THE LOGIC OF LARGE-FIRM ADVANTAGE IN REGULATION

“Larger” or “established” firms in an industry differ from their “smaller” and “newer” counterparts by greater revenues, more products, greater diversification of product lines, larger capitalization (more assets), and enhanced tolerance of financial exposure. They also enjoy relative advantages under numerous forms of regulation—ranging from price-setting institutions to content regulation (e.g., the Federal Communications Commission) to “consumer safety” measures. Such advantages appear in numerous industries, ranging from transportation (Rothenberg 1994) to telecommunications (Crandall and Flamm 1989; Noll 1973) to pharmaceuticals (Grabowski and Vernon 1983).

In few markets is the power of the state to reward firms greater than this last one, the pharmaceutical industry. Throughout the industrialized democracies, particularly in the United States, the very legal marketability of a drug product depends on prior approval by government regulators. As a result of the 1938 Food Drug and Cosmetic Act, the U.S. Food and Drug Administration (FDA) has sole authority to approve prescription drugs for marketing. The state remains the ultimate gatekeeper in the pharmaceutical industry, both here and overseas.

The existence of systematic regulatory advantages for larger firms in the pharmaceutical industry has been well documented. Grabowski and Vernon (1976) argue that “upward shifts in costs and risk produced by increased [pharmaceutical] regulation…operate to concentrate innovation in fewer and larger firms” (190). Thomas (1990) finds that the 1962 Amendments to the Food, Drug and Cosmetic Act—which required drug companies to show effectiveness in addition to safety before marketing—caused heavier reductions in product innovation among small firms than among large firms. More recently, Olson (1997) finds quicker approval times for firms with greater employment and more FDA product applications, and Carpenter (2002) and Kyle (2002) find that the speed and probability of regulated entry into pharmaceutical markets are (1) an increasing function of firm market experience and (2) a decreasing function of previous entrants. Employment and experience are characteristics highly correlated with firm size, and early entry into a market is per se correlated with age. Perhaps most crucial to these advantages is the finding that larger and older firms receive quicker approvals for their new drug applications (NDAs). Whatever advantages large firms have in product development seem to be magnified by FDA decision making.

The capture theory formulated by Bernstein (1955) and Huntington (1953) and rendered more explicit by Peltzman (1976) and Stigler (1975) offers a simple and powerful explanation for these advantages. Because larger firms have lower marginal costs and higher marginal benefits of political action, they can more easily capture the regulatory process, either by shifting legislation to their advantage (Stigler 1975, 123–28) or by shifting the administration of an otherwise costly law to their advantage (Stigler 1975, 162–66).¹

In the past two decades, scholars have refined the capture perspective by advancing “interest-group” and “rent-seeking” theories of regulation (Becker 1985; Rothenberg 1994). These views concede that producer interests are not monolithic—firm interests vary by specialization and other things—and that consumers can also organize. Yet even these theories rarely shy away from the premise that large firms in an industry tend to dominate regulation. As Rothenberg (1994, 4) argues, “The view that producer dominance is the modal description of regulatory politics remains perhaps the primary means of conceptualizing them for popular commentators and scholars alike.” And as the opening epigraph from Viscusi (1992) suggests, scholars are still apt to follow Stigler's advice for proper inference when large-firm advantage is observed (see also Bartel and Thomas 1987). For purposes of this paper, then, the proposition that lies at the heart of a “capture” account of regulation is that, by legislative structure or by cozy relations with the industry, regulators systematically favor large, established firms over newer entrants and smaller firms.

Stigler's argument is a powerful one and has usefully aided a generation of regulation scholars. Yet its counsel that regulatory results justify inference about regulatory mechanisms leaves open the possibility that numerous factors may explain observed regulatory outcomes. In particular, there may be cases where producers have not “captured” a legislature or agency but in which large-firm advantage prevails nonetheless.

Scholars have begun to recognize this possibility. Olson (1997) has employed “external signals” theory (Joskow 1974; Noll 1985) to suggest that regulators see firm experience as one of several “signals” of likely product quality—other signals being R&D and productivity—and so condition their reviews accordingly. Olson finds systematic evidence for this hypothesis in a regression analysis of drugs approved from 1990 to 1992. Controlling for clinical factors and FDA priority ratings, Olson finds that firms with greater R&D, more employees and more FDA product applications receive shorter approval times for their drug applications. A 1995 study by the General Accounting Office (GAO; 1995, 32) also finds that firms with more submissions receive shorter review times. Employment and submissions are highly correlated with firm size. Olson construes these results to mean not that larger firms capture drug regulation but, rather, that a rational agency interprets size-related firm characteristics as signals or correlates of unobserved product quality.

If Olson's argument is correct, then larger-firm advantage can exist even when regulators seek only to reduce their Type I and Type II errors. Yet is this result generalizable? Would mathematical formalization support the argument that an uncertain regulator will indeed act more quickly when firm characteristics shed light on product applications? What if the regulator is risk-neutral? and What if the agency is in fact influenced by the political clout of large firms? I am unaware of any attempt to construct a systematic formal theory that explores the firm-specific implications of approval regulation by a politically responsive, Bayesian agency.

A DYNAMIC MODEL OF APPROVAL REGULATION

Product approval presents the regulator with a learning problem. The regulator must review a new product application (with accompanying data) and decide when the apparent benefits of the product outweigh the costs or risks associated with its use.

Assumptions and Parameters

I stylize the model for the drug approval process as it operates in the United States and other countries. By regarding diseases as “markets” or “market niches,” curing probabilities as product quality, and danger as general “hazards” of a product, the model can be applied to other situations. Let all products be indexed by i, markets (diseases) by j, and firms by k. The model assumes an exogenous industry production process in which the agency expects products to be submitted at a constant rate λ_j>0 over time. (I discuss this under The Problem of Endogenous Submissions, below.) Products are assumed to have “niche specificity”; that is, drugs can treat one disease only.³

The one drug- one disease assumption can be relaxed without affecting the general results of the model. Clearly the exogeneity assumption here is violated in practice. Where the agency becomes more stringent in its product approval decisions, firms will develop fewer drugs and do so more slowly (Grabowski and Vernon 1983; Peltzman 1973; Thomas 1990). For a model that endogenizes submissions, see Carpenter and Ting 2004.

All products in the model are characterized by two parameters. First, let γ_ij(0[les ]γ_ij[les ]1) be product quality (for drugs, a curing probability of the drug, which can be interpreted as the fraction of people with disease j that drug i will cure). We assume that γ_ij is fixed and known with certainty throughout the agency's decision problem.⁴

One may also interpret γ_ij as the equilibrium market share upon the product's approval. The agency considers both safety and efficacy in this model, but for simplicity, I assume that only danger is learned and subject to uncertainty. The model can be adapted to the case where γ_ij is also learned, but the dynamics become quite complex when the agency updates more than two processes, and the simpler variant here adequately describes the review process.

Second, let μ_ij be the danger of the product, which can be thought of as the expected number of people that will be harmed or killed by the product over a given interval of time. Normalizing the interval to one, μ_ij may be thought of as the rate of harming consumers. The greater the danger of the product, the more its approval will harm the agency's reputation for protecting public safety.

I assume throughout that a product's danger is independent of its quality, which implies cov(μ_ij, γ_ij)=0. Because the product quality γ_ij is niche-specific, this independence is understood in a particular way. It is possible that a “good” cancer drug could cure a patient's cancer but induce such severe “side effects” in doing so that the patient dies of other causes (e.g., liver toxicity). I assume danger and quality are separable because quality is disease-specific (the cure of a specific ailment), whereas danger refers to literal “side effects,” namely, the harm a drug may cause to some organ or physiological process in the person's body other than that which was the drug's intended target.⁵

I thank an anonymous reviewer for this suggestion and for the clarifying language.

Learning about Danger in Continuous Time

The agency observes a series of experiments (e.g., clinical trials) in which a product either harms or does not harm the consumer. The sequence of binary outcomes—“harm” or “not harm”—becomes the “data” for the agency's decision. The model posits the evolution of these observations as a continuous-time Wiener process, or “Brownian motion.”⁶

As Dixit (1993, 2–3) shows, the Wiener-process representation can also be derived by taking limits toward zero on a discrete-time walk of binary outcomes.

Casually, we may think of Brownian motion as an “all-purpose” random process in continuous time whose movements are described by the Normal distribution.

Observed harm in regulatory review evolves according to a Wiener process X_it=X(t), a linear function of underlying danger (μ_ij) plus a random component, or

where μ_ij and σ_ij are constants and σ_ij>0, and where z(t) is a standard normal variable with mean zero and variance t. A more “dangerous” product—one with higher μ—will cause more harm, but harm will also be influenced by the random term z(t). A higher σ yields a more volatile review, namely, a series of experiments from which the agency will find it harder to learn what μ is.

Danger as the Cost of Approval

Observed harm is a Markov process, and the agency can learn about μ in a simple Bayesian fashion based on the observed history of X(t). Letting X(t) start arbitrarily at 0, then it is normally distributed with mean μt and variance σ²t. We assume that σ is the same across products but that μ differs across them, according to a normal distribution with mean m and variance s. For any product review of length t and accumulated harm X(t)=x, the dual [x, t] is a sufficient statistic for the history. In other words, the agency knowing the entire clinical history is no better off than if only [x, t] were revealed. Given these sufficient statistics, Bayesian estimates of μ are

Notice that

The posterior variance S(t) is the agency's expected uncertainty about the true value of μ. For this reason, the value of delaying another moment (another dt) is an increasing function of S(t).⁷

See Chernoff (1968) and later, Jovanovic (1979), for a similar modeling architecture. I rely throughout on the scale-invariance of Brownian diffusions (Miroschnichenko 1975, 389–91).

The agency's goal during product review is to estimate the danger of the product (μ). Although real-world agencies have the option to recall a bad product (or induce producers to recall it), the agency has no such option here. The reason concerns reputation protection. Once a product has done sufficient harm to warrant a recall, the agency cannot recover its reputational losses via recall because observers will infer that the agency has made a “bad” decision. Hence the decision to approve a product is reputationally irreversible.⁸

Other functional forms are possible. If the agency's reputation has the character of a durable asset subject to erosion, then perhaps the regulator's utility is dependent not on μ, but on exp(μ). This would render the stochastic process a “geometric” Brownian motion (Dixit 1993), but none of the substantive results of the model differ.

The political cost of approval, then, is the reputational loss that accrues to the agency from the danger of the product. Upon product approval, then, the agency loses μ, which can be learned only through preapproval review.

The Political Demand for Products and the Approval Payoff

One way for the agency to protect its reputation would be to review submissions forever (or reject them all). Yet there is often a political demand for products makes this delay costly. Drug approval provides a lucid example. In recent years the FDA has appeared to be highly responsive to the demands of (potential) drug consumers (Olson 1995; Vogel 1996). Firms also attempt to place pressure on the agency for quick approvals. Organized consumers and producers lobby the agency directly but also apply pressure indirectly through elected politicians.

For this section I define the agency's approval payoff from approving the N_j-th product as a function only of market attributes such as market size (disease prevalence), intensity of demand or “willingness to pay” (disease severity), public salience, and the political organization of consumers. The payoff may be written

Here L_j is the potential size of the market (in pharmaceutical settings, disease j's prevalence, or the number of persons with disease j).⁹

The constancy of L_j is a crucial assumption for the solution concept of smooth pasting. For now, simply assume that a rapidly growing disease has a high political multiplier. I also assume that indexation across j offers no information about danger; clinical risk and indications are assumed independent. In practice, this is almost certainly violated.

ψ_j is the political multiplier of market j, a positive parameter.¹⁰

The parameter can exceed one because citizens other than those directly afflicted by a disease—relatives, friends, or other allies in organized patient associations and the media—may lobby for drug approval.

ψ_j can be interpreted as the expected number of citizens, for every citizen in market j, who will apply pressure on the agency or the politicians governing it. (In pharmaceutical settings, this is the number of citizens per patient with disease j. In some [but not all] cases, this parameter may capture the political impact of disease severity). N_j is the number of marketed products that have already entered market j (in a pharmaceutical context, the number of available drugs that already treat disease j).

I postpone exact specification of A to the section on early-entrant protection. For now, a crucial aspect of equation (4) is that none of its constituent variables is indexed by k. That is, I assume (for now) that the approval payoff is unaffected by firm characteristics such as contributions, reputation, and rents. In other words, the agency is not “captured” in the sense that its preferences do not direct it to treat some firms differently from others per se.

The Agency's Optimal Policy

The problem facing the agency can be described as the optimal stopping of the process

, with the following objective (suppressing some subscripts):

where δ is the discount factor, t_app is a given approval time, μ* is the agency's estimate of danger at the optimal stopping time (as given in Eqs. [2]), ω denotes an elementary event in the probability space Ω, and y is a variable of integration.

The regulator's optimal policy is to divide the space of possible outcomes into two regions (DeGroot 1970)—a continuation region, where observed values of estimated danger

suggest waiting, and a termination region, where values of danger suggest “stopping” the review process and approval of the product. Figure 1 offers a sample division of the space. Proposition 1 characterizes the agency's best policy.

Simulated First Passage of the Danger Process $\hat{\mu}(\mu,t)$ through the Agency's Approval Barrier η(t)

Proposition 1

For the approval problem characterized by (5), there exists an optimal stopping policy characterized by the unique partitioning of

into a continuation region C and a termination region T, such that the approval barrier η*(t) represents the boundary of C, and the first passage of

through η*(t) represents the optimal approval of the product, where η*(t) is

Proof

Proofs of all propositions, lemmata, and corollaries are given in the Appendix.

By Eq. (3), the limit of η*(t) is δA, which

approaches from above. The fully discounted benefits of approval (which have remained constant throughout the decision) are equal to δA. Meanwhile, the cost of approval is μ. In the limit, the estimated danger

converges to the true danger μ, so that the product is approved asymptotically if μ<δA. In finite time, of course, the agency cannot make the decision in such terms. There is always a value to waiting for more information, regardless of the agency's attitude toward risk.

Characteristics of the Approval Distribution

Let G*(t) be the approval distribution, or the probability of approval at time t under the optimal policy. Proposition 2 notes an interesting property of G*(t): Even when the agency cannot formally reject a product, some products will still be “rejected” in the sense that they will never be approved.

Proposition 2. De Facto Rejection without a Rejection Option

In expectation, a nonzero fraction of products will never receive the agency's approval, as

.

In two ways, Proposition 2 coheres with actual regulatory behavior, for instance, the FDA. It is trivially consistent, of course, with the fact that not all drug submissions get approved. Yet it is also consistent with the fact that the FDA never formally rejects a drug. It simply deems a drug “not approvable,” and nothing prevents the producing company from submitting more data about the drug if it seeks approval at a later date (GAO 1995).

Comparative statics are based on the following result.

Proposition 3

Let E[t_app|t*_app<∞] be the expected approval time under the optimal policy. Then conditioned on approval, E[t_app|t*_app<∞] is strictly decreasing in the payoff A.

From Proposition 3, any variable in which the approval payoff is increasing is also a variable in which expected approval times are decreasing. Another prediction of the learning model concerns the speed of approval.

Proposition 4. The Scarcity of Quick Approval

Let g*(t) be the density of G*(t). For any product i[t_app(i)<∞] such that an optimal approval time exists, the approval hazard θ(t)=g*(t)/[1−G*(t)] has the following two properties:

While the scarcity of quick review does not always benefit larger, established firms, reflection suggests that, given a distribution of firms—existing and potential entrants to a market—tardy review will impose heavier relative costs on smaller firms. Smaller firms are usually newer ones, and delayed market entry can impede a cash flow crucial to firm survival and cripple the firm's relationship with potential investors who seek certainty over future product revenues. Larger firms are better able to survive this sort of exposure.

THE VALUE OF FAMILIARITY: FIRM EFFECTS AND FREQUENT SUBMISSIONS

The optimal stopping problem in Eq. (5) is repeated numerous times over the lifetime of a regulatory policy. In pharmaceutical regulation, for instance, the FDA has since 1950 received over 10,000 new drug applications (NDAs) and has approved over 1,500 new chemical entities, or drugs with new and complex molecular structures. State public utility commissions face repeated licensing applications for construction and operation of power plants from the same utility companies. For hydropower plants the Federal Energy Regulatory Commission (FERC) conducts such reviews, again repeatedly interacting with many companies. In numerous industries, regulators receive new products or license applications repeatedly from the same firms or interests. It is useful, then, to consider the dynamics of the agency's approval policy when the problem in Eq. (5) is repeated finitely.

If firms differ in the quality and care with which they produce products, then it is reasonable to believe that these differences are at least partially observable by the agency. Anecdotal evidence coheres with this intuition. The FDA employs hundreds of inspectors to survey production processes at pharmaceutical plants, and firms that are found in violation of federal regulations are frequently revisited.¹¹

To model the possibility of systematic firm differences in product quality or hazards, I introduce a firm-specific error term to the danger variable. Whereas before the specific danger of a product i(μ_i) was a random draw from a normal “production distribution” with mean m and variance s, I now respecify μ_i to depend on an independent firm effect, ξ_k, which is distributed standard normal across firms. The danger of any product is now

Notice that the parameter m is still the central tendency of the danger distribution, as E[μ′_ik]=m. By the addition of independent variances, however, Var[μ′_ik]=s+1. The rational agency learns about ξ_k, but this cannot be done within the confines of the problem in Eq. (5). It must be done across repetitions of the stopping exercise, that is, across products.

Because the danger effect ξ_k is distributed normally across firms, the agency can form a Bayesian estimate

.¹²

The agency's uncertainty about a firm's safety is the posterior variance

of the estimated firm effect

. The prior variance of the firm effect ξ_k is 1, which holds for all firms, but the posterior variance of the estimated firm effect depends on the firm's previous submissions, so that that

is a decreasing function of N_k.

where r(·) is the product-specific precision of

, and t_sub is the review time for all nonapproved products—or the elapsed time from the submission of those products to the present time (τ). When the regulator has had no experience with firm

simply reduces to 1, or the “prior” distribution of

. Given some experience with any firm (N_k>1), then at the beginning of review for any given product, the agency has better information about the firm-specific danger effect ξ_K for firm k=K. This fact forms the central intuition underlying the advantage of familiarity.

Proposition 5. The Advantage of Familiarity

For any set of firms k=1, 2…K such that ξ₁=ξ₂=···=ξ_K, the following two statements hold:

(a) the expected review time E[t_app(ik)|ξ_k] conditioned on the posterior variance

is a strictly decreasing function of N_k,

(b) the expected review time E[t_app(ik)|ξ_k] conditioned upon the sample variance of the danger estimate

is decreasing in N_k unless

, where

is the mean square error of

.

Proposition 5 specifies a relationship between the role of uncertainty about firms and drugs in the approval policy and the role of experience in reducing this uncertainty. If we form general expectations over regulatory behavior, then this uncertainty is understood in terms of the posterior variance of the danger estimate,

, and the relationship between expected approval times and regulatory learning is monotonic. If uncertainty is understood as the sample variance of the danger estimate, then for most realizations of μ_ik, expected approval times will decrease with regulatory learning.¹³

I thank an anonymous reviewer for pointing out this alternative notion of regulatory uncertainty.

Extreme realizations of μ_ik may increase regulatory uncertainty. However, the left-hand side of the inequality in Proposition 5(b) has zero expectation, so these jumps in regulatory uncertainty will not happen often. And they are less likely to happen when N_k is small.

Because the regulator's uncertainty about a firm is a generally decreasing function of regulatory experience with that firm (N_k), so is uncertainty about the drug under review, or S_i(k)(t). Since the optimal policy requires product approval when estimated danger

hits the barrier η*(t) from above, an increase in familiarity (N_k) pushes S(t) lower—and thus the barrier higher— for all t>0. Because the optimal stopping policy established in Proposition 1 invokes risk-neutrality, this is a strong result. Reduced prereview uncertainty about μ_ik translates into quicker approval times, conditioning on the event of approval.¹⁴

An important caveat merits remark. As Dixit (1993, 58) warns, it is not always the case that reduced uncertainty about a Wiener process translates into a quicker optimal stopping time. Recall that a crucial determinant of the agency's uncertainty about μ_ik is σ. As Dixit adroitly notes, increases in σ may actually reduce the stopping time because σ represents the volatility of X(t) and a more volatile process may more quickly reach the barrier. Proposition 4 rests on a different result and is robust to Dixit's cautionary note. Recall that σ does not differ across products but that the posterior variance of ε_k does. At the beginning of any product review (t=0), the agency knows the “prior” variance of μ by S(0)={s+Var^p[ξ_k]}⁻¹. In finite time, S(t) is always an increasing function of Var^p[ξ_k], independent of the value of σ.

Notice that the result in Proposition 5 makes no assumptions about how attributes of firms might affect the approval payoff. The only differences that prevail among firms are captured in N_k, the number of submissions. Over any given period of time, larger firms will submit more products than smaller ones, and this difference alone is sufficient to yield lower expected approval times. Hence there is protection—smaller and newer firms will, in expectation, wait longer to enter any given market—without capture.

The second result is more powerful. Because the result is conditioned on the firm-specific danger value

, Proposition 5 raises the possibility that firms with many submissions and with positive

(or below-average safety reputation) may actually receive quicker reviews, conditioning on the event that a given product is approved. In other words the advantage of familiarity outweighs poor firm reputation, to a point. I formalize this in Corollary 5.1.

Corollary 5.1

Consider two products with identical experimental histories and identical approval payoffs but submitted by different firms: one (k=k⁰) with no previous submissions, the other (k=k^B) with a “bad reputation”

but positive familiarity

. Let t_stop be the approval time for the drug submitted by the unknown firm (k⁰). Then when

or (conceptually) when drugs receive sufficiently “early approval” or the bad firm has a sufficiently low firm danger estimate low

, the product submitted by the bad but familiar firm receives quicker approval than the product submitted by the unknown firm.

One example of a perfect (one-for-one) tradeoff between the disadvantage of a bad reputation and the advantage of familiarity appears in Figure 2. Here a bad reputation—higher values of

—boost the starting point of the process

from m to

and slow approval. Yet to a limited extent, familiarity advantages can compensate and even outweigh these effects. Even as a bad reputation moves the process

away from the barrier, familiarity shifts the barrier back toward the process.¹⁵

If product submission were costless, then rational firms might attempt to take advantage of this familiarity result and load the regulator with frivolous submissions to increase their familiarity with the regulator. As DiMasi et al. (1991) show, however, pharmaceutical product development is enormously expensive even before the regulatory approval stage, so even if firm submissions were endogenous to the model, it is doubtful that this would be an equilibrium policy under any reasonable submission cost parameters.

Trade Off between Firm Danger Reputation and Regulatory Familiarity

EARLY–ENTRANT PROTECTION

Another characteristic of large firms is greater capitalization and an ability to enter new market niches at lower cost. The first new entities for AIDS, acute coronary syndrome, and erectile dysfunction were all developed by large companies (e.g., Burroughs–Wellcome with AZT). DiMasi et al. (1995, 213) present evidence consistent with the proposition that larger firms enter new markets earlier, as their drugs are more likely to receive FDA priority status and their per-drug sales are higher than those of small firms (see also Dranove and Meltzer 1994, 409). If earlier drugs for a disease receive quicker approvals, then larger and older firms benefit disproportionately if in fact they enter these new markets sooner than smaller firms do.

In this section I adjust the agency's approval function to consider the effects of the political demand of consumers on firm advantages. I focus on diseases as “market niches” because they play a crucial role in the pharmaceutical industry. In essence, there is no single “market” for pharmaceuticals; there are, rather, numerous markets bounded in size by the diseased population. This section of the argument is therefore tailored much more specifically to the case of drug regulation. (I consider other possible applications below.)

In the past two decades, the FDA has shown considerable responsiveness to pressure from organized patients and disease-specific advocacy groups.¹⁶

Before translating this principle into an equation for the approval payoff, I pause to note several features of this concept. I assume that the political demand for a drug differs from the economic demand for it in an important way. Political demand is greater for those individuals who have no therapeutic alternatives for their disease. In other words, if individuals are taking a drug that ameliorates their condition in some way, then they have less political demand for any more drugs for their disease, even if these drugs would improve their condition or would be available at a lower cost. In other words, the model rests on the assumption that, once individuals have adopted a drug, their contribution to the political demand pool drops to zero.¹⁷

In short, while individuals as consumers may maximize their health status and care about price, individuals as citizens satisfice with respect to the agency. The reason, I suggest, is that citizens have “traceability” constraints (Arnold 1990). They do not blame the agency for the high price of drugs—even if it is apparent that the regulatory process boosts drug prices. If the price of a drug is too high, or if citizens are weakly satisfied with a drug they are taking, they blame not the agency, but the producers.¹⁸

The Effect of Past and Future Drugs

Because political demand is a function primarily of those patients without therapeutic alternatives, the payoff is not a simple function of the curing power γ_i of the drug submitted. The reason is that drugs may already exist to treat disease j, and forthcoming drugs in the “pipeline” may also be expected to treat the disease. Existing drugs influence the approval payoff in the following way. If no drug has previously been approved for disease j, the total expected curing of drug i is γ_ijL_j. If one drug (“drug 1”) has already been approved, then the expected curing of drug 2 would be γ_2jL_j×(1−γ_1j). So for any series of drugs i=1, 2, …, N, the total curing of the Nth drug, given past approvals, is

Of course, pharmaceutical firms constantly introduce new drugs, and the rational agency will take into account this stream of future submissions. If the agency expects a sufficient number of high-quality drugs to be submitted for disease j in the very near future, then the approval payoff for the current drug is lower. To incorporate into the approval payoff the set of drugs that will be submitted and may be approved in the future, define χ as the time that the Nth drug (the one currently under consideration) is submitted. Let E_χ be the expectation operator computed at time χ. Let c_ij be the mean of the curing distribution f(γ_ij). Then the payoff for the Nth drug (A_N) is

The second term in brackets on the right-hand side is called the pipeline value and will be denoted A_π. The pipeline value is the expected fraction of present prevalence that remains after the curing of the full stream of future drugs, weighted by the probability of their approval under the optimal policy, and each fully discounted. The pipeline value can be zero only if the agency expects a cure-all to be approved with probability 1 immediately upon the submission of the drug. By Proposition 2, this event has negligible probability.

I also allow the approval payoff to be a function of the political influence of the firm submitting the drug, parameterized as ρ_k. (The first four parameters in Eq. (8) are assumed positive.) This parameter embeds the idea that better organized interests will receive preferential treatment. Capture theory may not always imply such a relationship, but Stigler's opening epigraph certainly does.

Intuitively, “early-entrant protection” (EEP) is the systematic advantage in expected time to approval for the first drugs approved for a disease relative to drugs developed later. The premise that drug consumption “removes” patients from politics is sufficient to demonstrate two forms of “protection” for the manufacturers that produce the first drugs for any disease.

Strong early-entrant protection: Expected approval times are a strictly increasing function of the order of entry, such that E[t_app(i=N+1)]>E[t_app(N)] for all N.

Weak early-entrant protection: One drug has a lower expected approval time than all subsequent drugs.

Analysis of the model suggests that early entrant protection depends critically upon how the curing power of therapies unfolds over a sequence of drugs targeting a given disease. An instructive result comes when we assume that this unfolding is stationary, that is, it has constant expectation or mean.

Proposition 6. Concave Early-Entrant Protection Under a Stationary Quality Distribution

For any disease j=J, consider a sequence of disease-specific drugs i=1, …N_j. Then γ_i,j is the curing probability for drug i for disease j. Let f(γ) be a stationary curing distribution for any disease, such that E[γ_i]=g for all i∈J. Then strong early entrant protection holds, such that E[t_app(N+1)]>E[t_app(N)] for all i. Moreover, the degree of protection — the approval time advantage of earlier as opposed to later entrants—is concave in N_j.

Concavity is far from trivial. It means not only that the first entrant to any disease niche receives the quickest review, but also that the second entrant has the largest expected delay relative to the previous entrant. In other words, the degree of protection given to the first entrant is the greatest.

Possible Interactions between Capture and Noncapture Factors: The Effect of Firm Clout

The case of stationarity is instructive for several reasons. The degree of early-entrant protection that a firm receives is an increasing function of all of the other parameters of the approval payoff. In particular, firm clout (ρ_k) boosts the effect of early entrant protection, and vice versa. This result is depicted in Figure 3, where the concavity of the protection function also appears. Where large firms have political rents and can influence the agency, it is a genuine advantage for such firms to enter new markets early, as their competitors will have disadvantages both in influence and in reduced “political demand” for new drugs.

Early Entrant Protection as a Function of Firm Political Clout

The Generality of Early-Entrant Protection

More generally, for any two drugs i=N and N+1 with identical levels of danger (μ_N=μ_N+1), the expected approval time for the Nth drug is always shorter unless the (N+1)th offers an improvement in curing power.

Proposition 7

Given two drugs with identical danger (μ_N=μ_N+1), then unless the (N+1)th drug improves upon the curing of the Nth, EEP must hold. The inverse is not true.

Early-Entrant Protection When Drugs Improve Over Time

To assume the stationarity of f(γ_ij) is restrictive; it implies that the curing power of drugs does not improve as new ones are developed. It is useful, then, to consider the case of stochastic improvement, where newer drugs will be expected to have superior curative power relative to older ones, a process driven by technological change.

Once we admit stochastic improvement in f(γ_ij), then early-entrant protection does not hold for all sequences of products. Intuitively, if the regulator expects the first drug for any disease to cure 10% of patients and firmly expects the second drug to cure 90%, the agency will consistently find it worthwhile to approve the second more quickly. A generalizable result, then, requires limits on the path of γ_ij. A realistic assumption is concave improvement, which would reflect decreasing marginal returns to technology or R&D.

Proposition 8. Early-Entrant Protection for Concave Improvement Functions Over the Curing Distribution

Let the stochastic improvement function satisfy (1) dE[γ]/dN>0 and (2) d²E[γ]/dN²<0, for all N. Then there are at least two forms of early-entrant protection.

1. For γ₁ sufficiently high, there is strong early-entrant protection.
2. For any γ₁<1, and with maximal improvement consistent with concavity, there is always weak early-entrant protection, or protection for some set of drugs submitted and approved early in the sequence.

Assuming decreasing marginal returns to technology, simple restrictions on the improvement of the second drug over the first (or the third drug over the second) are sufficient to derive shorter expected review times for the first drug or drugs approved, relative to all other drugs. Proposition 8 shows that early-entrant protection holds even when the regulator expects later entrants to a market niche to offer quality improvements.

THE PROBLEM OF ENDOGENOUS SUBMISSIONS

The principal constraint of this model is its assumption that firm submissions are exogenous. Endogenous submissions may pose two potential problems. First, the regulator may face trivial uncertainty if the act of regulatory submission is itself a signal that completely reveals product quality or hazards. If the regulator's standards were high enough and firms submitted only “good” products submitted in equilibrium, then there would be little point in extended regulatory review. So one question is: Would regulatory uncertainty disappear or substantially wane in a strategic setting? Second, even if considerable uncertainty remained, the regulator would undoubtedly know that some sorts of submissions were unlikely, and hence the distribution of product hazards for observed submissions would likely be truncated or characterized by an upper bound. Do the results here hold for product hazard distributions of this character?

BOUNDED RATIONALITY: ADAPTATION, PLACEBOS, AND FORGETTING BY TURNOVER

The model developed here assumes a good deal of rationality; the agency behaves as if it were solving a complex dynamic programming problem, it conducts real-time Bayesian updating on a continuous stochastic process, and it knows and uses both past regulatory submissions and (expected properties of) future regulatory submissions. It is worth reflecting, then, on whether the results derived here would hold under the more realistic scenario of a boundedly rational regulator. While each of the following ideas would require separate papers to elaborate, there are at least three ways that such bounded rationality could be approached.

A first possibility is that the real-world agency's learning is not perfectly Bayesian but rather “adaptive” (e.g., Bendor, Mookherjee, and Ray 2001). Instead of using Bayes' rule to update firm histories, for instance, the regulator might keep a running tally of experience with a firm in which only sufficiently “bad” interactions enter into memory. Or the regulator might be subject to “confirmatory bias” (Rabin and Schrag 1999), whereby the first interaction with the firm strongly conditions inference from subsequent interactions. The key to familiarity advantages is that, as long as repeated experience increases (even slightly) the precision with which the regulator can estimate firm danger, the familiarity result can still obtain. Hence the familiarity result probably extends to cases of non-Bayesian updating but may not hold under confirmatory bias.

Second, patients lobbying the agency may be subject to “placebo learning,” that is, overestimation of the quality of a product by virtue of placebo effects. FDA approval is sometimes influenced by the emotive appeals of patients who testify that the drug worked for them and that they had no other option. Even if the drug is effective, some of these patients' responses may be placebo-related. If the agency incorporates such lobbying without discounting for this possibility, it may overestimate the product quality (curing power) and underestimate product hazards. In other words, bounded rationality on the part of organized interests may translate to bounded rationality on the part of the regulator.

Third, personnel turnover may create a regulator with imperfect memory. The model here assumes a regulator with continuous and perfect memory; the regulator remembers not only the entire history of the present product review, but also the history of each firm that submits. In real-world settings, memory limitations strongly influence optimization decisions (Mullainathan 2002). In reality, new bureaucratic officials arrive to agencies without built-in memories (the histories and cases must be learned), and officials who depart the agency take valuable firm-specific and case-specific information with them (which may not be immediately or costlessly transferable to the officials who replace them). Given the pervasiveness of turnover in government workforces, turnover as “organizational memory” becomes a crucial policy issue. As a conjecture, it is quite possible that large firms whose reputations are well known to the general public and scientific community may suffer less when turnover is high, because their external reputations “compensate” or “substitute” at some level for the regulator's loss of information. Hence organizational memory might in fact exacerbate large-firm and older-firm advantage in regulatory settings.

CONCLUSION AND EMPIRICAL IMPLICATIONS

Conventional wisdom follows Stigler in arguing that, upon observing distinct policy or regulatory advantages for apparently wealthy or well-organized firms, one should infer that capture, rent-seeking, or some other form of distributive politics is necessarily at play. Yet Stigler was wrong: evidence of policy advantage is quite insufficient for a conclusion supportive of capture theory. A neutral learning regulator motivated only by reputation protection and constrained by political responsiveness to consumers would also provide advantages to larger and older firms.

While it generates some results that are observationally equivalent to those of capture theory, the reputation and learning model also has some divergent empirical implications. First, Proposition 5 and Corollary 5.1 suggest that observed large-firm advantage in approval regulation should attenuate (perhaps entirely) once previous submissions are controlled for. Evidence for a reduced apparent firm size effect, controlling for familiarity, would be evidence against the capture account and evidence in favor of the model here.²¹

While capture theory might also predict early-entrant advantage in approval regulation, there are numerous predictions made by the model for which capture theory cannot account. Early entrant protection should be conditioned not on the political organization of producers, as capture theory might predict, but on the political organization of consumers. Hence the model here uniquely predicts interaction effects between order of entry and consumer political organization. In addition, the corollaries of Proposition 8—the importance of the quality of the first product on the market and the regulators' adaptation toward the pipeline of future products—would also be inconsistent with capture theory.

Recent studies have arrived at empirical results supportive of the reputation and learning perspective advanced here. Olson (1997) interprets firm characteristics as potential signals to uncertain regulators about unobserved product quality. The formalization here generally supports Olson's argument and, conversely, Olson's results provide empirical support for this model. Kyle's (2002) results also provide support for early-entrant protection and familiarity advantages. The regulatory advantage of familiarity may best be interpreted not as political “coziness” between regulator and industry (Bartel and Thomas 1987; Stigler 1975, 162–63), but as the result of repeated observation of a sequence of firm products by a rational agency.

Familiarity advantages can hold even when an uncertain regulator encounters firms with a below-average safety record. While not all such firms have regulatory advantages, some of them will still receive quicker approvals when their products are accepted. As a result, the model here departs in some respects from Olson's logic; familiarity advantages are distinct from (though not independent of) firm-specific quality differentials. A test distinguishing between Olson's predictions and mine would require data on the actual quality of a product. The model here predicts that, even controlling for product quality, a familiarity effect should hold such that a firm's expected approval times should be decreasing in the number of previous product submissions.

The reputation and learning model may also explain patterns of large-firm advantage that prevail in areas of regulation outside of the pharmaceutical industry. These patterns include quicker license approvals for older and more established firms in the electricity industry and quicker permit granting for and quicker disposition of litigation for older and larger firms in the enforcement of environmental regulation (Gordon 1999; Spence 1999).

In conclusion, the reputation and learning model suggests that even an apparently “neutral” regulatory regime can bequeath systematic advantages to larger, more established firms. This result should give pause to students of institutional design who believe that rigorous political insulation of an agency can create an even playing field for all firms large and small. In this light, the model is similar to other formal efforts demonstrating the impossibility of neutral administrative arrangements (Hammond and Thomas 1989).

The lesson for students of political economy is one of proper inference. Observed advantages for larger and older firms in regulation may say little or nothing about politics because they may exist for many reasons, some related to capture and some not. Whether or not capture prevails in a given industry, the advantage of familiarity and early-entrant protection may interact with capture to magnify large-firm advantage in regulated industries.