Microsoft word - lichtenberg seer-medstat paper 2006-12-31.doc

Pharmaceutical innovation and U.S. cancer survival in the 1990s:
evidence from linked SEER-MEDSTAT data
Frank R. Lichtenberg
Pharmaceutical innovation and U.S. cancer survival in the 1990s:
evidence from linked SEER-MEDSTAT data
Abstract
This study examines the impact of pharmaceutical innovation and other factors on the survival of U.S. cancer patients during the 1990s. In particular, it investigates whether cancer survival rates increased more for those cancer sites that had the largest increases in the proportion of drug treatments that were “new” treatments. We control for “expected survival,” i.e. the survival of a comparable set of people that did not have cancer, thereby measuring the excess mortality that is associated with a cancer diagnosis. We also control for other types of medical innovation, i.e. innovation in surgical procedures, diagnostic radiology procedures, and radiation oncology procedures. Data on observed and expected survival rates, the number of people diagnosed, mean age at diagnosis, and stage distribution are obtained from the National Cancer Institute’s Surveillance, Epidemiology, and End Results (SEER) 1973-2003 Public-use Data. Estimates of rates of innovation in drugs and other treatment and diagnostic procedures were constructed from the MEDSTAT Marketscan database and other data sources. We compute weighted least-squares estimates of 12 versions of a survival model, based on different survival intervals, functional forms, and sets of weights. The drug vintage coefficient is positive and significant in almost every model. This indicates that the cancer sites whose drug vintage (measured by the share of post-1990 treatments) increased the most during the 1990s tended to have larger increases in observed survival rates, ceteris paribus. Estimates of the fraction of the 1992-1999 change in the observed survival rate that is attributable to the increased utilization of post-1990 drugs range from 12% to 121%. The estimated fraction is higher for shorter survival intervals, when observations are weighted by the number of MEDSTAT drug treatments, and for the logarithmic specification. The mean of the 12 estimates of the fraction of the 1992-1999 change in the observed survival rate that is attributable to the increased utilization of post-1990 drugs is 44%. Due to sampling and other measurement errors, these estimates may be conservative. The coefficients on measures of other types of medical innovation (in radiation oncology, diagnostic radiology, and surgery innovation) are generally not significant. However these measures may be less reliable than the drug innovation measure: they are based upon the year in which the AMA established a new procedure code, which may be a far less meaningful indicator of innovation than the year in which the FDA first approved a drug. This topic warrants further research. Frank R. Lichtenberg Graduate School of Business Columbia University 614 Uris Hall, 3022 Broadway New York, NY 10027 Phone: (212) 854-4408 frank.lichtenberg@columbia.edu Many clinical studies have compared the effects of newer and older drugs on cancer survival rates.1 The findings of these studies have been mixed. Some studies have found that use of newer cancer drugs increased survival rates. For example, Richardson et al (2005) compared bortezomib (FDA approved May 2003) with high-dose dexamethasone (FDA approved October 1958) in patients with relapsed multiple myeloma who had received one to three previous therapies. They found that patients treated with bortezomib had a longer survival than patients treated with dexamethasone: the one-year survival rate was 80 percent among patients taking bortezomib and 66 percent among patients taking dexamethasone (P=0.003), and the hazard ratio for overall survival with bortezomib was 0.57 (P=0.001). Similarly, Kantarjian et al (2005) concluded that imatinib mesylate (FDA approved May 2000) improved survival compared with other therapies in patients with accelerated-phase chronic myelogenous Other studies have found that use of newer cancer drugs did not increase survival rates. For example, von der Maase et al (2005) compared long-term survival in patients with locally advanced or metastatic transitional cell carcinoma of the urothelium treated with cisplatin and either gemcitabine (FDA approved May 1996) or methotrexate/vinblastine/doxorubicin (all of which were approved before 1975). A total of 405 patients were randomly assigned: 203 to the gemcitabine/cisplatin arm and 202 to the methotrexate/vinblastine/doxorubicin/cisplatin arm. Overall survival was similar in This paper will seek to determine the effect of pharmaceutical innovation—the use of newer drugs—in general on cancer survival rates. A reliable estimate of this effect can’t be obtained by simply surveying previous clinical studies of specific drugs and cancer sites, for two reasons. First, there is considerable variation in the methodology and metrics used in these studies, rendering comparison and aggregation difficult. 1 A PubMed search for (("Survival Rate") AND ("Antineoplastic Agents")) AND ("Comparative Study")) yields 387 items. Second, these studies may not provide a complete or representative picture; there may be little or no published evidence about the survival impact of some drugs.2 We will investigate whether cancer survival rates increased more for those cancer sites that had the largest increases in the proportion of drug treatments that were “new” treatments. We will control for “expected survival,” i.e. the survival of a comparable set of people that did not have cancer, thereby measuring the excess mortality that is associated with a cancer diagnosis. We will also control (imperfectly) for other types of medical innovation, i.e. innovation in surgical procedures, diagnostic radiology procedures, and radiation oncology procedures. Section I of this paper sketches a simple theory of cancer survival. Section II presents an econometric specification based on this theory. Section III describes the construction of data used to estimate this model. Estimation issues are discussed in Section IV. Empirical results are presented in Section V. Section VI contains a summary I. A simple theory of cancer survival
S = observed survival rate E = expected survival rate R = S / E = relative survival rate3 Q = treatment quality P = disease progression at time of diagnosis V = treatment vintage We postulate the following simple theory of cancer survival: R = S / E = f(Q, P) where f’Q > 0 and f’P < 0, or, more generally, where f’E > 0, f’Q > 0 and f’P < 0. 2 Johnson et al (2003) reported that only one-fourth of the oncology drug marketing applications approved by the FDA during the period January 1, 1990 to November 1, 2002 were based on direct evidence of survival benefits; 75% of approvals were based on surrogate end points (e.g. reduction in tumor size). 3 Ederer et al (1961). The observed survival rate is hypothesized to be an increasing function of expected survival and the quality of treatment, and a decreasing function of disease progression at time of diagnosis. Moreover, we hypothesize that treatment quality is an increasing where f’V > 0. Substituting (3) into (2), where f’E > 0, f’V > 0 and f’P < 0. The observed survival rate is hypothesized to be an increasing function of expected survival and treatment vintage, and a decreasing function of disease progression at time of diagnosis. Our primary objective is to estimate the effect of treatment vintage (V) on survival (S). Equation (4) indicates that if P is correlated with V, it is necessary to control adequately for P to obtain an unbiased estimate of the effect of treatment vintage on Measuring progression (or severity) of disease is often challenging in health economics. We will include five variables (or groups of variables) postulated to be indicators or determinants of the mean progression of disease: (1) Time dummies (“year effects”): control for changes in mean disease progression (2) Stage distribution of disease: the fraction of patients diagnosed with in situ (stage 0), localized/regional (stages 1 and 2),5 and distant (stage 4) cancer.6 (3) Vintage of diagnostic radiology procedures. Use of newer diagnostic radiology procedures may result in earlier detection, i.e. a reduction in P. (4) Number of people diagnosed and mean age at diagnosis. Improvements in diagnostic technology are likely to lead to earlier detection. This would result in an increase in the number of people diagnosed and a reduction (or below-average increase) in their mean age. 4 The vintage of a treatment is the year in which the treatment was first used. For example, the vintage of a drug is the year that the drug’s active ingredient was first approved by the FDA. 5 We combine stages 1 and 2 because these two stages are merged in the case of prostate cancer in SEER data. 6 The omitted stage category is “unstaged” (SEER Historic Stage A 9). All lymphomas and leukemias are considered unstaged II. Econometric specification
Based on this theory, we propose the following econometric model of observed f(Sit) = β1 drug_new%it + β2 rad_onc_new%it + β3 rad_diag_new%it + β4 surg_new%it + β5 f(Eit) + β6 ln(Nit) + β7 ageit + β8 in_situ%it + β9 loc_reg%it + β10 distant%it Sit = the observed survival rate of people diagnosed with cancer originating at site i in year t. The observed survival rate is the probability of surviving all causes of death for a specified time interval. Observed survival does not consider cause of death, it simply looks at who is alive and who is not. drug_new%it = % of drug treatments administered in year t associated with cancer originating at site i that used drugs approved by the FDA after 1990 rad_onc_new%it = % of radiation oncology procedures performed in year t associated with cancer originating at site i whose CPT codes were established by the American Medical Association (AMA) after 1990 rad_diag_new%it = % of diagnostic radiation procedures performed in year t associated with cancer originating at site i whose CPT codes were established by the AMA after 1990 surg_new%it = % of surgical procedures performed in year t associated with cancer originating at site i in year t whose CPT codes were established by the AMA after 1990 Eit = the expected survival rate of people diagnosed with cancer originating at site i in year t. The expected survival rate is the observed survival rate of a comparable (in terms of race, sex, and age) set of people who do not have cancer. ageit = the mean age of people diagnosed with cancer originating at site Nit = the number of people diagnosed with cancer originating at site i in_situ%it = the fraction of cancers originating at site i in year t that were loc_reg%it = the fraction of cancers originating at site i in year t that were diagnosed as localized or regional (stage 1 or 2) distant%it = the fraction of cancers originating at site i in year t that were Due to the presence of fixed cancer-site effects and year effects, this is a difference-in-differences model. A positive and significant estimate of β1 would signify that there were above-average increases in observed survival rates of cancer sites with above average increases in drug_new%, ceteris paribus. Since the expected survival rate is based on the age- (as well as race- and sex-) distribution of a comparable set of people who do not have cancer, controlling for mean age as well as expected survival may be redundant. III. Data construction
Survival data. Data on observed and expected survival rates, the number of people diagnosed, mean age at diagnosis, and stage distribution were obtained from the National Cancer Institute’s Surveillance, Epidemiology, and End Results (SEER) 1973-2003 Public-use Data. I used data from SEER 9 registries, which are Atlanta, Connecticut, Detroit, Hawaii, Iowa, New Mexico, San Francisco-Oakland, Seattle-Puget Sound, and Utah. In this data set, cases diagnosed from 1973 through 2003 are available for all registries except Seattle-Puget Sound (1974+) and Atlanta (1975+). The database contains one record for each of 3,260,176 tumors. However, the treatment innovation measures can only be constructed for the period 1992-2003. Cancer cases were classified using the SEER Cancer Causes of Death Recode 1969+ (9/17/2004).7 This classification includes 68 non-overlapping types of cancer. Survival rates may be calculated for a variety of time intervals (1-year, 2-year, etc.). We will estimate models of 1-year, 2-year, and 3-year survival rates. (The longer the time interval, the shorter the available time series.) Table 1 shows 1992 and 2000 2- year survival data for the top 30 (ranked by 1992 incidence) cancer sites, or groups of cancer sites. The 2-year relative survival rate for all cancer sites combined increased from 72% in 1992 to 75% in 2000. The 1992-2002 change in survival rates varied considerably across cancer sites. For example, the relative survival rate for Cervix Uteri declined from 86% to 81%, while the relative survival rate for Skin excluding Basal and Squamous increased from 83% to 97%. The relative survival rate for Leukemia declined 7 http://seer.cancer.gov/codrecode/1969+_d09172004/index.html from 59% to 57%, while the relative survival rate for Non-Hodgkin Lymphoma increased Treatment vintage data. First I will describe the construction of the drug innovation measure (drug_new%it). A similar approach was used to construct the other innovation measures. The drug innovation measure was defined as follows: drug_new%it = ∑p FREQpit POST1990p / ∑p FREQpit FREQpit = the number of times drug p was used to treat cancer originating at site i POST1990p = 1 if drug p was approved by the FDA after 1990 Data on utilization of medical procedures and products, by diagnosis and year (FREQpit), were obtained from the MEDSTAT Marketscan database. MEDSTAT contains data on outpatient and inpatient services (procedures) and outpatient prescriptions of hundreds of thousands, or even millions, of individuals. It is worth distinguishing between two types of drugs: self-administered drugs, and drugs administered by physicians and other medical providers (e.g., chemotherapy). Utilization of self-administered drugs is reported in outpatient prescription records (claims). These records generally don’t include any information about the patient’s diagnosis. In contrast, drugs that are administered by physicians and other medical providers are reported as outpatient and inpatient services (procedures). These records include information about the patient’s diagnosis. For most diseases other than cancer, the vast majority of drugs are self- administered, and determining the diagnosis associated with a particular drug’s use (hence measuring FREQpit) can be difficult. But an important fraction of drug treatments for cancer are administered by physicians and other medical providers. We will use data on provider-administered drugs only, since the number of times provider-administered drug p was used to treat cancer originating at site i in year t can be measured precisely. Each MEDSTAT outpatient and inpatient service record contains one procedure code and one or more ICD-9 diagnosis codes. Codes for drugs administered by providers are Healthcare Common Procedure Coding System (HCPCS) Level II Codes.8 Table 2 shows the top 40 (ranked by frequency) provider-administered drugs associated with all Only about a third of the drug treatments administered to cancer patients involve cancer drugs (antineoplastic agents). We will report estimates of two versions of eq. (5): one does not distinguish between cancer drugs and other drugs, and the other does. Table 3 shows a comparison of the top 40 provider-administered drugs associated with two major cancer sites (colon and breast) in 2003. We used Multum’s Lexicon database (http://www.multum.com/Lexicon.htm) to determine the active ingredients of the drugs corresponding to each of these HCPCS Level II Codes. We used data from the Drugs@FDA database (http://www.fda.gov/cder/drugsatfda/datafiles/default.htm) to determine the year in which each active ingredient was first approved by the FDA. The following table shows the mean value of drug_new%, and the number of provider-administered drugs upon which that statistic is based, for all cancer sites drug_new% Number of provider-administered Number of firms covered by 8 Level II of the HCPCS is a standardized coding system that is used primarily to identify products, supplies, and services not included in the CPT codes, such as ambulance services and durable medical equipment, prosthetics, orthotics, and supplies (DMEPOS) when used outside a physician's office. Because Medicare and other insurers cover a variety of services, supplies, and equipment that are not identified by CPT codes, the level II HCPCS codes were established for submitting claims for these items. The development and use of level II of the HCPCS began in the 1980's. Level II codes are also referred to as alpha-numeric codes because they consist of a single alphabetical letter followed by 4 numeric digits, while CPT codes are identified using 5 numeric digits. See http://www.cms.hhs.gov/medicare/hcpcs/codpayproc.asp The fraction of post-1990 drugs increased from 9% in 1992 to 33% in 2003. The 1992 figure is based on 17,731 observations (drug treatments), and the 2003 figure is based on 486,409 treatments. The increase in sample size is partly due to the fact that the number of firms covered by the MEDSTAT data increased from 45 in 1992 to 200 in A similar procedure was used to construct the radiology and surgery innovation measures (rad_onc_new%it, rad_diag_new%it, and surg_new%it). However, unlike drugs, radiology and surgical procedures are not subject to FDA approval,9 so FDA approval dates can’t be used to measure the vintage of these procedures. Radiology and surgical procedures are coded using Current Procedural Terminology (CPT) codes that are established and maintained by the American Medical Association.10 The AMA publishes a database (CPT Assistant Archives 1990-2003) that provides information about the year in which each CPT code was first established. The data are left-censored: if a code was established prior to 1990, we know only that it is a pre-1990 code. To construct the radiology and surgery innovation measures, we defined POST1990p = 1 if the CPT code for procedure p was established by the AMA after The radiology and surgery innovation measures are probably less reliable than the drug innovation measure, because FDA approval of a drug is more meaningful indicator than AMA establishment of a new CPT code. For example, measuring surgical innovation using CPT code changes may be problematic. Closer inspection of the data on surgical procedures reveals that some “new” procedures are probably just relabeled or reclassified old procedures, rather than true innovations. For example, the three procedures whose codes were added in 1997 which were most frequently performed in 9 Some new procedures may be closely related to medical device innovations, which are subject to FDA approval, but linking procedure innovations to FDA approvals of medical devices is difficult. 10 For a description of how CPT codes are maintained, the committees involved, and the entire CPT process, including the evolution of CPT, see http://www.ama-assn.org/ama/pub/category/3112.html. 1997 were 98940, 98941, and 98942, which correspond to different types of chiropractic manipulative treatment of the spine. Undoubtedly, this type of treatment was performed well before 1997. A new CPT code should therefore be considered a necessary condition for a medical innovation, but not a sufficient condition: all innovations have new CPT codes, but some new CPT codes are not innovations. The fraction of procedures with new CPT codes exceeds the fraction of truly innovative procedures, perhaps by a significant amount, and the degree of overstatement varies across diseases. In the future, I hope to develop improved measures of radiology and surgery innovation. Table 4 presents data on innovation measures in 2003, by cancer site, ranked by IV. Estimation issues

Several issues regarding the estimation of eq. (5) should be considered before we present the empirical results. These issues are: (1) functional form; (2) measurement error; (3) weighting; and (4) estimation by cancer stage vs. overall estimation. Functional form. The dependent variable of eq. (5) is specified to be an arbitrary function of the observed survival rate, f(S). Because the survival rate is bounded between zero and one, a linear function (e.g., f(S) = S) would not be an appropriate choice. We will estimate the model using two alternative functional forms. The first is the probit, i.e. f(S) = F-1(S), where F-1( ) is the inverse of the standard normal cumulative distribution. The second is the logarithmic, i.e. f(S) = ln(S).11 As shown in eq. (5), the same transformation is applied to the expected survival rate. Measurement error. As described above, the survival data and the treatment vintage data were obtained from different data sources and are based on different populations. The survival data were obtained from SEER 9 public-use data, which primarily covers elderly people in certain regions of the U.S.12 The treatment vintage data were obtained from the MEDSTAT Marketscan database, which primarily covers nonelderly people in other regions of the U.S.13 As rich as the MEDSTAT database is, it provides data on only a 11 F-1(S) is similar to ln(S / (1 – S)). 12 About two-thirds of cancer patients are 65 or over. 13 MEDSTAT has data on some patients in Medicare health plans, but these data were not available for this study. small proportion of all cancer treatments provided in the U.S. This may be illustrated with utilization data for a specific drug treatment, an ondansetron HCL injection (HCPCS code J2405). In 2003 MEDSTAT data, this procedure was performed in connection with a cancer diagnosis 11,845 times. According to Medicare Part B Physician/Supplier (http://www.cms.hhs.gov/MedicareFeeforSvcPartsAB/Downloads/LEVEL2SERV03.pdf) there were 6,381,294 allowed services of ondansetron HCL injection in 2003. The MEDSTAT frequency is only 0.2% of the Medicare frequency. We assume that treatment innovation indicators based on MEDSTAT data are useful, albeit noisy, indicators of the treatment innovation experienced by patients in SEER 9 registries. This sampling error is likely to bias the coefficients on the treatment Weighting. Eq. (5) is to be estimated using grouped data, where groups are defined by cancer site and year of diagnosis. These groups are very heterogeneous in terms of size, where size is measured either by number of SEER 9 patients or number of MEDSTAT treatments. For example, as shown in Table 4, in 2003 there were 140,122 drug treatments for breast cancer, and only 32 for cancer of the eye and orbit. We will estimate eq. (5) via weighted least-squares (WLS), where the weight is a measure of size. Consider two different measures of size: the number of SEER 9 patients, and the number of MEDSTAT drug treatments. In a given year, the correlation across cancer sites between these two measures is quite high. For example, in 2003 the correlation between the number of SEER 9 patients and the number of MEDSTAT drug treatments is about .85. This suggests that the choice between these two weights wouldn’t make much difference. As noted above, however, the MEDSTAT sample size increased dramatically over time, suggesting that the more recent innovation measures are far more reliable, and therefore deserve much greater weight.14 We will estimate eq. (5) with two different sets of weights--the number of SEER 9 patients, and the number of MEDSTAT drug treatments. We believe that the estimates based on the latter set of weights are more credible. 14 As shown in Table 1, the SEER 9 sample size barely increased over time. Estimation: overall vs. by stage. The SEER microdata contain information about the stage of cancer at time of diagnosis. Thus, it is feasible to calculate observed and expected survival rates by cancer site, year, and stage. However, we believe that our approach (controlling for stage distribution rather than analysis by stage) is preferable, for two reasons. First, there is no information about cancer stage in MEDSTAT. Hence, we can’t construct stage-specific treatment innovation measures. Even if we could construct such measures, due to a phenomenon known as stage migration, analysis by cancer stage is probably inappropriate. Changes in stage-specific survival may provide a distorted view of true survival change. In particular, the survival rate for every stage may improve even when overall survival does not change. The assignment of a given stage to a particular cancer may change over time due to advances in diagnostic technology. Stage migration occurs when diagnostic procedures change over time, resulting in an increase in the probability that a given cancer will be diagnosed in a more advanced stage. For example, certain distant metastases that would have been undetectable a few years ago can now be diagnosed by a computer tomography (CT) scan or by magnetic resonance imaging (MRI). Therefore, some patients who would have been diagnosed previously as having cancer in a localized or regional stage are now diagnosed as having cancer in a distant stage. The likely result would be to remove the worst survivors — those with previously undetected distant metastases — from the localized and regional categories and put them into the distant category. As a result, the stage-at-diagnosis distribution for a cancer may become less favorable over time, but the survival rates for each stage may improve: the early stage will lose cases that will survive shorter than those remaining in that category, while the advanced stage will gain cases that will survive longer than those already in that category. However, overall survival would not change (Feinstein et al., 1985). Stage migration is an important concept to understand when examining temporal trends in survival by stage at diagnosis as well as temporal trends in stage distributions; it could affect the analysis of virtually all solid 15 SEER Cancer Statistics Review 1973-1999 Overview, p. 12. V. Empirical results

We computed weighted least-squares estimates of 12 versions of eq. (5): three survival intervals (1-year, 2-year, and 3-year survival), two functional forms (probit and logarithmic), and two sets of weights (the number of MEDSTAT drug treatments, and the number of SEER 9 patients diagnosed); 3 * 2 * 2 = 12. Table 5 shows estimates of eq. (5) for all three survival intervals, based on the probit functional form and weighting by For all three survival intervals, the coefficient on drug_new% is positive and significant (p-value < .03). This indicates that the cancer sites whose drug vintage (measured by the share of post-1990 treatments) increased the most during the 1990s tended to have larger increases in observed survival rates, ceteris paribus. None of the coefficients on the diagnostic radiology innovation measure or the surgery innovation measure are significant. The coefficient on the radiology oncology innovation measure is positive and significant in the 2-year survival equation, but not in The coefficients on the expected survival variable F-1(E) are all positive and significant. This indicates that part of the increase in the observed survival rate of cancer patients can be attributed to factors that also increased the survival of people who did not The coefficients on the log of the number of SEER 9 patients diagnosed are also all positive and significant. This may be capturing the impact of improved (earlier) detection: one would expect above-average increases in the number of people diagnosed for cancer sites with the greatest improvements in detection. The age coefficient is negative and significant in the 1-year survival equation but insignificant in the other two equations. However, as noted above, the expected survival rate is based on the age- (as well as race- and sex-) distribution of a comparable set of people who do not have cancer, so controlling for mean age as well as expected survival Now let’s consider the stage distribution coefficients. Since survival is inversely related to disease progression, one might expect the in_situ% coefficient to exceed the loc_reg% coefficient, and the loc_reg% coefficient to exceed the distant% coefficient. This is the case in the 2-year and 3-year survival equations, but not in the 1-year equation. Due to stage migration, however, this ordering is not necessarily to be expected. Suppose that there was no change in the true stage distribution of any cancer site, but that some cancer sites had improved detection. These cancer sites would have the largest increase in distant%, and might also have above-average increases in survival. The estimates shown in Table 5 do not distinguish between cancer drugs and other drugs. Table 6 shows the effect of distinguishing between these two types of drugs. Line 1 of Table 6 shows the estimates of the drug coefficients when cancer drugs and other drugs are pooled. Lines 2 and 3 show the estimates of the drug coefficients when cancer drugs and other drugs are disaggregated. In the 1-year and 2-year survival models, the coefficient on cancer_drug_new% is positive and highly significant, whereas the coefficient on other_drug_new% is insignificant. This suggests that the gains in cancer survival are primarily attributable to cancer drugs as opposed to other drugs. In the 3-year survival model, neither of the coefficients in lines 2 and 3 are significant, although the coefficient on drug_new% (the utilization-weighted average of cancer_drug_new% and other_drug_new%) is positive and significant. This may be attributable to the fact that the 3-year estimates are based on 47% fewer observations (drug treatments) than the 1-year estimates and 24% fewer than the 2-year estimates. In the remainder of the paper we will consider models in which cancer drugs and other As noted above, we also estimated models using an alternative (logarithmic) functional form, and an alternative set of weights (the number of SEER 9 patients diagnosed). To conserve space, we will not present the full estimates of these other nine models. But to enable assessment of the robustness of the estimates, we present in Table 7 the t-statistics (indicating both the sign and statistical significance) on the treatment innovation measures, F-1(E), ln(N), and age from all 12 models. The drug vintage coefficient is positive and significant in 11 of the 12 models, and positive and marginally significant (p-value = .07) in the other model. The only other variable whose coefficient is generally significant with a consistent sign is the log of the number of SEER 9 patients diagnosed; this coefficient is positive and significant in 9 of the 12 models. The coefficient on the expected survival term is positive and significant in only 4 models. The radiation oncology, diagnostic radiology, and surgery innovation coefficients are positive and significant in 2, 1, and 0 models, respectively. From our estimates, we can calculate the fraction of the 1992-1999 change in the observed survival rate that is attributable to the increased utilization of post-1990 drugs, ceteris paribus. The fraction is equal to (β1 * (drug_new%1999 - drug_new%1992)) / (f(S1999) – f(S1992)) As noted above, the overall value of drug_new% increased from 9% in 1992 to 29% in 1999, so the above expression reduces to (β1 * 20%) / (f(S1999) – f(S1992)). The following table shows the observed survival rates for all cancer sites combined in the years 1992 The following table shows estimates from each of the 12 models of the fraction of the 1992-1999 change in the observed survival rate that is attributable to the increased Estimates of the fraction of the 1992-1999 change in the observed survival rate that is attributable to the increased utilization of post-1990 drugs range from 12% to 121%. The estimated fraction is higher for shorter survival intervals, when observations are weighted by the number of MEDSTAT drug treatments, and for the logarithmic specification. The mean of the 12 estimates of the fraction of the 1992-1999 change in the observed survival rate that is attributable to the increased utilization of post-1990 drugs is 44%. Summary and discussion
Previous studies (Brenner (2002)) have shown that long-term survival rates for many types of cancer have substantially improved in past decades because of advances in early detection and treatment. This study has examined the impact of pharmaceutical innovation and other factors on the survival of U.S. cancer patients during the 1990s. In particular, it investigated whether cancer survival rates increased more for those cancer sites that had the largest increases in the proportion of drug treatments that were “new” treatments. By controlling for “expected survival,” i.e. the survival of a comparable set of people that did not have cancer, we measured the excess mortality that is associated with a cancer diagnosis. We also controlled for other types of medical innovation, i.e. innovation in surgical procedures, diagnostic radiology procedures, and radiation Data on observed and expected survival rates, the number of people diagnosed, mean age at diagnosis, and stage distribution were obtained from the National Cancer Institute’s SEER public-use data. Estimates of rates of innovation in drugs and other treatment and diagnostic procedures were constructed from the MEDSTAT Marketscan database and other data sources. Treatment innovation indicators based on MEDSTAT data are likely to be useful, albeit noisy, indicators of the treatment innovation experienced by patients in SEER registries. This sampling error is likely to bias the coefficients on the treatment innovation measures towards zero. We computed weighted least-squares estimates of 12 versions of a survival model, based on different survival intervals, functional forms, and sets of weights. The drug vintage coefficient was positive and significant in almost every model. This indicates that the cancer sites whose drug vintage (measured by the share of post-1990 treatments) increased the most during the 1990s tended to have larger increases in observed survival rates, ceteris paribus. Estimates of the fraction of the 1992-1999 change in the observed survival rate that is attributable to the increased utilization of post-1990 drugs ranged from 12% to 121%. The estimated fraction is higher for shorter survival intervals, when observations are weighted by the number of MEDSTAT drug treatments, and for the logarithmic specification. The mean of the 12 estimates of the fraction of the 1992-1999 change in the observed survival rate that is attributable to the increased utilization of post-1990 drugs is 44%. Due to sampling and other measurement errors, these estimates may be The coefficients on measures of other types of medical innovation (in radiation oncology, diagnostic radiology, and surgery innovation) were generally not significant. However these measures may be less reliable than the drug innovation measure: they were based upon the year in which the AMA established a new procedure code, which may be a far less meaningful indicator of innovation than the year in which the FDA first approved a drug. This topic warrants further research.
References

Axtell LM, Cutler SJ. (1961), “The relative survival rate: A statistical methodology,” J
Natl Cancer Inst Monogr 6, pp. 101-121.
Brenner H. (2002), "Long-term survival rates of cancer patients achieved by the end of
the 20th century: a period analysis," Lancet Oct 12; 360(9340):1131-5.
Feinstein AR, Sosin DM, Wells CK (1985), “The Will Rogers phenomenon: Stage
migration and new diagnostic techniques as a source of misleading statistics for survival
of cancer,” New England Journal of Medicine 312, pp.1604-1608.
Johnson, John, Grant Williams, and Richard Pazdur, “End points and United States Food
and Drug Administration Approval of Oncology Drugs,” Journal of Clinical Oncology 21
(7), 1 April 2003, pp. 1404-11.
Kantarjian H, Talpaz M, O'Brien S, Giles F, Faderl S, Verstovsek S, Garcia-Manero G,
Shan J, Rios MB, Champlin R, de Lima M, Cortes J. (2005), “Survival benefit with
imatinib mesylate therapy in patients with accelerated-phase chronic myelogenous
leukemia--comparison with historic experience.” Cancer. May 15; 103(10): 2099-108.
Mosby (2004), Mosby's Drug Consult 2004, 14th Edition,
http://www.mosbysdrugconsult.com/
Richardson PG, et al (2005), “Bortezomib or high-dose dexamethasone for relapsed
multiple myeloma.” N Engl J Med. Jun 16;352(24):2487-98.
von der Maase H, Sengelov L, Roberts JT, Ricci S, Dogliotti L, Oliver T, Moore MJ, Zimmermann A, Arning M (2005), “Long-term survival results of a randomized trial comparing gemcitabine plus cisplatin, with methotrexate, vinblastine, doxorubicin, plus cisplatin in patients with bladder cancer,” J Clin Oncol. Jul 20; 23(21): 4602-8. 1992 and 2000 survival data, top 30 (ranked by 1992 incidence) cancer sites
Two-year survival rates
Observed
Expected
Relative
diagnosed
1992 2000 1992 2000 1992
Top 40 (ranked by frequency) provider-administered drugs
associated with cancer diagnoses in 2003
COUNT PERCENT
Comparison of top 20 (ranked by frequency) provider-administered drugs associated with two major cancer sites in 2003
Colon excluding Rectum
COUNT PERCENT
COUNT PERCENT
Innovation measures in 2003, by cancer site, ranked by number of drug treatments
Post-1990 procedures/total procedures
Number of procedures
Cancer site
drugs radiation diagnostic
radiation diagnostic surgery
oncology
radiology
oncology
radiology
Post-1990 procedures/total procedures
Number of procedures
Cancer site
drugs radiation diagnostic
radiation diagnostic surgery
oncology
radiology
oncology
radiology
Trachea, Mediastinum and Other Respiratory Organs Weighted least-squares estimates of eq. (5),
based on probit functional form and weighting by number of MEDSTAT drug treatments
1-year survival
2-year survival
3-year survival
Parameter
Estimate t-Value Prob > t Estimate t-Value Prob > t Estimate t-Value Prob > t
The dependent variable is F-1(Sit), where Sit is the observed survival rate of people diagnosed with cancer originating at site i in year t, and F-1( ) is the inverse of the standard normal cumulative distribution. All models include cancer-site fixed effects.
Distinguishing between cancer drugs and other drugs
1-year survival
2-year survival
3-year survival
Line Parameter
Estimate t-Value Prob > t Estimate t-Value Prob > t Estimate t-Value Prob > t
t-statistics (indicating signs and statistical significance) of estimated coefficients of 12 survival models
Survival interval
Functional form
MEDST MEDST MEDST SEER 9 SEER 9 SEER 9 MEDST MEDST MEDST SEER 9 SEER 9 SEER 9 AT drug AT drug AT drug patients patients patients AT drug AT drug AT drug patients patients patients treatment treatment treatment diagnose diagnose diagnose treatment treatment treatment diagnose diagnose diagnose Coefficient
drug_new%
SEER Cancer Causes of Death Recode 1969+ (9/17/2004) Cancer Causes of Death
All Malignant Cancers
Oral Cavity and Pharynx
Lip
Digestive System
Esophagus
Respiratory System
Nose, Nasal Cavity and Middle Ear
Bones and Joints
Skin excluding Basal and Squamous
Melanoma of the Skin
Female Genital System
Cervix Uteri
Male Genital System
Prostate
Urinary System
Urinary Bladder
Eye and Orbit
Brain and Other Nervous System
Endocrine System
Thyroid
Lymphoma
Hodgkin Lymphoma
Leukemia
Lymphocytic Leukemia
Acute Lymphocytic Leukemia
Myeloid and Monocytic LeukemiaAcute myeloid C46+
C26.1, C45.7+, C45.9+, C76-C80,
159.1, 195-199, 202.3, 202.5 C88, C96.0-C96.2, C96.7, C96.9, Miscellaneous Malignant Cancer
http://seer.cancer.gov/codrecode/1969+_d09172004/index.html

Source: http://economics.adelaide.edu.au/events/archive/2007/Pharmaceutical-innovation-and-US-cancer-survival-in-the-1990s-evidence-from-linked-SEER-MEDSTAT-data.pdf