Journal of Health Economics 22 (2003) 151–185
`
`The price of innovation: new estimates
`of drug development costs
`Joseph A. DiMasi a,∗
`, Ronald W. Hansen b, Henry G. Grabowski c
`a Tufts Center for the Study of Drug Development, Tufts University, 192 South Street,
`Suite 550, Boston, MA 02111, USA
`b William E. Simon Graduate School of Business Administration, University of Rochester, Rochester, NY, USA
`c Department of Economics, Duke University, Durham, NC, USA
`
`Received 17 January 2002; received in revised form 24 May 2002; accepted 28 October 2002
`
`Abstract
`
`The research and development costs of 68 randomly selected new drugs were obtained from a sur-
`vey of 10 pharmaceutical firms. These data were used to estimate the average pre-tax cost of new drug
`development. The costs of compounds abandoned during testing were linked to the costs of com-
`pounds that obtained marketing approval. The estimated average out-of-pocket cost per new drug is
`US$ 403 million (2000 dollars). Capitalizing out-of-pocket costs to the point of marketing approval
`at a real discount rate of 11% yields a total pre-approval cost estimate of US$ 802 million (2000 dol-
`lars). When compared to the results of an earlier study with a similar methodology, total capitalized
`costs were shown to have increased at an annual rate of 7.4% above general price inflation.
`© 2003 Elsevier Science B.V. All rights reserved.
`
`JEL classification: L65; O31
`
`Keywords: Innovation; R&D cost; Pharmaceutical industry; Discount rate; Technical success rates
`
`1. Introduction
`
`Innovations in the health sciences have resulted in dramatic changes in the ability to treat
`disease and improve the quality of life. Expenditures on pharmaceuticals have grown faster
`than other major components of the health care system since the late 1990s. Consequently,
`the debates on rising health care costs and the development of new medical technologies
`have focused increasingly on the pharmaceutical industry, which is both a major participant
`in the health care industry and a major source of advances in health care technologies.
`Corresponding author. Tel.: +1-617-636-2116.
`∗
`E-mail address: joseph.dimasi@tufts.edu (J.A. DiMasi).
`
`0167-6296/03/$ – see front matter © 2003 Elsevier Science B.V. All rights reserved.
`doi:10.1016/S0167-6296(02)00126-1
`
`

`
`152
`
`J.A. DiMasi et al. / Journal of Health Economics 22 (2003) 151–185
`
`One of the key components of the discussion is the role of private sector pharmaceutical
`industry investments in R&D and an understanding of the factors that affect this process.
`Although the industry engages in many forms of innovation, in general the most significant
`is the discovery and development of new chemical and biopharmaceutical entities that
`become new therapies. Our prior research (DiMasi et al., 1991) found that the discovery
`and development of new drugs is a very lengthy and costly process. In the research-based
`drug industry, R&D decisions have very long-term ramifications, and the impact of market
`or public policy changes may not be fully realized for many years. From both a policy
`perspective, as well as an industrial perspective, it is therefore important to continue to
`analyze the components of and trends in the costs of pharmaceutical innovation.
`In this paper we will build on research conducted by the current authors (DiMasi et al.,
`1991) and others on the economics of pharmaceutical R&D. As we described in our prior
`study, “Empirical analyses of the cost to discover and develop NCEs are interesting on
`several counts. First, knowledge of R&D costs is important for analyzing issues such as
`the returns on R&D investment. Second, the cost of a new drug has direct bearing on
`the organizational structure of innovation in pharmaceuticals. In this regard, higher real
`R&D costs have been cited as one of the main factors underlying the recent trend toward
`more mergers and industry consolidation. Third, R&D costs also influence the pattern
`of international resource allocation. Finally, the cost of R&D has become an important
`issue in its own right in the recent policy deliberations involving regulatory requirements
`and the economic performance of the pharmaceutical industry”. In the decade that has
`followed the publication of our earlier study, these issues remain paramount. In addition,
`the congressional debates on Medicare prescription drug coverage and various new state
`initiatives to fill gaps in coverage for the elderly and the uninsured have intensified the
`interest in the performance of the pharmaceutical industry.
`In the current study we are not attempting to directly answer the policy debates men-
`tioned above. Rather, our focus is on providing new estimates of economic parameters
`associated with the drug development process. In particular, we concentrate on estimates
`of the costs of pharmaceutical innovation. Our prior estimates have been used by the Office
`of Technology assessment (OTA), the Congressional Budget Office (CBO), and various
`researchers to analyze policy questions such as the effects on R&D activities of health care
`financing reform or changes in intellectual property legislation related to the pharmaceutical
`industry.
`The approach used in this paper follows our previous study (DiMasi et al., 1991) and
`the earlier work by Hansen (1979). Given the similarity in methodologies, we are able
`to compare our results in the current study with the estimates in the earlier studies to
`illustrate trends in development costs. All three studies used micro-level data on the cost
`and timing of development obtained through confidential surveys of pharmaceutical firms
`for a random sample of new drugs first investigated in humans by these firms. In the current
`study, the new drugs were first tested in humans anywhere in the world between 1983 and
`1994. The reported development costs ran through 2000.Ultimately, we are interested in
`the expected cost of development per approved new drug. The uncertainties in the research
`and development process result in expenditures on many development projects that are not
`successful in producing a marketed product. However, to produce an estimate of expected
`cost for a marketed product, we must allocate the costs of the unsuccessful projects to those
`
`

`
`J.A. DiMasi et al. / Journal of Health Economics 22 (2003) 151–185
`
`153
`
`that result in a marketed new product. The R&D process is lengthy, and as such it is important
`to know at what stage of development expenses occur. Viewed as an investment project, it
`is necessary to know both the amount of expenditures and the timing of these expenditures,
`since funds committed to R&D in advance of any returns from sales have both a direct
`and an opportunity cost. We used a unique database to estimate various cost parameters
`in the development process. Of particular concern is the estimation of the average pre-tax
`cost of new drug development, since we are interested in the resource costs of new drug
`development and how they have changed over time.
`
`1.1. Previous studies of the cost of pharmaceutical innovation
`
`A summary of early studies of the cost of drug development can be found in the authors’
`previous study (DiMasi et al., 1991) and in OTA (1993). In brief, the early studies were
`either based on a case study of a specific drug (usually ignoring the cost of failed projects)
`or relied on aggregate data. Since the R&D process often extends for a decade or more
`and the new drug development process often changes, it is difficult to estimate the cost
`of development from aggregated annual data. In contrast, the study by Hansen (1979)
`and the current authors’ previous study (DiMasi et al., 1991) estimated development cost
`based on data supplied by firms for a representative sample of drug development
`efforts.
`DiMasi et al. (1991) used data on self-originated new drugs to estimate the average cost
`of developing a new drug. They obtained data from 12 pharmaceutical firms on the research
`and development costs of 93 randomly selected new drugs that entered clinical trials be-
`tween 1970 and 1982. From these data they estimated the average pre-tax out-of-pocket
`cost per approved drug to be US$ 114 million (1987 dollars). Since these expenditures
`were spread out over nearly a dozen years, they capitalized these expenditures to the date
`of marketing approval using a 9% discount rate. This yielded an estimate of US$ 231
`million (1987 dollars). Measured in constant dollars, this value is more than double that
`obtained by Hansen for an earlier sample. DiMasi et al. (1991) also found that the average
`cost of the first two phases of clinical trials doubled between the first and second half of
`their sample. This led to the expectation that development costs would be higher in future
`samples.
`Based on an analysis by Myers and Shyam-Sunder performed for the OTA, the OTA
`(1993) report noted that the cost-of-capital for the industry was roughly 10% in the early
`1980s. This is moderately higher than the 9% used by DiMasi et al. (1991). The OTA also
`recalculated the DiMasi et al. (1991) numbers using an interest rate that varied over the life
`of the R&D cycle thereby raising the cost estimate by US$ 100 million in 1990 dollars.1
`The OTA presented both pre- and post-tax cost estimates.
`
`1 The OTA applied a range of discount rates that varied with the time to marketing approval. They chose 14%
`for the earliest stage R&D and 10% for development just prior to approval, with rates in between that declined
`linearly with time in development. This approach was meant to capture the essence of the risk-return staircase
`perspective expressed by Myers and others, and discussed below. The methodology described in Myers and Howe
`(1997) is actually quite different, but the OTA technique yielded results that would not be much different (for the
`same distribution of costs) than what one would have obtained with the correct methodology (Myers and Howe,
`1997, p. 33).
`
`

`
`154
`
`J.A. DiMasi et al. / Journal of Health Economics 22 (2003) 151–185
`
`Fig. 1. Inflation-adjusted industry R&D expenditures (2000 dollars) and US new chemical entity (NCE) approvals
`from 1963 to 2000. Source of data: PhRMA (2001) and Tufts CSDD Approved NCE Database.
`
`1.2. Aggregate data analyses
`
`There have been no recent comprehensive studies of the cost of developing new pharma-
`ceuticals from synthesis to marketing approval based on actual project-level data. However,
`aggregate data and data on parameters of the drug development process suggest that R&D
`costs have increased substantially since our earlier study. For example, the Pharmaceutical
`Research and Manufacturers of America (PhRMA, 2000) publishes an annual report on the
`R&D expenditures of its member firms that shows a continuous increase in outlays well in
`excess of inflation. Reports on specific components of the R&D process, such as the number
`of subjects in clinical trials (OTA, 1993; The Boston Consulting Group [BCG], 1993), also
`suggest an increase in the real cost of pharmaceutical innovation.
`Published aggregate industry data suggest that R&D costs have been increasing. Fig. 1
`shows reported aggregate annual domestic prescription drug R&D expenditures for mem-
`bers of the US pharmaceutical industry since 1963. The chart also shows the number of
`US new drug approvals by year. Given the much faster rate of growth of R&D expendi-
`tures, data such as these suggest that R&D costs have increased over time. However, they
`cannot be conclusive or precise. For one matter, the drug development process is known
`to be very lengthy. Thus, new drug approvals today are associated with R&D expenditures
`that were incurred many years prior. Ignoring the inherent lag structure underlying these
`data and simply dividing current R&D expenditures by the number of new drug approvals
`will in general yield inaccurate estimates.2 Given a substantial increasing trend in R&D
`
`2 The estimates would also vary widely from year-to-year. For example, if we divided each year’s real R&D
`expenditures by that year’s number of NCE approvals, we would obtain US$ 1 billion for 2000, US$ 743 million
`for 1999, US$ 839 million for 1998, US$ 568 million for 1997, US$ 400 million for 1996, US$ 635 million for
`1995, and US$ 878 million for 1994. While there is a general upward trend in such calculations, the year-to-year
`variability is not credible.
`
`

`
`J.A. DiMasi et al. / Journal of Health Economics 22 (2003) 151–185
`
`155
`
`expenditures, such calculations will result in greatly exaggerated estimates of out-of-pocket
`cost per approval.
`Secondly, even properly lagged time series would tend to be imprecise if aggregate in-
`dustry data were used as reported. The industry data include expenditures on improvements
`to existing products. Thus, they would overestimate pre-approval development costs. On
`the other hand, they also do not incorporate all of the R&D on licensed-in drugs since
`firms or other organizations that are not members of the US trade association would have
`conducted some of the work. On that account the data would tend to underestimate costs.
`Therefore, R&D cost estimates based on project-level data are needed to assure a reasonable
`level of confidence in the accuracy of the results. We present results based on such data in
`this study.
`The remainder of this paper is organized as follows. Section 2 describes the standard
`drug development paradigm, which serves as the structure through which the results are
`reported. Section 3 contains a description of the survey sample data and the population from
`which it was drawn. Section 4 describes the methodology used to derive R&D cost esti-
`mates. We present our base case pre-marketing approval R&D cost estimates in Section 5,
`as well as a comparison of our results with those of earlier studies to examine R&D cost
`trends. Section 6 provides sensitivity analyses for key parameters. Section 7 focuses on
`some extensions of the base case analyses: estimates of clinical development costs for ap-
`proved drugs by therapeutic significance, estimates of post-approval R&D costs, and a tax
`analysis. Section 8 contains data and analyses that corroborate our results. Finally, we offer
`some conclusions in Section 9.
`
`2. The new drug development process
`
`New drug development can proceed along varied pathways for different compounds, but
`a development paradigm has been articulated that has long served well as a general model.
`The paradigm is explained in some detail elsewhere (DiMasi et al., 1991; US Food and Drug
`Administration [FDA], 1999). In outline form, the paradigm portrays new drug discovery
`and development as proceeding in a sequence of (possibly overlapping) phases. Discovery
`programs result in the synthesis of compounds that are tested in assays and animal models.
`It was not possible to disaggregate our data into discovery and preclinical development
`testing costs,3 so for the purposes of this study discovery and preclinical development costs
`are grouped and referred to as preclinical costs.
`Clinical (human) testing typically proceeds through three successive phases. In phase I,
`a small number of usually healthy volunteers4 are tested to establish safe dosages and to
`gather information on the absorption, distribution, metabolic effects, excretion, and toxicity
`of the compound. To conduct clinical testing in the United States, a manufacturer must first
`
`3 The reported basic research expenditures by firm were highly variable, and suggest that different firms may
`categorize their pre-human testing expenditures somewhat differently. Thus, we report pre-human testing costs in
`one figure.
`4 In some therapeutic areas, testing is initially done on patients who have the disease or condition for which the
`compound is intended to be a treatment. This is ordinarily true in the cancer and AIDS areas.
`
`

`
`156
`
`J.A. DiMasi et al. / Journal of Health Economics 22 (2003) 151–185
`
`file an investigational new drug application (IND) with the FDA. However, initiation of
`human testing can, and often does, occur first outside the United States.
`Phase II trials are conducted with subjects who have the targeted disease or condition
`and are designed to obtain evidence on safety and preliminary data on efficacy. The number
`of subjects tested in this phase is larger than in phase I and may number in the hundreds.
`The final pre-approval clinical testing phase, phase III, typically consists of a number of
`large-scale (often multi-center) trials that are designed to firmly establish efficacy and to
`uncover side-effects that occur infrequently. The number of subjects in phase III trials for
`a compound can total in the thousands.
`Once drug developers believe that they have enough evidence of safety and efficacy,
`they will compile the results of their testing in an application to regulatory authorities
`for marketing approval. In the United States, manufacturers submit a new drug appli-
`cation (NDA) or a biological license application (BLA) to the FDA for review and
`approval.
`
`3. Data
`
`Ten multinational pharmaceutical firms, including both foreign and US-owned firms,
`provided data through a confidential survey of their new drug R&D costs.5 Data were
`collected on clinical phase costs for a randomly selected sample of the investigational drugs
`of the firms participating in the survey.6 The sample was taken from a Tufts Center for the
`Study of Drug Development (CSDD) database of investigational compounds. Cost and time
`data were also collected for expenditures on the kind of animal testing that often occurs
`concurrently with clinical trials.7 The compounds chosen were all self-originated; that is,
`their development up to initial regulatory marketing approval was conducted under the
`auspices of the surveyed firm.8 Licensed-in compounds were excluded because non-survey
`firms would have conducted portions of the R&D.9
`We also collected data from the cost survey participants on their aggregate annual phar-
`maceutical R&D expenditures for the period 1980–1999. The firms reported on total an-
`nual R&D expenditures broken down by expenditures on self-originated new drugs, on
`licensed-in or otherwise acquired new drugs, and on already-approved drugs. Annual ex-
`penditures on self-originated new drugs were further decomposed into expenditures during
`the pre-human and clinical periods.
`The National Institutes of Health (NIH) support through their own labs and through grants
`to researchers in academic and other non-profit institutions a substantial amount of research
`
`5 Using pharmaceutical sales to measure firm size, four of the survey firms are top 10 companies, another four
`are among the next 10 largest firms, and the remaining two are outside the top 20 (PJB, 2000).
`6 A copy of the survey instrument is available upon request.
`7 Long-term teratogenicity and carcinogenicity testing may be conducted after the initiation of clinical trials.
`8 This does not preclude situations in which the firm sponsors trials that are conducted by or in collaboration
`with a government agency, an individual or group in academia, a non-profit institute, or another firm.
`9 Large pharmaceutical firms much more often license-in than license-out new drug candidates. Firms that
`license-in compounds for further development pay a price for that right through up-front fees, milestone payments,
`and royalty arrangements.
`
`

`
`J.A. DiMasi et al. / Journal of Health Economics 22 (2003) 151–185
`
`157
`
`that expands fundamental knowledge about human biology (NIH, 2000; Scherer, 2000).
`This basic research sometimes results in leads that industrial researchers can capitalize
`on to assist them in discovering new therapeutic compounds.10 Some new compounds
`investigated by pharmaceutical firms, however, originated in government or academic labs.
`It is unclear whether the discovery and early development costs for such compounds are
`similar to those for compounds originating in industrial labs. These drugs, though, represent
`a very small portion of the total number developed. For example, NIH (2000) found that
`of 47 FDA-approved drugs that had reached at least US$ 500 million in US sales in 1999,
`the government had direct or indirect use or ownership patent rights to only four of them.11
`In addition, we used a Tufts CSDD database supplemented by commercial databases to
`determine that of the 284 new drugs approved in the United States from 1990 to 1999,12
`93.3% originated from industrial sources (either from the sponsoring firm or from another
`firm from which the compound was licensed or otherwise acquired). Government sources
`accounted for 3.2% of these approvals and academia and other non-profits accounted for
`the other 3.5%.13
`The survey firms accounted for 42% of pharmaceutical industry R&D expenditures.14
`The survey compounds were selected at random from data contained in the Tufts CSDD
`database of investigational compounds for the firms that agreed to participate in the R&D
`cost survey. Of the 68 compounds chosen, 61 are small molecule chemical entities, four are
`recombinant proteins, two are monoclonal antibodies, and one is a vaccine. Initial human
`testing anywhere in the world for these compounds occurred during the period 1983–1994.
`Development costs were obtained through 2000.15
`
`10 The NIH also supports the development of research tools that drug developers find useful. In addition, it funds
`training for many scientists, some of whom eventually are employed in the industrial sector.
`11 The four drugs were developed in part through the use of NIH-funded patented technologies. Three of the
`four products are recombinant proteins, with two being the same drug produced by two different companies. Each
`of the relevant patented technologies was developed at academic or non-profit institutions with financial support
`from the NIH.
`12 The definition of a new drug used for this analysis is a therapeutic new molecular entity approved by the FDA’s
`Center for Drug Evaluation and Research.
`13 The proportion of investigational drugs that derive from industrial sources is likely to be even higher, since
`acquired drugs have higher clinical approval success rates than do self-originated drugs (DiMasi, 2001b). Our
`cost survey firms were less reliant on licensing-in drugs from non-industrial sources than were firms as a whole;
`98.8% of their new drug approvals during 1990–1999 were from industrial sources. DiMasi (2000) found markedly
`greater market entry of small niche pharmaceutical firms in the 1990s relative to earlier periods as measured by
`sponsorship of new chemical entity (NCE) approvals. A disproportionate share of the approvals obtained by these
`new entrants was for drugs that originated in academia.
`14 The data used were aggregate firm pharmaceutical R&D expenditures for the cost survey firms, as reported
`on our questionnaire, in comparison to PhRMA member firm R&D expenditures (1994–1997) on ethical pharma-
`ceuticals, adjusted to global expenditure levels (PhRMA, 2001).
`15 Surveys were sent to 24 firms (some of whom have since merged). Twelve firms responded that they would
`participate in some form. The data that two firms ultimately provided were not useable. The 10 firms from which
`we used data provided information on 76 compounds. However, the data for eight of these compounds were not
`sufficiently comprehensive to use. The firms that did not participate in the survey cited a number of reasons for
`not doing so. The reasons included the extra demands that the transition effects of a relatively recent merger were
`placing on their relevant personnel, the time and expense of retrieving archival records in the manner required by
`the study, and difficulties in gathering the relevant data in a uniform manner because their accounting systems had
`changed significantly over the study period.
`
`

`
`158
`
`J.A. DiMasi et al. / Journal of Health Economics 22 (2003) 151–185
`
`We selected a stratified random sample of investigational compounds. Stratification was
`based on the time elapsed since the origination of clinical trials and the current status of
`that testing. Reported costs were weighted to reflect the characteristics of the population,
`so that knowledge of the population from which the sample was drawn was needed. The
`population is composed of all investigational compounds in the Tufts CSDD investigational
`drug database that met study criteria: the compounds were self-originated and first tested in
`humans anywhere in the world from 1983 to 1994, and we had the information necessary
`to classify them according our strata. We found 538 investigational drugs that met these
`criteria. Of these compounds, 82 (15.2%) have been approved for marketing, 9 (1.7%) had
`NDAs or BLAs that were submitted and are still active, 5 (0.9%) had NDAs or BLAs
`submitted but abandoned, 227 (42.2%) were terminated in 4 years or less from the initiation
`of clinical trials, 172 (32.0%) were terminated more than 4 years after the start of clinical
`testing, and 43 (8.0%) were still in active testing as of the most recent check (31 March
`2001).
`Some firms were not able to provide full phase cost data for every new drug sampled.
`For example, phase I cost data were available for 66 of the 68 new drugs. However, we
`had some phase cost data for every drug in the sample. In addition, five compounds were
`still active at the time of the study. For these drugs it is possible that there will be some
`future costs for the drug’s most recent phase. Thus, for this reason our cost estimates may
`be somewhat conservative. However, given the small number of drugs in this category and
`the fact that the impact would be on only one phase for each of these drugs, our overall cost
`estimates are not likely to be materially affected.
`
`4. Methodology for estimating new drug development costs
`
`The approach that we use to estimate development costs is similar to that described in
`our earlier work (DiMasi et al., 1991). We will outline here the general methodology for
`developing an overall cost estimate. In describing the approach, it will be clear that cost
`estimates for important components of the drug development process will also be derived
`along the way.
`The survey sample was stratified to reduce sampling error. Results from previous anal-
`yses suggested that the variability of drug costs tends to increase with the development
`phase or the amount of time that a drug spends in testing (Hansen, 1979; DiMasi et al.,
`1991). Costs for successful drugs (i.e. those that achieve regulatory approval) also tend to
`be higher and more variable than those for drug failures. Thus, we based our strata on the
`length of time that failed compounds were in clinical testing and whether or not a compound
`had reached the stage in which an application for marketing approval had been filed with
`the FDA.16
`
`16 Specifically, we used four strata: compounds that failed in 4 years or less of clinical testing; compounds that
`failed after more than 4 years had elapsed from initial human testing; compounds for which an NDA or a BLA had
`been submitted to the FDA; and compounds that were still in active testing (as of 30 March 2001). Compounds
`for which an application for marketing approval had been submitted or which had been abandoned after lengthy
`testing were deliberately oversampled. The reported sample values were then weighted, where the weights were
`determined so that the sample perfectly reflects the population in terms of the four strata.
`
`

`
`J.A. DiMasi et al. / Journal of Health Economics 22 (2003) 151–185
`
`159
`
`4.1. Expected costs in the clinical period
`
`Since new drug development is a risky process, with many compounds failing for every
`one that succeeds, it is necessary to analyze costs in expected value terms. The total clinical
`period cost for an individual drug can be viewed as the realization of a random variable, c.
`Given that it is not certain that development of a randomly selected investigational compound
`will proceed to a given phase, we may define expected clinical costs for a randomly selected
`investigational drug to be C = E(c) = pIµI|e+pIIµII|e+pIIIµIII|e+pAµA|e, where pI, pII,
`and pIII, are the probabilities that a randomly selected investigational compound will enter
`phases I–III, respectively, pA the probability that long-term animal testing will be conducted
`during the clinical trial period, and the µ’s are conditional expectations. Specifically, µI|e,
`µII|e, µIII|e, and µA|e are the population mean costs for drugs that enter phases I–III, and
`clinical period long-term animal testing, respectively.
`Weighted mean phase costs derived from the cost survey data were used to estimate the
`conditional expectations. A description of how the probabilities were estimated is presented
`in the next section. Assuming that the estimated mean phase costs and success probabilities
`are stochastically independent, the estimated expected value is an unbiased estimate of the
`population expected value.
`
`4.2. Clinical success and phase attrition rates
`
`An overall clinical approval success rate is the probability that a compound that enters the
`clinical testing pipeline will eventually be approved for marketing. Attrition rates describe
`the rate at which investigational drugs fall out of testing in the various clinical phases. A
`phase success rate is the probability that a drug will attain marketing approval if it enters the
`given phase. A phase transition probability is the likelihood that an investigational drug will
`proceed in testing from one phase to the next. All of these probabilities can be estimated
`from data in the Tufts CSDD database of investigational drugs from which our survey
`sample was drawn.
`The clinical approval success rate was estimated using a two-stage statistical estimation
`process that has been described in detail elsewhere (DiMasi et al., 1991; DiMasi, 2001b). The
`data used here consist of the investigational drugs in the Tufts CSDD database that were first
`tested in humans anywhere in the world from 1983 to 1994, with information on their status
`(approval or research abandonment) obtained through early 2001. Given that some of these
`investigational drugs were still in active testing at the end of the study period, some of the
`data are right-censored. Survival analysis can be applied in such a situation, where survival
`indicates that a drug has not reached its ultimate fate (either approval or abandonment).
`The Tufts CSDD database of investigational compounds contains information on the
`latest phase that an abandoned compound was in when it was terminated. These data were
`used to determine the distribution of research terminations by phases.17 These results,
`
`17 A small proportion of the compounds in the database were either still in clinical development (8.0%) or had
`an NDA or BLA filed but not yet approved (1.7%). For those drugs in these groups that will eventually fail, their
`abandonment will tend to occur in later testing phases. To deal with the potential bias in the estimated distribution
`of research terminations that would result from using just those compounds that had been abandoned by the end of
`
`

`
`160
`
`J.A. DiMasi et al. / Journal of Health Economics 22 (2003) 151–185
`
`together with the estimated overall clinical approval success rate were used to provide
`estimates of the probability that an investigational drug will enter a given phase, phase
`attrition rates, and phase transition probabilities. The estimated overall clinical approval
`success rate and the probabilities of entering various phases provide results with which
`estimates can be derived that include the cost of drugs that fail to make it through the
`development process. Specifically, we use the probabilities of entering a phase to estimate
`the expected out-of-pocket clinical cost per investigational drug. Adding the out-of-pocket
`preclinical cost estimate described below yields an estimate of total out-of-pocket cost per
`investigational drug. Dividing this estimate by the overall clinical success rate yields our
`estimate of out-of-pocket cost per approved drug.
`
`4.3. Out-of-pocket discovery and preclinical development costs
`
`Many costs incurred prior to clinical testing cannot be attributed to specific compounds.
`Thus, aggregate level data at the firm level were used to impute costs per drug for R&D
`incurred prior to human testing. Specifically, time series data for each surveyed firm on
`spending on pre-human R&D and on human testing for 1980–1999 were obtained, and a
`ratio of pre-human R&D expenditures to human testing expenditures was determined based
`on an appropriate lag structure (on average, pre-human R&D expenditures should occur
`years prior to the associated human testing costs). This ratio was then multiplied by an
`estimate of out-of-pocket clinical cost per drug, which is based on the project-level data, to
`yield an estimate of the pre-human R&D cost per new drug.18
`
`4.