclosure signals
     – consumers are heterogeneous in an oligopoly setting
     – consumers do not know the underlying distribution of
     – standard of certification is unclear
     – fear of obligation to disclose in the future
     – high quality sellers face price regulation or capacity
                  Theory (continued)

• Is it necessary to mandate seller disclosure?
        – if unraveling fails due to disclosure cost, mandatory
           disclosure is socially excessive
        – unintended consequences
              • sellers choose to be uninformed
              • sellers game the system
              • rationing of high quality goods
• Economics of certifiers
     – disclosure by consumer feedbacks has shortcomings
     – expert certification is not always honest and revealing
     – reputation plays a limited role
     – conflict of interest between certifiers and sellers
     – consumer guides are not perfect either
  Empirical Studies on Quality Disclosure

• Measurement and reporting
     – noise: sample size and mean reversion
     – risk adjustment
     – weighting
• Does unraveling hold in reality?
     – who discloses and who does not?
     – are disclosing firms always better than non-
        disclosing firms?
     – does self-policing work?
  Empirical Studies on Quality Disclosure

• Does disclosure improve consumer choice?
     – key identification issues
     – how to quantify value of information
• Does disclosure improve quality?
     – real improvement
     – selective improvement
     – gaming/cheating
• certifier behavior
      – grade differentiation
      – does competition help?
      – public vs. private certifiers

           Ginger Z. Jin
     Department of Economics
      University of Maryland

           Phillip Leslie
    Graduate School of Business
        Stanford University
• In November, 1997, CBS 2 News in Los Angeles aired
  “Behind the Kitchen Door”

• January 16, 1998, LA county restaurant inspectors start
  issuing hygiene grade cards

   –   A grade if score of 90 to 100
   –   B grade if score of 80 to 89
   –   C grade if score of 70 to 79
   –   score below 70 actual score shown

• Grade cards are prominently displayed in restaurant

• Score not shown on grade cards

• Full details of every restaurant inspection in LA
  between Jan 1996 and Jul 2002

• Quarterly revenue for each restaurant in LA for
  1996, 1997 and 1998

• Number of people admitted to hospital with specific
  illnesses, in each month, in each 3-digit-zip code,
  for all California from 1993 to 2000
• What is the impact of the grade cards on
  – consumers’ restaurant choices
  – restaurants’ hygiene quality
  – incidence of foodborne illness?

• Why did some restaurants have high hygiene
  scores before grade cards?

• Do grade cards change the behavior of restaurant
Impact of Grade Cards on
  Consumers’ Choices
Impact of Grade Cards on
  Consumers’ Choices
          Impact of Grade Cards on
            Consumers’ Choices
• Before grade cards, restaurant revenue is
  insensitive to changes in inspection scores

• After grade cards, revenue responds to grades
  – A grade: + 5.7%
  – B grade: + 0.7%
  – C grade: – 1.0%

• Total industry revenue increases by 3.3% ($250
  million increase in LA)
                  Impact of Grade Cards on
                  Average Inspection Scores
                      BEFORE            AFTER          DIFF
ALL restaurants                  81.6                88.7           7.1
Chains                           87.1              92.6            5.5
Zagat guide                      78.4              88.6          10.2
Chinese food                     78.4              86.3            7.9
Mexican food                     82.5              88.9            6.4
Pizza                            84.2              89.7            5.5
Low income areas                 80.5              88.5            8.0

   All entries are statistically different from the mean for all restaurants
       Impact of Grade Cards on
       Average Inspection Scores

                       C ef      t . r.
                                Sd E r

 a dt r   r d ad
M n a o yg a e c r s    .3 5
                       4 9 8     .4 4
                                1 0 6
 ou t r   r d ad
V l n a yg a e c r s    .2 5
                       3 5 8     .3 5
                                0 5 0
 n p ci n r t r a I
I s e to ciei I         .0 8
                       8 8 6     .9 0
                                0 9 7
 n p ci n r t r a I
I s e to ciei I I       0 1 8
                       1 .4 5    .3 4
                                1 5 2

O S                     9 9
                       6 ,9 1
 o f e t ua t
N . o r s a r ns        3 4
                       1 ,5 4
R2                      .5 7
                       0 8 4
           Impact of Grade Cards on
             Foodborne Illnesses

• We compare the number of food-related
  hospitalizations in LA with
  – non-food-related hospitalizations in LA
  – food-related hospitalizations outside LA

• We analyze hospitalizations for which 90% or more of
  cases are transmitted via food. This includes

                a na
                 l e
                S ol             hl i
                                  i l s
                Ab i             .o
                ue a
                 l m
                Tr i             rcl i
                                  ul s
                                 Be s
                it i i
                 s o
                  e s
                L rs             te ops i
                                  h f d ion
                                 Oro -o ng
Impact of Grade Cards on
  Foodborne Illnesses
Impact of Grade Cards on
  Foodborne Illnesses

• Grade cards cause restaurants to improve hygiene,
  which in turn reduces the number of food-related

• There is no real improvement by restaurants but
  consumers sort towards better-grade restaurants
To distinguish consumer sorting from
restaurant real improvement

• prob of getting sick at restaurants (g=1 if after grade

• prob of getting sick at home

• putting together:

•   Consumers become more sensitive to
    hygiene grades after LA adopts grade
•   Hygiene score improves after GC
•   # of foodborne-illness hospitalization
    reduces 20% after GC, due to hygiene
    improvement instead of consumer sorting
 Other work: can reputation explain score variations
 before Grade Cards? (Jin and Leslie AEJ-Micro 2009)

• 25% “A” restaurants
• 12% chain restaurants                Br
                                       e e
                                        o        ft

      o a - n hn
       m no ed a
      C p yw c i s                     89
                                        .        20

      r cs
       a h dhn
      Fn ie c i s                      88
                                        .        27

      n p d t ea a s
       e nn t r t
      Id e e rs un                     79
                                        .        97

• regional clustering
                                o f a al
                                 .   r  e
                                N o Vi b s   2

           e a r t hr c i tc
            s   a      e
           R t u nc aat rsi s   4
                                6            . 0
                                             0 16

            i F
           Cy E                  6
                                11           . 5
                                             0 94

           i oe E
           Z cd F                2
                                35           . 7
                                             0 69

           e ar t E
            s   a
           R t u nF             2, 1
                                 2 1
                                  2          . 4
                                             0 22
       Other work: Inspector Behavior
    (Becker, Jin and Leslie, working in progress)

• Grade cards make consumer demand
  sensitive to hygiene

• This creates incentives that may cause
  – restaurants to improve hygiene
  – inspectors “grade inflation”
     • restaurant managers challenge violations
     • inspectors may feel bad about the consequences of
       giving a B
Other Stated Problems with Grade Cards
• Misleading consumers since may not have A-
  grade hygiene after inspection
  – no different for inspections without grade cards

• Restaurant hygiene too complicated to summarize
  with a grade
  – applies objective scientific criteria

• Unfair punishment for violations that are quickly
  – owner-initiated inspections
Unanswered Questions
• Impact on restaurant prices?
• Impact on restaurants’ costs & profits?
• Impact on number of restaurants?
• Impact on illnesses not resulting in
• Long term effects?
  “Is More Information Better?” Dranove et
  al. 2003 JPE
It is difficult to measure the quality of expert services
New York and Pennsylvania mandate report cards on physician
and hospital coronary artery bypass graft (CABG) surgery mortality
 NY: hospital specific report cards (raw and risk-adjusted
 mortality) since 1991, doctor specific report cards since 1993
 PA: hospital and doctor specific mortality rates since 1993
Previous evidence from surveys:
 63% of cardiac surgeons reported that they accept only healthier
 patients due to report cards.
 59% of cardiologists reported that report cards made it more
 difficult to place severely ill candidates for CABG.
Economic concerns
Since CABG is elective, patients will search for CABG
surgeons. More so for sick patients.
Report cards may help patient search and therefore result in a
better match between patients and hospitals/doctors
However, risk adjustment may be incomplete. Doctors/hospitals
observe more info about patient risk than the compiler of the
report cards. → want to select healthier patients
Selection may occur even if the risk adjustment is complete:
because quality matters most for sick patients, low-quality
doctors want to select healthier patients
Selection may occur even if the risk adjustment is unbiased but
noisy: in that case, risk averse doctors want to select healthy
What happens if sick patients cannot get surgery? -- other
treatments (may be less effective), emergency rooms, death
    Empirical Strategy
    Treatment group: NY and PA
    Control: all other 48 states
    Medicare individual claims data
    Focus on AMI patients, with and without CABG surgery
    key assumptions:
      • AMI population does not change after report card
      • NY+PA are comparable with other states
    hospital level: distribution of patients receiving certain treatments within a

                                                                     Key coeff
    patient level: whether to have CABG or other treatments, sickness of
     patients, outcome of patients

   Illness severity: total hospital exp in the year before admission
   intensity of treatment: $ spent, number of days in hosp.
   outcomes: cardiac complications, subsequent heart failures, readmission
    Dranove et al. 2003 JPE: Conclusion

   Report card leads to doctors selecting
    healthier patients
   Report card leads to a better match
    between doctors and patients
   Report card leads to more medical
    expenditure and worse health outcomes
   Report card is welfare reducing
Theme #5: Entry and Exit
•   Approaches
    • use static market structure to infer equ. in entry and exit
      (Bresnahan and Reiss 1991, Sutton 1991)
    • use each firm’s entry/exit decision to infer expectation
      of future profits (Berry 1992)
    • macro distribution change in productivity drives
      entry/exit (Olley and Pakes 1996)
    • estimate market demand and entry/exit decision
            •    Berry and Waldfogel (1999)
            •    Ericson and Pakes (1995), Bajari, Benkard and
                 Levin (2007), Ryan (working paper)
    • endogenize product differentiation and entry/exit (Seim
Key factors in entry and exit
•   Data observations:
     •   # of firms
     •   who enters/exits
     •   order of entry/exit
     •   simultaneous entry/exit
•   Key factors determining entry and exit:
     •   market size
     •   # of potential entrants
     •   toughness of price competition (may vary by # of firms)
     •   sunk cost of entry (and advertising)
     •   firm heterogeneity in product attributes
     •   firm heterogeneity in productivity or cost efficiency
     •   multiple equilibria in theory?
Bresnahan and Reiss 1991 JPE
•   RQ: In a homogenous industry, why do some markets have one firm
    while other markets have two or more? How does price competition
    shape market structure?
•   Demand: Q = d(Z,P) * S(Y)

                                   market size

•   Supply: fixed cost F(W), marginal cost MC(q,W), avg cost AVC(q,W)
•   Monopolist: π1(S1)=[P1-AVC(q1,W)] * d(Z,P1)*S1-F
•   Monopoly threshold S1 denotes the smallest market a monopolist can
    survive (so that π1=0):
           S1 
                [ P1  AVC(q1 ,W )]d ( Z , P1 )
Bresnahan and Reiss 1991 JPE
•    Now consider the Nth entrant:      extra                 extra fixed
                                        variable              cost
 N  [ PN  AVC(q N ,W )  bN ]d N ( Z , PN )  FN  B N

      SN            FN  B N
 sN     
      N [ PN  AVC(q N ,W )  bN ]d N ( Z , PN )

 s N 1 FN 1  B N 1       ( PN  AVC N  bN )d N
                      
  sN     FN  B N        ( PN 1  AVC N 1  bN 1 )d N 1

        s N 1
 N  ,        1
Bresnahan and Reiss 1991 JPE
•   If we can measure entry threshold ratio (sN+1/sN), then we can make
    inference about barriers to entry (B and b)
•   Data: homogenous goods across many atomic, isolated markets
     •     yellow-pages  202 isolated local markets
     •     local services such as druggists, doctors, dentists, plumbers,
           barbers, etc.
     •     static cross-section data, with no explicit info on entries, exits, or
•   Estimation:

          N  S (Y ,  )  VN ( Z ,W ,  ,  )  FN (W ,  )  
     •     observe N firms (this is the dep var), we have πN>0 and πN+1<0
     •     ordered probit
Bresnahan and Reiss 1991 JPE
•   If we can measure entry threshold ratio (sN+1/sN), then we can make
    inference about barriers to entry (B and b)
•   Data: homogenous goods across many atomic, isolated markets
     •     yellow-pages  202 isolated local markets
     •     local services such as druggists, doctors, dentists, plumbers,
           barbers, etc.
     •     static cross-section data, with no explicit info on entries, exits, or
•   Estimation:

          N  S (Y ,  )  VN ( Z ,W ,  ,  )  FN (W ,  )  
     •     observe N firms (this is the dep var), we have πN>0 and πN+1<0
     •     ordered probit
Bresnahan and Reiss 1991 JPE
•   Conclusion:
     •   entry threshold is way over 1 in oligopolistic markets, suggesting
         that there is barrier to entry when # of firms is small
     •   whether the barrier to entry comes from extra fixed cost, extra
         variable cost, or a change in the intensity of price competition is
         unclear (unless we assume a lot of functional forms)
     •   ignore firm heterogeneity in cost efficiency (here assume
         homogenous good and homogenous firms)
     •   ignore product differentiation (which may alleviate price
     •   say something about barrier to entry without observing any
         dynamic entry!
     •   complementary data on price of tires (price goes down with N)
Sutton 1991
 •   Big RQ: why are markets concentrated? Is the concentration harmful or
     simply due to scale economy (large fixed cost) or firm efficiency?
 •   Sutton:
      •   fixed cost (and therefore scale economy) could be endogenous
      •   three stages:
               (1) entry (incur a fixed amount of entry cost)
               (2) choose fixed cost in other dimensions (advertising, R&D)
               (3) compete by price and product differentiation
      •   if ignore stage (2), entry varies by toughness of price competition.
          As market size goes up, concentration may go down infinitely.
      •   if include stage (2), entry varies by toughness of price competition
          and fixed cost spent in stage (2). As market size goes up, there is a
          lower bound of concentration.
Sutton 1991 – a concrete example
 •   Stage 1: n firms enter the market, each paying for an exogenous entry
 •   Stage 2: each firm can choose an advertising intensity A(u) to boost its
 •   Stage 3: consumer willingness to pay increases with A(u). (Advertising
     is always persuasive!). This is equivalent to a drop in marginal cost if we
     keep the demand constant.
 •   Now market size goes up,
      •   attract entry
      •   encourage more advertising, which entails a larger economy of
          scale thus limiting entry
      •   the balance of the two may keep the market concentrated
      •   the lower bound of the concentration is above zero if consumers
          keep on responding to A(u)
    Sutton 1991 – a concrete example
     •   Empirical strategy:
     •   Two groups of industries: one with intensive advertising, one without
     •   Examine the relationship between market concentration (C) and market
         size (S).
     •   Examine how C(S) varies by the toughness of price competition and the
         extent of endogenous sunk cost (ESC).
     •   The tougher the price competition gets, the slower C(S) goes to zero.
C    •   The more ESC the market allows, the higher the lower bound of C(S).
                                                                 high ESC

                                                                   no ESC

                                  S                                              S
Sutton 1991 – conclusion
 •   Firms’ choice leads to concentration.
 •   Combine product differentiation, strategic models and scale economy.
 •   Can we say anything beyond lower bound?
 •   Is market concentration good or bad, after we account for endogenous
     sunk cost?
 •   Extension: market size determines endogenous choice of product
     quality, see Steve Berry and Joel Waldfogel ―Product Quality and
     Market Size‖ (working paper, available on both authors’ websites)
Berry 1992 Econometrica
 •   RQ: how does airport presence affect an airline’s entry decision?
 •   Why interesting: airport presence may imply
      •   economy of scale (which is socially beneficial) or
      •   entry deterrence (which implies higher price thus socially harmful)
      •   This paper cannot distinguish these two
      •   related to deregulation of airline industry
      •   more broadly: entry determined by
               •    market size
               •    firm heterogeneity
Berry 1992 Econometrica
 •   Data: Origin and Destination Survey of Air Passenger Traffic
      •   market=city-pair,
      •   subject=airline as incumbent or potential entrant,
      •   time=two periods
 •   Conceptual Framework
      •   each incumbent or potential entrant takes period 1 setup as given
      •   each decides on whether to operate in a given city pair
      •   decisions across city pairs are assumed independent (no strategic or
          dynamic concerns across markets)
      •   enter if expected profit in period 2 is positive
Berry 1992 Econometrica
 •   Challenges:
      •   expected future profits depend on the equilibrium # of firms
      •   firms may be heterogeneous and this heterogeneity affects entry.
          This heterogeneity is also necessary otherwise we cannot predict
          which firm is more likely to enter.
      •   multiple Nash equilibria in the entry game
 •   Solutions:
      •   Firms only differ in fixed cost
      •   the post-entry game is symmetric so variable profits only depend
          on the number of entering firms (N), not the identity of entering
      •   If variable profits are strictly decreasing with N, then a pure
          strategy equilibrium involve a unique number of entering firms.
      •   Assume incumbents move first and most profitable firms move first
          within incumbents or potential entrants.
Berry 1992 Econometrica
 •   Firm k, market i:
                                   demand correlation within a

        ik ( N )  X i    ln( N )  uio  Zik     uik

                         variable profits       - fixed cost (firm
Berry 1992 Econometrica
 •   Dep Var: N and identity of entrants
 •   Simulated moment estimator:
      •   draw uio and uik
      •   Given parameters, simulate every firm’s entry decision
      •   Aggregate every firm’s entry decision to predict N
      •   repeat and take average of predicted N
      •   choose parameters to match predicted N and observed N, as well as
          the identity of actual entrants
 •   Alternative estimation:
      •   Take a market as an OBS, regress N on X – cannot account for firm
          heterogeneity and does not use info on the identity of entrants
      •   probit of each firm’s decision – ignore strategic reaction to N,
          assume no correlation in uio
      •   ordered probit on N – does not account for unobserved firm
Berry 1992 Conclusion
 •   City presence matters
 •   Competition reduces price
 •   Simulated estimator performs better than traditional procedures
      •   use more information (N and identity of entrants)
      •   allow for unobserved firm heterogeneity
      •   allow for unobserved market conditions
Berry and Waldfogel 1999 RAND
 •   RQ: When is free entry inefficient?
 •   Why important?
      •   directly related to welfare policy: shall we encourage free entry in
          all circumstances?
      •   justify theories on inefficient free entry
      •   see radio from a different perspective other than public goods
 •   Radio broadcasting
      •   local markets
      •   high fixed costs, low/zero variable cost
      •   free entry without regulations
      •   two sides markets: advertisers, listeners
      •   similar features may exist in newspapers, TV programs, cellular
          phones, etc.
Berry and Waldfogel 1999: Theory
 •   # of firms:
      •   free entry > social optimal > monopoly
      •   RQ: When is free entry inefficient?
 •   Intuition:
      •   economy of scale due to large fixed cost
      •   if free entry involves large business stealing effect, the large fixed
          cost incurred by each radio may be duplicative
      •   monopoly internalize business stealing among competitors, but
          does not internalize the negative externality between firm profits
          and consumer surplus
Berry and Waldfogel 1999: Estimation
 •   Assumptions:
      •   ignore product differentiation: consumers and advertisers do not
          benefit from product differentiation
      •   no strategic competition in advertising price (P changes with % of
          population listening to radios, but not with # of radios)
 •   Key to estimate:
      •   degree of business stealing among competing radios
      •   advertising revenue as a function of listeners
      •   fixed cost of setting up a radio station
Berry and Waldfogel 1999: Estimation
 •   Listeners’ demand for radio programs
      •   data: Aritron (1993) # of listeners by radio-market in spring 93
      •   method: Nested logit, match predicted market share with real
          market share (Berry 94)

                              ζ      listen

             listen                            D
                              A     B C ..

               •      Ujk=δjk + εjk= xjkβ+ξjk (note no p because radio is free)
               •      N of radio stations is endogenous. Use population and #
                      of radios outside the metro as IV.
               •      key parameter: ζ=1 if purely business stealing
Berry and Waldfogel 1999: Estimation
 •   Advertisers’ demand for # of listeners
      •   data: Duncan (1994) advertising revenue by station-market
      •   define p=advertising revenue per listener
      •   method: p = α (S(N))-ε
                                              demand elasticity
advertising           demand       total # of
price per             shifter      listeners
               •    N is endogenous
               •    use population and # of stations outside the metro as IV
                    for N
               •    key parameter: demand elasticity ε
Berry and Waldfogel 1999: Estimation
 •   entry cost
      •   assume every station has the same entry cost within a market
      •   ln (F) = xk μ + λ vk     λ: variance of unobserved fixed cost
      •   observe N stations 
                  •   variable profit (N+1) < fixed cost < variable profit (N)
      •   variable profits based on previous estimates of listener and
          advertising demand
      •   Similar to ordered probit  MLE
 •   joint estimation to ensure efficiency
Berry and Waldfogel 1999: Result
Berry and Waldfogel 1999: Result
Berry and Waldfogel 1999: Result
Berry and Waldfogel 1999: Simulation
Berry and Waldfogel 1999: Conclusion
 •   Incorporates two-side demands and endogenous entry
 •   Allows for policy simulation
 •   Rely on several simplification assumptions:
      •   Radios are homogenous (especially in advertising demand and
          entry equation, not necessarily in listeners’ demand)
      •   Fixed cost is symmetric across firms (no firm heterogeneity)
      •   Welfare calculation does not include listeners’ welfare from an
          increase of N
      •   Assume program features are exogenous. Open questions: How to
          model product attributes as firm choice?
      •   Minimize the strategic interaction between oligopolistic firms
 •   On-going work: Berry and Waldfogel 2007
 •   Review of entry liter: Berry and Tamer 2006 ―Identification in Models
     of Oligopoly Entry‖
    Goolsbee and Svyerson QJE: Southwest entry
•    RQ: how do incumbents respond to the threat of entry?
•    There is a large literature on entry preemption. Many economists argue that
     entry preemption can be rational, and, if it is successful, entry won’t occur.
•    It is difficult to test preemption empirically because preemption responds to
     the threat of entry, not the actual entry
•    Southwest airline provides a nice example
      •   A new player in the airline industry, famous for low fare
      •   Growing fast between 1993-2004
          • Volume and revenue almost triple
          • Add 22 new airports
      •   Threat of entry occurs if Southwest has operated (or announce to
          operate) in the airports of two cities but does not offer a route between
          the two cities. This may or may not become an actual entry.
   •       704 threats of entry, 533 actual entries
   •       25 quarter window around each Southwest threat of entry
            •   8 quarters before
            •   3+ quarters between threat of entry and actual entry
            •   3 quarters after actual entry
            •   Excluded periods as control

       •    Static view: should see no response before actual entry
       •    Dynamic view: should see price cut or extra capacity investment before
            actual entry
Robustness check
  •   Omitted cost changes drive entry and price cut?
  •   Results sensitive to the choice of time window?
  •   Impact on nearby airports in the same city
  •   Capacity investment
  •   Preemptive price cut

• Incumbent airlines start to cut price before
  actual entry, and the price cut is even
  deeper if the entry is uncertain
  – Evidence for preemption
  – Little evidence on strategic investment

• Why start to cut price long before threat of
  entry? Pretreatment time trends?

To top